diff options
Diffstat (limited to 'llvm/test')
| -rw-r--r-- | llvm/test/CodeGen/X86/aes-schedule.ll | 28 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/avx-schedule.ll | 302 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/avx2-schedule.ll | 266 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/bmi2-schedule.ll | 208 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/f16c-schedule.ll | 20 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/fma-schedule.ll | 128 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/recip-fastmath.ll | 118 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/recip-fastmath2.ll | 139 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/sse-schedule.ll | 172 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/sse2-schedule.ll | 322 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/sse3-schedule.ll | 30 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/sse41-schedule.ll | 98 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/sse42-schedule.ll | 26 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/ssse3-schedule.ll | 26 |
14 files changed, 1033 insertions, 850 deletions
diff --git a/llvm/test/CodeGen/X86/aes-schedule.ll b/llvm/test/CodeGen/X86/aes-schedule.ll index 53560a5fed6..c77f521ddf1 100644 --- a/llvm/test/CodeGen/X86/aes-schedule.ll +++ b/llvm/test/CodeGen/X86/aes-schedule.ll @@ -36,8 +36,8 @@ define <2 x i64> @test_aesdec(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_aesdec: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaesdec %xmm1, %xmm0, %xmm0 # sched: [7:1.00] -; SKYLAKE-NEXT: vaesdec (%rdi), %xmm0, %xmm0 # sched: [7:1.00] +; SKYLAKE-NEXT: vaesdec %xmm1, %xmm0, %xmm0 # sched: [4:1.00] +; SKYLAKE-NEXT: vaesdec (%rdi), %xmm0, %xmm0 # sched: [4:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_aesdec: @@ -85,8 +85,8 @@ define <2 x i64> @test_aesdeclast(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) ; ; SKYLAKE-LABEL: test_aesdeclast: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaesdeclast %xmm1, %xmm0, %xmm0 # sched: [7:1.00] -; SKYLAKE-NEXT: vaesdeclast (%rdi), %xmm0, %xmm0 # sched: [7:1.00] +; SKYLAKE-NEXT: vaesdeclast %xmm1, %xmm0, %xmm0 # sched: [4:1.00] +; SKYLAKE-NEXT: vaesdeclast (%rdi), %xmm0, %xmm0 # sched: [4:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_aesdeclast: @@ -134,8 +134,8 @@ define <2 x i64> @test_aesenc(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_aesenc: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaesenc %xmm1, %xmm0, %xmm0 # sched: [7:1.00] -; SKYLAKE-NEXT: vaesenc (%rdi), %xmm0, %xmm0 # sched: [7:1.00] +; SKYLAKE-NEXT: vaesenc %xmm1, %xmm0, %xmm0 # sched: [4:1.00] +; SKYLAKE-NEXT: vaesenc (%rdi), %xmm0, %xmm0 # sched: [4:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_aesenc: @@ -183,8 +183,8 @@ define <2 x i64> @test_aesenclast(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) ; ; SKYLAKE-LABEL: test_aesenclast: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaesenclast %xmm1, %xmm0, %xmm0 # sched: [7:1.00] -; SKYLAKE-NEXT: vaesenclast (%rdi), %xmm0, %xmm0 # sched: [7:1.00] +; SKYLAKE-NEXT: vaesenclast %xmm1, %xmm0, %xmm0 # sched: [4:1.00] +; SKYLAKE-NEXT: vaesenclast (%rdi), %xmm0, %xmm0 # sched: [4:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_aesenclast: @@ -236,9 +236,9 @@ define <2 x i64> @test_aesimc(<2 x i64> %a0, <2 x i64> *%a1) { ; ; SKYLAKE-LABEL: test_aesimc: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaesimc %xmm0, %xmm0 # sched: [14:2.00] -; SKYLAKE-NEXT: vaesimc (%rdi), %xmm1 # sched: [14:2.00] -; SKYLAKE-NEXT: vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vaesimc %xmm0, %xmm0 # sched: [8:2.00] +; SKYLAKE-NEXT: vaesimc (%rdi), %xmm1 # sched: [8:2.00] +; SKYLAKE-NEXT: vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_aesimc: @@ -293,9 +293,9 @@ define <2 x i64> @test_aeskeygenassist(<2 x i64> %a0, <2 x i64> *%a1) { ; ; SKYLAKE-LABEL: test_aeskeygenassist: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaeskeygenassist $7, %xmm0, %xmm0 # sched: [29:7.00] -; SKYLAKE-NEXT: vaeskeygenassist $7, (%rdi), %xmm1 # sched: [28:7.00] -; SKYLAKE-NEXT: vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vaeskeygenassist $7, %xmm0, %xmm0 # sched: [20:6.00] +; SKYLAKE-NEXT: vaeskeygenassist $7, (%rdi), %xmm1 # sched: [19:6.00] +; SKYLAKE-NEXT: vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_aeskeygenassist: diff --git a/llvm/test/CodeGen/X86/avx-schedule.ll b/llvm/test/CodeGen/X86/avx-schedule.ll index 28bd53c2534..12d41bc25e6 100644 --- a/llvm/test/CodeGen/X86/avx-schedule.ll +++ b/llvm/test/CodeGen/X86/avx-schedule.ll @@ -28,8 +28,8 @@ define <4 x double> @test_addpd(<4 x double> %a0, <4 x double> %a1, <4 x double> ; ; SKYLAKE-LABEL: test_addpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vaddpd (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_addpd: @@ -70,8 +70,8 @@ define <8 x float> @test_addps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a ; ; SKYLAKE-LABEL: test_addps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddps (%rdi), %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vaddps (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_addps: @@ -112,8 +112,8 @@ define <4 x double> @test_addsubpd(<4 x double> %a0, <4 x double> %a1, <4 x doub ; ; SKYLAKE-LABEL: test_addsubpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaddsubpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddsubpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddsubpd %ymm1, %ymm0, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vaddsubpd (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_addsubpd: @@ -155,8 +155,8 @@ define <8 x float> @test_addsubps(<8 x float> %a0, <8 x float> %a1, <8 x float> ; ; SKYLAKE-LABEL: test_addsubps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaddsubps %ymm1, %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddsubps (%rdi), %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddsubps %ymm1, %ymm0, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vaddsubps (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_addsubps: @@ -201,9 +201,9 @@ define <4 x double> @test_andnotpd(<4 x double> %a0, <4 x double> %a1, <4 x doub ; ; SKYLAKE-LABEL: test_andnotpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vandnpd %ymm1, %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vandnpd (%rdi), %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vandnpd %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vandnpd (%rdi), %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vaddpd %ymm0, %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_andnotpd: @@ -256,9 +256,9 @@ define <8 x float> @test_andnotps(<8 x float> %a0, <8 x float> %a1, <8 x float> ; ; SKYLAKE-LABEL: test_andnotps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vandnps %ymm1, %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vandnps (%rdi), %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vandnps %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vandnps (%rdi), %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_andnotps: @@ -311,9 +311,9 @@ define <4 x double> @test_andpd(<4 x double> %a0, <4 x double> %a1, <4 x double> ; ; SKYLAKE-LABEL: test_andpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vandpd %ymm1, %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vandpd (%rdi), %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vandpd %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vandpd (%rdi), %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vaddpd %ymm0, %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_andpd: @@ -364,9 +364,9 @@ define <8 x float> @test_andps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a ; ; SKYLAKE-LABEL: test_andps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vandps %ymm1, %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vandps (%rdi), %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vandps %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vandps (%rdi), %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_andps: @@ -417,8 +417,8 @@ define <4 x double> @test_blendpd(<4 x double> %a0, <4 x double> %a1, <4 x doubl ; ; SKYLAKE-LABEL: test_blendpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0],ymm1[1,2],ymm0[3] sched: [1:0.33] -; SKYLAKE-NEXT: vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0],ymm1[1,2],ymm0[3] sched: [1:0.50] +; SKYLAKE-NEXT: vaddpd %ymm0, %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0],mem[1,2],ymm0[3] sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -463,7 +463,7 @@ define <8 x float> @test_blendps(<8 x float> %a0, <8 x float> %a1, <8 x float> * ; ; SKYLAKE-LABEL: test_blendps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2],ymm0[3,4,5,6,7] sched: [1:0.33] +; SKYLAKE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2],ymm0[3,4,5,6,7] sched: [1:0.50] ; SKYLAKE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2],ymm0[3],mem[4,5,6],ymm0[7] sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -505,8 +505,8 @@ define <4 x double> @test_blendvpd(<4 x double> %a0, <4 x double> %a1, <4 x doub ; ; SKYLAKE-LABEL: test_blendvpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vblendvpd %ymm2, %ymm1, %ymm0, %ymm0 # sched: [2:2.00] -; SKYLAKE-NEXT: vblendvpd %ymm2, (%rdi), %ymm0, %ymm0 # sched: [2:2.00] +; SKYLAKE-NEXT: vblendvpd %ymm2, %ymm1, %ymm0, %ymm0 # sched: [2:0.67] +; SKYLAKE-NEXT: vblendvpd %ymm2, (%rdi), %ymm0, %ymm0 # sched: [2:0.67] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_blendvpd: @@ -548,8 +548,8 @@ define <8 x float> @test_blendvps(<8 x float> %a0, <8 x float> %a1, <8 x float> ; ; SKYLAKE-LABEL: test_blendvps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vblendvps %ymm2, %ymm1, %ymm0, %ymm0 # sched: [2:2.00] -; SKYLAKE-NEXT: vblendvps %ymm2, (%rdi), %ymm0, %ymm0 # sched: [2:2.00] +; SKYLAKE-NEXT: vblendvps %ymm2, %ymm1, %ymm0, %ymm0 # sched: [2:0.67] +; SKYLAKE-NEXT: vblendvps %ymm2, (%rdi), %ymm0, %ymm0 # sched: [2:0.67] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_blendvps: @@ -737,9 +737,9 @@ define <4 x double> @test_cmppd(<4 x double> %a0, <4 x double> %a1, <4 x double> ; ; SKYLAKE-LABEL: test_cmppd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcmpeqpd %ymm1, %ymm0, %ymm1 # sched: [3:1.00] -; SKYLAKE-NEXT: vcmpeqpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vorpd %ymm0, %ymm1, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vcmpeqpd %ymm1, %ymm0, %ymm1 # sched: [4:0.33] +; SKYLAKE-NEXT: vcmpeqpd (%rdi), %ymm0, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vorpd %ymm0, %ymm1, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cmppd: @@ -789,9 +789,9 @@ define <8 x float> @test_cmpps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a ; ; SKYLAKE-LABEL: test_cmpps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcmpeqps %ymm1, %ymm0, %ymm1 # sched: [3:1.00] -; SKYLAKE-NEXT: vcmpeqps (%rdi), %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vorps %ymm0, %ymm1, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vcmpeqps %ymm1, %ymm0, %ymm1 # sched: [4:0.33] +; SKYLAKE-NEXT: vcmpeqps (%rdi), %ymm0, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vorps %ymm0, %ymm1, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cmpps: @@ -841,9 +841,9 @@ define <4 x double> @test_cvtdq2pd(<4 x i32> %a0, <4 x i32> *%a1) { ; ; SKYLAKE-LABEL: test_cvtdq2pd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtdq2pd %xmm0, %ymm0 # sched: [6:1.00] -; SKYLAKE-NEXT: vcvtdq2pd (%rdi), %ymm1 # sched: [6:1.00] -; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtdq2pd %xmm0, %ymm0 # sched: [7:1.00] +; SKYLAKE-NEXT: vcvtdq2pd (%rdi), %ymm1 # sched: [7:1.00] +; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cvtdq2pd: @@ -892,9 +892,9 @@ define <8 x float> @test_cvtdq2ps(<8 x i32> %a0, <8 x i32> *%a1) { ; ; SKYLAKE-LABEL: test_cvtdq2ps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtdq2ps %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vcvtdq2ps (%rdi), %ymm1 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtdq2ps %ymm0, %ymm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vcvtdq2ps (%rdi), %ymm1 # sched: [4:0.50] +; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cvtdq2ps: @@ -941,8 +941,8 @@ define <8 x i32> @test_cvtpd2dq(<4 x double> %a0, <4 x double> *%a1) { ; ; SKYLAKE-LABEL: test_cvtpd2dq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvttpd2dq %ymm0, %xmm0 # sched: [6:1.00] -; SKYLAKE-NEXT: vcvttpd2dqy (%rdi), %xmm1 # sched: [7:1.00] +; SKYLAKE-NEXT: vcvttpd2dq %ymm0, %xmm0 # sched: [7:1.00] +; SKYLAKE-NEXT: vcvttpd2dqy (%rdi), %xmm1 # sched: [8:1.00] ; SKYLAKE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0 # sched: [3:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -990,8 +990,8 @@ define <8 x float> @test_cvtpd2ps(<4 x double> %a0, <4 x double> *%a1) { ; ; SKYLAKE-LABEL: test_cvtpd2ps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtpd2ps %ymm0, %xmm0 # sched: [6:1.00] -; SKYLAKE-NEXT: vcvtpd2psy (%rdi), %xmm1 # sched: [7:1.00] +; SKYLAKE-NEXT: vcvtpd2ps %ymm0, %xmm0 # sched: [7:1.00] +; SKYLAKE-NEXT: vcvtpd2psy (%rdi), %xmm1 # sched: [8:1.00] ; SKYLAKE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0 # sched: [3:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1039,9 +1039,9 @@ define <8 x i32> @test_cvtps2dq(<8 x float> %a0, <8 x float> *%a1) { ; ; SKYLAKE-LABEL: test_cvtps2dq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvttps2dq %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vcvttps2dq (%rdi), %ymm1 # sched: [3:1.00] -; SKYLAKE-NEXT: vorps %ymm1, %ymm0, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vcvttps2dq %ymm0, %ymm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vcvttps2dq (%rdi), %ymm1 # sched: [4:0.50] +; SKYLAKE-NEXT: vorps %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cvtps2dq: @@ -1085,8 +1085,8 @@ define <4 x double> @test_divpd(<4 x double> %a0, <4 x double> %a1, <4 x double> ; ; SKYLAKE-LABEL: test_divpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vdivpd %ymm1, %ymm0, %ymm0 # sched: [35:2.00] -; SKYLAKE-NEXT: vdivpd (%rdi), %ymm0, %ymm0 # sched: [35:2.00] +; SKYLAKE-NEXT: vdivpd %ymm1, %ymm0, %ymm0 # sched: [14:1.00] +; SKYLAKE-NEXT: vdivpd (%rdi), %ymm0, %ymm0 # sched: [14:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_divpd: @@ -1127,8 +1127,8 @@ define <8 x float> @test_divps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a ; ; SKYLAKE-LABEL: test_divps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vdivps %ymm1, %ymm0, %ymm0 # sched: [21:2.00] -; SKYLAKE-NEXT: vdivps (%rdi), %ymm0, %ymm0 # sched: [21:2.00] +; SKYLAKE-NEXT: vdivps %ymm1, %ymm0, %ymm0 # sched: [11:1.00] +; SKYLAKE-NEXT: vdivps (%rdi), %ymm0, %ymm0 # sched: [11:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_divps: @@ -1169,8 +1169,8 @@ define <8 x float> @test_dpps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a2 ; ; SKYLAKE-LABEL: test_dpps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vdpps $7, %ymm1, %ymm0, %ymm0 # sched: [14:2.00] -; SKYLAKE-NEXT: vdpps $7, (%rdi), %ymm0, %ymm0 # sched: [14:2.00] +; SKYLAKE-NEXT: vdpps $7, %ymm1, %ymm0, %ymm0 # sched: [13:1.33] +; SKYLAKE-NEXT: vdpps $7, (%rdi), %ymm0, %ymm0 # sched: [13:1.33] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_dpps: @@ -1259,8 +1259,8 @@ define <4 x double> @test_haddpd(<4 x double> %a0, <4 x double> %a1, <4 x double ; ; SKYLAKE-LABEL: test_haddpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vhaddpd %ymm1, %ymm0, %ymm0 # sched: [5:2.00] -; SKYLAKE-NEXT: vhaddpd (%rdi), %ymm0, %ymm0 # sched: [5:2.00] +; SKYLAKE-NEXT: vhaddpd %ymm1, %ymm0, %ymm0 # sched: [6:2.00] +; SKYLAKE-NEXT: vhaddpd (%rdi), %ymm0, %ymm0 # sched: [6:2.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_haddpd: @@ -1302,8 +1302,8 @@ define <8 x float> @test_haddps(<8 x float> %a0, <8 x float> %a1, <8 x float> *% ; ; SKYLAKE-LABEL: test_haddps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vhaddps %ymm1, %ymm0, %ymm0 # sched: [5:2.00] -; SKYLAKE-NEXT: vhaddps (%rdi), %ymm0, %ymm0 # sched: [5:2.00] +; SKYLAKE-NEXT: vhaddps %ymm1, %ymm0, %ymm0 # sched: [6:2.00] +; SKYLAKE-NEXT: vhaddps (%rdi), %ymm0, %ymm0 # sched: [6:2.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_haddps: @@ -1345,8 +1345,8 @@ define <4 x double> @test_hsubpd(<4 x double> %a0, <4 x double> %a1, <4 x double ; ; SKYLAKE-LABEL: test_hsubpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vhsubpd %ymm1, %ymm0, %ymm0 # sched: [5:2.00] -; SKYLAKE-NEXT: vhsubpd (%rdi), %ymm0, %ymm0 # sched: [5:2.00] +; SKYLAKE-NEXT: vhsubpd %ymm1, %ymm0, %ymm0 # sched: [6:2.00] +; SKYLAKE-NEXT: vhsubpd (%rdi), %ymm0, %ymm0 # sched: [6:2.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_hsubpd: @@ -1388,8 +1388,8 @@ define <8 x float> @test_hsubps(<8 x float> %a0, <8 x float> %a1, <8 x float> *% ; ; SKYLAKE-LABEL: test_hsubps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vhsubps %ymm1, %ymm0, %ymm0 # sched: [5:2.00] -; SKYLAKE-NEXT: vhsubps (%rdi), %ymm0, %ymm0 # sched: [5:2.00] +; SKYLAKE-NEXT: vhsubps %ymm1, %ymm0, %ymm0 # sched: [6:2.00] +; SKYLAKE-NEXT: vhsubps (%rdi), %ymm0, %ymm0 # sched: [6:2.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_hsubps: @@ -1436,7 +1436,7 @@ define <8 x float> @test_insertf128(<8 x float> %a0, <4 x float> %a1, <4 x float ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 # sched: [3:1.00] ; SKYLAKE-NEXT: vinsertf128 $1, (%rdi), %ymm0, %ymm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_insertf128: @@ -1520,8 +1520,8 @@ define <2 x double> @test_maskmovpd(i8* %a0, <2 x i64> %a1, <2 x double> %a2) { ; ; SKYLAKE-LABEL: test_maskmovpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmaskmovpd (%rdi), %xmm0, %xmm2 # sched: [2:2.00] -; SKYLAKE-NEXT: vmaskmovpd %xmm1, %xmm0, (%rdi) # sched: [4:1.00] +; SKYLAKE-NEXT: vmaskmovpd (%rdi), %xmm0, %xmm2 # sched: [1:0.50] +; SKYLAKE-NEXT: vmaskmovpd %xmm1, %xmm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: vmovapd %xmm2, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1569,8 +1569,8 @@ define <4 x double> @test_maskmovpd_ymm(i8* %a0, <4 x i64> %a1, <4 x double> %a2 ; ; SKYLAKE-LABEL: test_maskmovpd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmaskmovpd (%rdi), %ymm0, %ymm2 # sched: [2:2.00] -; SKYLAKE-NEXT: vmaskmovpd %ymm1, %ymm0, (%rdi) # sched: [4:1.00] +; SKYLAKE-NEXT: vmaskmovpd (%rdi), %ymm0, %ymm2 # sched: [1:0.50] +; SKYLAKE-NEXT: vmaskmovpd %ymm1, %ymm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: vmovapd %ymm2, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1618,8 +1618,8 @@ define <4 x float> @test_maskmovps(i8* %a0, <4 x i32> %a1, <4 x float> %a2) { ; ; SKYLAKE-LABEL: test_maskmovps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmaskmovps (%rdi), %xmm0, %xmm2 # sched: [2:2.00] -; SKYLAKE-NEXT: vmaskmovps %xmm1, %xmm0, (%rdi) # sched: [4:1.00] +; SKYLAKE-NEXT: vmaskmovps (%rdi), %xmm0, %xmm2 # sched: [1:0.50] +; SKYLAKE-NEXT: vmaskmovps %xmm1, %xmm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: vmovaps %xmm2, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1667,8 +1667,8 @@ define <8 x float> @test_maskmovps_ymm(i8* %a0, <8 x i32> %a1, <8 x float> %a2) ; ; SKYLAKE-LABEL: test_maskmovps_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmaskmovps (%rdi), %ymm0, %ymm2 # sched: [2:2.00] -; SKYLAKE-NEXT: vmaskmovps %ymm1, %ymm0, (%rdi) # sched: [4:1.00] +; SKYLAKE-NEXT: vmaskmovps (%rdi), %ymm0, %ymm2 # sched: [1:0.50] +; SKYLAKE-NEXT: vmaskmovps %ymm1, %ymm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: vmovaps %ymm2, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1713,8 +1713,8 @@ define <4 x double> @test_maxpd(<4 x double> %a0, <4 x double> %a1, <4 x double> ; ; SKYLAKE-LABEL: test_maxpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmaxpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vmaxpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vmaxpd %ymm1, %ymm0, %ymm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vmaxpd (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_maxpd: @@ -1756,8 +1756,8 @@ define <8 x float> @test_maxps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a ; ; SKYLAKE-LABEL: test_maxps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmaxps %ymm1, %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vmaxps (%rdi), %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vmaxps %ymm1, %ymm0, %ymm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vmaxps (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_maxps: @@ -1799,8 +1799,8 @@ define <4 x double> @test_minpd(<4 x double> %a0, <4 x double> %a1, <4 x double> ; ; SKYLAKE-LABEL: test_minpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vminpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vminpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vminpd %ymm1, %ymm0, %ymm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vminpd (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_minpd: @@ -1842,8 +1842,8 @@ define <8 x float> @test_minps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a ; ; SKYLAKE-LABEL: test_minps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vminps %ymm1, %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vminps (%rdi), %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vminps %ymm1, %ymm0, %ymm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vminps (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_minps: @@ -1889,7 +1889,7 @@ define <4 x double> @test_movapd(<4 x double> *%a0, <4 x double> *%a1) { ; SKYLAKE-LABEL: test_movapd: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovapd (%rdi), %ymm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vaddpd %ymm0, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %ymm0, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovapd %ymm0, (%rsi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1937,7 +1937,7 @@ define <8 x float> @test_movaps(<8 x float> *%a0, <8 x float> *%a1) { ; SKYLAKE-LABEL: test_movaps: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovaps (%rdi), %ymm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vaddps %ymm0, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %ymm0, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovaps %ymm0, (%rsi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1986,7 +1986,7 @@ define <4 x double> @test_movddup(<4 x double> %a0, <4 x double> *%a1) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovddup {{.*#+}} ymm0 = ymm0[0,0,2,2] sched: [1:1.00] ; SKYLAKE-NEXT: vmovddup {{.*#+}} ymm1 = mem[0,0,2,2] sched: [1:0.50] -; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_movddup: @@ -2030,7 +2030,7 @@ define i32 @test_movmskpd(<4 x double> %a0) { ; ; SKYLAKE-LABEL: test_movmskpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmovmskpd %ymm0, %eax # sched: [3:1.00] +; SKYLAKE-NEXT: vmovmskpd %ymm0, %eax # sched: [2:1.00] ; SKYLAKE-NEXT: vzeroupper # sched: [4:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2070,7 +2070,7 @@ define i32 @test_movmskps(<8 x float> %a0) { ; ; SKYLAKE-LABEL: test_movmskps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmovmskps %ymm0, %eax # sched: [3:1.00] +; SKYLAKE-NEXT: vmovmskps %ymm0, %eax # sched: [2:1.00] ; SKYLAKE-NEXT: vzeroupper # sched: [4:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2110,7 +2110,7 @@ define <4 x double> @test_movntpd(<4 x double> %a0, <4 x double> *%a1) { ; ; SKYLAKE-LABEL: test_movntpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaddpd %ymm0, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %ymm0, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovntpd %ymm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2151,7 +2151,7 @@ define <8 x float> @test_movntps(<8 x float> %a0, <8 x float> *%a1) { ; ; SKYLAKE-LABEL: test_movntps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaddps %ymm0, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %ymm0, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovntps %ymm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2197,7 +2197,7 @@ define <8 x float> @test_movshdup(<8 x float> %a0, <8 x float> *%a1) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovshdup {{.*#+}} ymm0 = ymm0[1,1,3,3,5,5,7,7] sched: [1:1.00] ; SKYLAKE-NEXT: vmovshdup {{.*#+}} ymm1 = mem[1,1,3,3,5,5,7,7] sched: [1:0.50] -; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_movshdup: @@ -2246,7 +2246,7 @@ define <8 x float> @test_movsldup(<8 x float> %a0, <8 x float> *%a1) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovsldup {{.*#+}} ymm0 = ymm0[0,0,2,2,4,4,6,6] sched: [1:1.00] ; SKYLAKE-NEXT: vmovsldup {{.*#+}} ymm1 = mem[0,0,2,2,4,4,6,6] sched: [1:0.50] -; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_movsldup: @@ -2296,7 +2296,7 @@ define <4 x double> @test_movupd(<4 x double> *%a0, <4 x double> *%a1) { ; SKYLAKE-LABEL: test_movupd: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovupd (%rdi), %ymm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vaddpd %ymm0, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %ymm0, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovupd %ymm0, (%rsi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2346,7 +2346,7 @@ define <8 x float> @test_movups(<8 x float> *%a0, <8 x float> *%a1) { ; SKYLAKE-LABEL: test_movups: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovups (%rdi), %ymm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vaddps %ymm0, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %ymm0, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovups %ymm0, (%rsi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2390,8 +2390,8 @@ define <4 x double> @test_mulpd(<4 x double> %a0, <4 x double> %a1, <4 x double> ; ; SKYLAKE-LABEL: test_mulpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmulpd %ymm1, %ymm0, %ymm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vmulpd (%rdi), %ymm0, %ymm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vmulpd %ymm1, %ymm0, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vmulpd (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_mulpd: @@ -2432,8 +2432,8 @@ define <8 x float> @test_mulps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a ; ; SKYLAKE-LABEL: test_mulps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmulps %ymm1, %ymm0, %ymm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vmulps (%rdi), %ymm0, %ymm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vmulps %ymm1, %ymm0, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vmulps (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_mulps: @@ -2477,9 +2477,9 @@ define <4 x double> @orpd(<4 x double> %a0, <4 x double> %a1, <4 x double> *%a2) ; ; SKYLAKE-LABEL: orpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vorpd %ymm1, %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vorpd (%rdi), %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vorpd %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vorpd (%rdi), %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vaddpd %ymm0, %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: orpd: @@ -2530,9 +2530,9 @@ define <8 x float> @test_orps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a2 ; ; SKYLAKE-LABEL: test_orps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vorps %ymm1, %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vorps (%rdi), %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vorps %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vorps (%rdi), %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_orps: @@ -2585,7 +2585,7 @@ define <4 x double> @test_perm2f128(<4 x double> %a0, <4 x double> %a1, <4 x dou ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm0[2,3],ymm1[0,1] sched: [3:1.00] ; SKYLAKE-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],mem[0,1] sched: [3:1.00] -; SKYLAKE-NEXT: vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %ymm0, %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_perm2f128: @@ -2634,7 +2634,7 @@ define <2 x double> @test_permilpd(<2 x double> %a0, <2 x double> *%a1) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpermilpd {{.*#+}} xmm0 = xmm0[1,0] sched: [1:1.00] ; SKYLAKE-NEXT: vpermilpd {{.*#+}} xmm1 = mem[1,0] sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_permilpd: @@ -2683,7 +2683,7 @@ define <4 x double> @test_permilpd_ymm(<4 x double> %a0, <4 x double> *%a1) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpermilpd {{.*#+}} ymm0 = ymm0[1,0,2,3] sched: [1:1.00] ; SKYLAKE-NEXT: vpermilpd {{.*#+}} ymm1 = mem[1,0,2,3] sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_permilpd_ymm: @@ -2732,7 +2732,7 @@ define <4 x float> @test_permilps(<4 x float> %a0, <4 x float> *%a1) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpermilps {{.*#+}} xmm0 = xmm0[3,2,1,0] sched: [1:1.00] ; SKYLAKE-NEXT: vpermilps {{.*#+}} xmm1 = mem[3,2,1,0] sched: [1:1.00] -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_permilps: @@ -2781,7 +2781,7 @@ define <8 x float> @test_permilps_ymm(<8 x float> %a0, <8 x float> *%a1) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpermilps {{.*#+}} ymm0 = ymm0[3,2,1,0,7,6,5,4] sched: [1:1.00] ; SKYLAKE-NEXT: vpermilps {{.*#+}} ymm1 = mem[3,2,1,0,7,6,5,4] sched: [1:1.00] -; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_permilps_ymm: @@ -3000,9 +3000,9 @@ define <8 x float> @test_rcpps(<8 x float> %a0, <8 x float> *%a1) { ; ; SKYLAKE-LABEL: test_rcpps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vrcpps (%rdi), %ymm1 # sched: [11:2.00] -; SKYLAKE-NEXT: vrcpps %ymm0, %ymm0 # sched: [11:2.00] -; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vrcpps %ymm0, %ymm0 # sched: [4:1.00] +; SKYLAKE-NEXT: vrcpps (%rdi), %ymm1 # sched: [4:1.00] +; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_rcpps: @@ -3050,9 +3050,9 @@ define <4 x double> @test_roundpd(<4 x double> %a0, <4 x double> *%a1) { ; ; SKYLAKE-LABEL: test_roundpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vroundpd $7, %ymm0, %ymm0 # sched: [5:1.25] -; SKYLAKE-NEXT: vroundpd $7, (%rdi), %ymm1 # sched: [6:2.00] -; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vroundpd $7, %ymm0, %ymm0 # sched: [8:0.67] +; SKYLAKE-NEXT: vroundpd $7, (%rdi), %ymm1 # sched: [8:0.67] +; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_roundpd: @@ -3100,9 +3100,9 @@ define <8 x float> @test_roundps(<8 x float> %a0, <8 x float> *%a1) { ; ; SKYLAKE-LABEL: test_roundps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vroundps $7, %ymm0, %ymm0 # sched: [5:1.25] -; SKYLAKE-NEXT: vroundps $7, (%rdi), %ymm1 # sched: [6:2.00] -; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vroundps $7, %ymm0, %ymm0 # sched: [8:0.67] +; SKYLAKE-NEXT: vroundps $7, (%rdi), %ymm1 # sched: [8:0.67] +; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_roundps: @@ -3150,9 +3150,9 @@ define <8 x float> @test_rsqrtps(<8 x float> %a0, <8 x float> *%a1) { ; ; SKYLAKE-LABEL: test_rsqrtps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vrsqrtps (%rdi), %ymm1 # sched: [11:2.00] -; SKYLAKE-NEXT: vrsqrtps %ymm0, %ymm0 # sched: [11:2.00] -; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vrsqrtps %ymm0, %ymm0 # sched: [4:1.00] +; SKYLAKE-NEXT: vrsqrtps (%rdi), %ymm1 # sched: [4:1.00] +; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_rsqrtps: @@ -3202,7 +3202,7 @@ define <4 x double> @test_shufpd(<4 x double> %a0, <4 x double> %a1, <4 x double ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[1],ymm1[0],ymm0[2],ymm1[3] sched: [1:1.00] ; SKYLAKE-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[1],mem[0],ymm1[2],mem[3] sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_shufpd: @@ -3291,9 +3291,9 @@ define <4 x double> @test_sqrtpd(<4 x double> %a0, <4 x double> *%a1) { ; ; SKYLAKE-LABEL: test_sqrtpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vsqrtpd (%rdi), %ymm1 # sched: [35:2.00] -; SKYLAKE-NEXT: vsqrtpd %ymm0, %ymm0 # sched: [35:2.00] -; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vsqrtpd %ymm0, %ymm0 # sched: [18:1.00] +; SKYLAKE-NEXT: vsqrtpd (%rdi), %ymm1 # sched: [18:1.00] +; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_sqrtpd: @@ -3341,9 +3341,9 @@ define <8 x float> @test_sqrtps(<8 x float> %a0, <8 x float> *%a1) { ; ; SKYLAKE-LABEL: test_sqrtps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vsqrtps (%rdi), %ymm1 # sched: [21:2.00] -; SKYLAKE-NEXT: vsqrtps %ymm0, %ymm0 # sched: [21:2.00] -; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vsqrtps %ymm0, %ymm0 # sched: [12:1.00] +; SKYLAKE-NEXT: vsqrtps (%rdi), %ymm1 # sched: [12:1.00] +; SKYLAKE-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_sqrtps: @@ -3388,8 +3388,8 @@ define <4 x double> @test_subpd(<4 x double> %a0, <4 x double> %a1, <4 x double> ; ; SKYLAKE-LABEL: test_subpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vsubpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vsubpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vsubpd %ymm1, %ymm0, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vsubpd (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_subpd: @@ -3430,8 +3430,8 @@ define <8 x float> @test_subps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a ; ; SKYLAKE-LABEL: test_subps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vsubps %ymm1, %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vsubps (%rdi), %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vsubps %ymm1, %ymm0, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vsubps (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_subps: @@ -3482,10 +3482,10 @@ define i32 @test_testpd(<2 x double> %a0, <2 x double> %a1, <2 x double> *%a2) { ; SKYLAKE-LABEL: test_testpd: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: xorl %eax, %eax # sched: [1:0.25] -; SKYLAKE-NEXT: vtestpd %xmm1, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: setb %al # sched: [1:0.50] -; SKYLAKE-NEXT: vtestpd (%rdi), %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: adcl $0, %eax # sched: [2:0.50] +; SKYLAKE-NEXT: vtestpd %xmm1, %xmm0 # sched: [2:1.00] +; SKYLAKE-NEXT: setb %al # sched: [1:1.00] +; SKYLAKE-NEXT: vtestpd (%rdi), %xmm0 # sched: [2:1.00] +; SKYLAKE-NEXT: adcl $0, %eax # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_testpd: @@ -3547,10 +3547,10 @@ define i32 @test_testpd_ymm(<4 x double> %a0, <4 x double> %a1, <4 x double> *%a ; SKYLAKE-LABEL: test_testpd_ymm: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: xorl %eax, %eax # sched: [1:0.25] -; SKYLAKE-NEXT: vtestpd %ymm1, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: setb %al # sched: [1:0.50] -; SKYLAKE-NEXT: vtestpd (%rdi), %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: adcl $0, %eax # sched: [2:0.50] +; SKYLAKE-NEXT: vtestpd %ymm1, %ymm0 # sched: [2:1.00] +; SKYLAKE-NEXT: setb %al # sched: [1:1.00] +; SKYLAKE-NEXT: vtestpd (%rdi), %ymm0 # sched: [2:1.00] +; SKYLAKE-NEXT: adcl $0, %eax # sched: [1:1.00] ; SKYLAKE-NEXT: vzeroupper # sched: [4:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -3611,10 +3611,10 @@ define i32 @test_testps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) { ; SKYLAKE-LABEL: test_testps: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: xorl %eax, %eax # sched: [1:0.25] -; SKYLAKE-NEXT: vtestps %xmm1, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: setb %al # sched: [1:0.50] -; SKYLAKE-NEXT: vtestps (%rdi), %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: adcl $0, %eax # sched: [2:0.50] +; SKYLAKE-NEXT: vtestps %xmm1, %xmm0 # sched: [2:1.00] +; SKYLAKE-NEXT: setb %al # sched: [1:1.00] +; SKYLAKE-NEXT: vtestps (%rdi), %xmm0 # sched: [2:1.00] +; SKYLAKE-NEXT: adcl $0, %eax # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_testps: @@ -3676,10 +3676,10 @@ define i32 @test_testps_ymm(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a2) ; SKYLAKE-LABEL: test_testps_ymm: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: xorl %eax, %eax # sched: [1:0.25] -; SKYLAKE-NEXT: vtestps %ymm1, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: setb %al # sched: [1:0.50] -; SKYLAKE-NEXT: vtestps (%rdi), %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: adcl $0, %eax # sched: [2:0.50] +; SKYLAKE-NEXT: vtestps %ymm1, %ymm0 # sched: [2:1.00] +; SKYLAKE-NEXT: setb %al # sched: [1:1.00] +; SKYLAKE-NEXT: vtestps (%rdi), %ymm0 # sched: [2:1.00] +; SKYLAKE-NEXT: adcl $0, %eax # sched: [1:1.00] ; SKYLAKE-NEXT: vzeroupper # sched: [4:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -3735,7 +3735,7 @@ define <4 x double> @test_unpckhpd(<4 x double> %a0, <4 x double> %a1, <4 x doub ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3] sched: [1:1.00] ; SKYLAKE-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],mem[1],ymm1[3],mem[3] sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_unpckhpd: @@ -3826,7 +3826,7 @@ define <4 x double> @test_unpcklpd(<4 x double> %a0, <4 x double> %a1, <4 x doub ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] sched: [1:1.00] ; SKYLAKE-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],mem[0],ymm1[2],mem[2] sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_unpcklpd: @@ -3915,9 +3915,9 @@ define <4 x double> @test_xorpd(<4 x double> %a0, <4 x double> %a1, <4 x double> ; ; SKYLAKE-LABEL: test_xorpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vxorpd %ymm1, %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vxorpd (%rdi), %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vxorpd %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vxorpd (%rdi), %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vaddpd %ymm0, %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_xorpd: @@ -3968,9 +3968,9 @@ define <8 x float> @test_xorps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a ; ; SKYLAKE-LABEL: test_xorps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vxorps %ymm1, %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vxorps (%rdi), %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vxorps %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vxorps (%rdi), %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_xorps: @@ -4015,7 +4015,7 @@ define void @test_zeroall() { ; ; SKYLAKE-LABEL: test_zeroall: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vzeroall # sched: [16:16.00] +; SKYLAKE-NEXT: vzeroall # sched: [16:4.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_zeroall: diff --git a/llvm/test/CodeGen/X86/avx2-schedule.ll b/llvm/test/CodeGen/X86/avx2-schedule.ll index f579b372a25..a35aec79061 100644 --- a/llvm/test/CodeGen/X86/avx2-schedule.ll +++ b/llvm/test/CodeGen/X86/avx2-schedule.ll @@ -50,7 +50,7 @@ define <4 x double> @test_broadcastsd_ymm(<2 x double> %a0) { ; SKYLAKE-LABEL: test_broadcastsd_ymm: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vbroadcastsd %xmm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddpd %ymm0, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %ymm0, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_broadcastsd_ymm: @@ -79,7 +79,7 @@ define <4 x float> @test_broadcastss(<4 x float> %a0) { ; SKYLAKE-LABEL: test_broadcastss: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vbroadcastss %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vaddps %xmm0, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %xmm0, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_broadcastss: @@ -108,7 +108,7 @@ define <8 x float> @test_broadcastss_ymm(<4 x float> %a0) { ; SKYLAKE-LABEL: test_broadcastss_ymm: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vbroadcastss %xmm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddps %ymm0, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %ymm0, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_broadcastss_ymm: @@ -178,7 +178,7 @@ define <2 x double> @test_gatherdpd(<2 x double> %a0, i8* %a1, <4 x i32> %a2, <2 ; ; SKYLAKE-LABEL: test_gatherdpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vgatherdpd %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [1:?] +; SKYLAKE-NEXT: vgatherdpd %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [17:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_gatherdpd: @@ -203,7 +203,7 @@ define <4 x double> @test_gatherdpd_ymm(<4 x double> %a0, i8* %a1, <4 x i32> %a2 ; ; SKYLAKE-LABEL: test_gatherdpd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vgatherdpd %ymm2, (%rdi,%xmm1,8), %ymm0 # sched: [1:?] +; SKYLAKE-NEXT: vgatherdpd %ymm2, (%rdi,%xmm1,8), %ymm0 # sched: [20:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_gatherdpd_ymm: @@ -228,7 +228,7 @@ define <4 x float> @test_gatherdps(<4 x float> %a0, i8* %a1, <4 x i32> %a2, <4 x ; ; SKYLAKE-LABEL: test_gatherdps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vgatherdps %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [1:?] +; SKYLAKE-NEXT: vgatherdps %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [17:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_gatherdps: @@ -253,7 +253,7 @@ define <8 x float> @test_gatherdps_ymm(<8 x float> %a0, i8* %a1, <8 x i32> %a2, ; ; SKYLAKE-LABEL: test_gatherdps_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vgatherdps %ymm2, (%rdi,%ymm1,4), %ymm0 # sched: [1:?] +; SKYLAKE-NEXT: vgatherdps %ymm2, (%rdi,%ymm1,4), %ymm0 # sched: [20:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_gatherdps_ymm: @@ -278,7 +278,7 @@ define <2 x double> @test_gatherqpd(<2 x double> %a0, i8* %a1, <2 x i64> %a2, <2 ; ; SKYLAKE-LABEL: test_gatherqpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vgatherqpd %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [1:?] +; SKYLAKE-NEXT: vgatherqpd %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [17:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_gatherqpd: @@ -303,7 +303,7 @@ define <4 x double> @test_gatherqpd_ymm(<4 x double> %a0, i8* %a1, <4 x i64> %a2 ; ; SKYLAKE-LABEL: test_gatherqpd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vgatherqpd %ymm2, (%rdi,%ymm1,8), %ymm0 # sched: [1:?] +; SKYLAKE-NEXT: vgatherqpd %ymm2, (%rdi,%ymm1,8), %ymm0 # sched: [20:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_gatherqpd_ymm: @@ -328,7 +328,7 @@ define <4 x float> @test_gatherqps(<4 x float> %a0, i8* %a1, <2 x i64> %a2, <4 x ; ; SKYLAKE-LABEL: test_gatherqps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vgatherqps %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [1:?] +; SKYLAKE-NEXT: vgatherqps %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [17:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_gatherqps: @@ -355,7 +355,7 @@ define <4 x float> @test_gatherqps_ymm(<4 x float> %a0, i8* %a1, <4 x i64> %a2, ; ; SKYLAKE-LABEL: test_gatherqps_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vgatherqps %xmm2, (%rdi,%ymm1,4), %xmm0 # sched: [1:?] +; SKYLAKE-NEXT: vgatherqps %xmm2, (%rdi,%ymm1,4), %xmm0 # sched: [20:1.00] ; SKYLAKE-NEXT: vzeroupper # sched: [4:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -446,8 +446,8 @@ define <16 x i16> @test_mpsadbw(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_mpsadbw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmpsadbw $7, %ymm1, %ymm0, %ymm0 # sched: [7:2.00] -; SKYLAKE-NEXT: vmpsadbw $7, (%rdi), %ymm0, %ymm0 # sched: [7:2.00] +; SKYLAKE-NEXT: vmpsadbw $7, %ymm1, %ymm0, %ymm0 # sched: [4:2.00] +; SKYLAKE-NEXT: vmpsadbw $7, (%rdi), %ymm0, %ymm0 # sched: [4:2.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_mpsadbw: @@ -480,9 +480,9 @@ define <32 x i8> @test_pabsb(<32 x i8> %a0, <32 x i8> *%a1) { ; ; SKYLAKE-LABEL: test_pabsb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpabsb %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpabsb %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpabsb (%rdi), %ymm1 # sched: [1:0.50] -; SKYLAKE-NEXT: vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pabsb: @@ -516,9 +516,9 @@ define <8 x i32> @test_pabsd(<8 x i32> %a0, <8 x i32> *%a1) { ; ; SKYLAKE-LABEL: test_pabsd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpabsd %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpabsd %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpabsd (%rdi), %ymm1 # sched: [1:0.50] -; SKYLAKE-NEXT: vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pabsd: @@ -552,9 +552,9 @@ define <16 x i16> @test_pabsw(<16 x i16> %a0, <16 x i16> *%a1) { ; ; SKYLAKE-LABEL: test_pabsw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpabsw %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpabsw %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpabsw (%rdi), %ymm1 # sched: [1:0.50] -; SKYLAKE-NEXT: vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pabsw: @@ -804,7 +804,7 @@ define <32 x i8> @test_paddsb(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_paddsb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpaddsb %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpaddsb %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpaddsb (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -835,7 +835,7 @@ define <16 x i16> @test_paddsw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_paddsw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpaddsw %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpaddsw %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpaddsw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -866,7 +866,7 @@ define <32 x i8> @test_paddusb(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_paddusb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpaddusb %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpaddusb %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpaddusb (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -897,7 +897,7 @@ define <16 x i16> @test_paddusw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_paddusw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpaddusw %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpaddusw %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpaddusw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -990,7 +990,7 @@ define <4 x i64> @test_pand(<4 x i64> %a0, <4 x i64> %a1, <4 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_pand: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpand %ymm1, %ymm0, %ymm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpand %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpand (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpaddq %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] @@ -1025,7 +1025,7 @@ define <4 x i64> @test_pandn(<4 x i64> %a0, <4 x i64> %a1, <4 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_pandn: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpandn %ymm1, %ymm0, %ymm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpandn %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpandn (%rdi), %ymm0, %ymm1 # sched: [1:0.50] ; SKYLAKE-NEXT: vpaddq %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] @@ -1060,7 +1060,7 @@ define <32 x i8> @test_pavgb(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pavgb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpavgb %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpavgb %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpavgb (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1100,7 +1100,7 @@ define <16 x i16> @test_pavgw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_pavgw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpavgw %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpavgw %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpavgw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1142,7 +1142,7 @@ define <4 x i32> @test_pblendd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pblendd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm0[3] sched: [1:0.33] +; SKYLAKE-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm0[3] sched: [1:0.50] ; SKYLAKE-NEXT: vpblendd {{.*#+}} xmm1 = mem[0],xmm1[1],mem[2],xmm1[3] sched: [1:0.50] ; SKYLAKE-NEXT: vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] @@ -1177,7 +1177,7 @@ define <8 x i32> @test_pblendd_ymm(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) ; ; SKYLAKE-LABEL: test_pblendd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] sched: [1:0.33] +; SKYLAKE-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] sched: [1:0.50] ; SKYLAKE-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],mem[1,2],ymm1[3,4,5,6,7] sched: [1:0.50] ; SKYLAKE-NEXT: vpaddd %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] @@ -1210,8 +1210,8 @@ define <32 x i8> @test_pblendvb(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> %a2, <32 ; ; SKYLAKE-LABEL: test_pblendvb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpblendvb %ymm2, %ymm1, %ymm0, %ymm0 # sched: [2:2.00] -; SKYLAKE-NEXT: vpblendvb %ymm3, (%rdi), %ymm0, %ymm0 # sched: [2:2.00] +; SKYLAKE-NEXT: vpblendvb %ymm2, %ymm1, %ymm0, %ymm0 # sched: [2:0.67] +; SKYLAKE-NEXT: vpblendvb %ymm3, (%rdi), %ymm0, %ymm0 # sched: [2:0.67] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pblendvb: @@ -1242,7 +1242,7 @@ define <16 x i16> @test_pblendw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; SKYLAKE-LABEL: test_pblendw: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7,8,9],ymm1[10,11,12],ymm0[13,14,15] sched: [1:1.00] -; SKYLAKE-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1],mem[2],ymm0[3],mem[4],ymm0[5],mem[6],ymm0[7],mem[8],ymm0[9],mem[10],ymm0[11],mem[12],ymm0[13],mem[14],ymm0[15] sched: [4:1.00] +; SKYLAKE-NEXT: vpblendw {{.*#+}} ymm0 = mem[0],ymm0[1],mem[2],ymm0[3],mem[4],ymm0[5],mem[6],ymm0[7],mem[8],ymm0[9],mem[10],ymm0[11],mem[12],ymm0[13],mem[14],ymm0[15] sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pblendw: @@ -1274,7 +1274,7 @@ define <16 x i8> @test_pbroadcastb(<16 x i8> %a0, <16 x i8> *%a1) { ; SKYLAKE-LABEL: test_pbroadcastb: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpbroadcastb %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vpbroadcastb (%rdi), %xmm1 # sched: [4:1.00] +; SKYLAKE-NEXT: vpbroadcastb (%rdi), %xmm1 # sched: [1:1.00] ; SKYLAKE-NEXT: vpaddb %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1309,7 +1309,7 @@ define <32 x i8> @test_pbroadcastb_ymm(<32 x i8> %a0, <32 x i8> *%a1) { ; SKYLAKE-LABEL: test_pbroadcastb_ymm: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpbroadcastb %xmm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vpbroadcastb (%rdi), %ymm1 # sched: [4:1.00] +; SKYLAKE-NEXT: vpbroadcastb (%rdi), %ymm1 # sched: [1:1.00] ; SKYLAKE-NEXT: vpaddb %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1484,7 +1484,7 @@ define <8 x i16> @test_pbroadcastw(<8 x i16> %a0, <8 x i16> *%a1) { ; SKYLAKE-LABEL: test_pbroadcastw: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpbroadcastw %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vpbroadcastw (%rdi), %xmm1 # sched: [4:1.00] +; SKYLAKE-NEXT: vpbroadcastw (%rdi), %xmm1 # sched: [1:1.00] ; SKYLAKE-NEXT: vpaddw %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1519,7 +1519,7 @@ define <16 x i16> @test_pbroadcastw_ymm(<16 x i16> %a0, <16 x i16> *%a1) { ; SKYLAKE-LABEL: test_pbroadcastw_ymm: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpbroadcastw %xmm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vpbroadcastw (%rdi), %ymm1 # sched: [4:1.00] +; SKYLAKE-NEXT: vpbroadcastw (%rdi), %ymm1 # sched: [1:1.00] ; SKYLAKE-NEXT: vpaddw %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1551,7 +1551,7 @@ define <32 x i8> @test_pcmpeqb(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpeqb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpcmpeqb (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1583,7 +1583,7 @@ define <8 x i32> @test_pcmpeqd(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpeqd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpeqd %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpeqd %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpcmpeqd (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1615,7 +1615,7 @@ define <4 x i64> @test_pcmpeqq(<4 x i64> %a0, <4 x i64> %a1, <4 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpeqq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpeqq %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpeqq %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpcmpeqq (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1647,7 +1647,7 @@ define <16 x i16> @test_pcmpeqw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_pcmpeqw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpcmpeqw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1679,7 +1679,7 @@ define <32 x i8> @test_pcmpgtb(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpgtb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpgtb %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpgtb %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpcmpgtb (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1711,7 +1711,7 @@ define <8 x i32> @test_pcmpgtd(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpgtd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpgtd %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpgtd %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpcmpgtd (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1743,8 +1743,8 @@ define <4 x i64> @test_pcmpgtq(<4 x i64> %a0, <4 x i64> %a1, <4 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpgtq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpcmpgtq (%rdi), %ymm0, %ymm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vpcmpgtq (%rdi), %ymm0, %ymm0 # sched: [3:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pcmpgtq: @@ -1775,7 +1775,7 @@ define <16 x i16> @test_pcmpgtw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_pcmpgtw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpcmpgtw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1882,7 +1882,7 @@ define <4 x double> @test_permpd(<4 x double> %a0, <4 x double> *%a1) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,2,2,3] sched: [3:1.00] ; SKYLAKE-NEXT: vpermpd {{.*#+}} ymm1 = mem[0,2,2,3] sched: [3:1.00] -; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %ymm1, %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_permpd: @@ -1917,7 +1917,7 @@ define <8 x float> @test_permps(<8 x i32> %a0, <8 x float> %a1, <8 x float> *%a2 ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpermps %ymm1, %ymm0, %ymm1 # sched: [3:1.00] ; SKYLAKE-NEXT: vpermps (%rdi), %ymm0, %ymm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_permps: @@ -1982,7 +1982,7 @@ define <4 x i32> @test_pgatherdd(<4 x i32> %a0, i8* %a1, <4 x i32> %a2, <4 x i32 ; ; SKYLAKE-LABEL: test_pgatherdd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpgatherdd %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [1:?] +; SKYLAKE-NEXT: vpgatherdd %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [17:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pgatherdd: @@ -2007,7 +2007,7 @@ define <8 x i32> @test_pgatherdd_ymm(<8 x i32> %a0, i8* %a1, <8 x i32> %a2, <8 x ; ; SKYLAKE-LABEL: test_pgatherdd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpgatherdd %ymm2, (%rdi,%ymm1,2), %ymm0 # sched: [1:?] +; SKYLAKE-NEXT: vpgatherdd %ymm2, (%rdi,%ymm1,2), %ymm0 # sched: [20:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pgatherdd_ymm: @@ -2032,7 +2032,7 @@ define <2 x i64> @test_pgatherdq(<2 x i64> %a0, i8* %a1, <4 x i32> %a2, <2 x i64 ; ; SKYLAKE-LABEL: test_pgatherdq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpgatherdq %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [1:?] +; SKYLAKE-NEXT: vpgatherdq %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [17:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pgatherdq: @@ -2057,7 +2057,7 @@ define <4 x i64> @test_pgatherdq_ymm(<4 x i64> %a0, i8* %a1, <4 x i32> %a2, <4 x ; ; SKYLAKE-LABEL: test_pgatherdq_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpgatherdq %ymm2, (%rdi,%xmm1,2), %ymm0 # sched: [1:?] +; SKYLAKE-NEXT: vpgatherdq %ymm2, (%rdi,%xmm1,2), %ymm0 # sched: [20:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pgatherdq_ymm: @@ -2082,7 +2082,7 @@ define <4 x i32> @test_pgatherqd(<4 x i32> %a0, i8* %a1, <2 x i64> %a2, <4 x i32 ; ; SKYLAKE-LABEL: test_pgatherqd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpgatherqd %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [1:?] +; SKYLAKE-NEXT: vpgatherqd %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [17:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pgatherqd: @@ -2109,7 +2109,7 @@ define <4 x i32> @test_pgatherqd_ymm(<4 x i32> %a0, i8* %a1, <4 x i64> %a2, <4 x ; ; SKYLAKE-LABEL: test_pgatherqd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpgatherqd %xmm2, (%rdi,%ymm1,2), %xmm0 # sched: [1:?] +; SKYLAKE-NEXT: vpgatherqd %xmm2, (%rdi,%ymm1,2), %xmm0 # sched: [20:1.00] ; SKYLAKE-NEXT: vzeroupper # sched: [4:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2136,7 +2136,7 @@ define <2 x i64> @test_pgatherqq(<2 x i64> %a0, i8 *%a1, <2 x i64> %a2, <2 x i64 ; ; SKYLAKE-LABEL: test_pgatherqq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpgatherqq %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [1:?] +; SKYLAKE-NEXT: vpgatherqq %xmm2, (%rdi,%xmm1,2), %xmm0 # sched: [17:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pgatherqq: @@ -2161,7 +2161,7 @@ define <4 x i64> @test_pgatherqq_ymm(<4 x i64> %a0, i8 *%a1, <4 x i64> %a2, <4 x ; ; SKYLAKE-LABEL: test_pgatherqq_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpgatherqq %ymm2, (%rdi,%ymm1,2), %ymm0 # sched: [1:?] +; SKYLAKE-NEXT: vpgatherqq %ymm2, (%rdi,%ymm1,2), %ymm0 # sched: [20:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pgatherqq_ymm: @@ -2374,8 +2374,8 @@ define <16 x i16> @test_pmaddubsw(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) ; ; SKYLAKE-LABEL: test_pmaddubsw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaddubsw %ymm1, %ymm0, %ymm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmaddubsw (%rdi), %ymm0, %ymm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmaddubsw %ymm1, %ymm0, %ymm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmaddubsw (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pmaddubsw: @@ -2406,8 +2406,8 @@ define <8 x i32> @test_pmaddwd(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_pmaddwd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaddwd %ymm1, %ymm0, %ymm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmaddwd (%rdi), %ymm0, %ymm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmaddwd %ymm1, %ymm0, %ymm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmaddwd (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pmaddwd: @@ -2440,8 +2440,8 @@ define <4 x i32> @test_pmaskmovd(i8* %a0, <4 x i32> %a1, <4 x i32> %a2) { ; ; SKYLAKE-LABEL: test_pmaskmovd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaskmovd (%rdi), %xmm0, %xmm2 # sched: [2:2.00] -; SKYLAKE-NEXT: vpmaskmovd %xmm1, %xmm0, (%rdi) # sched: [4:1.00] +; SKYLAKE-NEXT: vpmaskmovd (%rdi), %xmm0, %xmm2 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaskmovd %xmm1, %xmm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: vmovdqa %xmm2, %xmm0 # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2475,8 +2475,8 @@ define <8 x i32> @test_pmaskmovd_ymm(i8* %a0, <8 x i32> %a1, <8 x i32> %a2) { ; ; SKYLAKE-LABEL: test_pmaskmovd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaskmovd (%rdi), %ymm0, %ymm2 # sched: [2:2.00] -; SKYLAKE-NEXT: vpmaskmovd %ymm1, %ymm0, (%rdi) # sched: [4:1.00] +; SKYLAKE-NEXT: vpmaskmovd (%rdi), %ymm0, %ymm2 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaskmovd %ymm1, %ymm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: vmovdqa %ymm2, %ymm0 # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2510,8 +2510,8 @@ define <2 x i64> @test_pmaskmovq(i8* %a0, <2 x i64> %a1, <2 x i64> %a2) { ; ; SKYLAKE-LABEL: test_pmaskmovq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaskmovq (%rdi), %xmm0, %xmm2 # sched: [2:2.00] -; SKYLAKE-NEXT: vpmaskmovq %xmm1, %xmm0, (%rdi) # sched: [4:1.00] +; SKYLAKE-NEXT: vpmaskmovq (%rdi), %xmm0, %xmm2 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaskmovq %xmm1, %xmm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: vmovdqa %xmm2, %xmm0 # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2545,8 +2545,8 @@ define <4 x i64> @test_pmaskmovq_ymm(i8* %a0, <4 x i64> %a1, <4 x i64> %a2) { ; ; SKYLAKE-LABEL: test_pmaskmovq_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaskmovq (%rdi), %ymm0, %ymm2 # sched: [2:2.00] -; SKYLAKE-NEXT: vpmaskmovq %ymm1, %ymm0, (%rdi) # sched: [4:1.00] +; SKYLAKE-NEXT: vpmaskmovq (%rdi), %ymm0, %ymm2 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaskmovq %ymm1, %ymm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: vmovdqa %ymm2, %ymm0 # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2578,7 +2578,7 @@ define <32 x i8> @test_pmaxsb(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pmaxsb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaxsb %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaxsb %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpmaxsb (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2609,7 +2609,7 @@ define <8 x i32> @test_pmaxsd(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pmaxsd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpmaxsd (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2640,7 +2640,7 @@ define <16 x i16> @test_pmaxsw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_pmaxsw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpmaxsw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2671,7 +2671,7 @@ define <32 x i8> @test_pmaxub(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pmaxub: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaxub %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaxub %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpmaxub (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2702,7 +2702,7 @@ define <8 x i32> @test_pmaxud(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pmaxud: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaxud %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaxud %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpmaxud (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2733,7 +2733,7 @@ define <16 x i16> @test_pmaxuw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_pmaxuw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaxuw %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaxuw %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpmaxuw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2764,7 +2764,7 @@ define <32 x i8> @test_pminsb(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pminsb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpminsb %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpminsb %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpminsb (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2795,7 +2795,7 @@ define <8 x i32> @test_pminsd(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pminsd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpminsd %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpminsd %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpminsd (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2826,7 +2826,7 @@ define <16 x i16> @test_pminsw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_pminsw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpminsw %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpminsw %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpminsw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2857,7 +2857,7 @@ define <32 x i8> @test_pminub(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pminub: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpminub %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpminub %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpminub (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2888,7 +2888,7 @@ define <8 x i32> @test_pminud(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pminud: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpminud %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpminud %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpminud (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2919,7 +2919,7 @@ define <16 x i16> @test_pminuw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_pminuw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpminuw %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpminuw %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpminuw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2950,7 +2950,7 @@ define i32 @test_pmovmskb(<32 x i8> %a0) { ; ; SKYLAKE-LABEL: test_pmovmskb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmovmskb %ymm0, %eax # sched: [3:1.00] +; SKYLAKE-NEXT: vpmovmskb %ymm0, %eax # sched: [2:1.00] ; SKYLAKE-NEXT: vzeroupper # sched: [4:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -3411,8 +3411,8 @@ define <4 x i64> @test_pmuldq(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pmuldq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmuldq %ymm1, %ymm0, %ymm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmuldq (%rdi), %ymm0, %ymm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmuldq %ymm1, %ymm0, %ymm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmuldq (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pmuldq: @@ -3443,8 +3443,8 @@ define <16 x i16> @test_pmulhrsw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2 ; ; SKYLAKE-LABEL: test_pmulhrsw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmulhrsw %ymm1, %ymm0, %ymm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmulhrsw (%rdi), %ymm0, %ymm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmulhrsw %ymm1, %ymm0, %ymm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmulhrsw (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pmulhrsw: @@ -3474,8 +3474,8 @@ define <16 x i16> @test_pmulhuw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_pmulhuw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmulhuw %ymm1, %ymm0, %ymm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmulhuw (%rdi), %ymm0, %ymm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmulhuw %ymm1, %ymm0, %ymm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmulhuw (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pmulhuw: @@ -3505,8 +3505,8 @@ define <16 x i16> @test_pmulhw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_pmulhw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmulhw %ymm1, %ymm0, %ymm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmulhw (%rdi), %ymm0, %ymm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmulhw %ymm1, %ymm0, %ymm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmulhw (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pmulhw: @@ -3536,8 +3536,8 @@ define <8 x i32> @test_pmulld(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pmulld: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmulld %ymm1, %ymm0, %ymm0 # sched: [10:2.00] -; SKYLAKE-NEXT: vpmulld (%rdi), %ymm0, %ymm0 # sched: [10:2.00] +; SKYLAKE-NEXT: vpmulld %ymm1, %ymm0, %ymm0 # sched: [8:0.67] +; SKYLAKE-NEXT: vpmulld (%rdi), %ymm0, %ymm0 # sched: [8:0.67] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pmulld: @@ -3566,8 +3566,8 @@ define <16 x i16> @test_pmullw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_pmullw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmullw %ymm1, %ymm0, %ymm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmullw (%rdi), %ymm0, %ymm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmullw %ymm1, %ymm0, %ymm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmullw (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pmullw: @@ -3596,8 +3596,8 @@ define <4 x i64> @test_pmuludq(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pmuludq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmuludq %ymm1, %ymm0, %ymm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmuludq (%rdi), %ymm0, %ymm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmuludq %ymm1, %ymm0, %ymm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmuludq (%rdi), %ymm0, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pmuludq: @@ -3630,7 +3630,7 @@ define <4 x i64> @test_por(<4 x i64> %a0, <4 x i64> %a1, <4 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_por: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpor (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpaddq %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] @@ -3663,8 +3663,8 @@ define <4 x i64> @test_psadbw(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_psadbw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsadbw %ymm1, %ymm0, %ymm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpsadbw (%rdi), %ymm0, %ymm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpsadbw %ymm1, %ymm0, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vpsadbw (%rdi), %ymm0, %ymm0 # sched: [3:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_psadbw: @@ -3765,7 +3765,7 @@ define <16 x i16> @test_pshufhw(<16 x i16> %a0, <16 x i16> *%a1) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,7,6,5,4,8,9,10,11,15,14,13,12] sched: [1:1.00] ; SKYLAKE-NEXT: vpshufhw {{.*#+}} ymm1 = mem[0,1,2,3,5,4,7,6,8,9,10,11,13,12,15,14] sched: [1:1.00] -; SKYLAKE-NEXT: vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pshufhw: @@ -3800,7 +3800,7 @@ define <16 x i16> @test_pshuflw(<16 x i16> %a0, <16 x i16> *%a1) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[3,2,1,0,4,5,6,7,11,10,9,8,12,13,14,15] sched: [1:1.00] ; SKYLAKE-NEXT: vpshuflw {{.*#+}} ymm1 = mem[1,0,3,2,4,5,6,7,9,8,11,10,12,13,14,15] sched: [1:1.00] -; SKYLAKE-NEXT: vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_pshuflw: @@ -3831,7 +3831,7 @@ define <32 x i8> @test_psignb(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_psignb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsignb %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpsignb %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsignb (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -3862,7 +3862,7 @@ define <8 x i32> @test_psignd(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_psignd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsignd %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpsignd %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsignd (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -3893,7 +3893,7 @@ define <16 x i16> @test_psignw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_psignw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsignw %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpsignw %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsignw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -3927,7 +3927,7 @@ define <8 x i32> @test_pslld(<8 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; SKYLAKE-LABEL: test_pslld: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpslld %xmm1, %ymm0, %ymm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vpslld (%rdi), %ymm0, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpslld (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpslld $2, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -3987,7 +3987,7 @@ define <4 x i64> @test_psllq(<4 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; SKYLAKE-LABEL: test_psllq: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsllq %xmm1, %ymm0, %ymm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vpsllq (%rdi), %ymm0, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsllq (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpsllq $2, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4020,8 +4020,8 @@ define <4 x i32> @test_psllvd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_psllvd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsllvd %xmm1, %xmm0, %xmm0 # sched: [3:2.00] -; SKYLAKE-NEXT: vpsllvd (%rdi), %xmm0, %xmm0 # sched: [3:2.00] +; SKYLAKE-NEXT: vpsllvd %xmm1, %xmm0, %xmm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsllvd (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_psllvd: @@ -4051,8 +4051,8 @@ define <8 x i32> @test_psllvd_ymm(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) ; ; SKYLAKE-LABEL: test_psllvd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsllvd %ymm1, %ymm0, %ymm0 # sched: [3:2.00] -; SKYLAKE-NEXT: vpsllvd (%rdi), %ymm0, %ymm0 # sched: [3:2.00] +; SKYLAKE-NEXT: vpsllvd %ymm1, %ymm0, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsllvd (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_psllvd_ymm: @@ -4083,7 +4083,7 @@ define <2 x i64> @test_psllvq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; SKYLAKE-LABEL: test_psllvq: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsllvq %xmm1, %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vpsllvq (%rdi), %xmm0, %xmm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsllvq (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_psllvq: @@ -4114,7 +4114,7 @@ define <4 x i64> @test_psllvq_ymm(<4 x i64> %a0, <4 x i64> %a1, <4 x i64> *%a2) ; SKYLAKE-LABEL: test_psllvq_ymm: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsllvq %ymm1, %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vpsllvq (%rdi), %ymm0, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsllvq (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_psllvq_ymm: @@ -4147,7 +4147,7 @@ define <16 x i16> @test_psllw(<16 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; SKYLAKE-LABEL: test_psllw: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsllw %xmm1, %ymm0, %ymm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vpsllw (%rdi), %ymm0, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsllw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpsllw $2, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4183,7 +4183,7 @@ define <8 x i32> @test_psrad(<8 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; SKYLAKE-LABEL: test_psrad: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsrad %xmm1, %ymm0, %ymm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vpsrad (%rdi), %ymm0, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsrad (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpsrad $2, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4216,8 +4216,8 @@ define <4 x i32> @test_psravd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_psravd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsravd %xmm1, %xmm0, %xmm0 # sched: [3:2.00] -; SKYLAKE-NEXT: vpsravd (%rdi), %xmm0, %xmm0 # sched: [3:2.00] +; SKYLAKE-NEXT: vpsravd %xmm1, %xmm0, %xmm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsravd (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_psravd: @@ -4247,8 +4247,8 @@ define <8 x i32> @test_psravd_ymm(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) ; ; SKYLAKE-LABEL: test_psravd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsravd %ymm1, %ymm0, %ymm0 # sched: [3:2.00] -; SKYLAKE-NEXT: vpsravd (%rdi), %ymm0, %ymm0 # sched: [3:2.00] +; SKYLAKE-NEXT: vpsravd %ymm1, %ymm0, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsravd (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_psravd_ymm: @@ -4281,7 +4281,7 @@ define <16 x i16> @test_psraw(<16 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; SKYLAKE-LABEL: test_psraw: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsraw %xmm1, %ymm0, %ymm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vpsraw (%rdi), %ymm0, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsraw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpsraw $2, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4317,7 +4317,7 @@ define <8 x i32> @test_psrld(<8 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; SKYLAKE-LABEL: test_psrld: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsrld %xmm1, %ymm0, %ymm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vpsrld (%rdi), %ymm0, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsrld (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpsrld $2, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4377,7 +4377,7 @@ define <4 x i64> @test_psrlq(<4 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; SKYLAKE-LABEL: test_psrlq: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsrlq %xmm1, %ymm0, %ymm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vpsrlq (%rdi), %ymm0, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsrlq (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpsrlq $2, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4410,8 +4410,8 @@ define <4 x i32> @test_psrlvd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_psrlvd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsrlvd %xmm1, %xmm0, %xmm0 # sched: [3:2.00] -; SKYLAKE-NEXT: vpsrlvd (%rdi), %xmm0, %xmm0 # sched: [3:2.00] +; SKYLAKE-NEXT: vpsrlvd %xmm1, %xmm0, %xmm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsrlvd (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_psrlvd: @@ -4441,8 +4441,8 @@ define <8 x i32> @test_psrlvd_ymm(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) ; ; SKYLAKE-LABEL: test_psrlvd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsrlvd %ymm1, %ymm0, %ymm0 # sched: [3:2.00] -; SKYLAKE-NEXT: vpsrlvd (%rdi), %ymm0, %ymm0 # sched: [3:2.00] +; SKYLAKE-NEXT: vpsrlvd %ymm1, %ymm0, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsrlvd (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_psrlvd_ymm: @@ -4473,7 +4473,7 @@ define <2 x i64> @test_psrlvq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; SKYLAKE-LABEL: test_psrlvq: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsrlvq %xmm1, %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vpsrlvq (%rdi), %xmm0, %xmm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsrlvq (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_psrlvq: @@ -4504,7 +4504,7 @@ define <4 x i64> @test_psrlvq_ymm(<4 x i64> %a0, <4 x i64> %a1, <4 x i64> *%a2) ; SKYLAKE-LABEL: test_psrlvq_ymm: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsrlvq %ymm1, %ymm0, %ymm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vpsrlvq (%rdi), %ymm0, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsrlvq (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_psrlvq_ymm: @@ -4537,7 +4537,7 @@ define <16 x i16> @test_psrlw(<16 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; SKYLAKE-LABEL: test_psrlw: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsrlw %xmm1, %ymm0, %ymm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vpsrlw (%rdi), %ymm0, %ymm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vpsrlw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpsrlw $2, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4660,7 +4660,7 @@ define <32 x i8> @test_psubsb(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_psubsb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsubsb %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpsubsb %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsubsb (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4691,7 +4691,7 @@ define <16 x i16> @test_psubsw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_psubsw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsubsw %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpsubsw %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsubsw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4722,7 +4722,7 @@ define <32 x i8> @test_psubusb(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_psubusb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsubusb %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpsubusb %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsubusb (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4753,7 +4753,7 @@ define <16 x i16> @test_psubusw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) ; ; SKYLAKE-LABEL: test_psubusw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsubusw %ymm1, %ymm0, %ymm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpsubusw %ymm1, %ymm0, %ymm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsubusw (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4850,7 +4850,7 @@ define <8 x i32> @test_punpckhdq(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpunpckhdq {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] sched: [1:1.00] ; SKYLAKE-NEXT: vpunpckhdq {{.*#+}} ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7] sched: [1:1.00] -; SKYLAKE-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsubd %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4984,7 +4984,7 @@ define <8 x i32> @test_punpckldq(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5] sched: [1:1.00] ; SKYLAKE-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[4],mem[4],ymm0[5],mem[5] sched: [1:1.00] -; SKYLAKE-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsubd %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -5084,7 +5084,7 @@ define <4 x i64> @test_pxor(<4 x i64> %a0, <4 x i64> %a1, <4 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_pxor: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpxor %ymm1, %ymm0, %ymm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpxor %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpxor (%rdi), %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpaddq %ymm1, %ymm0, %ymm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] diff --git a/llvm/test/CodeGen/X86/bmi2-schedule.ll b/llvm/test/CodeGen/X86/bmi2-schedule.ll index 22c8b154732..172a460b9d0 100644 --- a/llvm/test/CodeGen/X86/bmi2-schedule.ll +++ b/llvm/test/CodeGen/X86/bmi2-schedule.ll @@ -230,12 +230,26 @@ define i32 @test_rorx_i32(i32 %a0, i32 %a1, i32 *%a2) { ; GENERIC-NEXT: addl %ecx, %eax # sched: [1:0.33] ; GENERIC-NEXT: retq # sched: [1:1.00] ; -; COMMON-LABEL: test_rorx_i32: -; COMMON: # BB#0: -; COMMON-NEXT: rorxl $5, %edi, %ecx # sched: [1:0.50] -; COMMON-NEXT: rorxl $5, (%rdx), %eax # sched: [1:0.50] -; COMMON-NEXT: addl %ecx, %eax # sched: [1:0.25] -; COMMON-NEXT: retq # sched: [2:1.00] +; HASWELL-LABEL: test_rorx_i32: +; HASWELL: # BB#0: +; HASWELL-NEXT: rorxl $5, %edi, %ecx # sched: [1:0.50] +; HASWELL-NEXT: rorxl $5, (%rdx), %eax # sched: [1:0.50] +; HASWELL-NEXT: addl %ecx, %eax # sched: [1:0.25] +; HASWELL-NEXT: retq # sched: [2:1.00] +; +; SKYLAKE-LABEL: test_rorx_i32: +; SKYLAKE: # BB#0: +; SKYLAKE-NEXT: rorxl $5, %edi, %ecx # sched: [1:1.00] +; SKYLAKE-NEXT: rorxl $5, (%rdx), %eax # sched: [1:0.50] +; SKYLAKE-NEXT: addl %ecx, %eax # sched: [1:0.25] +; SKYLAKE-NEXT: retq # sched: [2:1.00] +; +; KNL-LABEL: test_rorx_i32: +; KNL: # BB#0: +; KNL-NEXT: rorxl $5, %edi, %ecx # sched: [1:0.50] +; KNL-NEXT: rorxl $5, (%rdx), %eax # sched: [1:0.50] +; KNL-NEXT: addl %ecx, %eax # sched: [1:0.25] +; KNL-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_rorx_i32: ; ZNVER1: # BB#0: @@ -262,12 +276,26 @@ define i64 @test_rorx_i64(i64 %a0, i64 %a1, i64 *%a2) { ; GENERIC-NEXT: addq %rcx, %rax # sched: [1:0.33] ; GENERIC-NEXT: retq # sched: [1:1.00] ; -; COMMON-LABEL: test_rorx_i64: -; COMMON: # BB#0: -; COMMON-NEXT: rorxq $5, %rdi, %rcx # sched: [1:0.50] -; COMMON-NEXT: rorxq $5, (%rdx), %rax # sched: [1:0.50] -; COMMON-NEXT: addq %rcx, %rax # sched: [1:0.25] -; COMMON-NEXT: retq # sched: [2:1.00] +; HASWELL-LABEL: test_rorx_i64: +; HASWELL: # BB#0: +; HASWELL-NEXT: rorxq $5, %rdi, %rcx # sched: [1:0.50] +; HASWELL-NEXT: rorxq $5, (%rdx), %rax # sched: [1:0.50] +; HASWELL-NEXT: addq %rcx, %rax # sched: [1:0.25] +; HASWELL-NEXT: retq # sched: [2:1.00] +; +; SKYLAKE-LABEL: test_rorx_i64: +; SKYLAKE: # BB#0: +; SKYLAKE-NEXT: rorxq $5, %rdi, %rcx # sched: [1:1.00] +; SKYLAKE-NEXT: rorxq $5, (%rdx), %rax # sched: [1:0.50] +; SKYLAKE-NEXT: addq %rcx, %rax # sched: [1:0.25] +; SKYLAKE-NEXT: retq # sched: [2:1.00] +; +; KNL-LABEL: test_rorx_i64: +; KNL: # BB#0: +; KNL-NEXT: rorxq $5, %rdi, %rcx # sched: [1:0.50] +; KNL-NEXT: rorxq $5, (%rdx), %rax # sched: [1:0.50] +; KNL-NEXT: addq %rcx, %rax # sched: [1:0.25] +; KNL-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_rorx_i64: ; ZNVER1: # BB#0: @@ -294,12 +322,26 @@ define i32 @test_sarx_i32(i32 %a0, i32 %a1, i32 *%a2) { ; GENERIC-NEXT: addl %ecx, %eax # sched: [1:0.33] ; GENERIC-NEXT: retq # sched: [1:1.00] ; -; COMMON-LABEL: test_sarx_i32: -; COMMON: # BB#0: -; COMMON-NEXT: sarxl %esi, %edi, %ecx # sched: [1:0.50] -; COMMON-NEXT: sarxl %esi, (%rdx), %eax # sched: [1:0.50] -; COMMON-NEXT: addl %ecx, %eax # sched: [1:0.25] -; COMMON-NEXT: retq # sched: [2:1.00] +; HASWELL-LABEL: test_sarx_i32: +; HASWELL: # BB#0: +; HASWELL-NEXT: sarxl %esi, %edi, %ecx # sched: [1:0.50] +; HASWELL-NEXT: sarxl %esi, (%rdx), %eax # sched: [1:0.50] +; HASWELL-NEXT: addl %ecx, %eax # sched: [1:0.25] +; HASWELL-NEXT: retq # sched: [2:1.00] +; +; SKYLAKE-LABEL: test_sarx_i32: +; SKYLAKE: # BB#0: +; SKYLAKE-NEXT: sarxl %esi, %edi, %ecx # sched: [1:1.00] +; SKYLAKE-NEXT: sarxl %esi, (%rdx), %eax # sched: [1:0.50] +; SKYLAKE-NEXT: addl %ecx, %eax # sched: [1:0.25] +; SKYLAKE-NEXT: retq # sched: [2:1.00] +; +; KNL-LABEL: test_sarx_i32: +; KNL: # BB#0: +; KNL-NEXT: sarxl %esi, %edi, %ecx # sched: [1:0.50] +; KNL-NEXT: sarxl %esi, (%rdx), %eax # sched: [1:0.50] +; KNL-NEXT: addl %ecx, %eax # sched: [1:0.25] +; KNL-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_sarx_i32: ; ZNVER1: # BB#0: @@ -322,12 +364,26 @@ define i64 @test_sarx_i64(i64 %a0, i64 %a1, i64 *%a2) { ; GENERIC-NEXT: addq %rcx, %rax # sched: [1:0.33] ; GENERIC-NEXT: retq # sched: [1:1.00] ; -; COMMON-LABEL: test_sarx_i64: -; COMMON: # BB#0: -; COMMON-NEXT: sarxq %rsi, %rdi, %rcx # sched: [1:0.50] -; COMMON-NEXT: sarxq %rsi, (%rdx), %rax # sched: [1:0.50] -; COMMON-NEXT: addq %rcx, %rax # sched: [1:0.25] -; COMMON-NEXT: retq # sched: [2:1.00] +; HASWELL-LABEL: test_sarx_i64: +; HASWELL: # BB#0: +; HASWELL-NEXT: sarxq %rsi, %rdi, %rcx # sched: [1:0.50] +; HASWELL-NEXT: sarxq %rsi, (%rdx), %rax # sched: [1:0.50] +; HASWELL-NEXT: addq %rcx, %rax # sched: [1:0.25] +; HASWELL-NEXT: retq # sched: [2:1.00] +; +; SKYLAKE-LABEL: test_sarx_i64: +; SKYLAKE: # BB#0: +; SKYLAKE-NEXT: sarxq %rsi, %rdi, %rcx # sched: [1:1.00] +; SKYLAKE-NEXT: sarxq %rsi, (%rdx), %rax # sched: [1:0.50] +; SKYLAKE-NEXT: addq %rcx, %rax # sched: [1:0.25] +; SKYLAKE-NEXT: retq # sched: [2:1.00] +; +; KNL-LABEL: test_sarx_i64: +; KNL: # BB#0: +; KNL-NEXT: sarxq %rsi, %rdi, %rcx # sched: [1:0.50] +; KNL-NEXT: sarxq %rsi, (%rdx), %rax # sched: [1:0.50] +; KNL-NEXT: addq %rcx, %rax # sched: [1:0.25] +; KNL-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_sarx_i64: ; ZNVER1: # BB#0: @@ -350,12 +406,26 @@ define i32 @test_shlx_i32(i32 %a0, i32 %a1, i32 *%a2) { ; GENERIC-NEXT: addl %ecx, %eax # sched: [1:0.33] ; GENERIC-NEXT: retq # sched: [1:1.00] ; -; COMMON-LABEL: test_shlx_i32: -; COMMON: # BB#0: -; COMMON-NEXT: shlxl %esi, %edi, %ecx # sched: [1:0.50] -; COMMON-NEXT: shlxl %esi, (%rdx), %eax # sched: [1:0.50] -; COMMON-NEXT: addl %ecx, %eax # sched: [1:0.25] -; COMMON-NEXT: retq # sched: [2:1.00] +; HASWELL-LABEL: test_shlx_i32: +; HASWELL: # BB#0: +; HASWELL-NEXT: shlxl %esi, %edi, %ecx # sched: [1:0.50] +; HASWELL-NEXT: shlxl %esi, (%rdx), %eax # sched: [1:0.50] +; HASWELL-NEXT: addl %ecx, %eax # sched: [1:0.25] +; HASWELL-NEXT: retq # sched: [2:1.00] +; +; SKYLAKE-LABEL: test_shlx_i32: +; SKYLAKE: # BB#0: +; SKYLAKE-NEXT: shlxl %esi, %edi, %ecx # sched: [1:1.00] +; SKYLAKE-NEXT: shlxl %esi, (%rdx), %eax # sched: [1:0.50] +; SKYLAKE-NEXT: addl %ecx, %eax # sched: [1:0.25] +; SKYLAKE-NEXT: retq # sched: [2:1.00] +; +; KNL-LABEL: test_shlx_i32: +; KNL: # BB#0: +; KNL-NEXT: shlxl %esi, %edi, %ecx # sched: [1:0.50] +; KNL-NEXT: shlxl %esi, (%rdx), %eax # sched: [1:0.50] +; KNL-NEXT: addl %ecx, %eax # sched: [1:0.25] +; KNL-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_shlx_i32: ; ZNVER1: # BB#0: @@ -378,12 +448,26 @@ define i64 @test_shlx_i64(i64 %a0, i64 %a1, i64 *%a2) { ; GENERIC-NEXT: addq %rcx, %rax # sched: [1:0.33] ; GENERIC-NEXT: retq # sched: [1:1.00] ; -; COMMON-LABEL: test_shlx_i64: -; COMMON: # BB#0: -; COMMON-NEXT: shlxq %rsi, %rdi, %rcx # sched: [1:0.50] -; COMMON-NEXT: shlxq %rsi, (%rdx), %rax # sched: [1:0.50] -; COMMON-NEXT: addq %rcx, %rax # sched: [1:0.25] -; COMMON-NEXT: retq # sched: [2:1.00] +; HASWELL-LABEL: test_shlx_i64: +; HASWELL: # BB#0: +; HASWELL-NEXT: shlxq %rsi, %rdi, %rcx # sched: [1:0.50] +; HASWELL-NEXT: shlxq %rsi, (%rdx), %rax # sched: [1:0.50] +; HASWELL-NEXT: addq %rcx, %rax # sched: [1:0.25] +; HASWELL-NEXT: retq # sched: [2:1.00] +; +; SKYLAKE-LABEL: test_shlx_i64: +; SKYLAKE: # BB#0: +; SKYLAKE-NEXT: shlxq %rsi, %rdi, %rcx # sched: [1:1.00] +; SKYLAKE-NEXT: shlxq %rsi, (%rdx), %rax # sched: [1:0.50] +; SKYLAKE-NEXT: addq %rcx, %rax # sched: [1:0.25] +; SKYLAKE-NEXT: retq # sched: [2:1.00] +; +; KNL-LABEL: test_shlx_i64: +; KNL: # BB#0: +; KNL-NEXT: shlxq %rsi, %rdi, %rcx # sched: [1:0.50] +; KNL-NEXT: shlxq %rsi, (%rdx), %rax # sched: [1:0.50] +; KNL-NEXT: addq %rcx, %rax # sched: [1:0.25] +; KNL-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_shlx_i64: ; ZNVER1: # BB#0: @@ -406,12 +490,26 @@ define i32 @test_shrx_i32(i32 %a0, i32 %a1, i32 *%a2) { ; GENERIC-NEXT: addl %ecx, %eax # sched: [1:0.33] ; GENERIC-NEXT: retq # sched: [1:1.00] ; -; COMMON-LABEL: test_shrx_i32: -; COMMON: # BB#0: -; COMMON-NEXT: shrxl %esi, %edi, %ecx # sched: [1:0.50] -; COMMON-NEXT: shrxl %esi, (%rdx), %eax # sched: [1:0.50] -; COMMON-NEXT: addl %ecx, %eax # sched: [1:0.25] -; COMMON-NEXT: retq # sched: [2:1.00] +; HASWELL-LABEL: test_shrx_i32: +; HASWELL: # BB#0: +; HASWELL-NEXT: shrxl %esi, %edi, %ecx # sched: [1:0.50] +; HASWELL-NEXT: shrxl %esi, (%rdx), %eax # sched: [1:0.50] +; HASWELL-NEXT: addl %ecx, %eax # sched: [1:0.25] +; HASWELL-NEXT: retq # sched: [2:1.00] +; +; SKYLAKE-LABEL: test_shrx_i32: +; SKYLAKE: # BB#0: +; SKYLAKE-NEXT: shrxl %esi, %edi, %ecx # sched: [1:1.00] +; SKYLAKE-NEXT: shrxl %esi, (%rdx), %eax # sched: [1:0.50] +; SKYLAKE-NEXT: addl %ecx, %eax # sched: [1:0.25] +; SKYLAKE-NEXT: retq # sched: [2:1.00] +; +; KNL-LABEL: test_shrx_i32: +; KNL: # BB#0: +; KNL-NEXT: shrxl %esi, %edi, %ecx # sched: [1:0.50] +; KNL-NEXT: shrxl %esi, (%rdx), %eax # sched: [1:0.50] +; KNL-NEXT: addl %ecx, %eax # sched: [1:0.25] +; KNL-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_shrx_i32: ; ZNVER1: # BB#0: @@ -434,12 +532,26 @@ define i64 @test_shrx_i64(i64 %a0, i64 %a1, i64 *%a2) { ; GENERIC-NEXT: addq %rcx, %rax # sched: [1:0.33] ; GENERIC-NEXT: retq # sched: [1:1.00] ; -; COMMON-LABEL: test_shrx_i64: -; COMMON: # BB#0: -; COMMON-NEXT: shrxq %rsi, %rdi, %rcx # sched: [1:0.50] -; COMMON-NEXT: shrxq %rsi, (%rdx), %rax # sched: [1:0.50] -; COMMON-NEXT: addq %rcx, %rax # sched: [1:0.25] -; COMMON-NEXT: retq # sched: [2:1.00] +; HASWELL-LABEL: test_shrx_i64: +; HASWELL: # BB#0: +; HASWELL-NEXT: shrxq %rsi, %rdi, %rcx # sched: [1:0.50] +; HASWELL-NEXT: shrxq %rsi, (%rdx), %rax # sched: [1:0.50] +; HASWELL-NEXT: addq %rcx, %rax # sched: [1:0.25] +; HASWELL-NEXT: retq # sched: [2:1.00] +; +; SKYLAKE-LABEL: test_shrx_i64: +; SKYLAKE: # BB#0: +; SKYLAKE-NEXT: shrxq %rsi, %rdi, %rcx # sched: [1:1.00] +; SKYLAKE-NEXT: shrxq %rsi, (%rdx), %rax # sched: [1:0.50] +; SKYLAKE-NEXT: addq %rcx, %rax # sched: [1:0.25] +; SKYLAKE-NEXT: retq # sched: [2:1.00] +; +; KNL-LABEL: test_shrx_i64: +; KNL: # BB#0: +; KNL-NEXT: shrxq %rsi, %rdi, %rcx # sched: [1:0.50] +; KNL-NEXT: shrxq %rsi, (%rdx), %rax # sched: [1:0.50] +; KNL-NEXT: addq %rcx, %rax # sched: [1:0.25] +; KNL-NEXT: retq # sched: [2:1.00] ; ; ZNVER1-LABEL: test_shrx_i64: ; ZNVER1: # BB#0: diff --git a/llvm/test/CodeGen/X86/f16c-schedule.ll b/llvm/test/CodeGen/X86/f16c-schedule.ll index 166e491b271..3ecfc5426a2 100644 --- a/llvm/test/CodeGen/X86/f16c-schedule.ll +++ b/llvm/test/CodeGen/X86/f16c-schedule.ll @@ -30,9 +30,9 @@ define <4 x float> @test_vcvtph2ps_128(<8 x i16> %a0, <8 x i16> *%a1) { ; ; SKYLAKE-LABEL: test_vcvtph2ps_128: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtph2ps (%rdi), %xmm1 # sched: [1:1.00] -; SKYLAKE-NEXT: vcvtph2ps %xmm0, %xmm0 # sched: [2:1.00] -; SKYLAKE-NEXT: vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtph2ps (%rdi), %xmm1 # sched: [4:0.50] +; SKYLAKE-NEXT: vcvtph2ps %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vaddps %xmm0, %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_vcvtph2ps_128: @@ -80,9 +80,9 @@ define <8 x float> @test_vcvtph2ps_256(<8 x i16> %a0, <8 x i16> *%a1) { ; ; SKYLAKE-LABEL: test_vcvtph2ps_256: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtph2ps (%rdi), %ymm1 # sched: [1:1.00] -; SKYLAKE-NEXT: vcvtph2ps %xmm0, %ymm0 # sched: [2:1.00] -; SKYLAKE-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtph2ps (%rdi), %ymm1 # sched: [4:0.50] +; SKYLAKE-NEXT: vcvtph2ps %xmm0, %ymm0 # sched: [7:1.00] +; SKYLAKE-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_vcvtph2ps_256: @@ -127,8 +127,8 @@ define <8 x i16> @test_vcvtps2ph_128(<4 x float> %a0, <4 x float> %a1, <4 x i16> ; ; SKYLAKE-LABEL: test_vcvtps2ph_128: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtps2ph $0, %xmm0, %xmm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vcvtps2ph $0, %xmm1, (%rdi) # sched: [4:1.00] +; SKYLAKE-NEXT: vcvtps2ph $0, %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vcvtps2ph $0, %xmm1, (%rdi) # sched: [5:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_vcvtps2ph_128: @@ -174,8 +174,8 @@ define <8 x i16> @test_vcvtps2ph_256(<8 x float> %a0, <8 x float> %a1, <8 x i16> ; ; SKYLAKE-LABEL: test_vcvtps2ph_256: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtps2ph $0, %ymm0, %xmm0 # sched: [6:1.00] -; SKYLAKE-NEXT: vcvtps2ph $0, %ymm1, (%rdi) # sched: [6:1.00] +; SKYLAKE-NEXT: vcvtps2ph $0, %ymm0, %xmm0 # sched: [7:1.00] +; SKYLAKE-NEXT: vcvtps2ph $0, %ymm1, (%rdi) # sched: [7:1.00] ; SKYLAKE-NEXT: vzeroupper # sched: [4:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; diff --git a/llvm/test/CodeGen/X86/fma-schedule.ll b/llvm/test/CodeGen/X86/fma-schedule.ll index f482659e63e..c127451b8cb 100644 --- a/llvm/test/CodeGen/X86/fma-schedule.ll +++ b/llvm/test/CodeGen/X86/fma-schedule.ll @@ -28,8 +28,8 @@ define <2 x double> @test_vfmadd213pd(<2 x double> %a0, <2 x double> %a1, <2 x d ; ; SKYLAKE-LABEL: test_vfmadd213pd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmadd213pd %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmadd213pd (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmadd213pd %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmadd213pd (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmadd213pd: @@ -64,8 +64,8 @@ define <4 x double> @test_vfmadd213pd_ymm(<4 x double> %a0, <4 x double> %a1, <4 ; ; SKYLAKE-LABEL: test_vfmadd213pd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmadd213pd %ymm2, %ymm1, %ymm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmadd213pd (%rdi), %ymm1, %ymm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmadd213pd %ymm2, %ymm1, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmadd213pd (%rdi), %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmadd213pd_ymm: @@ -100,8 +100,8 @@ define <4 x float> @test_vfmadd213ps(<4 x float> %a0, <4 x float> %a1, <4 x floa ; ; SKYLAKE-LABEL: test_vfmadd213ps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmadd213ps %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmadd213ps (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmadd213ps %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmadd213ps (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmadd213ps: @@ -136,8 +136,8 @@ define <8 x float> @test_vfmadd213ps_ymm(<8 x float> %a0, <8 x float> %a1, <8 x ; ; SKYLAKE-LABEL: test_vfmadd213ps_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmadd213ps %ymm2, %ymm1, %ymm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmadd213ps (%rdi), %ymm1, %ymm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmadd213ps %ymm2, %ymm1, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmadd213ps (%rdi), %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmadd213ps_ymm: @@ -172,8 +172,8 @@ define <2 x double> @test_vfmadd213sd(<2 x double> %a0, <2 x double> %a1, <2 x d ; ; SKYLAKE-LABEL: test_vfmadd213sd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmadd213sd %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmadd213sd (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmadd213sd %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmadd213sd (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmadd213sd: @@ -208,8 +208,8 @@ define <4 x float> @test_vfmadd213ss(<4 x float> %a0, <4 x float> %a1, <4 x floa ; ; SKYLAKE-LABEL: test_vfmadd213ss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmadd213ss %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmadd213ss (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmadd213ss %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmadd213ss (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmadd213ss: @@ -256,8 +256,8 @@ define <2 x double> @test_vfmaddsubpd(<2 x double> %a0, <2 x double> %a1, <2 x d ; ; SKYLAKE-LABEL: test_vfmaddsubpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmaddsub213pd %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmaddsub213pd (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmaddsub213pd %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmaddsub213pd (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmaddsubpd: @@ -292,8 +292,8 @@ define <4 x double> @test_vfmaddsubpd_ymm(<4 x double> %a0, <4 x double> %a1, <4 ; ; SKYLAKE-LABEL: test_vfmaddsubpd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmaddsub213pd %ymm2, %ymm1, %ymm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmaddsub213pd (%rdi), %ymm1, %ymm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmaddsub213pd %ymm2, %ymm1, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmaddsub213pd (%rdi), %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmaddsubpd_ymm: @@ -328,8 +328,8 @@ define <4 x float> @test_vfmaddsubps(<4 x float> %a0, <4 x float> %a1, <4 x floa ; ; SKYLAKE-LABEL: test_vfmaddsubps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmaddsub213ps %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmaddsub213ps (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmaddsub213ps %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmaddsub213ps (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmaddsubps: @@ -364,8 +364,8 @@ define <8 x float> @test_vfmaddsubps_ymm(<8 x float> %a0, <8 x float> %a1, <8 x ; ; SKYLAKE-LABEL: test_vfmaddsubps_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmaddsub213ps %ymm2, %ymm1, %ymm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmaddsub213ps (%rdi), %ymm1, %ymm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmaddsub213ps %ymm2, %ymm1, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmaddsub213ps (%rdi), %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmaddsubps_ymm: @@ -412,8 +412,8 @@ define <2 x double> @test_vfmsubaddpd(<2 x double> %a0, <2 x double> %a1, <2 x d ; ; SKYLAKE-LABEL: test_vfmsubaddpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmsubadd213pd %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmsubadd213pd (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmsubadd213pd %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmsubadd213pd (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmsubaddpd: @@ -448,8 +448,8 @@ define <4 x double> @test_vfmsubaddpd_ymm(<4 x double> %a0, <4 x double> %a1, <4 ; ; SKYLAKE-LABEL: test_vfmsubaddpd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmsubadd213pd %ymm2, %ymm1, %ymm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmsubadd213pd (%rdi), %ymm1, %ymm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmsubadd213pd %ymm2, %ymm1, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmsubadd213pd (%rdi), %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmsubaddpd_ymm: @@ -484,8 +484,8 @@ define <4 x float> @test_vfmsubaddps(<4 x float> %a0, <4 x float> %a1, <4 x floa ; ; SKYLAKE-LABEL: test_vfmsubaddps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmsubadd213ps %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmsubadd213ps (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmsubadd213ps %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmsubadd213ps (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmsubaddps: @@ -520,8 +520,8 @@ define <8 x float> @test_vfmsubaddps_ymm(<8 x float> %a0, <8 x float> %a1, <8 x ; ; SKYLAKE-LABEL: test_vfmsubaddps_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmsubadd213ps %ymm2, %ymm1, %ymm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmsubadd213ps (%rdi), %ymm1, %ymm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmsubadd213ps %ymm2, %ymm1, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmsubadd213ps (%rdi), %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmsubaddps_ymm: @@ -568,8 +568,8 @@ define <2 x double> @test_vfmsub213pd(<2 x double> %a0, <2 x double> %a1, <2 x d ; ; SKYLAKE-LABEL: test_vfmsub213pd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmsub213pd %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmsub213pd (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmsub213pd %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmsub213pd (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmsub213pd: @@ -604,8 +604,8 @@ define <4 x double> @test_vfmsub213pd_ymm(<4 x double> %a0, <4 x double> %a1, <4 ; ; SKYLAKE-LABEL: test_vfmsub213pd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmsub213pd %ymm2, %ymm1, %ymm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmsub213pd (%rdi), %ymm1, %ymm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmsub213pd %ymm2, %ymm1, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmsub213pd (%rdi), %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmsub213pd_ymm: @@ -640,8 +640,8 @@ define <4 x float> @test_vfmsub213ps(<4 x float> %a0, <4 x float> %a1, <4 x floa ; ; SKYLAKE-LABEL: test_vfmsub213ps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmsub213ps %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmsub213ps (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmsub213ps %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmsub213ps (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmsub213ps: @@ -676,8 +676,8 @@ define <8 x float> @test_vfmsub213ps_ymm(<8 x float> %a0, <8 x float> %a1, <8 x ; ; SKYLAKE-LABEL: test_vfmsub213ps_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmsub213ps %ymm2, %ymm1, %ymm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmsub213ps (%rdi), %ymm1, %ymm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmsub213ps %ymm2, %ymm1, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmsub213ps (%rdi), %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmsub213ps_ymm: @@ -712,8 +712,8 @@ define <2 x double> @test_vfmsub213sd(<2 x double> %a0, <2 x double> %a1, <2 x d ; ; SKYLAKE-LABEL: test_vfmsub213sd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmsub213sd %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmsub213sd (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmsub213sd %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmsub213sd (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmsub213sd: @@ -748,8 +748,8 @@ define <4 x float> @test_vfmsub213ss(<4 x float> %a0, <4 x float> %a1, <4 x floa ; ; SKYLAKE-LABEL: test_vfmsub213ss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfmsub213ss %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfmsub213ss (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfmsub213ss %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfmsub213ss (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfmsub213ss: @@ -796,8 +796,8 @@ define <2 x double> @test_vfnmadd213pd(<2 x double> %a0, <2 x double> %a1, <2 x ; ; SKYLAKE-LABEL: test_vfnmadd213pd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfnmadd213pd %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfnmadd213pd (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfnmadd213pd %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfnmadd213pd (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfnmadd213pd: @@ -832,8 +832,8 @@ define <4 x double> @test_vfnmadd213pd_ymm(<4 x double> %a0, <4 x double> %a1, < ; ; SKYLAKE-LABEL: test_vfnmadd213pd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfnmadd213pd %ymm2, %ymm1, %ymm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfnmadd213pd (%rdi), %ymm1, %ymm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfnmadd213pd %ymm2, %ymm1, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfnmadd213pd (%rdi), %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfnmadd213pd_ymm: @@ -868,8 +868,8 @@ define <4 x float> @test_vfnmadd213ps(<4 x float> %a0, <4 x float> %a1, <4 x flo ; ; SKYLAKE-LABEL: test_vfnmadd213ps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfnmadd213ps %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfnmadd213ps (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfnmadd213ps %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfnmadd213ps (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfnmadd213ps: @@ -904,8 +904,8 @@ define <8 x float> @test_vfnmadd213ps_ymm(<8 x float> %a0, <8 x float> %a1, <8 x ; ; SKYLAKE-LABEL: test_vfnmadd213ps_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfnmadd213ps %ymm2, %ymm1, %ymm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfnmadd213ps (%rdi), %ymm1, %ymm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfnmadd213ps %ymm2, %ymm1, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfnmadd213ps (%rdi), %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfnmadd213ps_ymm: @@ -940,8 +940,8 @@ define <2 x double> @test_vfnmadd213sd(<2 x double> %a0, <2 x double> %a1, <2 x ; ; SKYLAKE-LABEL: test_vfnmadd213sd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfnmadd213sd %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfnmadd213sd (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfnmadd213sd %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfnmadd213sd (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfnmadd213sd: @@ -976,8 +976,8 @@ define <4 x float> @test_vfnmadd213ss(<4 x float> %a0, <4 x float> %a1, <4 x flo ; ; SKYLAKE-LABEL: test_vfnmadd213ss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfnmadd213ss %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfnmadd213ss (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfnmadd213ss %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfnmadd213ss (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfnmadd213ss: @@ -1024,8 +1024,8 @@ define <2 x double> @test_vfnmsub213pd(<2 x double> %a0, <2 x double> %a1, <2 x ; ; SKYLAKE-LABEL: test_vfnmsub213pd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfnmsub213pd %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfnmsub213pd (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfnmsub213pd %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfnmsub213pd (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfnmsub213pd: @@ -1060,8 +1060,8 @@ define <4 x double> @test_vfnmsub213pd_ymm(<4 x double> %a0, <4 x double> %a1, < ; ; SKYLAKE-LABEL: test_vfnmsub213pd_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfnmsub213pd %ymm2, %ymm1, %ymm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfnmsub213pd (%rdi), %ymm1, %ymm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfnmsub213pd %ymm2, %ymm1, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfnmsub213pd (%rdi), %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfnmsub213pd_ymm: @@ -1096,8 +1096,8 @@ define <4 x float> @test_vfnmsub213ps(<4 x float> %a0, <4 x float> %a1, <4 x flo ; ; SKYLAKE-LABEL: test_vfnmsub213ps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfnmsub213ps %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfnmsub213ps (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfnmsub213ps %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfnmsub213ps (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfnmsub213ps: @@ -1132,8 +1132,8 @@ define <8 x float> @test_vfnmsub213ps_ymm(<8 x float> %a0, <8 x float> %a1, <8 x ; ; SKYLAKE-LABEL: test_vfnmsub213ps_ymm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfnmsub213ps %ymm2, %ymm1, %ymm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfnmsub213ps (%rdi), %ymm1, %ymm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfnmsub213ps %ymm2, %ymm1, %ymm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfnmsub213ps (%rdi), %ymm1, %ymm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfnmsub213ps_ymm: @@ -1168,8 +1168,8 @@ define <2 x double> @test_vfnmsub213sd(<2 x double> %a0, <2 x double> %a1, <2 x ; ; SKYLAKE-LABEL: test_vfnmsub213sd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfnmsub213sd %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfnmsub213sd (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfnmsub213sd %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfnmsub213sd (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfnmsub213sd: @@ -1204,8 +1204,8 @@ define <4 x float> @test_vfnmsub213ss(<4 x float> %a0, <4 x float> %a1, <4 x flo ; ; SKYLAKE-LABEL: test_vfnmsub213ss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vfnmsub213ss %xmm2, %xmm1, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vfnmsub213ss (%rdi), %xmm1, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vfnmsub213ss %xmm2, %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vfnmsub213ss (%rdi), %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; KNL-LABEL: test_vfnmsub213ss: diff --git a/llvm/test/CodeGen/X86/recip-fastmath.ll b/llvm/test/CodeGen/X86/recip-fastmath.ll index e33c688b0a0..5fb531da706 100644 --- a/llvm/test/CodeGen/X86/recip-fastmath.ll +++ b/llvm/test/CodeGen/X86/recip-fastmath.ll @@ -61,11 +61,17 @@ define float @f32_no_estimate(float %x) #0 { ; HASWELL-NO-FMA-NEXT: vdivss %xmm0, %xmm1, %xmm0 ; HASWELL-NO-FMA-NEXT: retq ; -; AVX512-LABEL: f32_no_estimate: -; AVX512: # BB#0: -; AVX512-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [1:0.50] -; AVX512-NEXT: vdivss %xmm0, %xmm1, %xmm0 # sched: [13:1.00] -; AVX512-NEXT: retq # sched: [2:1.00] +; KNL-LABEL: f32_no_estimate: +; KNL: # BB#0: +; KNL-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [1:0.50] +; KNL-NEXT: vdivss %xmm0, %xmm1, %xmm0 # sched: [13:1.00] +; KNL-NEXT: retq # sched: [2:1.00] +; +; SKX-LABEL: f32_no_estimate: +; SKX: # BB#0: +; SKX-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [1:0.50] +; SKX-NEXT: vdivss %xmm0, %xmm1, %xmm0 # sched: [11:1.00] +; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast float 1.0, %x ret float %div } @@ -136,12 +142,19 @@ define float @f32_one_step(float %x) #1 { ; HASWELL-NO-FMA-NEXT: vaddss %xmm0, %xmm1, %xmm0 ; HASWELL-NO-FMA-NEXT: retq ; -; AVX512-LABEL: f32_one_step: -; AVX512: # BB#0: -; AVX512-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 -; AVX512-NEXT: vfnmadd213ss {{.*}}(%rip), %xmm1, %xmm0 # sched: [5:0.50] -; AVX512-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm0 # sched: [5:0.50] -; AVX512-NEXT: retq # sched: [2:1.00] +; KNL-LABEL: f32_one_step: +; KNL: # BB#0: +; KNL-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 +; KNL-NEXT: vfnmadd213ss {{.*}}(%rip), %xmm1, %xmm0 # sched: [5:0.50] +; KNL-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm0 # sched: [5:0.50] +; KNL-NEXT: retq # sched: [2:1.00] +; +; SKX-LABEL: f32_one_step: +; SKX: # BB#0: +; SKX-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 +; SKX-NEXT: vfnmadd213ss {{.*}}(%rip), %xmm1, %xmm0 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm0 # sched: [4:0.50] +; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast float 1.0, %x ret float %div } @@ -242,16 +255,27 @@ define float @f32_two_step(float %x) #2 { ; HASWELL-NO-FMA-NEXT: vaddss %xmm0, %xmm1, %xmm0 ; HASWELL-NO-FMA-NEXT: retq ; -; AVX512-LABEL: f32_two_step: -; AVX512: # BB#0: -; AVX512-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 -; AVX512-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [1:0.50] -; AVX512-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:1.00] -; AVX512-NEXT: vfnmadd213ss %xmm2, %xmm0, %xmm3 # sched: [5:0.50] -; AVX512-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm3 # sched: [5:0.50] -; AVX512-NEXT: vfnmadd213ss %xmm2, %xmm3, %xmm0 # sched: [5:0.50] -; AVX512-NEXT: vfmadd132ss %xmm3, %xmm3, %xmm0 # sched: [5:0.50] -; AVX512-NEXT: retq # sched: [2:1.00] +; KNL-LABEL: f32_two_step: +; KNL: # BB#0: +; KNL-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 +; KNL-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [1:0.50] +; KNL-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:1.00] +; KNL-NEXT: vfnmadd213ss %xmm2, %xmm0, %xmm3 # sched: [5:0.50] +; KNL-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm3 # sched: [5:0.50] +; KNL-NEXT: vfnmadd213ss %xmm2, %xmm3, %xmm0 # sched: [5:0.50] +; KNL-NEXT: vfmadd132ss %xmm3, %xmm3, %xmm0 # sched: [5:0.50] +; KNL-NEXT: retq # sched: [2:1.00] +; +; SKX-LABEL: f32_two_step: +; SKX: # BB#0: +; SKX-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 +; SKX-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [1:0.50] +; SKX-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:1.00] +; SKX-NEXT: vfnmadd213ss %xmm2, %xmm0, %xmm3 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm3 # sched: [4:0.50] +; SKX-NEXT: vfnmadd213ss %xmm2, %xmm3, %xmm0 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ss %xmm3, %xmm3, %xmm0 # sched: [4:0.50] +; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast float 1.0, %x ret float %div } @@ -300,11 +324,17 @@ define <4 x float> @v4f32_no_estimate(<4 x float> %x) #0 { ; HASWELL-NO-FMA-NEXT: vdivps %xmm0, %xmm1, %xmm0 ; HASWELL-NO-FMA-NEXT: retq ; -; AVX512-LABEL: v4f32_no_estimate: -; AVX512: # BB#0: -; AVX512-NEXT: vbroadcastss {{.*#+}} xmm1 = [1,1,1,1] sched: [1:0.50] -; AVX512-NEXT: vdivps %xmm0, %xmm1, %xmm0 # sched: [13:1.00] -; AVX512-NEXT: retq # sched: [2:1.00] +; KNL-LABEL: v4f32_no_estimate: +; KNL: # BB#0: +; KNL-NEXT: vbroadcastss {{.*#+}} xmm1 = [1,1,1,1] sched: [1:0.50] +; KNL-NEXT: vdivps %xmm0, %xmm1, %xmm0 # sched: [13:1.00] +; KNL-NEXT: retq # sched: [2:1.00] +; +; SKX-LABEL: v4f32_no_estimate: +; SKX: # BB#0: +; SKX-NEXT: vbroadcastss {{.*#+}} xmm1 = [1,1,1,1] sched: [1:0.50] +; SKX-NEXT: vdivps %xmm0, %xmm1, %xmm0 # sched: [11:1.00] +; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>, %x ret <4 x float> %div } @@ -388,7 +418,7 @@ define <4 x float> @v4f32_one_step(<4 x float> %x) #1 { ; SKX: # BB#0: ; SKX-NEXT: vrcp14ps %xmm0, %xmm1 ; SKX-NEXT: vfnmadd213ps {{.*}}(%rip){1to4}, %xmm1, %xmm0 -; SKX-NEXT: vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [5:0.50] +; SKX-NEXT: vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [4:0.50] ; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>, %x ret <4 x float> %div @@ -506,10 +536,10 @@ define <4 x float> @v4f32_two_step(<4 x float> %x) #2 { ; SKX-NEXT: vrcp14ps %xmm0, %xmm1 ; SKX-NEXT: vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [1:0.50] ; SKX-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:1.00] -; SKX-NEXT: vfnmadd213ps %xmm2, %xmm0, %xmm3 # sched: [5:0.50] -; SKX-NEXT: vfmadd132ps %xmm1, %xmm1, %xmm3 # sched: [5:0.50] -; SKX-NEXT: vfnmadd213ps %xmm2, %xmm3, %xmm0 # sched: [5:0.50] -; SKX-NEXT: vfmadd132ps %xmm3, %xmm3, %xmm0 # sched: [5:0.50] +; SKX-NEXT: vfnmadd213ps %xmm2, %xmm0, %xmm3 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ps %xmm1, %xmm1, %xmm3 # sched: [4:0.50] +; SKX-NEXT: vfnmadd213ps %xmm2, %xmm3, %xmm0 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ps %xmm3, %xmm3, %xmm0 # sched: [4:0.50] ; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>, %x ret <4 x float> %div @@ -562,11 +592,17 @@ define <8 x float> @v8f32_no_estimate(<8 x float> %x) #0 { ; HASWELL-NO-FMA-NEXT: vdivps %ymm0, %ymm1, %ymm0 ; HASWELL-NO-FMA-NEXT: retq ; -; AVX512-LABEL: v8f32_no_estimate: -; AVX512: # BB#0: -; AVX512-NEXT: vbroadcastss {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1] sched: [1:0.50] -; AVX512-NEXT: vdivps %ymm0, %ymm1, %ymm0 # sched: [21:2.00] -; AVX512-NEXT: retq # sched: [2:1.00] +; KNL-LABEL: v8f32_no_estimate: +; KNL: # BB#0: +; KNL-NEXT: vbroadcastss {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1] sched: [1:0.50] +; KNL-NEXT: vdivps %ymm0, %ymm1, %ymm0 # sched: [21:2.00] +; KNL-NEXT: retq # sched: [2:1.00] +; +; SKX-LABEL: v8f32_no_estimate: +; SKX: # BB#0: +; SKX-NEXT: vbroadcastss {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1] sched: [1:0.50] +; SKX-NEXT: vdivps %ymm0, %ymm1, %ymm0 # sched: [11:1.00] +; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast <8 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x ret <8 x float> %div } @@ -657,7 +693,7 @@ define <8 x float> @v8f32_one_step(<8 x float> %x) #1 { ; SKX: # BB#0: ; SKX-NEXT: vrcp14ps %ymm0, %ymm1 ; SKX-NEXT: vfnmadd213ps {{.*}}(%rip){1to8}, %ymm1, %ymm0 -; SKX-NEXT: vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [5:0.50] +; SKX-NEXT: vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [4:0.50] ; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast <8 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x ret <8 x float> %div @@ -788,10 +824,10 @@ define <8 x float> @v8f32_two_step(<8 x float> %x) #2 { ; SKX-NEXT: vrcp14ps %ymm0, %ymm1 ; SKX-NEXT: vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [1:0.50] ; SKX-NEXT: vmovaps %ymm1, %ymm3 # sched: [1:1.00] -; SKX-NEXT: vfnmadd213ps %ymm2, %ymm0, %ymm3 # sched: [5:0.50] -; SKX-NEXT: vfmadd132ps %ymm1, %ymm1, %ymm3 # sched: [5:0.50] -; SKX-NEXT: vfnmadd213ps %ymm2, %ymm3, %ymm0 # sched: [5:0.50] -; SKX-NEXT: vfmadd132ps %ymm3, %ymm3, %ymm0 # sched: [5:0.50] +; SKX-NEXT: vfnmadd213ps %ymm2, %ymm0, %ymm3 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ps %ymm1, %ymm1, %ymm3 # sched: [4:0.50] +; SKX-NEXT: vfnmadd213ps %ymm2, %ymm3, %ymm0 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ps %ymm3, %ymm3, %ymm0 # sched: [4:0.50] ; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast <8 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x ret <8 x float> %div diff --git a/llvm/test/CodeGen/X86/recip-fastmath2.ll b/llvm/test/CodeGen/X86/recip-fastmath2.ll index ac63974d3ad..057ef903565 100644 --- a/llvm/test/CodeGen/X86/recip-fastmath2.ll +++ b/llvm/test/CodeGen/X86/recip-fastmath2.ll @@ -54,11 +54,17 @@ define float @f32_no_step_2(float %x) #3 { ; HASWELL-NO-FMA-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm0 # sched: [5:0.50] ; HASWELL-NO-FMA-NEXT: retq # sched: [2:1.00] ; -; AVX512-LABEL: f32_no_step_2: -; AVX512: # BB#0: -; AVX512-NEXT: vrcp14ss %xmm0, %xmm0, %xmm0 -; AVX512-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm0 # sched: [5:0.50] -; AVX512-NEXT: retq # sched: [2:1.00] +; KNL-LABEL: f32_no_step_2: +; KNL: # BB#0: +; KNL-NEXT: vrcp14ss %xmm0, %xmm0, %xmm0 +; KNL-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm0 # sched: [5:0.50] +; KNL-NEXT: retq # sched: [2:1.00] +; +; SKX-LABEL: f32_no_step_2: +; SKX: # BB#0: +; SKX-NEXT: vrcp14ss %xmm0, %xmm0, %xmm0 +; SKX-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm0 # sched: [4:0.50] +; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast float 1234.0, %x ret float %div } @@ -136,13 +142,21 @@ define float @f32_one_step_2(float %x) #1 { ; HASWELL-NO-FMA-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm0 # sched: [5:0.50] ; HASWELL-NO-FMA-NEXT: retq # sched: [2:1.00] ; -; AVX512-LABEL: f32_one_step_2: -; AVX512: # BB#0: -; AVX512-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 -; AVX512-NEXT: vfnmadd213ss {{.*}}(%rip), %xmm1, %xmm0 # sched: [5:0.50] -; AVX512-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm0 # sched: [5:0.50] -; AVX512-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm0 # sched: [5:0.50] -; AVX512-NEXT: retq # sched: [2:1.00] +; KNL-LABEL: f32_one_step_2: +; KNL: # BB#0: +; KNL-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 +; KNL-NEXT: vfnmadd213ss {{.*}}(%rip), %xmm1, %xmm0 # sched: [5:0.50] +; KNL-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm0 # sched: [5:0.50] +; KNL-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm0 # sched: [5:0.50] +; KNL-NEXT: retq # sched: [2:1.00] +; +; SKX-LABEL: f32_one_step_2: +; SKX: # BB#0: +; SKX-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 +; SKX-NEXT: vfnmadd213ss {{.*}}(%rip), %xmm1, %xmm0 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm0 # sched: [4:0.50] +; SKX-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm0 # sched: [4:0.50] +; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast float 3456.0, %x ret float %div } @@ -227,14 +241,23 @@ define float @f32_one_step_2_divs(float %x) #1 { ; HASWELL-NO-FMA-NEXT: vmulss %xmm0, %xmm1, %xmm0 # sched: [5:0.50] ; HASWELL-NO-FMA-NEXT: retq # sched: [2:1.00] ; -; AVX512-LABEL: f32_one_step_2_divs: -; AVX512: # BB#0: -; AVX512-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 -; AVX512-NEXT: vfnmadd213ss {{.*}}(%rip), %xmm1, %xmm0 # sched: [5:0.50] -; AVX512-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm0 # sched: [5:0.50] -; AVX512-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm1 # sched: [5:0.50] -; AVX512-NEXT: vmulss %xmm0, %xmm1, %xmm0 # sched: [5:0.50] -; AVX512-NEXT: retq # sched: [2:1.00] +; KNL-LABEL: f32_one_step_2_divs: +; KNL: # BB#0: +; KNL-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 +; KNL-NEXT: vfnmadd213ss {{.*}}(%rip), %xmm1, %xmm0 # sched: [5:0.50] +; KNL-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm0 # sched: [5:0.50] +; KNL-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm1 # sched: [5:0.50] +; KNL-NEXT: vmulss %xmm0, %xmm1, %xmm0 # sched: [5:0.50] +; KNL-NEXT: retq # sched: [2:1.00] +; +; SKX-LABEL: f32_one_step_2_divs: +; SKX: # BB#0: +; SKX-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 +; SKX-NEXT: vfnmadd213ss {{.*}}(%rip), %xmm1, %xmm0 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm0 # sched: [4:0.50] +; SKX-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm1 # sched: [4:0.50] +; SKX-NEXT: vmulss %xmm0, %xmm1, %xmm0 # sched: [4:0.50] +; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast float 3456.0, %x %div2 = fdiv fast float %div, %x ret float %div2 @@ -343,17 +366,29 @@ define float @f32_two_step_2(float %x) #2 { ; HASWELL-NO-FMA-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm0 # sched: [5:0.50] ; HASWELL-NO-FMA-NEXT: retq # sched: [2:1.00] ; -; AVX512-LABEL: f32_two_step_2: -; AVX512: # BB#0: -; AVX512-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 -; AVX512-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [1:0.50] -; AVX512-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:1.00] -; AVX512-NEXT: vfnmadd213ss %xmm2, %xmm0, %xmm3 # sched: [5:0.50] -; AVX512-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm3 # sched: [5:0.50] -; AVX512-NEXT: vfnmadd213ss %xmm2, %xmm3, %xmm0 # sched: [5:0.50] -; AVX512-NEXT: vfmadd132ss %xmm3, %xmm3, %xmm0 # sched: [5:0.50] -; AVX512-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm0 # sched: [5:0.50] -; AVX512-NEXT: retq # sched: [2:1.00] +; KNL-LABEL: f32_two_step_2: +; KNL: # BB#0: +; KNL-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 +; KNL-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [1:0.50] +; KNL-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:1.00] +; KNL-NEXT: vfnmadd213ss %xmm2, %xmm0, %xmm3 # sched: [5:0.50] +; KNL-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm3 # sched: [5:0.50] +; KNL-NEXT: vfnmadd213ss %xmm2, %xmm3, %xmm0 # sched: [5:0.50] +; KNL-NEXT: vfmadd132ss %xmm3, %xmm3, %xmm0 # sched: [5:0.50] +; KNL-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm0 # sched: [5:0.50] +; KNL-NEXT: retq # sched: [2:1.00] +; +; SKX-LABEL: f32_two_step_2: +; SKX: # BB#0: +; SKX-NEXT: vrcp14ss %xmm0, %xmm0, %xmm1 +; SKX-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [1:0.50] +; SKX-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:1.00] +; SKX-NEXT: vfnmadd213ss %xmm2, %xmm0, %xmm3 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ss %xmm1, %xmm1, %xmm3 # sched: [4:0.50] +; SKX-NEXT: vfnmadd213ss %xmm2, %xmm3, %xmm0 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ss %xmm3, %xmm3, %xmm0 # sched: [4:0.50] +; SKX-NEXT: vmulss {{.*}}(%rip), %xmm0, %xmm0 # sched: [4:0.50] +; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast float 6789.0, %x ret float %div } @@ -445,8 +480,8 @@ define <4 x float> @v4f32_one_step2(<4 x float> %x) #1 { ; SKX: # BB#0: ; SKX-NEXT: vrcp14ps %xmm0, %xmm1 ; SKX-NEXT: vfnmadd213ps {{.*}}(%rip){1to4}, %xmm1, %xmm0 -; SKX-NEXT: vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [5:0.50] -; SKX-NEXT: vmulps {{.*}}(%rip), %xmm0, %xmm0 # sched: [5:0.50] +; SKX-NEXT: vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [4:0.50] +; SKX-NEXT: vmulps {{.*}}(%rip), %xmm0, %xmm0 # sched: [4:0.50] ; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, %x ret <4 x float> %div @@ -547,9 +582,9 @@ define <4 x float> @v4f32_one_step_2_divs(<4 x float> %x) #1 { ; SKX: # BB#0: ; SKX-NEXT: vrcp14ps %xmm0, %xmm1 ; SKX-NEXT: vfnmadd213ps {{.*}}(%rip){1to4}, %xmm1, %xmm0 -; SKX-NEXT: vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [5:0.50] -; SKX-NEXT: vmulps {{.*}}(%rip), %xmm0, %xmm1 # sched: [5:0.50] -; SKX-NEXT: vmulps %xmm0, %xmm1, %xmm0 # sched: [5:0.50] +; SKX-NEXT: vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [4:0.50] +; SKX-NEXT: vmulps {{.*}}(%rip), %xmm0, %xmm1 # sched: [4:0.50] +; SKX-NEXT: vmulps %xmm0, %xmm1, %xmm0 # sched: [4:0.50] ; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, %x %div2 = fdiv fast <4 x float> %div, %x @@ -676,11 +711,11 @@ define <4 x float> @v4f32_two_step2(<4 x float> %x) #2 { ; SKX-NEXT: vrcp14ps %xmm0, %xmm1 ; SKX-NEXT: vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [1:0.50] ; SKX-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:1.00] -; SKX-NEXT: vfnmadd213ps %xmm2, %xmm0, %xmm3 # sched: [5:0.50] -; SKX-NEXT: vfmadd132ps %xmm1, %xmm1, %xmm3 # sched: [5:0.50] -; SKX-NEXT: vfnmadd213ps %xmm2, %xmm3, %xmm0 # sched: [5:0.50] -; SKX-NEXT: vfmadd132ps %xmm3, %xmm3, %xmm0 # sched: [5:0.50] -; SKX-NEXT: vmulps {{.*}}(%rip), %xmm0, %xmm0 # sched: [5:0.50] +; SKX-NEXT: vfnmadd213ps %xmm2, %xmm0, %xmm3 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ps %xmm1, %xmm1, %xmm3 # sched: [4:0.50] +; SKX-NEXT: vfnmadd213ps %xmm2, %xmm3, %xmm0 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ps %xmm3, %xmm3, %xmm0 # sched: [4:0.50] +; SKX-NEXT: vmulps {{.*}}(%rip), %xmm0, %xmm0 # sched: [4:0.50] ; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, %x ret <4 x float> %div @@ -781,8 +816,8 @@ define <8 x float> @v8f32_one_step2(<8 x float> %x) #1 { ; SKX: # BB#0: ; SKX-NEXT: vrcp14ps %ymm0, %ymm1 ; SKX-NEXT: vfnmadd213ps {{.*}}(%rip){1to8}, %ymm1, %ymm0 -; SKX-NEXT: vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [5:0.50] -; SKX-NEXT: vmulps {{.*}}(%rip), %ymm0, %ymm0 # sched: [5:0.50] +; SKX-NEXT: vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [4:0.50] +; SKX-NEXT: vmulps {{.*}}(%rip), %ymm0, %ymm0 # sched: [4:0.50] ; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast <8 x float> <float 1.0, float 2.0, float 3.0, float 4.0, float 5.0, float 6.0, float 7.0, float 8.0>, %x ret <8 x float> %div @@ -892,9 +927,9 @@ define <8 x float> @v8f32_one_step_2_divs(<8 x float> %x) #1 { ; SKX: # BB#0: ; SKX-NEXT: vrcp14ps %ymm0, %ymm1 ; SKX-NEXT: vfnmadd213ps {{.*}}(%rip){1to8}, %ymm1, %ymm0 -; SKX-NEXT: vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [5:0.50] -; SKX-NEXT: vmulps {{.*}}(%rip), %ymm0, %ymm1 # sched: [5:0.50] -; SKX-NEXT: vmulps %ymm0, %ymm1, %ymm0 # sched: [5:0.50] +; SKX-NEXT: vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [4:0.50] +; SKX-NEXT: vmulps {{.*}}(%rip), %ymm0, %ymm1 # sched: [4:0.50] +; SKX-NEXT: vmulps %ymm0, %ymm1, %ymm0 # sched: [4:0.50] ; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast <8 x float> <float 1.0, float 2.0, float 3.0, float 4.0, float 5.0, float 6.0, float 7.0, float 8.0>, %x %div2 = fdiv fast <8 x float> %div, %x @@ -1035,11 +1070,11 @@ define <8 x float> @v8f32_two_step2(<8 x float> %x) #2 { ; SKX-NEXT: vrcp14ps %ymm0, %ymm1 ; SKX-NEXT: vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [1:0.50] ; SKX-NEXT: vmovaps %ymm1, %ymm3 # sched: [1:1.00] -; SKX-NEXT: vfnmadd213ps %ymm2, %ymm0, %ymm3 # sched: [5:0.50] -; SKX-NEXT: vfmadd132ps %ymm1, %ymm1, %ymm3 # sched: [5:0.50] -; SKX-NEXT: vfnmadd213ps %ymm2, %ymm3, %ymm0 # sched: [5:0.50] -; SKX-NEXT: vfmadd132ps %ymm3, %ymm3, %ymm0 # sched: [5:0.50] -; SKX-NEXT: vmulps {{.*}}(%rip), %ymm0, %ymm0 # sched: [5:0.50] +; SKX-NEXT: vfnmadd213ps %ymm2, %ymm0, %ymm3 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ps %ymm1, %ymm1, %ymm3 # sched: [4:0.50] +; SKX-NEXT: vfnmadd213ps %ymm2, %ymm3, %ymm0 # sched: [4:0.50] +; SKX-NEXT: vfmadd132ps %ymm3, %ymm3, %ymm0 # sched: [4:0.50] +; SKX-NEXT: vmulps {{.*}}(%rip), %ymm0, %ymm0 # sched: [4:0.50] ; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast <8 x float> <float 1.0, float 2.0, float 3.0, float 4.0, float 5.0, float 6.0, float 7.0, float 8.0>, %x ret <8 x float> %div @@ -1149,7 +1184,7 @@ define <8 x float> @v8f32_no_step2(<8 x float> %x) #3 { ; SKX-LABEL: v8f32_no_step2: ; SKX: # BB#0: ; SKX-NEXT: vrcp14ps %ymm0, %ymm0 -; SKX-NEXT: vmulps {{.*}}(%rip), %ymm0, %ymm0 # sched: [5:0.50] +; SKX-NEXT: vmulps {{.*}}(%rip), %ymm0, %ymm0 # sched: [4:0.50] ; SKX-NEXT: retq # sched: [2:1.00] %div = fdiv fast <8 x float> <float 1.0, float 2.0, float 3.0, float 4.0, float 5.0, float 6.0, float 7.0, float 8.0>, %x ret <8 x float> %div diff --git a/llvm/test/CodeGen/X86/sse-schedule.ll b/llvm/test/CodeGen/X86/sse-schedule.ll index e4ead498782..9100809d373 100644 --- a/llvm/test/CodeGen/X86/sse-schedule.ll +++ b/llvm/test/CodeGen/X86/sse-schedule.ll @@ -42,8 +42,8 @@ define <4 x float> @test_addps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a ; ; SKYLAKE-LABEL: test_addps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddps (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vaddps (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_addps: @@ -96,8 +96,8 @@ define float @test_addss(float %a0, float %a1, float *%a2) { ; ; SKYLAKE-LABEL: test_addss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddss (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddss %xmm1, %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vaddss (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_addss: @@ -154,8 +154,8 @@ define <4 x float> @test_andps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a ; ; SKYLAKE-LABEL: test_andps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vandps %xmm1, %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vandps (%rdi), %xmm0, %xmm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vandps %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vandps (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_andps: @@ -216,8 +216,8 @@ define <4 x float> @test_andnotps(<4 x float> %a0, <4 x float> %a1, <4 x float> ; ; SKYLAKE-LABEL: test_andnotps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vandnps %xmm1, %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vandnps (%rdi), %xmm0, %xmm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vandnps %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vandnps (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_andnotps: @@ -281,9 +281,9 @@ define <4 x float> @test_cmpps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a ; ; SKYLAKE-LABEL: test_cmpps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcmpeqps %xmm1, %xmm0, %xmm1 # sched: [3:1.00] -; SKYLAKE-NEXT: vcmpeqps (%rdi), %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vorps %xmm0, %xmm1, %xmm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vcmpeqps %xmm1, %xmm0, %xmm1 # sched: [4:0.33] +; SKYLAKE-NEXT: vcmpeqps (%rdi), %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vorps %xmm0, %xmm1, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cmpps: @@ -342,7 +342,7 @@ define float @test_cmpss(float %a0, float %a1, float *%a2) { ; SKYLAKE-LABEL: test_cmpss: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vcmpeqss %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vcmpeqss (%rdi), %xmm0, %xmm0 # sched: [7:1.00] +; SKYLAKE-NEXT: vcmpeqss (%rdi), %xmm0, %xmm0 # sched: [8:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cmpss: @@ -441,12 +441,12 @@ define i32 @test_comiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) { ; SKYLAKE-LABEL: test_comiss: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vcomiss %xmm1, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: setnp %al # sched: [1:0.50] -; SKYLAKE-NEXT: sete %cl # sched: [1:0.50] +; SKYLAKE-NEXT: setnp %al # sched: [1:1.00] +; SKYLAKE-NEXT: sete %cl # sched: [1:1.00] ; SKYLAKE-NEXT: andb %al, %cl # sched: [1:0.25] -; SKYLAKE-NEXT: vcomiss (%rdi), %xmm0 # sched: [7:1.00] -; SKYLAKE-NEXT: setnp %al # sched: [1:0.50] -; SKYLAKE-NEXT: sete %dl # sched: [1:0.50] +; SKYLAKE-NEXT: vcomiss (%rdi), %xmm0 # sched: [8:1.00] +; SKYLAKE-NEXT: setnp %al # sched: [1:1.00] +; SKYLAKE-NEXT: sete %dl # sched: [1:1.00] ; SKYLAKE-NEXT: andb %al, %dl # sched: [1:0.25] ; SKYLAKE-NEXT: orb %cl, %dl # sched: [1:0.25] ; SKYLAKE-NEXT: movzbl %dl, %eax # sched: [1:0.25] @@ -525,9 +525,9 @@ define float @test_cvtsi2ss(i32 %a0, i32 *%a1) { ; ; SKYLAKE-LABEL: test_cvtsi2ss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtsi2ssl %edi, %xmm0, %xmm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vcvtsi2ssl (%rsi), %xmm1, %xmm1 # sched: [8:1.00] -; SKYLAKE-NEXT: vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtsi2ssl %edi, %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vcvtsi2ssl (%rsi), %xmm1, %xmm1 # sched: [9:1.00] +; SKYLAKE-NEXT: vaddss %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cvtsi2ss: @@ -588,9 +588,9 @@ define float @test_cvtsi2ssq(i64 %a0, i64 *%a1) { ; ; SKYLAKE-LABEL: test_cvtsi2ssq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtsi2ssq %rdi, %xmm0, %xmm0 # sched: [5:2.00] -; SKYLAKE-NEXT: vcvtsi2ssq (%rsi), %xmm1, %xmm1 # sched: [8:1.00] -; SKYLAKE-NEXT: vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtsi2ssq %rdi, %xmm0, %xmm0 # sched: [6:2.00] +; SKYLAKE-NEXT: vcvtsi2ssq (%rsi), %xmm1, %xmm1 # sched: [9:1.00] +; SKYLAKE-NEXT: vaddss %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cvtsi2ssq: @@ -651,8 +651,8 @@ define i32 @test_cvtss2si(float %a0, float *%a1) { ; ; SKYLAKE-LABEL: test_cvtss2si: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtss2si %xmm0, %ecx # sched: [4:1.00] -; SKYLAKE-NEXT: vcvtss2si (%rdi), %eax # sched: [4:1.00] +; SKYLAKE-NEXT: vcvtss2si %xmm0, %ecx # sched: [6:1.00] +; SKYLAKE-NEXT: vcvtss2si (%rdi), %eax # sched: [6:1.00] ; SKYLAKE-NEXT: addl %ecx, %eax # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -717,8 +717,8 @@ define i64 @test_cvtss2siq(float %a0, float *%a1) { ; ; SKYLAKE-LABEL: test_cvtss2siq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtss2si %xmm0, %rcx # sched: [4:1.00] -; SKYLAKE-NEXT: vcvtss2si (%rdi), %rax # sched: [4:1.00] +; SKYLAKE-NEXT: vcvtss2si %xmm0, %rcx # sched: [6:1.00] +; SKYLAKE-NEXT: vcvtss2si (%rdi), %rax # sched: [6:1.00] ; SKYLAKE-NEXT: addq %rcx, %rax # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -783,8 +783,8 @@ define i32 @test_cvttss2si(float %a0, float *%a1) { ; ; SKYLAKE-LABEL: test_cvttss2si: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvttss2si %xmm0, %ecx # sched: [4:1.00] -; SKYLAKE-NEXT: vcvttss2si (%rdi), %eax # sched: [4:1.00] +; SKYLAKE-NEXT: vcvttss2si %xmm0, %ecx # sched: [7:1.00] +; SKYLAKE-NEXT: vcvttss2si (%rdi), %eax # sched: [6:1.00] ; SKYLAKE-NEXT: addl %ecx, %eax # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -846,8 +846,8 @@ define i64 @test_cvttss2siq(float %a0, float *%a1) { ; ; SKYLAKE-LABEL: test_cvttss2siq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvttss2si %xmm0, %rcx # sched: [4:1.00] -; SKYLAKE-NEXT: vcvttss2si (%rdi), %rax # sched: [4:1.00] +; SKYLAKE-NEXT: vcvttss2si %xmm0, %rcx # sched: [7:1.00] +; SKYLAKE-NEXT: vcvttss2si (%rdi), %rax # sched: [6:1.00] ; SKYLAKE-NEXT: addq %rcx, %rax # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -904,8 +904,8 @@ define <4 x float> @test_divps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a ; ; SKYLAKE-LABEL: test_divps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vdivps %xmm1, %xmm0, %xmm0 # sched: [13:1.00] -; SKYLAKE-NEXT: vdivps (%rdi), %xmm0, %xmm0 # sched: [13:1.00] +; SKYLAKE-NEXT: vdivps %xmm1, %xmm0, %xmm0 # sched: [11:1.00] +; SKYLAKE-NEXT: vdivps (%rdi), %xmm0, %xmm0 # sched: [11:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_divps: @@ -958,8 +958,8 @@ define float @test_divss(float %a0, float %a1, float *%a2) { ; ; SKYLAKE-LABEL: test_divss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vdivss %xmm1, %xmm0, %xmm0 # sched: [13:1.00] -; SKYLAKE-NEXT: vdivss (%rdi), %xmm0, %xmm0 # sched: [13:1.00] +; SKYLAKE-NEXT: vdivss %xmm1, %xmm0, %xmm0 # sched: [11:1.00] +; SKYLAKE-NEXT: vdivss (%rdi), %xmm0, %xmm0 # sched: [11:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_divss: @@ -1068,8 +1068,8 @@ define <4 x float> @test_maxps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a ; ; SKYLAKE-LABEL: test_maxps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmaxps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vmaxps (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vmaxps %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vmaxps (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_maxps: @@ -1123,8 +1123,8 @@ define <4 x float> @test_maxss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a ; ; SKYLAKE-LABEL: test_maxss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmaxss %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vmaxss (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vmaxss %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vmaxss (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_maxss: @@ -1178,8 +1178,8 @@ define <4 x float> @test_minps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a ; ; SKYLAKE-LABEL: test_minps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vminps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vminps (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vminps %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vminps (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_minps: @@ -1233,8 +1233,8 @@ define <4 x float> @test_minss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a ; ; SKYLAKE-LABEL: test_minss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vminss %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vminss (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vminss %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vminss (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_minss: @@ -1294,7 +1294,7 @@ define void @test_movaps(<4 x float> *%a0, <4 x float> *%a1) { ; SKYLAKE-LABEL: test_movaps: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovaps (%rdi), %xmm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vaddps %xmm0, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %xmm0, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovaps %xmm0, (%rsi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1412,7 +1412,7 @@ define void @test_movhps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) { ; SKYLAKE-LABEL: test_movhps: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [1:1.00] -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vpextrq $1, %xmm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1475,7 +1475,7 @@ define <4 x float> @test_movlhps(<4 x float> %a0, <4 x float> %a1) { ; SKYLAKE-LABEL: test_movlhps: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00] -; SKYLAKE-NEXT: vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %xmm0, %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_movlhps: @@ -1533,7 +1533,7 @@ define void @test_movlps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) { ; SKYLAKE-LABEL: test_movlps: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [1:1.00] -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovlps %xmm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1590,7 +1590,7 @@ define i32 @test_movmskps(<4 x float> %a0) { ; ; SKYLAKE-LABEL: test_movmskps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmovmskps %xmm0, %eax # sched: [3:1.00] +; SKYLAKE-NEXT: vmovmskps %xmm0, %eax # sched: [2:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_movmskps: @@ -1696,7 +1696,7 @@ define void @test_movss_mem(float* %a0, float* %a1) { ; SKYLAKE-LABEL: test_movss_mem: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [1:0.50] -; SKYLAKE-NEXT: vaddss %xmm0, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddss %xmm0, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovss %xmm0, (%rsi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1753,7 +1753,7 @@ define <4 x float> @test_movss_reg(<4 x float> %a0, <4 x float> %a1) { ; ; SKYLAKE-LABEL: test_movss_reg: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.33] +; SKYLAKE-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_movss_reg: @@ -1808,7 +1808,7 @@ define void @test_movups(<4 x float> *%a0, <4 x float> *%a1) { ; SKYLAKE-LABEL: test_movups: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovups (%rdi), %xmm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vaddps %xmm0, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %xmm0, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovups %xmm0, (%rsi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1864,8 +1864,8 @@ define <4 x float> @test_mulps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a ; ; SKYLAKE-LABEL: test_mulps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmulps %xmm1, %xmm0, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vmulps (%rdi), %xmm0, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vmulps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vmulps (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_mulps: @@ -1918,8 +1918,8 @@ define float @test_mulss(float %a0, float %a1, float *%a2) { ; ; SKYLAKE-LABEL: test_mulss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmulss %xmm1, %xmm0, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vmulss (%rdi), %xmm0, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vmulss %xmm1, %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vmulss (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_mulss: @@ -1976,8 +1976,8 @@ define <4 x float> @test_orps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2 ; ; SKYLAKE-LABEL: test_orps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vorps %xmm1, %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vorps (%rdi), %xmm0, %xmm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vorps %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vorps (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_orps: @@ -2092,9 +2092,9 @@ define <4 x float> @test_rcpps(<4 x float> %a0, <4 x float> *%a1) { ; ; SKYLAKE-LABEL: test_rcpps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vrcpps %xmm0, %xmm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vrcpps (%rdi), %xmm1 # sched: [5:1.00] -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vrcpps %xmm0, %xmm0 # sched: [4:1.00] +; SKYLAKE-NEXT: vrcpps (%rdi), %xmm1 # sched: [4:1.00] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_rcpps: @@ -2163,10 +2163,10 @@ define <4 x float> @test_rcpss(float %a0, float *%a1) { ; ; SKYLAKE-LABEL: test_rcpss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vrcpss %xmm0, %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vrcpss %xmm0, %xmm0, %xmm0 # sched: [4:1.00] ; SKYLAKE-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [1:0.50] -; SKYLAKE-NEXT: vrcpss %xmm1, %xmm1, %xmm1 # sched: [5:1.00] -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vrcpss %xmm1, %xmm1, %xmm1 # sched: [4:1.00] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_rcpss: @@ -2234,9 +2234,9 @@ define <4 x float> @test_rsqrtps(<4 x float> %a0, <4 x float> *%a1) { ; ; SKYLAKE-LABEL: test_rsqrtps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vrsqrtps %xmm0, %xmm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vrsqrtps (%rdi), %xmm1 # sched: [5:1.00] -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vrsqrtps %xmm0, %xmm0 # sched: [4:1.00] +; SKYLAKE-NEXT: vrsqrtps (%rdi), %xmm1 # sched: [4:1.00] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_rsqrtps: @@ -2305,10 +2305,10 @@ define <4 x float> @test_rsqrtss(float %a0, float *%a1) { ; ; SKYLAKE-LABEL: test_rsqrtss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vrsqrtss %xmm0, %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vrsqrtss %xmm0, %xmm0, %xmm0 # sched: [4:1.00] ; SKYLAKE-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [1:0.50] -; SKYLAKE-NEXT: vrsqrtss %xmm1, %xmm1, %xmm1 # sched: [5:1.00] -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vrsqrtss %xmm1, %xmm1, %xmm1 # sched: [4:1.00] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_rsqrtss: @@ -2484,9 +2484,9 @@ define <4 x float> @test_sqrtps(<4 x float> %a0, <4 x float> *%a1) { ; ; SKYLAKE-LABEL: test_sqrtps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vsqrtps %xmm0, %xmm0 # sched: [14:1.00] -; SKYLAKE-NEXT: vsqrtps (%rdi), %xmm1 # sched: [14:1.00] -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vsqrtps %xmm0, %xmm0 # sched: [12:1.00] +; SKYLAKE-NEXT: vsqrtps (%rdi), %xmm1 # sched: [12:1.00] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_sqrtps: @@ -2555,10 +2555,10 @@ define <4 x float> @test_sqrtss(<4 x float> %a0, <4 x float> *%a1) { ; ; SKYLAKE-LABEL: test_sqrtss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vsqrtss %xmm0, %xmm0, %xmm0 # sched: [14:1.00] +; SKYLAKE-NEXT: vsqrtss %xmm0, %xmm0, %xmm0 # sched: [12:1.00] ; SKYLAKE-NEXT: vmovaps (%rdi), %xmm1 # sched: [1:0.50] -; SKYLAKE-NEXT: vsqrtss %xmm1, %xmm1, %xmm1 # sched: [14:1.00] -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vsqrtss %xmm1, %xmm1, %xmm1 # sched: [12:1.00] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_sqrtss: @@ -2673,8 +2673,8 @@ define <4 x float> @test_subps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a ; ; SKYLAKE-LABEL: test_subps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vsubps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vsubps (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vsubps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vsubps (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_subps: @@ -2727,8 +2727,8 @@ define float @test_subss(float %a0, float %a1, float *%a2) { ; ; SKYLAKE-LABEL: test_subss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vsubss %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vsubss (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vsubss %xmm1, %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vsubss (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_subss: @@ -2822,12 +2822,12 @@ define i32 @test_ucomiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) { ; SKYLAKE-LABEL: test_ucomiss: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vucomiss %xmm1, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: setnp %al # sched: [1:0.50] -; SKYLAKE-NEXT: sete %cl # sched: [1:0.50] +; SKYLAKE-NEXT: setnp %al # sched: [1:1.00] +; SKYLAKE-NEXT: sete %cl # sched: [1:1.00] ; SKYLAKE-NEXT: andb %al, %cl # sched: [1:0.25] -; SKYLAKE-NEXT: vucomiss (%rdi), %xmm0 # sched: [7:1.00] -; SKYLAKE-NEXT: setnp %al # sched: [1:0.50] -; SKYLAKE-NEXT: sete %dl # sched: [1:0.50] +; SKYLAKE-NEXT: vucomiss (%rdi), %xmm0 # sched: [8:1.00] +; SKYLAKE-NEXT: setnp %al # sched: [1:1.00] +; SKYLAKE-NEXT: sete %dl # sched: [1:1.00] ; SKYLAKE-NEXT: andb %al, %dl # sched: [1:0.25] ; SKYLAKE-NEXT: orb %cl, %dl # sched: [1:0.25] ; SKYLAKE-NEXT: movzbl %dl, %eax # sched: [1:0.25] @@ -3021,8 +3021,8 @@ define <4 x float> @test_xorps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a ; ; SKYLAKE-LABEL: test_xorps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vxorps %xmm1, %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vxorps (%rdi), %xmm0, %xmm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vxorps %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vxorps (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_xorps: diff --git a/llvm/test/CodeGen/X86/sse2-schedule.ll b/llvm/test/CodeGen/X86/sse2-schedule.ll index 9e46219f34a..8e9649de507 100644 --- a/llvm/test/CodeGen/X86/sse2-schedule.ll +++ b/llvm/test/CodeGen/X86/sse2-schedule.ll @@ -42,8 +42,8 @@ define <2 x double> @test_addpd(<2 x double> %a0, <2 x double> %a1, <2 x double> ; ; SKYLAKE-LABEL: test_addpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vaddpd (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_addpd: @@ -96,8 +96,8 @@ define double @test_addsd(double %a0, double %a1, double *%a2) { ; ; SKYLAKE-LABEL: test_addsd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaddsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddsd (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddsd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vaddsd (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_addsd: @@ -155,9 +155,9 @@ define <2 x double> @test_andpd(<2 x double> %a0, <2 x double> %a1, <2 x double> ; ; SKYLAKE-LABEL: test_andpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vandpd %xmm1, %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vandpd (%rdi), %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vandpd %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vandpd (%rdi), %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vaddpd %xmm0, %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_andpd: @@ -222,9 +222,9 @@ define <2 x double> @test_andnotpd(<2 x double> %a0, <2 x double> %a1, <2 x doub ; ; SKYLAKE-LABEL: test_andnotpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vandnpd %xmm1, %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vandnpd (%rdi), %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vandnpd %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vandnpd (%rdi), %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vaddpd %xmm0, %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_andnotpd: @@ -291,9 +291,9 @@ define <2 x double> @test_cmppd(<2 x double> %a0, <2 x double> %a1, <2 x double> ; ; SKYLAKE-LABEL: test_cmppd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcmpeqpd %xmm1, %xmm0, %xmm1 # sched: [3:1.00] -; SKYLAKE-NEXT: vcmpeqpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vorpd %xmm0, %xmm1, %xmm0 # sched: [1:1.00] +; SKYLAKE-NEXT: vcmpeqpd %xmm1, %xmm0, %xmm1 # sched: [4:0.33] +; SKYLAKE-NEXT: vcmpeqpd (%rdi), %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vorpd %xmm0, %xmm1, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cmppd: @@ -352,7 +352,7 @@ define double @test_cmpsd(double %a0, double %a1, double *%a2) { ; SKYLAKE-LABEL: test_cmpsd: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vcmpeqsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vcmpeqsd (%rdi), %xmm0, %xmm0 # sched: [7:1.00] +; SKYLAKE-NEXT: vcmpeqsd (%rdi), %xmm0, %xmm0 # sched: [8:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cmpsd: @@ -451,12 +451,12 @@ define i32 @test_comisd(<2 x double> %a0, <2 x double> %a1, <2 x double> *%a2) { ; SKYLAKE-LABEL: test_comisd: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vcomisd %xmm1, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: setnp %al # sched: [1:0.50] -; SKYLAKE-NEXT: sete %cl # sched: [1:0.50] +; SKYLAKE-NEXT: setnp %al # sched: [1:1.00] +; SKYLAKE-NEXT: sete %cl # sched: [1:1.00] ; SKYLAKE-NEXT: andb %al, %cl # sched: [1:0.25] -; SKYLAKE-NEXT: vcomisd (%rdi), %xmm0 # sched: [7:1.00] -; SKYLAKE-NEXT: setnp %al # sched: [1:0.50] -; SKYLAKE-NEXT: sete %dl # sched: [1:0.50] +; SKYLAKE-NEXT: vcomisd (%rdi), %xmm0 # sched: [8:1.00] +; SKYLAKE-NEXT: setnp %al # sched: [1:1.00] +; SKYLAKE-NEXT: sete %dl # sched: [1:1.00] ; SKYLAKE-NEXT: andb %al, %dl # sched: [1:0.25] ; SKYLAKE-NEXT: orb %cl, %dl # sched: [1:0.25] ; SKYLAKE-NEXT: movzbl %dl, %eax # sched: [1:0.25] @@ -535,9 +535,9 @@ define <2 x double> @test_cvtdq2pd(<4 x i32> %a0, <4 x i32> *%a1) { ; ; SKYLAKE-LABEL: test_cvtdq2pd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtdq2pd %xmm0, %xmm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vcvtdq2pd (%rdi), %xmm1 # sched: [4:1.00] -; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtdq2pd %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vcvtdq2pd (%rdi), %xmm1 # sched: [5:1.00] +; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cvtdq2pd: @@ -601,9 +601,9 @@ define <4 x float> @test_cvtdq2ps(<4 x i32> %a0, <4 x i32> *%a1) { ; ; SKYLAKE-LABEL: test_cvtdq2ps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtdq2ps %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vcvtdq2ps (%rdi), %xmm1 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtdq2ps %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vcvtdq2ps (%rdi), %xmm1 # sched: [4:0.50] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cvtdq2ps: @@ -665,8 +665,8 @@ define <4 x i32> @test_cvtpd2dq(<2 x double> %a0, <2 x double> *%a1) { ; ; SKYLAKE-LABEL: test_cvtpd2dq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtpd2dq %xmm0, %xmm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vcvtpd2dqx (%rdi), %xmm1 # sched: [7:1.00] +; SKYLAKE-NEXT: vcvtpd2dq %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vcvtpd2dqx (%rdi), %xmm1 # sched: [8:1.00] ; SKYLAKE-NEXT: vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -730,9 +730,9 @@ define <4 x float> @test_cvtpd2ps(<2 x double> %a0, <2 x double> *%a1) { ; ; SKYLAKE-LABEL: test_cvtpd2ps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtpd2ps %xmm0, %xmm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vcvtpd2psx (%rdi), %xmm1 # sched: [7:1.00] -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtpd2ps %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vcvtpd2psx (%rdi), %xmm1 # sched: [8:1.00] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cvtpd2ps: @@ -795,8 +795,8 @@ define <4 x i32> @test_cvtps2dq(<4 x float> %a0, <4 x float> *%a1) { ; ; SKYLAKE-LABEL: test_cvtps2dq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtps2dq %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vcvtps2dq (%rdi), %xmm1 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtps2dq %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vcvtps2dq (%rdi), %xmm1 # sched: [4:0.50] ; SKYLAKE-NEXT: vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -860,9 +860,9 @@ define <2 x double> @test_cvtps2pd(<4 x float> %a0, <4 x float> *%a1) { ; ; SKYLAKE-LABEL: test_cvtps2pd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtps2pd %xmm0, %xmm0 # sched: [2:1.00] -; SKYLAKE-NEXT: vcvtps2pd (%rdi), %xmm1 # sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtps2pd %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vcvtps2pd (%rdi), %xmm1 # sched: [4:0.50] +; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cvtps2pd: @@ -925,8 +925,8 @@ define i32 @test_cvtsd2si(double %a0, double *%a1) { ; ; SKYLAKE-LABEL: test_cvtsd2si: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtsd2si %xmm0, %ecx # sched: [4:1.00] -; SKYLAKE-NEXT: vcvtsd2si (%rdi), %eax # sched: [4:1.00] +; SKYLAKE-NEXT: vcvtsd2si %xmm0, %ecx # sched: [6:1.00] +; SKYLAKE-NEXT: vcvtsd2si (%rdi), %eax # sched: [6:1.00] ; SKYLAKE-NEXT: addl %ecx, %eax # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -991,8 +991,8 @@ define i64 @test_cvtsd2siq(double %a0, double *%a1) { ; ; SKYLAKE-LABEL: test_cvtsd2siq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtsd2si %xmm0, %rcx # sched: [4:1.00] -; SKYLAKE-NEXT: vcvtsd2si (%rdi), %rax # sched: [4:1.00] +; SKYLAKE-NEXT: vcvtsd2si %xmm0, %rcx # sched: [6:1.00] +; SKYLAKE-NEXT: vcvtsd2si (%rdi), %rax # sched: [6:1.00] ; SKYLAKE-NEXT: addq %rcx, %rax # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1063,10 +1063,10 @@ define float @test_cvtsd2ss(double %a0, double *%a1) { ; ; SKYLAKE-LABEL: test_cvtsd2ss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtsd2ss %xmm0, %xmm0, %xmm0 # sched: [4:1.00] +; SKYLAKE-NEXT: vcvtsd2ss %xmm0, %xmm0, %xmm0 # sched: [5:1.00] ; SKYLAKE-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero sched: [1:0.50] -; SKYLAKE-NEXT: vcvtsd2ss %xmm1, %xmm1, %xmm1 # sched: [4:1.00] -; SKYLAKE-NEXT: vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtsd2ss %xmm1, %xmm1, %xmm1 # sched: [5:1.00] +; SKYLAKE-NEXT: vaddss %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cvtsd2ss: @@ -1129,9 +1129,9 @@ define double @test_cvtsi2sd(i32 %a0, i32 *%a1) { ; ; SKYLAKE-LABEL: test_cvtsi2sd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtsi2sdl %edi, %xmm0, %xmm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vcvtsi2sdl (%rsi), %xmm1, %xmm1 # sched: [8:1.00] -; SKYLAKE-NEXT: vaddsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtsi2sdl %edi, %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vcvtsi2sdl (%rsi), %xmm1, %xmm1 # sched: [9:1.00] +; SKYLAKE-NEXT: vaddsd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cvtsi2sd: @@ -1192,9 +1192,9 @@ define double @test_cvtsi2sdq(i64 %a0, i64 *%a1) { ; ; SKYLAKE-LABEL: test_cvtsi2sdq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtsi2sdq %rdi, %xmm0, %xmm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vcvtsi2sdq (%rsi), %xmm1, %xmm1 # sched: [8:1.00] -; SKYLAKE-NEXT: vaddsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtsi2sdq %rdi, %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vcvtsi2sdq (%rsi), %xmm1, %xmm1 # sched: [9:1.00] +; SKYLAKE-NEXT: vaddsd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cvtsi2sdq: @@ -1263,10 +1263,10 @@ define double @test_cvtss2sd(float %a0, float *%a1) { ; ; SKYLAKE-LABEL: test_cvtss2sd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvtss2sd %xmm0, %xmm0, %xmm0 # sched: [2:1.00] +; SKYLAKE-NEXT: vcvtss2sd %xmm0, %xmm0, %xmm0 # sched: [5:1.00] ; SKYLAKE-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [1:0.50] -; SKYLAKE-NEXT: vcvtss2sd %xmm1, %xmm1, %xmm1 # sched: [2:1.00] -; SKYLAKE-NEXT: vaddsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvtss2sd %xmm1, %xmm1, %xmm1 # sched: [5:1.00] +; SKYLAKE-NEXT: vaddsd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_cvtss2sd: @@ -1330,8 +1330,8 @@ define <4 x i32> @test_cvttpd2dq(<2 x double> %a0, <2 x double> *%a1) { ; ; SKYLAKE-LABEL: test_cvttpd2dq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvttpd2dq %xmm0, %xmm0 # sched: [4:1.00] -; SKYLAKE-NEXT: vcvttpd2dqx (%rdi), %xmm1 # sched: [7:1.00] +; SKYLAKE-NEXT: vcvttpd2dq %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vcvttpd2dqx (%rdi), %xmm1 # sched: [8:1.00] ; SKYLAKE-NEXT: vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1396,8 +1396,8 @@ define <4 x i32> @test_cvttps2dq(<4 x float> %a0, <4 x float> *%a1) { ; ; SKYLAKE-LABEL: test_cvttps2dq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvttps2dq %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vcvttps2dq (%rdi), %xmm1 # sched: [3:1.00] +; SKYLAKE-NEXT: vcvttps2dq %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vcvttps2dq (%rdi), %xmm1 # sched: [4:0.50] ; SKYLAKE-NEXT: vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1459,8 +1459,8 @@ define i32 @test_cvttsd2si(double %a0, double *%a1) { ; ; SKYLAKE-LABEL: test_cvttsd2si: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvttsd2si %xmm0, %ecx # sched: [4:1.00] -; SKYLAKE-NEXT: vcvttsd2si (%rdi), %eax # sched: [4:1.00] +; SKYLAKE-NEXT: vcvttsd2si %xmm0, %ecx # sched: [6:1.00] +; SKYLAKE-NEXT: vcvttsd2si (%rdi), %eax # sched: [6:1.00] ; SKYLAKE-NEXT: addl %ecx, %eax # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1522,8 +1522,8 @@ define i64 @test_cvttsd2siq(double %a0, double *%a1) { ; ; SKYLAKE-LABEL: test_cvttsd2siq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vcvttsd2si %xmm0, %rcx # sched: [4:1.00] -; SKYLAKE-NEXT: vcvttsd2si (%rdi), %rax # sched: [4:1.00] +; SKYLAKE-NEXT: vcvttsd2si %xmm0, %rcx # sched: [6:1.00] +; SKYLAKE-NEXT: vcvttsd2si (%rdi), %rax # sched: [6:1.00] ; SKYLAKE-NEXT: addq %rcx, %rax # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1580,8 +1580,8 @@ define <2 x double> @test_divpd(<2 x double> %a0, <2 x double> %a1, <2 x double> ; ; SKYLAKE-LABEL: test_divpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vdivpd %xmm1, %xmm0, %xmm0 # sched: [20:1.00] -; SKYLAKE-NEXT: vdivpd (%rdi), %xmm0, %xmm0 # sched: [20:1.00] +; SKYLAKE-NEXT: vdivpd %xmm1, %xmm0, %xmm0 # sched: [14:1.00] +; SKYLAKE-NEXT: vdivpd (%rdi), %xmm0, %xmm0 # sched: [14:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_divpd: @@ -1634,8 +1634,8 @@ define double @test_divsd(double %a0, double %a1, double *%a2) { ; ; SKYLAKE-LABEL: test_divsd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vdivsd %xmm1, %xmm0, %xmm0 # sched: [20:1.00] -; SKYLAKE-NEXT: vdivsd (%rdi), %xmm0, %xmm0 # sched: [20:1.00] +; SKYLAKE-NEXT: vdivsd %xmm1, %xmm0, %xmm0 # sched: [14:1.00] +; SKYLAKE-NEXT: vdivsd (%rdi), %xmm0, %xmm0 # sched: [14:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_divsd: @@ -1839,8 +1839,8 @@ define <2 x double> @test_maxpd(<2 x double> %a0, <2 x double> %a1, <2 x double> ; ; SKYLAKE-LABEL: test_maxpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmaxpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vmaxpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vmaxpd %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vmaxpd (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_maxpd: @@ -1894,8 +1894,8 @@ define <2 x double> @test_maxsd(<2 x double> %a0, <2 x double> %a1, <2 x double> ; ; SKYLAKE-LABEL: test_maxsd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmaxsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vmaxsd (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vmaxsd %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vmaxsd (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_maxsd: @@ -1949,8 +1949,8 @@ define <2 x double> @test_minpd(<2 x double> %a0, <2 x double> %a1, <2 x double> ; ; SKYLAKE-LABEL: test_minpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vminpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vminpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vminpd %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vminpd (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_minpd: @@ -2004,8 +2004,8 @@ define <2 x double> @test_minsd(<2 x double> %a0, <2 x double> %a1, <2 x double> ; ; SKYLAKE-LABEL: test_minsd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vminsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vminsd (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vminsd %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vminsd (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_minsd: @@ -2065,7 +2065,7 @@ define void @test_movapd(<2 x double> *%a0, <2 x double> *%a1) { ; SKYLAKE-LABEL: test_movapd: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovapd (%rdi), %xmm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vaddpd %xmm0, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %xmm0, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovapd %xmm0, (%rsi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2269,7 +2269,7 @@ define i32 @test_movd(<4 x i32> %a0, i32 %a1, i32 *%a2) { ; SKYLAKE-NEXT: vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [1:0.50] ; SKYLAKE-NEXT: vpaddd %xmm1, %xmm0, %xmm1 # sched: [1:0.50] ; SKYLAKE-NEXT: vpaddd %xmm2, %xmm0, %xmm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vmovd %xmm0, %eax # sched: [1:1.00] +; SKYLAKE-NEXT: vmovd %xmm0, %eax # sched: [2:1.00] ; SKYLAKE-NEXT: vmovd %xmm1, (%rsi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2360,7 +2360,7 @@ define i64 @test_movd_64(<2 x i64> %a0, i64 %a1, i64 *%a2) { ; SKYLAKE-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero sched: [1:0.50] ; SKYLAKE-NEXT: vpaddq %xmm1, %xmm0, %xmm1 # sched: [1:0.50] ; SKYLAKE-NEXT: vpaddq %xmm2, %xmm0, %xmm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vmovq %xmm0, %rax # sched: [1:1.00] +; SKYLAKE-NEXT: vmovq %xmm0, %rax # sched: [2:1.00] ; SKYLAKE-NEXT: vmovq %xmm1, (%rsi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2433,7 +2433,7 @@ define void @test_movhpd(<2 x double> %a0, <2 x double> %a1, x86_mmx *%a2) { ; SKYLAKE-LABEL: test_movhpd: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovhpd %xmm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2498,7 +2498,7 @@ define void @test_movlpd(<2 x double> %a0, <2 x double> %a1, x86_mmx *%a2) { ; SKYLAKE-LABEL: test_movlpd: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovlpd %xmm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2554,7 +2554,7 @@ define i32 @test_movmskpd(<2 x double> %a0) { ; ; SKYLAKE-LABEL: test_movmskpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmovmskpd %xmm0, %eax # sched: [3:1.00] +; SKYLAKE-NEXT: vmovmskpd %xmm0, %eax # sched: [2:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_movmskpd: @@ -2659,7 +2659,7 @@ define void @test_movntpd(<2 x double> %a0, <2 x double> *%a1) { ; ; SKYLAKE-LABEL: test_movntpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaddpd %xmm0, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %xmm0, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovntpd %xmm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2778,7 +2778,7 @@ define <2 x i64> @test_movq_reg(<2 x i64> %a0, <2 x i64> %a1) { ; ; SKYLAKE-LABEL: test_movq_reg: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero sched: [1:0.33] +; SKYLAKE-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero sched: [1:0.50] ; SKYLAKE-NEXT: vpaddq %xmm0, %xmm1, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2837,7 +2837,7 @@ define void @test_movsd_mem(double* %a0, double* %a1) { ; SKYLAKE-LABEL: test_movsd_mem: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero sched: [1:0.50] -; SKYLAKE-NEXT: vaddsd %xmm0, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddsd %xmm0, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovsd %xmm0, (%rsi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2950,7 +2950,7 @@ define void @test_movupd(<2 x double> *%a0, <2 x double> *%a1) { ; SKYLAKE-LABEL: test_movupd: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovupd (%rdi), %xmm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vaddpd %xmm0, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %xmm0, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vmovupd %xmm0, (%rsi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -3006,8 +3006,8 @@ define <2 x double> @test_mulpd(<2 x double> %a0, <2 x double> %a1, <2 x double> ; ; SKYLAKE-LABEL: test_mulpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmulpd %xmm1, %xmm0, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vmulpd (%rdi), %xmm0, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vmulpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vmulpd (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_mulpd: @@ -3060,8 +3060,8 @@ define double @test_mulsd(double %a0, double %a1, double *%a2) { ; ; SKYLAKE-LABEL: test_mulsd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmulsd %xmm1, %xmm0, %xmm0 # sched: [5:0.50] -; SKYLAKE-NEXT: vmulsd (%rdi), %xmm0, %xmm0 # sched: [5:0.50] +; SKYLAKE-NEXT: vmulsd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vmulsd (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_mulsd: @@ -3119,9 +3119,9 @@ define <2 x double> @test_orpd(<2 x double> %a0, <2 x double> %a1, <2 x double> ; ; SKYLAKE-LABEL: test_orpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vorpd %xmm1, %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vorpd (%rdi), %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vorpd %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vorpd (%rdi), %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vaddpd %xmm0, %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_orpd: @@ -3535,7 +3535,7 @@ define <16 x i8> @test_paddsb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_paddsb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpaddsb %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpaddsb %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpaddsb (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -3594,7 +3594,7 @@ define <8 x i16> @test_paddsw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_paddsw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpaddsw %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpaddsw %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpaddsw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -3653,7 +3653,7 @@ define <16 x i8> @test_paddusb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_paddusb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpaddusb %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpaddusb %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpaddusb (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -3712,7 +3712,7 @@ define <8 x i16> @test_paddusw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_paddusw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpaddusw %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpaddusw %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpaddusw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -3830,7 +3830,7 @@ define <2 x i64> @test_pand(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_pand: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpand %xmm1, %xmm0, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpand %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpand (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] @@ -3899,7 +3899,7 @@ define <2 x i64> @test_pandn(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_pandn: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpandn %xmm1, %xmm0, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpandn %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpandn (%rdi), %xmm0, %xmm1 # sched: [1:0.50] ; SKYLAKE-NEXT: vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] @@ -3963,7 +3963,7 @@ define <16 x i8> @test_pavgb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pavgb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpavgb %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpavgb %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpavgb (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4031,7 +4031,7 @@ define <8 x i16> @test_pavgw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_pavgw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpavgw %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpavgw %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpavgw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4102,9 +4102,9 @@ define <16 x i8> @test_pcmpeqb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpeqb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm1 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm1 # sched: [1:1.00] ; SKYLAKE-NEXT: vpcmpeqb (%rdi), %xmm0, %xmm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pcmpeqb: @@ -4168,9 +4168,9 @@ define <4 x i32> @test_pcmpeqd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpeqd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm1 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm1 # sched: [1:1.00] ; SKYLAKE-NEXT: vpcmpeqd (%rdi), %xmm0, %xmm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pcmpeqd: @@ -4234,9 +4234,9 @@ define <8 x i16> @test_pcmpeqw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpeqw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm1 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm1 # sched: [1:1.00] ; SKYLAKE-NEXT: vpcmpeqw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pcmpeqw: @@ -4301,9 +4301,9 @@ define <16 x i8> @test_pcmpgtb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpgtb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm1 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm1 # sched: [1:1.00] ; SKYLAKE-NEXT: vpcmpgtb (%rdi), %xmm0, %xmm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pcmpgtb: @@ -4368,9 +4368,9 @@ define <4 x i32> @test_pcmpgtd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpgtd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm1 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm1 # sched: [1:1.00] ; SKYLAKE-NEXT: vpcmpeqd (%rdi), %xmm0, %xmm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pcmpgtd: @@ -4435,9 +4435,9 @@ define <8 x i16> @test_pcmpgtw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpgtw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm1 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm1 # sched: [1:1.00] ; SKYLAKE-NEXT: vpcmpgtw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] -; SKYLAKE-NEXT: vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pcmpgtw: @@ -4494,7 +4494,7 @@ define i16 @test_pextrw(<8 x i16> %a0) { ; ; SKYLAKE-LABEL: test_pextrw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpextrw $6, %xmm0, %eax # sched: [2:1.00] +; SKYLAKE-NEXT: vpextrw $6, %xmm0, %eax # sched: [3:1.00] ; SKYLAKE-NEXT: # kill: %AX<def> %AX<kill> %EAX<kill> ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4612,8 +4612,8 @@ define <4 x i32> @test_pmaddwd(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_pmaddwd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaddwd %xmm1, %xmm0, %xmm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmaddwd (%rdi), %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmaddwd %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmaddwd (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pmaddwd: @@ -4672,7 +4672,7 @@ define <8 x i16> @test_pmaxsw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_pmaxsw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaxsw %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaxsw %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpmaxsw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4731,7 +4731,7 @@ define <16 x i8> @test_pmaxub(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pmaxub: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaxub %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaxub %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpmaxub (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4790,7 +4790,7 @@ define <8 x i16> @test_pminsw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_pminsw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpminsw %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpminsw %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpminsw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4849,7 +4849,7 @@ define <16 x i8> @test_pminub(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pminub: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpminub %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpminub %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpminub (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -4901,7 +4901,7 @@ define i32 @test_pmovmskb(<16 x i8> %a0) { ; ; SKYLAKE-LABEL: test_pmovmskb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmovmskb %xmm0, %eax # sched: [3:1.00] +; SKYLAKE-NEXT: vpmovmskb %xmm0, %eax # sched: [2:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pmovmskb: @@ -4951,8 +4951,8 @@ define <8 x i16> @test_pmulhuw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_pmulhuw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmulhuw %xmm1, %xmm0, %xmm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmulhuw (%rdi), %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmulhuw %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmulhuw (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pmulhuw: @@ -5006,8 +5006,8 @@ define <8 x i16> @test_pmulhw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_pmulhw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmulhw %xmm1, %xmm0, %xmm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmulhw (%rdi), %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmulhw %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmulhw (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pmulhw: @@ -5061,8 +5061,8 @@ define <8 x i16> @test_pmullw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_pmullw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmullw %xmm1, %xmm0, %xmm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmullw (%rdi), %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmullw %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmullw (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pmullw: @@ -5123,8 +5123,8 @@ define <2 x i64> @test_pmuludq(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pmuludq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmuludq %xmm1, %xmm0, %xmm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmuludq (%rdi), %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmuludq %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmuludq (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pmuludq: @@ -5184,7 +5184,7 @@ define <2 x i64> @test_por(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_por: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpor (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] @@ -5250,8 +5250,8 @@ define <2 x i64> @test_psadbw(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_psadbw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsadbw %xmm1, %xmm0, %xmm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpsadbw (%rdi), %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpsadbw %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vpsadbw (%rdi), %xmm0, %xmm0 # sched: [3:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_psadbw: @@ -5507,7 +5507,7 @@ define <4 x i32> @test_pslld(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; SKYLAKE-LABEL: test_pslld: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpslld %xmm1, %xmm0, %xmm0 # sched: [2:1.00] -; SKYLAKE-NEXT: vpslld (%rdi), %xmm0, %xmm0 # sched: [2:1.00] +; SKYLAKE-NEXT: vpslld (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpslld $2, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -5622,7 +5622,7 @@ define <2 x i64> @test_psllq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; SKYLAKE-LABEL: test_psllq: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsllq %xmm1, %xmm0, %xmm0 # sched: [2:1.00] -; SKYLAKE-NEXT: vpsllq (%rdi), %xmm0, %xmm0 # sched: [2:1.00] +; SKYLAKE-NEXT: vpsllq (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpsllq $2, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -5687,7 +5687,7 @@ define <8 x i16> @test_psllw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; SKYLAKE-LABEL: test_psllw: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsllw %xmm1, %xmm0, %xmm0 # sched: [2:1.00] -; SKYLAKE-NEXT: vpsllw (%rdi), %xmm0, %xmm0 # sched: [2:1.00] +; SKYLAKE-NEXT: vpsllw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpsllw $2, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -5752,7 +5752,7 @@ define <4 x i32> @test_psrad(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; SKYLAKE-LABEL: test_psrad: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsrad %xmm1, %xmm0, %xmm0 # sched: [2:1.00] -; SKYLAKE-NEXT: vpsrad (%rdi), %xmm0, %xmm0 # sched: [2:1.00] +; SKYLAKE-NEXT: vpsrad (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpsrad $2, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -5817,7 +5817,7 @@ define <8 x i16> @test_psraw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; SKYLAKE-LABEL: test_psraw: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsraw %xmm1, %xmm0, %xmm0 # sched: [2:1.00] -; SKYLAKE-NEXT: vpsraw (%rdi), %xmm0, %xmm0 # sched: [2:1.00] +; SKYLAKE-NEXT: vpsraw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpsraw $2, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -5882,7 +5882,7 @@ define <4 x i32> @test_psrld(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; SKYLAKE-LABEL: test_psrld: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsrld %xmm1, %xmm0, %xmm0 # sched: [2:1.00] -; SKYLAKE-NEXT: vpsrld (%rdi), %xmm0, %xmm0 # sched: [2:1.00] +; SKYLAKE-NEXT: vpsrld (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpsrld $2, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -5997,7 +5997,7 @@ define <2 x i64> @test_psrlq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; SKYLAKE-LABEL: test_psrlq: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsrlq %xmm1, %xmm0, %xmm0 # sched: [2:1.00] -; SKYLAKE-NEXT: vpsrlq (%rdi), %xmm0, %xmm0 # sched: [2:1.00] +; SKYLAKE-NEXT: vpsrlq (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpsrlq $2, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -6062,7 +6062,7 @@ define <8 x i16> @test_psrlw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; SKYLAKE-LABEL: test_psrlw: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpsrlw %xmm1, %xmm0, %xmm0 # sched: [2:1.00] -; SKYLAKE-NEXT: vpsrlw (%rdi), %xmm0, %xmm0 # sched: [2:1.00] +; SKYLAKE-NEXT: vpsrlw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpsrlw $2, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -6295,7 +6295,7 @@ define <16 x i8> @test_psubsb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_psubsb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsubsb %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpsubsb %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsubsb (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -6354,7 +6354,7 @@ define <8 x i16> @test_psubsw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_psubsw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsubsw %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpsubsw %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsubsw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -6413,7 +6413,7 @@ define <16 x i8> @test_psubusb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_psubusb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsubusb %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpsubusb %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsubusb (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -6472,7 +6472,7 @@ define <8 x i16> @test_psubusw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_psubusw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsubusw %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpsubusw %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsubusw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -7078,7 +7078,7 @@ define <2 x i64> @test_pxor(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_pxor: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpxor %xmm1, %xmm0, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpxor %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpxor (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] @@ -7143,7 +7143,7 @@ define <2 x double> @test_shufpd(<2 x double> %a0, <2 x double> %a1, <2 x double ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vshufpd {{.*#+}} xmm0 = xmm0[1],xmm1[0] sched: [1:1.00] ; SKYLAKE-NEXT: vshufpd {{.*#+}} xmm1 = xmm1[1],mem[0] sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_shufpd: @@ -7205,9 +7205,9 @@ define <2 x double> @test_sqrtpd(<2 x double> %a0, <2 x double> *%a1) { ; ; SKYLAKE-LABEL: test_sqrtpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vsqrtpd %xmm0, %xmm0 # sched: [21:1.00] -; SKYLAKE-NEXT: vsqrtpd (%rdi), %xmm1 # sched: [21:1.00] -; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vsqrtpd %xmm0, %xmm0 # sched: [18:1.00] +; SKYLAKE-NEXT: vsqrtpd (%rdi), %xmm1 # sched: [18:1.00] +; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_sqrtpd: @@ -7276,10 +7276,10 @@ define <2 x double> @test_sqrtsd(<2 x double> %a0, <2 x double> *%a1) { ; ; SKYLAKE-LABEL: test_sqrtsd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vsqrtsd %xmm0, %xmm0, %xmm0 # sched: [21:1.00] +; SKYLAKE-NEXT: vsqrtsd %xmm0, %xmm0, %xmm0 # sched: [18:1.00] ; SKYLAKE-NEXT: vmovapd (%rdi), %xmm1 # sched: [1:0.50] -; SKYLAKE-NEXT: vsqrtsd %xmm1, %xmm1, %xmm1 # sched: [21:1.00] -; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vsqrtsd %xmm1, %xmm1, %xmm1 # sched: [18:1.00] +; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_sqrtsd: @@ -7338,8 +7338,8 @@ define <2 x double> @test_subpd(<2 x double> %a0, <2 x double> %a1, <2 x double> ; ; SKYLAKE-LABEL: test_subpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vsubpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vsubpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vsubpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vsubpd (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_subpd: @@ -7392,8 +7392,8 @@ define double @test_subsd(double %a0, double %a1, double *%a2) { ; ; SKYLAKE-LABEL: test_subsd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vsubsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vsubsd (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vsubsd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vsubsd (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_subsd: @@ -7487,12 +7487,12 @@ define i32 @test_ucomisd(<2 x double> %a0, <2 x double> %a1, <2 x double> *%a2) ; SKYLAKE-LABEL: test_ucomisd: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vucomisd %xmm1, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: setnp %al # sched: [1:0.50] -; SKYLAKE-NEXT: sete %cl # sched: [1:0.50] +; SKYLAKE-NEXT: setnp %al # sched: [1:1.00] +; SKYLAKE-NEXT: sete %cl # sched: [1:1.00] ; SKYLAKE-NEXT: andb %al, %cl # sched: [1:0.25] -; SKYLAKE-NEXT: vucomisd (%rdi), %xmm0 # sched: [7:1.00] -; SKYLAKE-NEXT: setnp %al # sched: [1:0.50] -; SKYLAKE-NEXT: sete %dl # sched: [1:0.50] +; SKYLAKE-NEXT: vucomisd (%rdi), %xmm0 # sched: [8:1.00] +; SKYLAKE-NEXT: setnp %al # sched: [1:1.00] +; SKYLAKE-NEXT: sete %dl # sched: [1:1.00] ; SKYLAKE-NEXT: andb %al, %dl # sched: [1:0.25] ; SKYLAKE-NEXT: orb %cl, %dl # sched: [1:0.25] ; SKYLAKE-NEXT: movzbl %dl, %eax # sched: [1:0.25] @@ -7573,7 +7573,7 @@ define <2 x double> @test_unpckhpd(<2 x double> %a0, <2 x double> %a1, <2 x doub ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1] sched: [1:1.00] ; SKYLAKE-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],mem[1] sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_unpckhpd: @@ -7642,7 +7642,7 @@ define <2 x double> @test_unpcklpd(<2 x double> %a0, <2 x double> %a1, <2 x doub ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00] ; SKYLAKE-NEXT: vunpcklpd {{.*#+}} xmm1 = xmm0[0],mem[0] sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_unpcklpd: @@ -7703,9 +7703,9 @@ define <2 x double> @test_xorpd(<2 x double> %a0, <2 x double> %a1, <2 x double> ; ; SKYLAKE-LABEL: test_xorpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vxorpd %xmm1, %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vxorpd (%rdi), %xmm0, %xmm0 # sched: [1:1.00] -; SKYLAKE-NEXT: vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vxorpd %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vxorpd (%rdi), %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vaddpd %xmm0, %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_xorpd: diff --git a/llvm/test/CodeGen/X86/sse3-schedule.ll b/llvm/test/CodeGen/X86/sse3-schedule.ll index 7f6d0221f55..e8bd6fe5c60 100644 --- a/llvm/test/CodeGen/X86/sse3-schedule.ll +++ b/llvm/test/CodeGen/X86/sse3-schedule.ll @@ -42,8 +42,8 @@ define <2 x double> @test_addsubpd(<2 x double> %a0, <2 x double> %a1, <2 x doub ; ; SKYLAKE-LABEL: test_addsubpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaddsubpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddsubpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddsubpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vaddsubpd (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_addsubpd: @@ -97,8 +97,8 @@ define <4 x float> @test_addsubps(<4 x float> %a0, <4 x float> %a1, <4 x float> ; ; SKYLAKE-LABEL: test_addsubps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vaddsubps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] -; SKYLAKE-NEXT: vaddsubps (%rdi), %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddsubps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vaddsubps (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_addsubps: @@ -152,8 +152,8 @@ define <2 x double> @test_haddpd(<2 x double> %a0, <2 x double> %a1, <2 x double ; ; SKYLAKE-LABEL: test_haddpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vhaddpd %xmm1, %xmm0, %xmm0 # sched: [5:2.00] -; SKYLAKE-NEXT: vhaddpd (%rdi), %xmm0, %xmm0 # sched: [5:2.00] +; SKYLAKE-NEXT: vhaddpd %xmm1, %xmm0, %xmm0 # sched: [6:2.00] +; SKYLAKE-NEXT: vhaddpd (%rdi), %xmm0, %xmm0 # sched: [6:2.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_haddpd: @@ -207,8 +207,8 @@ define <4 x float> @test_haddps(<4 x float> %a0, <4 x float> %a1, <4 x float> *% ; ; SKYLAKE-LABEL: test_haddps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vhaddps %xmm1, %xmm0, %xmm0 # sched: [5:2.00] -; SKYLAKE-NEXT: vhaddps (%rdi), %xmm0, %xmm0 # sched: [5:2.00] +; SKYLAKE-NEXT: vhaddps %xmm1, %xmm0, %xmm0 # sched: [6:2.00] +; SKYLAKE-NEXT: vhaddps (%rdi), %xmm0, %xmm0 # sched: [6:2.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_haddps: @@ -262,8 +262,8 @@ define <2 x double> @test_hsubpd(<2 x double> %a0, <2 x double> %a1, <2 x double ; ; SKYLAKE-LABEL: test_hsubpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vhsubpd %xmm1, %xmm0, %xmm0 # sched: [5:2.00] -; SKYLAKE-NEXT: vhsubpd (%rdi), %xmm0, %xmm0 # sched: [5:2.00] +; SKYLAKE-NEXT: vhsubpd %xmm1, %xmm0, %xmm0 # sched: [6:2.00] +; SKYLAKE-NEXT: vhsubpd (%rdi), %xmm0, %xmm0 # sched: [6:2.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_hsubpd: @@ -317,8 +317,8 @@ define <4 x float> @test_hsubps(<4 x float> %a0, <4 x float> %a1, <4 x float> *% ; ; SKYLAKE-LABEL: test_hsubps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vhsubps %xmm1, %xmm0, %xmm0 # sched: [5:2.00] -; SKYLAKE-NEXT: vhsubps (%rdi), %xmm0, %xmm0 # sched: [5:2.00] +; SKYLAKE-NEXT: vhsubps %xmm1, %xmm0, %xmm0 # sched: [6:2.00] +; SKYLAKE-NEXT: vhsubps (%rdi), %xmm0, %xmm0 # sched: [6:2.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_hsubps: @@ -488,7 +488,7 @@ define <2 x double> @test_movddup(<2 x double> %a0, <2 x double> *%a1) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovddup {{.*#+}} xmm0 = xmm0[0,0] sched: [1:1.00] ; SKYLAKE-NEXT: vmovddup {{.*#+}} xmm1 = mem[0,0] sched: [1:0.50] -; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_movddup: @@ -552,7 +552,7 @@ define <4 x float> @test_movshdup(<4 x float> %a0, <4 x float> *%a1) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovshdup {{.*#+}} xmm0 = xmm0[1,1,3,3] sched: [1:1.00] ; SKYLAKE-NEXT: vmovshdup {{.*#+}} xmm1 = mem[1,1,3,3] sched: [1:0.50] -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_movshdup: @@ -616,7 +616,7 @@ define <4 x float> @test_movsldup(<4 x float> %a0, <4 x float> *%a1) { ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vmovsldup {{.*#+}} xmm0 = xmm0[0,0,2,2] sched: [1:1.00] ; SKYLAKE-NEXT: vmovsldup {{.*#+}} xmm1 = mem[0,0,2,2] sched: [1:0.50] -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_movsldup: diff --git a/llvm/test/CodeGen/X86/sse41-schedule.ll b/llvm/test/CodeGen/X86/sse41-schedule.ll index 6d1430d98b1..824b688eb98 100644 --- a/llvm/test/CodeGen/X86/sse41-schedule.ll +++ b/llvm/test/CodeGen/X86/sse41-schedule.ll @@ -39,8 +39,8 @@ define <2 x double> @test_blendpd(<2 x double> %a0, <2 x double> %a1, <2 x doubl ; ; SKYLAKE-LABEL: test_blendpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vblendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1] sched: [1:0.33] -; SKYLAKE-NEXT: vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vblendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1] sched: [1:0.50] +; SKYLAKE-NEXT: vaddpd %xmm0, %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: vblendpd {{.*#+}} xmm0 = xmm0[0],mem[1] sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -91,7 +91,7 @@ define <4 x float> @test_blendps(<4 x float> %a0, <4 x float> %a1, <4 x float> * ; ; SKYLAKE-LABEL: test_blendps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2],xmm0[3] sched: [1:0.33] +; SKYLAKE-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2],xmm0[3] sched: [1:0.50] ; SKYLAKE-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2,3] sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -145,8 +145,8 @@ define <2 x double> @test_blendvpd(<2 x double> %a0, <2 x double> %a1, <2 x doub ; ; SKYLAKE-LABEL: test_blendvpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vblendvpd %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:2.00] -; SKYLAKE-NEXT: vblendvpd %xmm2, (%rdi), %xmm0, %xmm0 # sched: [2:2.00] +; SKYLAKE-NEXT: vblendvpd %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:0.67] +; SKYLAKE-NEXT: vblendvpd %xmm2, (%rdi), %xmm0, %xmm0 # sched: [2:0.67] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_blendvpd: @@ -200,8 +200,8 @@ define <4 x float> @test_blendvps(<4 x float> %a0, <4 x float> %a1, <4 x float> ; ; SKYLAKE-LABEL: test_blendvps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vblendvps %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:2.00] -; SKYLAKE-NEXT: vblendvps %xmm2, (%rdi), %xmm0, %xmm0 # sched: [2:2.00] +; SKYLAKE-NEXT: vblendvps %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:0.67] +; SKYLAKE-NEXT: vblendvps %xmm2, (%rdi), %xmm0, %xmm0 # sched: [2:0.67] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_blendvps: @@ -298,8 +298,8 @@ define <4 x float> @test_dpps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2 ; ; SKYLAKE-LABEL: test_dpps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vdpps $7, %xmm1, %xmm0, %xmm0 # sched: [14:2.00] -; SKYLAKE-NEXT: vdpps $7, (%rdi), %xmm0, %xmm0 # sched: [14:2.00] +; SKYLAKE-NEXT: vdpps $7, %xmm1, %xmm0, %xmm0 # sched: [13:1.33] +; SKYLAKE-NEXT: vdpps $7, (%rdi), %xmm0, %xmm0 # sched: [13:1.33] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_dpps: @@ -436,8 +436,8 @@ define <8 x i16> @test_mpsadbw(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_mpsadbw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vmpsadbw $7, %xmm1, %xmm0, %xmm0 # sched: [7:2.00] -; SKYLAKE-NEXT: vmpsadbw $7, (%rdi), %xmm0, %xmm0 # sched: [7:2.00] +; SKYLAKE-NEXT: vmpsadbw $7, %xmm1, %xmm0, %xmm0 # sched: [4:2.00] +; SKYLAKE-NEXT: vmpsadbw $7, (%rdi), %xmm0, %xmm0 # sched: [4:2.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_mpsadbw: @@ -542,8 +542,8 @@ define <16 x i8> @test_pblendvb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> %a2, <16 ; ; SKYLAKE-LABEL: test_pblendvb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpblendvb %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:2.00] -; SKYLAKE-NEXT: vpblendvb %xmm2, (%rdi), %xmm0, %xmm0 # sched: [2:2.00] +; SKYLAKE-NEXT: vpblendvb %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:0.67] +; SKYLAKE-NEXT: vpblendvb %xmm2, (%rdi), %xmm0, %xmm0 # sched: [2:0.67] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pblendvb: @@ -592,7 +592,7 @@ define <8 x i16> @test_pblendw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; SKYLAKE-LABEL: test_pblendw: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5],xmm0[6],xmm1[7] sched: [1:1.00] -; SKYLAKE-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],mem[2,3],xmm0[4,5,6],mem[7] sched: [4:1.00] +; SKYLAKE-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],mem[2,3],xmm0[4,5,6],mem[7] sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pblendw: @@ -639,7 +639,7 @@ define <2 x i64> @test_pcmpeqq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpeqq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpeqq %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpcmpeqq %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpcmpeqq (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -689,7 +689,7 @@ define i32 @test_pextrb(<16 x i8> %a0, i8 *%a1) { ; ; SKYLAKE-LABEL: test_pextrb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpextrb $3, %xmm0, %eax # sched: [2:1.00] +; SKYLAKE-NEXT: vpextrb $3, %xmm0, %eax # sched: [3:1.00] ; SKYLAKE-NEXT: vpextrb $1, %xmm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -738,7 +738,7 @@ define i32 @test_pextrd(<4 x i32> %a0, i32 *%a1) { ; ; SKYLAKE-LABEL: test_pextrd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpextrd $3, %xmm0, %eax # sched: [2:1.00] +; SKYLAKE-NEXT: vpextrd $3, %xmm0, %eax # sched: [3:1.00] ; SKYLAKE-NEXT: vpextrd $1, %xmm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -786,7 +786,7 @@ define i64 @test_pextrq(<2 x i64> %a0, <2 x i64> %a1, i64 *%a2) { ; ; SKYLAKE-LABEL: test_pextrq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpextrq $1, %xmm0, %rax # sched: [2:1.00] +; SKYLAKE-NEXT: vpextrq $1, %xmm0, %rax # sched: [3:1.00] ; SKYLAKE-NEXT: vpextrq $1, %xmm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -834,7 +834,7 @@ define i32 @test_pextrw(<8 x i16> %a0, i16 *%a1) { ; ; SKYLAKE-LABEL: test_pextrw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpextrw $3, %xmm0, %eax # sched: [2:1.00] +; SKYLAKE-NEXT: vpextrw $3, %xmm0, %eax # sched: [3:1.00] ; SKYLAKE-NEXT: vpextrw $1, %xmm0, (%rdi) # sched: [1:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -883,8 +883,8 @@ define <8 x i16> @test_phminposuw(<8 x i16> *%a0) { ; ; SKYLAKE-LABEL: test_phminposuw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vphminposuw (%rdi), %xmm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vphminposuw %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vphminposuw (%rdi), %xmm0 # sched: [4:0.50] +; SKYLAKE-NEXT: vphminposuw %xmm0, %xmm0 # sched: [4:0.33] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_phminposuw: @@ -1084,7 +1084,7 @@ define <16 x i8> @test_pmaxsb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pmaxsb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaxsb %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaxsb %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpmaxsb (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1133,7 +1133,7 @@ define <4 x i32> @test_pmaxsd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pmaxsd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpmaxsd (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1182,7 +1182,7 @@ define <4 x i32> @test_pmaxud(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pmaxud: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaxud %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaxud %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpmaxud (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1231,7 +1231,7 @@ define <8 x i16> @test_pmaxuw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_pmaxuw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaxuw %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpmaxuw %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpmaxuw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1280,7 +1280,7 @@ define <16 x i8> @test_pminsb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pminsb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpminsb %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpminsb %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpminsb (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1329,7 +1329,7 @@ define <4 x i32> @test_pminsd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pminsd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpminsd %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpminsd %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpminsd (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1378,7 +1378,7 @@ define <4 x i32> @test_pminud(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pminud: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpminud %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpminud %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpminud (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -1427,7 +1427,7 @@ define <8 x i16> @test_pminuw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_pminuw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpminuw %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpminuw %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpminuw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -2172,8 +2172,8 @@ define <2 x i64> @test_pmuldq(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pmuldq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmuldq %xmm1, %xmm0, %xmm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmuldq (%rdi), %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmuldq %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmuldq (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pmuldq: @@ -2222,8 +2222,8 @@ define <4 x i32> @test_pmulld(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_pmulld: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmulld %xmm1, %xmm0, %xmm0 # sched: [10:2.00] -; SKYLAKE-NEXT: vpmulld (%rdi), %xmm0, %xmm0 # sched: [10:2.00] +; SKYLAKE-NEXT: vpmulld %xmm1, %xmm0, %xmm0 # sched: [8:0.67] +; SKYLAKE-NEXT: vpmulld (%rdi), %xmm0, %xmm0 # sched: [8:0.67] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pmulld: @@ -2286,10 +2286,10 @@ define i32 @test_ptest(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_ptest: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vptest %xmm1, %xmm0 # sched: [2:1.00] -; SKYLAKE-NEXT: setb %al # sched: [1:0.50] -; SKYLAKE-NEXT: vptest (%rdi), %xmm0 # sched: [2:1.00] -; SKYLAKE-NEXT: setb %cl # sched: [1:0.50] +; SKYLAKE-NEXT: vptest %xmm1, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: setb %al # sched: [1:1.00] +; SKYLAKE-NEXT: vptest (%rdi), %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: setb %cl # sched: [1:1.00] ; SKYLAKE-NEXT: andb %al, %cl # sched: [1:0.25] ; SKYLAKE-NEXT: movzbl %cl, %eax # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] @@ -2353,9 +2353,9 @@ define <2 x double> @test_roundpd(<2 x double> %a0, <2 x double> *%a1) { ; ; SKYLAKE-LABEL: test_roundpd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vroundpd $7, %xmm0, %xmm0 # sched: [5:1.25] -; SKYLAKE-NEXT: vroundpd $7, (%rdi), %xmm1 # sched: [6:2.00] -; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vroundpd $7, %xmm0, %xmm0 # sched: [8:0.67] +; SKYLAKE-NEXT: vroundpd $7, (%rdi), %xmm1 # sched: [8:0.67] +; SKYLAKE-NEXT: vaddpd %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_roundpd: @@ -2411,9 +2411,9 @@ define <4 x float> @test_roundps(<4 x float> %a0, <4 x float> *%a1) { ; ; SKYLAKE-LABEL: test_roundps: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vroundps $7, %xmm0, %xmm0 # sched: [5:1.25] -; SKYLAKE-NEXT: vroundps $7, (%rdi), %xmm1 # sched: [6:2.00] -; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vroundps $7, %xmm0, %xmm0 # sched: [8:0.67] +; SKYLAKE-NEXT: vroundps $7, (%rdi), %xmm1 # sched: [8:0.67] +; SKYLAKE-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_roundps: @@ -2470,9 +2470,9 @@ define <2 x double> @test_roundsd(<2 x double> %a0, <2 x double> %a1, <2 x doubl ; ; SKYLAKE-LABEL: test_roundsd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vroundsd $7, %xmm1, %xmm0, %xmm1 # sched: [5:1.25] -; SKYLAKE-NEXT: vroundsd $7, (%rdi), %xmm0, %xmm0 # sched: [6:2.00] -; SKYLAKE-NEXT: vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vroundsd $7, %xmm1, %xmm0, %xmm1 # sched: [8:0.67] +; SKYLAKE-NEXT: vroundsd $7, (%rdi), %xmm0, %xmm0 # sched: [8:0.67] +; SKYLAKE-NEXT: vaddpd %xmm0, %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_roundsd: @@ -2529,9 +2529,9 @@ define <4 x float> @test_roundss(<4 x float> %a0, <4 x float> %a1, <4 x float> * ; ; SKYLAKE-LABEL: test_roundss: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vroundss $7, %xmm1, %xmm0, %xmm1 # sched: [5:1.25] -; SKYLAKE-NEXT: vroundss $7, (%rdi), %xmm0, %xmm0 # sched: [6:2.00] -; SKYLAKE-NEXT: vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vroundss $7, %xmm1, %xmm0, %xmm1 # sched: [8:0.67] +; SKYLAKE-NEXT: vroundss $7, (%rdi), %xmm0, %xmm0 # sched: [8:0.67] +; SKYLAKE-NEXT: vaddps %xmm0, %xmm1, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_roundss: diff --git a/llvm/test/CodeGen/X86/sse42-schedule.ll b/llvm/test/CodeGen/X86/sse42-schedule.ll index f625dccead2..7b9471c7048 100644 --- a/llvm/test/CodeGen/X86/sse42-schedule.ll +++ b/llvm/test/CodeGen/X86/sse42-schedule.ll @@ -40,7 +40,7 @@ define i32 @crc32_32_8(i32 %a0, i8 %a1, i8 *%a2) { ; SKYLAKE-LABEL: crc32_32_8: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: crc32b %sil, %edi # sched: [3:1.00] -; SKYLAKE-NEXT: crc32b (%rdx), %edi # sched: [7:1.00] +; SKYLAKE-NEXT: crc32b (%rdx), %edi # sched: [8:1.00] ; SKYLAKE-NEXT: movl %edi, %eax # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -96,7 +96,7 @@ define i32 @crc32_32_16(i32 %a0, i16 %a1, i16 *%a2) { ; SKYLAKE-LABEL: crc32_32_16: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: crc32w %si, %edi # sched: [3:1.00] -; SKYLAKE-NEXT: crc32w (%rdx), %edi # sched: [7:1.00] +; SKYLAKE-NEXT: crc32w (%rdx), %edi # sched: [8:1.00] ; SKYLAKE-NEXT: movl %edi, %eax # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -152,7 +152,7 @@ define i32 @crc32_32_32(i32 %a0, i32 %a1, i32 *%a2) { ; SKYLAKE-LABEL: crc32_32_32: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: crc32l %esi, %edi # sched: [3:1.00] -; SKYLAKE-NEXT: crc32l (%rdx), %edi # sched: [7:1.00] +; SKYLAKE-NEXT: crc32l (%rdx), %edi # sched: [8:1.00] ; SKYLAKE-NEXT: movl %edi, %eax # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -208,7 +208,7 @@ define i64 @crc32_64_8(i64 %a0, i8 %a1, i8 *%a2) nounwind { ; SKYLAKE-LABEL: crc32_64_8: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: crc32b %sil, %edi # sched: [3:1.00] -; SKYLAKE-NEXT: crc32b (%rdx), %edi # sched: [7:1.00] +; SKYLAKE-NEXT: crc32b (%rdx), %edi # sched: [8:1.00] ; SKYLAKE-NEXT: movq %rdi, %rax # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -264,7 +264,7 @@ define i64 @crc32_64_64(i64 %a0, i64 %a1, i64 *%a2) { ; SKYLAKE-LABEL: crc32_64_64: ; SKYLAKE: # BB#0: ; SKYLAKE-NEXT: crc32q %rsi, %rdi # sched: [3:1.00] -; SKYLAKE-NEXT: crc32q (%rdx), %rdi # sched: [7:1.00] +; SKYLAKE-NEXT: crc32q (%rdx), %rdi # sched: [8:1.00] ; SKYLAKE-NEXT: movq %rdi, %rax # sched: [1:0.25] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -503,9 +503,9 @@ define i32 @test_pcmpistri(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpistri: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpistri $7, %xmm1, %xmm0 # sched: [11:3.00] +; SKYLAKE-NEXT: vpcmpistri $7, %xmm1, %xmm0 # sched: [10:3.00] ; SKYLAKE-NEXT: movl %ecx, %eax # sched: [1:0.25] -; SKYLAKE-NEXT: vpcmpistri $7, (%rdi), %xmm0 # sched: [11:3.00] +; SKYLAKE-NEXT: vpcmpistri $7, (%rdi), %xmm0 # sched: [10:3.00] ; SKYLAKE-NEXT: # kill: %ECX<def> %ECX<kill> %RCX<def> ; SKYLAKE-NEXT: leal (%rcx,%rax), %eax # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] @@ -562,8 +562,8 @@ define <16 x i8> @test_pcmpistrm(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpistrm: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpistrm $7, %xmm1, %xmm0 # sched: [11:3.00] -; SKYLAKE-NEXT: vpcmpistrm $7, (%rdi), %xmm0 # sched: [11:3.00] +; SKYLAKE-NEXT: vpcmpistrm $7, %xmm1, %xmm0 # sched: [10:3.00] +; SKYLAKE-NEXT: vpcmpistrm $7, (%rdi), %xmm0 # sched: [10:3.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pcmpistrm: @@ -611,8 +611,8 @@ define <2 x i64> @test_pcmpgtq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_pcmpgtq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpcmpgtq (%rdi), %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm0 # sched: [3:1.00] +; SKYLAKE-NEXT: vpcmpgtq (%rdi), %xmm0, %xmm0 # sched: [3:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pcmpgtq: @@ -661,8 +661,8 @@ define <2 x i64> @test_pclmulqdq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) { ; ; SKYLAKE-LABEL: test_pclmulqdq: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpclmulqdq $0, %xmm1, %xmm0, %xmm0 # sched: [11:2.00] -; SKYLAKE-NEXT: vpclmulqdq $0, (%rdi), %xmm0, %xmm0 # sched: [11:2.00] +; SKYLAKE-NEXT: vpclmulqdq $0, %xmm1, %xmm0, %xmm0 # sched: [6:1.00] +; SKYLAKE-NEXT: vpclmulqdq $0, (%rdi), %xmm0, %xmm0 # sched: [6:1.00] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pclmulqdq: diff --git a/llvm/test/CodeGen/X86/ssse3-schedule.ll b/llvm/test/CodeGen/X86/ssse3-schedule.ll index cfdac02dc1c..1ddcaf813b5 100644 --- a/llvm/test/CodeGen/X86/ssse3-schedule.ll +++ b/llvm/test/CodeGen/X86/ssse3-schedule.ll @@ -48,9 +48,9 @@ define <16 x i8> @test_pabsb(<16 x i8> %a0, <16 x i8> *%a1) { ; ; SKYLAKE-LABEL: test_pabsb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpabsb %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpabsb %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpabsb (%rdi), %xmm1 # sched: [1:0.50] -; SKYLAKE-NEXT: vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pabsb: @@ -113,9 +113,9 @@ define <4 x i32> @test_pabsd(<4 x i32> %a0, <4 x i32> *%a1) { ; ; SKYLAKE-LABEL: test_pabsd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpabsd %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpabsd %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpabsd (%rdi), %xmm1 # sched: [1:0.50] -; SKYLAKE-NEXT: vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pabsd: @@ -178,9 +178,9 @@ define <8 x i16> @test_pabsw(<8 x i16> %a0, <8 x i16> *%a1) { ; ; SKYLAKE-LABEL: test_pabsw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpabsw %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpabsw %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpabsw (%rdi), %xmm1 # sched: [1:0.50] -; SKYLAKE-NEXT: vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33] +; SKYLAKE-NEXT: vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pabsw: @@ -626,8 +626,8 @@ define <8 x i16> @test_pmaddubsw(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_pmaddubsw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmaddubsw %xmm1, %xmm0, %xmm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmaddubsw (%rdi), %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmaddubsw %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmaddubsw (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pmaddubsw: @@ -682,8 +682,8 @@ define <8 x i16> @test_pmulhrsw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_pmulhrsw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpmulhrsw %xmm1, %xmm0, %xmm0 # sched: [5:1.00] -; SKYLAKE-NEXT: vpmulhrsw (%rdi), %xmm0, %xmm0 # sched: [5:1.00] +; SKYLAKE-NEXT: vpmulhrsw %xmm1, %xmm0, %xmm0 # sched: [4:0.33] +; SKYLAKE-NEXT: vpmulhrsw (%rdi), %xmm0, %xmm0 # sched: [4:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; ; BTVER2-LABEL: test_pmulhrsw: @@ -796,7 +796,7 @@ define <16 x i8> @test_psignb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) { ; ; SKYLAKE-LABEL: test_psignb: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsignb %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpsignb %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsignb (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -855,7 +855,7 @@ define <4 x i32> @test_psignd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) { ; ; SKYLAKE-LABEL: test_psignd: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsignd %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpsignd %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsignd (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; @@ -914,7 +914,7 @@ define <8 x i16> @test_psignw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) { ; ; SKYLAKE-LABEL: test_psignw: ; SKYLAKE: # BB#0: -; SKYLAKE-NEXT: vpsignw %xmm1, %xmm0, %xmm0 # sched: [1:0.50] +; SKYLAKE-NEXT: vpsignw %xmm1, %xmm0, %xmm0 # sched: [1:1.00] ; SKYLAKE-NEXT: vpsignw (%rdi), %xmm0, %xmm0 # sched: [1:0.50] ; SKYLAKE-NEXT: retq # sched: [2:1.00] ; |

