diff options
Diffstat (limited to 'llvm/test/CodeGen/X86/sse-schedule.ll')
| -rw-r--r-- | llvm/test/CodeGen/X86/sse-schedule.ll | 2440 | 
1 files changed, 2422 insertions, 18 deletions
| diff --git a/llvm/test/CodeGen/X86/sse-schedule.ll b/llvm/test/CodeGen/X86/sse-schedule.ll index 91b6ce880aa..75343257deb 100644 --- a/llvm/test/CodeGen/X86/sse-schedule.ll +++ b/llvm/test/CodeGen/X86/sse-schedule.ll @@ -1,15 +1,25 @@  ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=x86-64 | FileCheck %s --check-prefix=CHECK --check-prefix=GENERIC -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=atom | FileCheck %s --check-prefix=CHECK --check-prefix=ATOM -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=slm | FileCheck %s --check-prefix=CHECK --check-prefix=SLM -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=sandybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=ivybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=broadwell | FileCheck %s --check-prefix=CHECK --check-prefix=BROADWELL -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=SKYLAKE -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skx | FileCheck %s --check-prefix=CHECK --check-prefix=SKX -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=btver2 | FileCheck %s --check-prefix=CHECK --check-prefix=BTVER2 -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1 | FileCheck %s --check-prefix=CHECK --check-prefix=ZNVER1 +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=x86-64 | FileCheck %s --check-prefixes=CHECK,GENERIC +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=atom -mattr=-sse3 | FileCheck %s --check-prefixes=CHECK,ATOM +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=slm -mattr=-sse3 | FileCheck %s --check-prefixes=CHECK,SLM +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=sandybridge -mattr=-sse3 | FileCheck %s --check-prefixes=CHECK,SANDY-SSE +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=sandybridge -mattr=-avx2 | FileCheck %s --check-prefixes=CHECK,SANDY +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=ivybridge -mattr=-sse3 | FileCheck %s --check-prefixes=CHECK,SANDY-SSE +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=ivybridge -mattr=-avx2 | FileCheck %s --check-prefixes=CHECK,SANDY +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell -mattr=-sse3 | FileCheck %s --check-prefixes=CHECK,HASWELL-SSE +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell -mattr=-avx2 | FileCheck %s --check-prefixes=CHECK,HASWELL +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=broadwell -mattr=-sse3 | FileCheck %s --check-prefixes=CHECK,BROADWELL-SSE +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=broadwell -mattr=-avx2 | FileCheck %s --check-prefixes=CHECK,BROADWELL +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake -mattr=-sse3 | FileCheck %s --check-prefixes=CHECK,SKYLAKE-SSE +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake -mattr=-avx2 | FileCheck %s --check-prefixes=CHECK,SKYLAKE +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skx -mattr=-sse3 | FileCheck %s --check-prefixes=CHECK,SKX-SSE +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skx -mattr=-avx2 | FileCheck %s --check-prefixes=CHECK,SKX +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=btver2 -mattr=-sse3 | FileCheck %s --check-prefixes=CHECK,BTVER2-SSE +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=btver2 -mattr=-avx2 | FileCheck %s --check-prefixes=CHECK,BTVER2 +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1 -mattr=-sse3 | FileCheck %s --check-prefixes=CHECK,ZNVER1-SSE +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1 -mattr=-avx2 | FileCheck %s --check-prefixes=CHECK,ZNVER1 + +; FIXME: we should really use -mattr=-sse2 here but some of the comparison tests don't work without access to legal <4 x i32> types.  define <4 x float> @test_addps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; GENERIC-LABEL: test_addps: @@ -30,42 +40,84 @@ define <4 x float> @test_addps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; SLM-NEXT:    addps (%rdi), %xmm0 # sched: [6:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_addps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    addps (%rdi), %xmm0 # sched: [9:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_addps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    vaddps (%rdi), %xmm0, %xmm0 # sched: [9:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_addps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    addps (%rdi), %xmm0 # sched: [9:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_addps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    vaddps (%rdi), %xmm0, %xmm0 # sched: [9:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_addps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    addps (%rdi), %xmm0 # sched: [8:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_addps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    vaddps (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_addps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    addps (%rdi), %xmm0 # sched: [10:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_addps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    vaddps (%rdi), %xmm0, %xmm0 # sched: [10:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_addps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    addps (%rdi), %xmm0 # sched: [10:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_addps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    vaddps (%rdi), %xmm0, %xmm0 # sched: [10:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_addps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    addps (%rdi), %xmm0 # sched: [8:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_addps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    vaddps (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_addps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    addps (%rdi), %xmm0 # sched: [10:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_addps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] @@ -96,42 +148,84 @@ define float @test_addss(float %a0, float %a1, float *%a2) {  ; SLM-NEXT:    addss (%rdi), %xmm0 # sched: [6:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_addss: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    addss (%rdi), %xmm0 # sched: [9:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_addss:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    vaddss (%rdi), %xmm0, %xmm0 # sched: [9:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_addss: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    addss (%rdi), %xmm0 # sched: [8:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_addss:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    vaddss (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_addss: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    addss (%rdi), %xmm0 # sched: [8:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_addss:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    vaddss (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_addss: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    addss (%rdi), %xmm0 # sched: [9:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_addss:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    vaddss (%rdi), %xmm0, %xmm0 # sched: [9:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_addss: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    addss (%rdi), %xmm0 # sched: [9:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_addss:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    vaddss (%rdi), %xmm0, %xmm0 # sched: [9:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_addss: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    addss (%rdi), %xmm0 # sched: [8:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_addss:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    vaddss (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_addss: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    addss (%rdi), %xmm0 # sched: [10:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_addss:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00] @@ -166,42 +260,84 @@ define <4 x float> @test_andps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; SLM-NEXT:    andps (%rdi), %xmm0 # sched: [4:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_andps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    andps %xmm1, %xmm0 # sched: [1:1.00] +; SANDY-SSE-NEXT:    andps (%rdi), %xmm0 # sched: [7:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_andps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vandps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]  ; SANDY-NEXT:    vandps (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_andps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    andps %xmm1, %xmm0 # sched: [1:1.00] +; HASWELL-SSE-NEXT:    andps (%rdi), %xmm0 # sched: [7:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_andps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vandps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]  ; HASWELL-NEXT:    vandps (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_andps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    andps %xmm1, %xmm0 # sched: [1:1.00] +; BROADWELL-SSE-NEXT:    andps (%rdi), %xmm0 # sched: [6:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_andps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vandps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]  ; BROADWELL-NEXT:    vandps (%rdi), %xmm0, %xmm0 # sched: [6:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_andps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    andps %xmm1, %xmm0 # sched: [1:0.33] +; SKYLAKE-SSE-NEXT:    andps (%rdi), %xmm0 # sched: [7:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_andps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vandps %xmm1, %xmm0, %xmm0 # sched: [1:0.33]  ; SKYLAKE-NEXT:    vandps (%rdi), %xmm0, %xmm0 # sched: [7:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_andps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    andps %xmm1, %xmm0 # sched: [1:0.33] +; SKX-SSE-NEXT:    andps (%rdi), %xmm0 # sched: [7:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_andps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vandps %xmm1, %xmm0, %xmm0 # sched: [1:0.33]  ; SKX-NEXT:    vandps (%rdi), %xmm0, %xmm0 # sched: [7:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_andps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    andps %xmm1, %xmm0 # sched: [1:0.50] +; BTVER2-SSE-NEXT:    andps (%rdi), %xmm0 # sched: [6:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_andps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vandps %xmm1, %xmm0, %xmm0 # sched: [1:0.50]  ; BTVER2-NEXT:    vandps (%rdi), %xmm0, %xmm0 # sched: [6:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_andps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    andps %xmm1, %xmm0 # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    andps (%rdi), %xmm0 # sched: [8:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_andps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vandps %xmm1, %xmm0, %xmm0 # sched: [1:0.25] @@ -240,42 +376,84 @@ define <4 x float> @test_andnotps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; SLM-NEXT:    andnps (%rdi), %xmm0 # sched: [4:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_andnotps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    andnps %xmm1, %xmm0 # sched: [1:1.00] +; SANDY-SSE-NEXT:    andnps (%rdi), %xmm0 # sched: [7:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_andnotps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vandnps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]  ; SANDY-NEXT:    vandnps (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_andnotps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    andnps %xmm1, %xmm0 # sched: [1:1.00] +; HASWELL-SSE-NEXT:    andnps (%rdi), %xmm0 # sched: [7:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_andnotps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vandnps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]  ; HASWELL-NEXT:    vandnps (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_andnotps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    andnps %xmm1, %xmm0 # sched: [1:1.00] +; BROADWELL-SSE-NEXT:    andnps (%rdi), %xmm0 # sched: [6:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_andnotps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vandnps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]  ; BROADWELL-NEXT:    vandnps (%rdi), %xmm0, %xmm0 # sched: [6:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_andnotps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    andnps %xmm1, %xmm0 # sched: [1:0.33] +; SKYLAKE-SSE-NEXT:    andnps (%rdi), %xmm0 # sched: [7:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_andnotps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vandnps %xmm1, %xmm0, %xmm0 # sched: [1:0.33]  ; SKYLAKE-NEXT:    vandnps (%rdi), %xmm0, %xmm0 # sched: [7:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_andnotps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    andnps %xmm1, %xmm0 # sched: [1:0.33] +; SKX-SSE-NEXT:    andnps (%rdi), %xmm0 # sched: [7:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_andnotps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vandnps %xmm1, %xmm0, %xmm0 # sched: [1:0.33]  ; SKX-NEXT:    vandnps (%rdi), %xmm0, %xmm0 # sched: [7:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_andnotps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    andnps %xmm1, %xmm0 # sched: [1:0.50] +; BTVER2-SSE-NEXT:    andnps (%rdi), %xmm0 # sched: [6:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_andnotps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vandnps %xmm1, %xmm0, %xmm0 # sched: [1:0.50]  ; BTVER2-NEXT:    vandnps (%rdi), %xmm0, %xmm0 # sched: [6:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_andnotps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    andnps %xmm1, %xmm0 # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    andnps (%rdi), %xmm0 # sched: [8:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_andnotps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vandnps %xmm1, %xmm0, %xmm0 # sched: [1:0.25] @@ -315,6 +493,13 @@ define <4 x float> @test_cmpps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; SLM-NEXT:    orps %xmm1, %xmm0 # sched: [1:0.50]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_cmpps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    cmpeqps %xmm0, %xmm1 # sched: [3:1.00] +; SANDY-SSE-NEXT:    cmpeqps (%rdi), %xmm0 # sched: [9:1.00] +; SANDY-SSE-NEXT:    orps %xmm1, %xmm0 # sched: [1:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_cmpps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm1 # sched: [3:1.00] @@ -322,6 +507,13 @@ define <4 x float> @test_cmpps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; SANDY-NEXT:    vorps %xmm0, %xmm1, %xmm0 # sched: [1:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_cmpps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    cmpeqps %xmm0, %xmm1 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    cmpeqps (%rdi), %xmm0 # sched: [9:1.00] +; HASWELL-SSE-NEXT:    orps %xmm1, %xmm0 # sched: [1:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_cmpps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm1 # sched: [3:1.00] @@ -329,6 +521,13 @@ define <4 x float> @test_cmpps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; HASWELL-NEXT:    vorps %xmm0, %xmm1, %xmm0 # sched: [1:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_cmpps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    cmpeqps %xmm0, %xmm1 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    cmpeqps (%rdi), %xmm0 # sched: [8:1.00] +; BROADWELL-SSE-NEXT:    orps %xmm1, %xmm0 # sched: [1:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_cmpps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm1 # sched: [3:1.00] @@ -336,6 +535,13 @@ define <4 x float> @test_cmpps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; BROADWELL-NEXT:    vorps %xmm0, %xmm1, %xmm0 # sched: [1:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_cmpps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    cmpeqps %xmm0, %xmm1 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    cmpeqps (%rdi), %xmm0 # sched: [10:0.50] +; SKYLAKE-SSE-NEXT:    orps %xmm1, %xmm0 # sched: [1:0.33] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_cmpps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm1 # sched: [4:0.50] @@ -343,14 +549,27 @@ define <4 x float> @test_cmpps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; SKYLAKE-NEXT:    vorps %xmm0, %xmm1, %xmm0 # sched: [1:0.33]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_cmpps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    cmpeqps %xmm0, %xmm1 # sched: [4:0.33] +; SKX-SSE-NEXT:    cmpeqps (%rdi), %xmm0 # sched: [10:0.50] +; SKX-SSE-NEXT:    orps %xmm1, %xmm0 # sched: [1:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_cmpps:  ; SKX:       # %bb.0: -; SKX-NEXT:    vcmpeqps %xmm1, %xmm0, %k0 # sched: [3:1.00] -; SKX-NEXT:    vcmpeqps (%rdi), %xmm0, %k1 # sched: [9:1.00] -; SKX-NEXT:    korw %k1, %k0, %k0 # sched: [1:1.00] -; SKX-NEXT:    vpmovm2d %k0, %xmm0 # sched: [1:0.25] +; SKX-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm1 # sched: [4:0.33] +; SKX-NEXT:    vcmpeqps (%rdi), %xmm0, %xmm0 # sched: [10:0.50] +; SKX-NEXT:    vorps %xmm0, %xmm1, %xmm0 # sched: [1:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_cmpps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    cmpeqps %xmm0, %xmm1 # sched: [2:1.00] +; BTVER2-SSE-NEXT:    cmpeqps (%rdi), %xmm0 # sched: [7:1.00] +; BTVER2-SSE-NEXT:    orps %xmm1, %xmm0 # sched: [1:0.50] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_cmpps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm1 # sched: [2:1.00] @@ -358,6 +577,13 @@ define <4 x float> @test_cmpps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; BTVER2-NEXT:    vorps %xmm0, %xmm1, %xmm0 # sched: [1:0.50]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_cmpps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    cmpeqps %xmm0, %xmm1 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    cmpeqps (%rdi), %xmm0 # sched: [10:1.00] +; ZNVER1-SSE-NEXT:    orps %xmm1, %xmm0 # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_cmpps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm1 # sched: [3:1.00] @@ -392,42 +618,84 @@ define float @test_cmpss(float %a0, float %a1, float *%a2) {  ; SLM-NEXT:    cmpeqss (%rdi), %xmm0 # sched: [6:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_cmpss: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    cmpeqss %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    cmpeqss (%rdi), %xmm0 # sched: [9:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_cmpss:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vcmpeqss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    vcmpeqss (%rdi), %xmm0, %xmm0 # sched: [9:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_cmpss: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    cmpeqss %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    cmpeqss (%rdi), %xmm0 # sched: [8:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_cmpss:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vcmpeqss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    vcmpeqss (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_cmpss: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    cmpeqss %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    cmpeqss (%rdi), %xmm0 # sched: [8:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_cmpss:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vcmpeqss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    vcmpeqss (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_cmpss: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    cmpeqss %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    cmpeqss (%rdi), %xmm0 # sched: [9:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_cmpss:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vcmpeqss %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    vcmpeqss (%rdi), %xmm0, %xmm0 # sched: [9:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_cmpss: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    cmpeqss %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    cmpeqss (%rdi), %xmm0 # sched: [9:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_cmpss:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vcmpeqss %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    vcmpeqss (%rdi), %xmm0, %xmm0 # sched: [9:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_cmpss: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    cmpeqss %xmm1, %xmm0 # sched: [2:1.00] +; BTVER2-SSE-NEXT:    cmpeqss (%rdi), %xmm0 # sched: [7:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_cmpss:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vcmpeqss %xmm1, %xmm0, %xmm0 # sched: [2:1.00]  ; BTVER2-NEXT:    vcmpeqss (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_cmpss: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    cmpeqss %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    cmpeqss (%rdi), %xmm0 # sched: [10:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_cmpss:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vcmpeqss %xmm1, %xmm0, %xmm0 # sched: [3:1.00] @@ -487,6 +755,20 @@ define i32 @test_comiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; SLM-NEXT:    movzbl %dl, %eax # sched: [1:0.50]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_comiss: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    comiss %xmm1, %xmm0 # sched: [2:1.00] +; SANDY-SSE-NEXT:    setnp %al # sched: [1:0.50] +; SANDY-SSE-NEXT:    sete %cl # sched: [1:0.50] +; SANDY-SSE-NEXT:    andb %al, %cl # sched: [1:0.33] +; SANDY-SSE-NEXT:    comiss (%rdi), %xmm0 # sched: [8:1.00] +; SANDY-SSE-NEXT:    setnp %al # sched: [1:0.50] +; SANDY-SSE-NEXT:    sete %dl # sched: [1:0.50] +; SANDY-SSE-NEXT:    andb %al, %dl # sched: [1:0.33] +; SANDY-SSE-NEXT:    orb %cl, %dl # sched: [1:0.33] +; SANDY-SSE-NEXT:    movzbl %dl, %eax # sched: [1:0.33] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_comiss:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vcomiss %xmm1, %xmm0 # sched: [2:1.00] @@ -501,6 +783,20 @@ define i32 @test_comiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; SANDY-NEXT:    movzbl %dl, %eax # sched: [1:0.33]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_comiss: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    comiss %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    setnp %al # sched: [1:0.50] +; HASWELL-SSE-NEXT:    sete %cl # sched: [1:0.50] +; HASWELL-SSE-NEXT:    andb %al, %cl # sched: [1:0.25] +; HASWELL-SSE-NEXT:    comiss (%rdi), %xmm0 # sched: [8:1.00] +; HASWELL-SSE-NEXT:    setnp %al # sched: [1:0.50] +; HASWELL-SSE-NEXT:    sete %dl # sched: [1:0.50] +; HASWELL-SSE-NEXT:    andb %al, %dl # sched: [1:0.25] +; HASWELL-SSE-NEXT:    orb %cl, %dl # sched: [1:0.25] +; HASWELL-SSE-NEXT:    movzbl %dl, %eax # sched: [1:0.25] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_comiss:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vcomiss %xmm1, %xmm0 # sched: [3:1.00] @@ -515,6 +811,20 @@ define i32 @test_comiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; HASWELL-NEXT:    movzbl %dl, %eax # sched: [1:0.25]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_comiss: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    comiss %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    setnp %al # sched: [1:0.50] +; BROADWELL-SSE-NEXT:    sete %cl # sched: [1:0.50] +; BROADWELL-SSE-NEXT:    andb %al, %cl # sched: [1:0.25] +; BROADWELL-SSE-NEXT:    comiss (%rdi), %xmm0 # sched: [8:1.00] +; BROADWELL-SSE-NEXT:    setnp %al # sched: [1:0.50] +; BROADWELL-SSE-NEXT:    sete %dl # sched: [1:0.50] +; BROADWELL-SSE-NEXT:    andb %al, %dl # sched: [1:0.25] +; BROADWELL-SSE-NEXT:    orb %cl, %dl # sched: [1:0.25] +; BROADWELL-SSE-NEXT:    movzbl %dl, %eax # sched: [1:0.25] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_comiss:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vcomiss %xmm1, %xmm0 # sched: [3:1.00] @@ -529,6 +839,20 @@ define i32 @test_comiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; BROADWELL-NEXT:    movzbl %dl, %eax # sched: [1:0.25]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_comiss: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    comiss %xmm1, %xmm0 # sched: [2:1.00] +; SKYLAKE-SSE-NEXT:    setnp %al # sched: [1:0.50] +; SKYLAKE-SSE-NEXT:    sete %cl # sched: [1:0.50] +; SKYLAKE-SSE-NEXT:    andb %al, %cl # sched: [1:0.25] +; SKYLAKE-SSE-NEXT:    comiss (%rdi), %xmm0 # sched: [7:1.00] +; SKYLAKE-SSE-NEXT:    setnp %al # sched: [1:0.50] +; SKYLAKE-SSE-NEXT:    sete %dl # sched: [1:0.50] +; SKYLAKE-SSE-NEXT:    andb %al, %dl # sched: [1:0.25] +; SKYLAKE-SSE-NEXT:    orb %cl, %dl # sched: [1:0.25] +; SKYLAKE-SSE-NEXT:    movzbl %dl, %eax # sched: [1:0.25] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_comiss:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vcomiss %xmm1, %xmm0 # sched: [2:1.00] @@ -543,6 +867,20 @@ define i32 @test_comiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; SKYLAKE-NEXT:    movzbl %dl, %eax # sched: [1:0.25]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_comiss: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    comiss %xmm1, %xmm0 # sched: [2:1.00] +; SKX-SSE-NEXT:    setnp %al # sched: [1:0.50] +; SKX-SSE-NEXT:    sete %cl # sched: [1:0.50] +; SKX-SSE-NEXT:    andb %al, %cl # sched: [1:0.25] +; SKX-SSE-NEXT:    comiss (%rdi), %xmm0 # sched: [7:1.00] +; SKX-SSE-NEXT:    setnp %al # sched: [1:0.50] +; SKX-SSE-NEXT:    sete %dl # sched: [1:0.50] +; SKX-SSE-NEXT:    andb %al, %dl # sched: [1:0.25] +; SKX-SSE-NEXT:    orb %cl, %dl # sched: [1:0.25] +; SKX-SSE-NEXT:    movzbl %dl, %eax # sched: [1:0.25] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_comiss:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vcomiss %xmm1, %xmm0 # sched: [2:1.00] @@ -557,6 +895,20 @@ define i32 @test_comiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; SKX-NEXT:    movzbl %dl, %eax # sched: [1:0.25]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_comiss: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    comiss %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    setnp %al # sched: [1:0.50] +; BTVER2-SSE-NEXT:    sete %cl # sched: [1:0.50] +; BTVER2-SSE-NEXT:    andb %al, %cl # sched: [1:0.50] +; BTVER2-SSE-NEXT:    comiss (%rdi), %xmm0 # sched: [8:1.00] +; BTVER2-SSE-NEXT:    setnp %al # sched: [1:0.50] +; BTVER2-SSE-NEXT:    sete %dl # sched: [1:0.50] +; BTVER2-SSE-NEXT:    andb %al, %dl # sched: [1:0.50] +; BTVER2-SSE-NEXT:    orb %cl, %dl # sched: [1:0.50] +; BTVER2-SSE-NEXT:    movzbl %dl, %eax # sched: [1:0.50] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_comiss:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vcomiss %xmm1, %xmm0 # sched: [3:1.00] @@ -571,6 +923,20 @@ define i32 @test_comiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; BTVER2-NEXT:    movzbl %dl, %eax # sched: [1:0.50]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_comiss: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    comiss %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    setnp %al # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    sete %cl # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    andb %al, %cl # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    comiss (%rdi), %xmm0 # sched: [10:1.00] +; ZNVER1-SSE-NEXT:    setnp %al # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    sete %dl # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    andb %al, %dl # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    orb %cl, %dl # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    movzbl %dl, %eax # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_comiss:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vcomiss %xmm1, %xmm0 # sched: [3:1.00] @@ -614,6 +980,13 @@ define float @test_cvtsi2ss(i32 %a0, i32 *%a1) {  ; SLM-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_cvtsi2ss: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    cvtsi2ssl %edi, %xmm1 # sched: [5:2.00] +; SANDY-SSE-NEXT:    cvtsi2ssl (%rsi), %xmm0 # sched: [10:1.00] +; SANDY-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_cvtsi2ss:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vcvtsi2ssl %edi, %xmm0, %xmm0 # sched: [5:2.00] @@ -621,6 +994,13 @@ define float @test_cvtsi2ss(i32 %a0, i32 *%a1) {  ; SANDY-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_cvtsi2ss: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    cvtsi2ssl %edi, %xmm1 # sched: [4:1.00] +; HASWELL-SSE-NEXT:    cvtsi2ssl (%rsi), %xmm0 # sched: [9:1.00] +; HASWELL-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_cvtsi2ss:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vcvtsi2ssl %edi, %xmm0, %xmm0 # sched: [4:1.00] @@ -628,6 +1008,13 @@ define float @test_cvtsi2ss(i32 %a0, i32 *%a1) {  ; HASWELL-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_cvtsi2ss: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    cvtsi2ssl %edi, %xmm1 # sched: [4:1.00] +; BROADWELL-SSE-NEXT:    cvtsi2ssl (%rsi), %xmm0 # sched: [9:1.00] +; BROADWELL-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_cvtsi2ss:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vcvtsi2ssl %edi, %xmm0, %xmm0 # sched: [4:1.00] @@ -635,6 +1022,13 @@ define float @test_cvtsi2ss(i32 %a0, i32 *%a1) {  ; BROADWELL-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_cvtsi2ss: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    cvtsi2ssl %edi, %xmm1 # sched: [5:1.00] +; SKYLAKE-SSE-NEXT:    cvtsi2ssl (%rsi), %xmm0 # sched: [9:1.00] +; SKYLAKE-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_cvtsi2ss:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vcvtsi2ssl %edi, %xmm0, %xmm0 # sched: [5:1.00] @@ -642,6 +1036,13 @@ define float @test_cvtsi2ss(i32 %a0, i32 *%a1) {  ; SKYLAKE-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_cvtsi2ss: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    cvtsi2ssl %edi, %xmm1 # sched: [5:1.00] +; SKX-SSE-NEXT:    cvtsi2ssl (%rsi), %xmm0 # sched: [9:1.00] +; SKX-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_cvtsi2ss:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vcvtsi2ssl %edi, %xmm0, %xmm0 # sched: [5:1.00] @@ -649,6 +1050,13 @@ define float @test_cvtsi2ss(i32 %a0, i32 *%a1) {  ; SKX-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_cvtsi2ss: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    cvtsi2ssl (%rsi), %xmm0 # sched: [8:1.00] +; BTVER2-SSE-NEXT:    cvtsi2ssl %edi, %xmm1 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_cvtsi2ss:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vcvtsi2ssl %edi, %xmm0, %xmm0 # sched: [3:1.00] @@ -656,6 +1064,13 @@ define float @test_cvtsi2ss(i32 %a0, i32 *%a1) {  ; BTVER2-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_cvtsi2ss: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    cvtsi2ssl (%rsi), %xmm0 # sched: [12:1.00] +; ZNVER1-SSE-NEXT:    cvtsi2ssl %edi, %xmm1 # sched: [5:1.00] +; ZNVER1-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_cvtsi2ss:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vcvtsi2ssl %edi, %xmm0, %xmm0 # sched: [5:1.00] @@ -691,6 +1106,13 @@ define float @test_cvtsi2ssq(i64 %a0, i64 *%a1) {  ; SLM-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_cvtsi2ssq: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    cvtsi2ssq %rdi, %xmm1 # sched: [5:2.00] +; SANDY-SSE-NEXT:    cvtsi2ssq (%rsi), %xmm0 # sched: [10:1.00] +; SANDY-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_cvtsi2ssq:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vcvtsi2ssq %rdi, %xmm0, %xmm0 # sched: [5:2.00] @@ -698,6 +1120,13 @@ define float @test_cvtsi2ssq(i64 %a0, i64 *%a1) {  ; SANDY-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_cvtsi2ssq: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    cvtsi2ssq %rdi, %xmm1 # sched: [5:2.00] +; HASWELL-SSE-NEXT:    cvtsi2ssq (%rsi), %xmm0 # sched: [9:1.00] +; HASWELL-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_cvtsi2ssq:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vcvtsi2ssq %rdi, %xmm0, %xmm0 # sched: [5:2.00] @@ -705,6 +1134,13 @@ define float @test_cvtsi2ssq(i64 %a0, i64 *%a1) {  ; HASWELL-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_cvtsi2ssq: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    cvtsi2ssq %rdi, %xmm1 # sched: [5:2.00] +; BROADWELL-SSE-NEXT:    cvtsi2ssq (%rsi), %xmm0 # sched: [9:1.00] +; BROADWELL-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_cvtsi2ssq:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vcvtsi2ssq %rdi, %xmm0, %xmm0 # sched: [5:2.00] @@ -712,6 +1148,13 @@ define float @test_cvtsi2ssq(i64 %a0, i64 *%a1) {  ; BROADWELL-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_cvtsi2ssq: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    cvtsi2ssq %rdi, %xmm1 # sched: [6:2.00] +; SKYLAKE-SSE-NEXT:    cvtsi2ssq (%rsi), %xmm0 # sched: [9:1.00] +; SKYLAKE-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_cvtsi2ssq:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vcvtsi2ssq %rdi, %xmm0, %xmm0 # sched: [6:2.00] @@ -719,6 +1162,13 @@ define float @test_cvtsi2ssq(i64 %a0, i64 *%a1) {  ; SKYLAKE-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_cvtsi2ssq: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    cvtsi2ssq %rdi, %xmm1 # sched: [6:2.00] +; SKX-SSE-NEXT:    cvtsi2ssq (%rsi), %xmm0 # sched: [9:1.00] +; SKX-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_cvtsi2ssq:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vcvtsi2ssq %rdi, %xmm0, %xmm0 # sched: [6:2.00] @@ -726,6 +1176,13 @@ define float @test_cvtsi2ssq(i64 %a0, i64 *%a1) {  ; SKX-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_cvtsi2ssq: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    cvtsi2ssq (%rsi), %xmm0 # sched: [8:1.00] +; BTVER2-SSE-NEXT:    cvtsi2ssq %rdi, %xmm1 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_cvtsi2ssq:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vcvtsi2ssq %rdi, %xmm0, %xmm0 # sched: [3:1.00] @@ -733,6 +1190,13 @@ define float @test_cvtsi2ssq(i64 %a0, i64 *%a1) {  ; BTVER2-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_cvtsi2ssq: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    cvtsi2ssq (%rsi), %xmm0 # sched: [12:1.00] +; ZNVER1-SSE-NEXT:    cvtsi2ssq %rdi, %xmm1 # sched: [5:1.00] +; ZNVER1-SSE-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_cvtsi2ssq:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vcvtsi2ssq %rdi, %xmm0, %xmm0 # sched: [5:1.00] @@ -768,6 +1232,13 @@ define i32 @test_cvtss2si(float %a0, float *%a1) {  ; SLM-NEXT:    addl %ecx, %eax # sched: [1:0.50]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_cvtss2si: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    cvtss2si %xmm0, %ecx # sched: [5:1.00] +; SANDY-SSE-NEXT:    cvtss2si (%rdi), %eax # sched: [9:1.00] +; SANDY-SSE-NEXT:    addl %ecx, %eax # sched: [1:0.33] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_cvtss2si:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vcvtss2si %xmm0, %ecx # sched: [5:1.00] @@ -775,6 +1246,13 @@ define i32 @test_cvtss2si(float %a0, float *%a1) {  ; SANDY-NEXT:    addl %ecx, %eax # sched: [1:0.33]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_cvtss2si: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    cvtss2si %xmm0, %ecx # sched: [4:1.00] +; HASWELL-SSE-NEXT:    cvtss2si (%rdi), %eax # sched: [9:1.00] +; HASWELL-SSE-NEXT:    addl %ecx, %eax # sched: [1:0.25] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_cvtss2si:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vcvtss2si %xmm0, %ecx # sched: [4:1.00] @@ -782,6 +1260,13 @@ define i32 @test_cvtss2si(float %a0, float *%a1) {  ; HASWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_cvtss2si: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    cvtss2si (%rdi), %eax # sched: [9:1.00] +; BROADWELL-SSE-NEXT:    cvtss2si %xmm0, %ecx # sched: [4:1.00] +; BROADWELL-SSE-NEXT:    addl %ecx, %eax # sched: [1:0.25] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_cvtss2si:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vcvtss2si (%rdi), %eax # sched: [9:1.00] @@ -789,6 +1274,13 @@ define i32 @test_cvtss2si(float %a0, float *%a1) {  ; BROADWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_cvtss2si: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    cvtss2si %xmm0, %ecx # sched: [6:1.00] +; SKYLAKE-SSE-NEXT:    cvtss2si (%rdi), %eax # sched: [11:1.00] +; SKYLAKE-SSE-NEXT:    addl %ecx, %eax # sched: [1:0.25] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_cvtss2si:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vcvtss2si %xmm0, %ecx # sched: [6:1.00] @@ -796,6 +1288,13 @@ define i32 @test_cvtss2si(float %a0, float *%a1) {  ; SKYLAKE-NEXT:    addl %ecx, %eax # sched: [1:0.25]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_cvtss2si: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    cvtss2si %xmm0, %ecx # sched: [6:1.00] +; SKX-SSE-NEXT:    cvtss2si (%rdi), %eax # sched: [11:1.00] +; SKX-SSE-NEXT:    addl %ecx, %eax # sched: [1:0.25] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_cvtss2si:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vcvtss2si %xmm0, %ecx # sched: [6:1.00] @@ -803,6 +1302,13 @@ define i32 @test_cvtss2si(float %a0, float *%a1) {  ; SKX-NEXT:    addl %ecx, %eax # sched: [1:0.25]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_cvtss2si: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    cvtss2si (%rdi), %eax # sched: [8:1.00] +; BTVER2-SSE-NEXT:    cvtss2si %xmm0, %ecx # sched: [3:1.00] +; BTVER2-SSE-NEXT:    addl %ecx, %eax # sched: [1:0.50] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_cvtss2si:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vcvtss2si (%rdi), %eax # sched: [8:1.00] @@ -810,6 +1316,13 @@ define i32 @test_cvtss2si(float %a0, float *%a1) {  ; BTVER2-NEXT:    addl %ecx, %eax # sched: [1:0.50]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_cvtss2si: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    cvtss2si (%rdi), %eax # sched: [12:1.00] +; ZNVER1-SSE-NEXT:    cvtss2si %xmm0, %ecx # sched: [5:1.00] +; ZNVER1-SSE-NEXT:    addl %ecx, %eax # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_cvtss2si:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vcvtss2si (%rdi), %eax # sched: [12:1.00] @@ -848,6 +1361,13 @@ define i64 @test_cvtss2siq(float %a0, float *%a1) {  ; SLM-NEXT:    addq %rcx, %rax # sched: [1:0.50]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_cvtss2siq: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    cvtss2si %xmm0, %rcx # sched: [5:1.00] +; SANDY-SSE-NEXT:    cvtss2si (%rdi), %rax # sched: [9:1.00] +; SANDY-SSE-NEXT:    addq %rcx, %rax # sched: [1:0.33] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_cvtss2siq:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vcvtss2si %xmm0, %rcx # sched: [5:1.00] @@ -855,6 +1375,13 @@ define i64 @test_cvtss2siq(float %a0, float *%a1) {  ; SANDY-NEXT:    addq %rcx, %rax # sched: [1:0.33]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_cvtss2siq: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    cvtss2si %xmm0, %rcx # sched: [4:1.00] +; HASWELL-SSE-NEXT:    cvtss2si (%rdi), %rax # sched: [9:1.00] +; HASWELL-SSE-NEXT:    addq %rcx, %rax # sched: [1:0.25] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_cvtss2siq:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vcvtss2si %xmm0, %rcx # sched: [4:1.00] @@ -862,6 +1389,13 @@ define i64 @test_cvtss2siq(float %a0, float *%a1) {  ; HASWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_cvtss2siq: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    cvtss2si (%rdi), %rax # sched: [9:1.00] +; BROADWELL-SSE-NEXT:    cvtss2si %xmm0, %rcx # sched: [4:1.00] +; BROADWELL-SSE-NEXT:    addq %rcx, %rax # sched: [1:0.25] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_cvtss2siq:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vcvtss2si (%rdi), %rax # sched: [9:1.00] @@ -869,6 +1403,13 @@ define i64 @test_cvtss2siq(float %a0, float *%a1) {  ; BROADWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_cvtss2siq: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    cvtss2si %xmm0, %rcx # sched: [6:1.00] +; SKYLAKE-SSE-NEXT:    cvtss2si (%rdi), %rax # sched: [11:1.00] +; SKYLAKE-SSE-NEXT:    addq %rcx, %rax # sched: [1:0.25] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_cvtss2siq:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vcvtss2si %xmm0, %rcx # sched: [6:1.00] @@ -876,6 +1417,13 @@ define i64 @test_cvtss2siq(float %a0, float *%a1) {  ; SKYLAKE-NEXT:    addq %rcx, %rax # sched: [1:0.25]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_cvtss2siq: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    cvtss2si %xmm0, %rcx # sched: [6:1.00] +; SKX-SSE-NEXT:    cvtss2si (%rdi), %rax # sched: [11:1.00] +; SKX-SSE-NEXT:    addq %rcx, %rax # sched: [1:0.25] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_cvtss2siq:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vcvtss2si %xmm0, %rcx # sched: [6:1.00] @@ -883,6 +1431,13 @@ define i64 @test_cvtss2siq(float %a0, float *%a1) {  ; SKX-NEXT:    addq %rcx, %rax # sched: [1:0.25]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_cvtss2siq: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    cvtss2si (%rdi), %rax # sched: [8:1.00] +; BTVER2-SSE-NEXT:    cvtss2si %xmm0, %rcx # sched: [3:1.00] +; BTVER2-SSE-NEXT:    addq %rcx, %rax # sched: [1:0.50] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_cvtss2siq:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vcvtss2si (%rdi), %rax # sched: [8:1.00] @@ -890,6 +1445,13 @@ define i64 @test_cvtss2siq(float %a0, float *%a1) {  ; BTVER2-NEXT:    addq %rcx, %rax # sched: [1:0.50]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_cvtss2siq: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    cvtss2si (%rdi), %rax # sched: [12:1.00] +; ZNVER1-SSE-NEXT:    cvtss2si %xmm0, %rcx # sched: [5:1.00] +; ZNVER1-SSE-NEXT:    addq %rcx, %rax # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_cvtss2siq:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vcvtss2si (%rdi), %rax # sched: [12:1.00] @@ -928,6 +1490,13 @@ define i32 @test_cvttss2si(float %a0, float *%a1) {  ; SLM-NEXT:    addl %ecx, %eax # sched: [1:0.50]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_cvttss2si: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    cvttss2si %xmm0, %ecx # sched: [5:1.00] +; SANDY-SSE-NEXT:    cvttss2si (%rdi), %eax # sched: [9:1.00] +; SANDY-SSE-NEXT:    addl %ecx, %eax # sched: [1:0.33] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_cvttss2si:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vcvttss2si %xmm0, %ecx # sched: [5:1.00] @@ -935,6 +1504,13 @@ define i32 @test_cvttss2si(float %a0, float *%a1) {  ; SANDY-NEXT:    addl %ecx, %eax # sched: [1:0.33]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_cvttss2si: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    cvttss2si %xmm0, %ecx # sched: [4:1.00] +; HASWELL-SSE-NEXT:    cvttss2si (%rdi), %eax # sched: [9:1.00] +; HASWELL-SSE-NEXT:    addl %ecx, %eax # sched: [1:0.25] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_cvttss2si:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vcvttss2si %xmm0, %ecx # sched: [4:1.00] @@ -942,6 +1518,13 @@ define i32 @test_cvttss2si(float %a0, float *%a1) {  ; HASWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_cvttss2si: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    cvttss2si (%rdi), %eax # sched: [9:1.00] +; BROADWELL-SSE-NEXT:    cvttss2si %xmm0, %ecx # sched: [4:1.00] +; BROADWELL-SSE-NEXT:    addl %ecx, %eax # sched: [1:0.25] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_cvttss2si:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vcvttss2si (%rdi), %eax # sched: [9:1.00] @@ -949,6 +1532,13 @@ define i32 @test_cvttss2si(float %a0, float *%a1) {  ; BROADWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_cvttss2si: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    cvttss2si %xmm0, %ecx # sched: [7:1.00] +; SKYLAKE-SSE-NEXT:    cvttss2si (%rdi), %eax # sched: [11:1.00] +; SKYLAKE-SSE-NEXT:    addl %ecx, %eax # sched: [1:0.25] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_cvttss2si:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vcvttss2si %xmm0, %ecx # sched: [7:1.00] @@ -956,6 +1546,13 @@ define i32 @test_cvttss2si(float %a0, float *%a1) {  ; SKYLAKE-NEXT:    addl %ecx, %eax # sched: [1:0.25]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_cvttss2si: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    cvttss2si %xmm0, %ecx # sched: [7:1.00] +; SKX-SSE-NEXT:    cvttss2si (%rdi), %eax # sched: [11:1.00] +; SKX-SSE-NEXT:    addl %ecx, %eax # sched: [1:0.25] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_cvttss2si:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vcvttss2si %xmm0, %ecx # sched: [7:1.00] @@ -963,6 +1560,13 @@ define i32 @test_cvttss2si(float %a0, float *%a1) {  ; SKX-NEXT:    addl %ecx, %eax # sched: [1:0.25]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_cvttss2si: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    cvttss2si (%rdi), %eax # sched: [8:1.00] +; BTVER2-SSE-NEXT:    cvttss2si %xmm0, %ecx # sched: [3:1.00] +; BTVER2-SSE-NEXT:    addl %ecx, %eax # sched: [1:0.50] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_cvttss2si:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vcvttss2si (%rdi), %eax # sched: [8:1.00] @@ -970,6 +1574,13 @@ define i32 @test_cvttss2si(float %a0, float *%a1) {  ; BTVER2-NEXT:    addl %ecx, %eax # sched: [1:0.50]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_cvttss2si: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    cvttss2si (%rdi), %eax # sched: [12:1.00] +; ZNVER1-SSE-NEXT:    cvttss2si %xmm0, %ecx # sched: [5:1.00] +; ZNVER1-SSE-NEXT:    addl %ecx, %eax # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_cvttss2si:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vcvttss2si (%rdi), %eax # sched: [12:1.00] @@ -1005,6 +1616,13 @@ define i64 @test_cvttss2siq(float %a0, float *%a1) {  ; SLM-NEXT:    addq %rcx, %rax # sched: [1:0.50]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_cvttss2siq: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    cvttss2si %xmm0, %rcx # sched: [5:1.00] +; SANDY-SSE-NEXT:    cvttss2si (%rdi), %rax # sched: [9:1.00] +; SANDY-SSE-NEXT:    addq %rcx, %rax # sched: [1:0.33] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_cvttss2siq:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vcvttss2si %xmm0, %rcx # sched: [5:1.00] @@ -1012,6 +1630,13 @@ define i64 @test_cvttss2siq(float %a0, float *%a1) {  ; SANDY-NEXT:    addq %rcx, %rax # sched: [1:0.33]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_cvttss2siq: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    cvttss2si (%rdi), %rax # sched: [10:1.00] +; HASWELL-SSE-NEXT:    cvttss2si %xmm0, %rcx # sched: [4:1.00] +; HASWELL-SSE-NEXT:    addq %rcx, %rax # sched: [1:0.25] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_cvttss2siq:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vcvttss2si %xmm0, %rcx # sched: [4:1.00] @@ -1019,6 +1644,13 @@ define i64 @test_cvttss2siq(float %a0, float *%a1) {  ; HASWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_cvttss2siq: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    cvttss2si (%rdi), %rax # sched: [10:1.00] +; BROADWELL-SSE-NEXT:    cvttss2si %xmm0, %rcx # sched: [4:1.00] +; BROADWELL-SSE-NEXT:    addq %rcx, %rax # sched: [1:0.25] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_cvttss2siq:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vcvttss2si (%rdi), %rax # sched: [9:1.00] @@ -1026,6 +1658,13 @@ define i64 @test_cvttss2siq(float %a0, float *%a1) {  ; BROADWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_cvttss2siq: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    cvttss2si %xmm0, %rcx # sched: [7:1.00] +; SKYLAKE-SSE-NEXT:    cvttss2si (%rdi), %rax # sched: [12:1.00] +; SKYLAKE-SSE-NEXT:    addq %rcx, %rax # sched: [1:0.25] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_cvttss2siq:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vcvttss2si %xmm0, %rcx # sched: [7:1.00] @@ -1033,6 +1672,13 @@ define i64 @test_cvttss2siq(float %a0, float *%a1) {  ; SKYLAKE-NEXT:    addq %rcx, %rax # sched: [1:0.25]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_cvttss2siq: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    cvttss2si %xmm0, %rcx # sched: [7:1.00] +; SKX-SSE-NEXT:    cvttss2si (%rdi), %rax # sched: [12:1.00] +; SKX-SSE-NEXT:    addq %rcx, %rax # sched: [1:0.25] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_cvttss2siq:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vcvttss2si %xmm0, %rcx # sched: [7:1.00] @@ -1040,6 +1686,13 @@ define i64 @test_cvttss2siq(float %a0, float *%a1) {  ; SKX-NEXT:    addq %rcx, %rax # sched: [1:0.25]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_cvttss2siq: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    cvttss2si (%rdi), %rax # sched: [8:1.00] +; BTVER2-SSE-NEXT:    cvttss2si %xmm0, %rcx # sched: [3:1.00] +; BTVER2-SSE-NEXT:    addq %rcx, %rax # sched: [1:0.50] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_cvttss2siq:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vcvttss2si (%rdi), %rax # sched: [8:1.00] @@ -1047,6 +1700,13 @@ define i64 @test_cvttss2siq(float %a0, float *%a1) {  ; BTVER2-NEXT:    addq %rcx, %rax # sched: [1:0.50]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_cvttss2siq: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    cvttss2si (%rdi), %rax # sched: [12:1.00] +; ZNVER1-SSE-NEXT:    cvttss2si %xmm0, %rcx # sched: [5:1.00] +; ZNVER1-SSE-NEXT:    addq %rcx, %rax # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_cvttss2siq:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vcvttss2si (%rdi), %rax # sched: [12:1.00] @@ -1079,42 +1739,84 @@ define <4 x float> @test_divps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; SLM-NEXT:    divps (%rdi), %xmm0 # sched: [37:34.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_divps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    divps %xmm1, %xmm0 # sched: [14:1.00] +; SANDY-SSE-NEXT:    divps (%rdi), %xmm0 # sched: [20:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_divps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vdivps %xmm1, %xmm0, %xmm0 # sched: [14:1.00]  ; SANDY-NEXT:    vdivps (%rdi), %xmm0, %xmm0 # sched: [20:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_divps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    divps %xmm1, %xmm0 # sched: [11:1.00] +; HASWELL-SSE-NEXT:    divps (%rdi), %xmm0 # sched: [17:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_divps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vdivps %xmm1, %xmm0, %xmm0 # sched: [13:1.00]  ; HASWELL-NEXT:    vdivps (%rdi), %xmm0, %xmm0 # sched: [19:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_divps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    divps %xmm1, %xmm0 # sched: [11:1.00] +; BROADWELL-SSE-NEXT:    divps (%rdi), %xmm0 # sched: [16:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_divps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vdivps %xmm1, %xmm0, %xmm0 # sched: [11:1.00]  ; BROADWELL-NEXT:    vdivps (%rdi), %xmm0, %xmm0 # sched: [16:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_divps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    divps %xmm1, %xmm0 # sched: [11:1.00] +; SKYLAKE-SSE-NEXT:    divps (%rdi), %xmm0 # sched: [17:1.00] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_divps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vdivps %xmm1, %xmm0, %xmm0 # sched: [11:1.00]  ; SKYLAKE-NEXT:    vdivps (%rdi), %xmm0, %xmm0 # sched: [17:1.00]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_divps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    divps %xmm1, %xmm0 # sched: [11:1.00] +; SKX-SSE-NEXT:    divps (%rdi), %xmm0 # sched: [17:1.00] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_divps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vdivps %xmm1, %xmm0, %xmm0 # sched: [11:1.00]  ; SKX-NEXT:    vdivps (%rdi), %xmm0, %xmm0 # sched: [17:1.00]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_divps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    divps %xmm1, %xmm0 # sched: [19:19.00] +; BTVER2-SSE-NEXT:    divps (%rdi), %xmm0 # sched: [24:19.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_divps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vdivps %xmm1, %xmm0, %xmm0 # sched: [19:19.00]  ; BTVER2-NEXT:    vdivps (%rdi), %xmm0, %xmm0 # sched: [24:19.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_divps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    divps %xmm1, %xmm0 # sched: [15:1.00] +; ZNVER1-SSE-NEXT:    divps (%rdi), %xmm0 # sched: [22:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_divps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vdivps %xmm1, %xmm0, %xmm0 # sched: [15:1.00] @@ -1145,42 +1847,84 @@ define float @test_divss(float %a0, float %a1, float *%a2) {  ; SLM-NEXT:    divss (%rdi), %xmm0 # sched: [37:34.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_divss: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    divss %xmm1, %xmm0 # sched: [14:1.00] +; SANDY-SSE-NEXT:    divss (%rdi), %xmm0 # sched: [20:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_divss:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vdivss %xmm1, %xmm0, %xmm0 # sched: [14:1.00]  ; SANDY-NEXT:    vdivss (%rdi), %xmm0, %xmm0 # sched: [20:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_divss: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    divss %xmm1, %xmm0 # sched: [11:1.00] +; HASWELL-SSE-NEXT:    divss (%rdi), %xmm0 # sched: [16:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_divss:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vdivss %xmm1, %xmm0, %xmm0 # sched: [13:1.00]  ; HASWELL-NEXT:    vdivss (%rdi), %xmm0, %xmm0 # sched: [18:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_divss: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    divss %xmm1, %xmm0 # sched: [11:1.00] +; BROADWELL-SSE-NEXT:    divss (%rdi), %xmm0 # sched: [16:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_divss:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vdivss %xmm1, %xmm0, %xmm0 # sched: [11:1.00]  ; BROADWELL-NEXT:    vdivss (%rdi), %xmm0, %xmm0 # sched: [16:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_divss: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    divss %xmm1, %xmm0 # sched: [11:1.00] +; SKYLAKE-SSE-NEXT:    divss (%rdi), %xmm0 # sched: [16:1.00] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_divss:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vdivss %xmm1, %xmm0, %xmm0 # sched: [11:1.00]  ; SKYLAKE-NEXT:    vdivss (%rdi), %xmm0, %xmm0 # sched: [16:1.00]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_divss: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    divss %xmm1, %xmm0 # sched: [11:1.00] +; SKX-SSE-NEXT:    divss (%rdi), %xmm0 # sched: [16:1.00] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_divss:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vdivss %xmm1, %xmm0, %xmm0 # sched: [11:1.00]  ; SKX-NEXT:    vdivss (%rdi), %xmm0, %xmm0 # sched: [16:1.00]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_divss: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    divss %xmm1, %xmm0 # sched: [19:19.00] +; BTVER2-SSE-NEXT:    divss (%rdi), %xmm0 # sched: [24:19.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_divss:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vdivss %xmm1, %xmm0, %xmm0 # sched: [19:19.00]  ; BTVER2-NEXT:    vdivss (%rdi), %xmm0, %xmm0 # sched: [24:19.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_divss: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    divss %xmm1, %xmm0 # sched: [15:1.00] +; ZNVER1-SSE-NEXT:    divss (%rdi), %xmm0 # sched: [22:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_divss:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vdivss %xmm1, %xmm0, %xmm0 # sched: [15:1.00] @@ -1211,42 +1955,84 @@ define void @test_ldmxcsr(i32 %a0) {  ; SLM-NEXT:    ldmxcsr -{{[0-9]+}}(%rsp) # sched: [3:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_ldmxcsr: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [5:1.00] +; SANDY-SSE-NEXT:    ldmxcsr -{{[0-9]+}}(%rsp) # sched: [5:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_ldmxcsr:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [5:1.00]  ; SANDY-NEXT:    vldmxcsr -{{[0-9]+}}(%rsp) # sched: [5:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_ldmxcsr: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:1.00] +; HASWELL-SSE-NEXT:    ldmxcsr -{{[0-9]+}}(%rsp) # sched: [7:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_ldmxcsr:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:1.00]  ; HASWELL-NEXT:    vldmxcsr -{{[0-9]+}}(%rsp) # sched: [7:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_ldmxcsr: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:1.00] +; BROADWELL-SSE-NEXT:    ldmxcsr -{{[0-9]+}}(%rsp) # sched: [7:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_ldmxcsr:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:1.00]  ; BROADWELL-NEXT:    vldmxcsr -{{[0-9]+}}(%rsp) # sched: [7:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_ldmxcsr: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:1.00] +; SKYLAKE-SSE-NEXT:    ldmxcsr -{{[0-9]+}}(%rsp) # sched: [7:1.00] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_ldmxcsr:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:1.00]  ; SKYLAKE-NEXT:    vldmxcsr -{{[0-9]+}}(%rsp) # sched: [7:1.00]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_ldmxcsr: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:1.00] +; SKX-SSE-NEXT:    ldmxcsr -{{[0-9]+}}(%rsp) # sched: [7:1.00] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_ldmxcsr:  ; SKX:       # %bb.0:  ; SKX-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:1.00]  ; SKX-NEXT:    vldmxcsr -{{[0-9]+}}(%rsp) # sched: [7:1.00]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_ldmxcsr: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:1.00] +; BTVER2-SSE-NEXT:    ldmxcsr -{{[0-9]+}}(%rsp) # sched: [5:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_ldmxcsr:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:1.00]  ; BTVER2-NEXT:    vldmxcsr -{{[0-9]+}}(%rsp) # sched: [5:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_ldmxcsr: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:0.50] +; ZNVER1-SSE-NEXT:    ldmxcsr -{{[0-9]+}}(%rsp) # sched: [100:?] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_ldmxcsr:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:0.50] @@ -1279,42 +2065,84 @@ define <4 x float> @test_maxps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; SLM-NEXT:    maxps (%rdi), %xmm0 # sched: [6:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_maxps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    maxps %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    maxps (%rdi), %xmm0 # sched: [9:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_maxps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    vmaxps (%rdi), %xmm0, %xmm0 # sched: [9:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_maxps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    maxps %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    maxps (%rdi), %xmm0 # sched: [9:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_maxps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    vmaxps (%rdi), %xmm0, %xmm0 # sched: [9:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_maxps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    maxps %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    maxps (%rdi), %xmm0 # sched: [8:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_maxps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    vmaxps (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_maxps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    maxps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    maxps (%rdi), %xmm0 # sched: [10:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_maxps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    vmaxps (%rdi), %xmm0, %xmm0 # sched: [10:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_maxps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    maxps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    maxps (%rdi), %xmm0 # sched: [10:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_maxps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    vmaxps (%rdi), %xmm0, %xmm0 # sched: [10:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_maxps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    maxps %xmm1, %xmm0 # sched: [2:1.00] +; BTVER2-SSE-NEXT:    maxps (%rdi), %xmm0 # sched: [7:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_maxps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 # sched: [2:1.00]  ; BTVER2-NEXT:    vmaxps (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_maxps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    maxps %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    maxps (%rdi), %xmm0 # sched: [10:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_maxps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] @@ -1346,42 +2174,84 @@ define <4 x float> @test_maxss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; SLM-NEXT:    maxss (%rdi), %xmm0 # sched: [6:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_maxss: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    maxss %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    maxss (%rdi), %xmm0 # sched: [9:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_maxss:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vmaxss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    vmaxss (%rdi), %xmm0, %xmm0 # sched: [9:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_maxss: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    maxss %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    maxss (%rdi), %xmm0 # sched: [8:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_maxss:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vmaxss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    vmaxss (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_maxss: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    maxss %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    maxss (%rdi), %xmm0 # sched: [8:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_maxss:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vmaxss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    vmaxss (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_maxss: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    maxss %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    maxss (%rdi), %xmm0 # sched: [9:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_maxss:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vmaxss %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    vmaxss (%rdi), %xmm0, %xmm0 # sched: [9:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_maxss: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    maxss %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    maxss (%rdi), %xmm0 # sched: [9:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_maxss:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vmaxss %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    vmaxss (%rdi), %xmm0, %xmm0 # sched: [9:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_maxss: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    maxss %xmm1, %xmm0 # sched: [2:1.00] +; BTVER2-SSE-NEXT:    maxss (%rdi), %xmm0 # sched: [7:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_maxss:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmaxss %xmm1, %xmm0, %xmm0 # sched: [2:1.00]  ; BTVER2-NEXT:    vmaxss (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_maxss: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    maxss %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    maxss (%rdi), %xmm0 # sched: [10:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_maxss:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmaxss %xmm1, %xmm0, %xmm0 # sched: [3:1.00] @@ -1413,42 +2283,84 @@ define <4 x float> @test_minps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; SLM-NEXT:    minps (%rdi), %xmm0 # sched: [6:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_minps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    minps %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    minps (%rdi), %xmm0 # sched: [9:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_minps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vminps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    vminps (%rdi), %xmm0, %xmm0 # sched: [9:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_minps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    minps %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    minps (%rdi), %xmm0 # sched: [9:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_minps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vminps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    vminps (%rdi), %xmm0, %xmm0 # sched: [9:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_minps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    minps %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    minps (%rdi), %xmm0 # sched: [8:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_minps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vminps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    vminps (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_minps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    minps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    minps (%rdi), %xmm0 # sched: [10:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_minps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vminps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    vminps (%rdi), %xmm0, %xmm0 # sched: [10:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_minps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    minps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    minps (%rdi), %xmm0 # sched: [10:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_minps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vminps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    vminps (%rdi), %xmm0, %xmm0 # sched: [10:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_minps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    minps %xmm1, %xmm0 # sched: [2:1.00] +; BTVER2-SSE-NEXT:    minps (%rdi), %xmm0 # sched: [7:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_minps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vminps %xmm1, %xmm0, %xmm0 # sched: [2:1.00]  ; BTVER2-NEXT:    vminps (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_minps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    minps %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    minps (%rdi), %xmm0 # sched: [10:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_minps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vminps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] @@ -1480,42 +2392,84 @@ define <4 x float> @test_minss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; SLM-NEXT:    minss (%rdi), %xmm0 # sched: [6:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_minss: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    minss %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    minss (%rdi), %xmm0 # sched: [9:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_minss:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vminss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    vminss (%rdi), %xmm0, %xmm0 # sched: [9:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_minss: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    minss %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    minss (%rdi), %xmm0 # sched: [8:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_minss:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vminss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    vminss (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_minss: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    minss %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    minss (%rdi), %xmm0 # sched: [8:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_minss:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vminss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    vminss (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_minss: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    minss %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    minss (%rdi), %xmm0 # sched: [9:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_minss:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vminss %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    vminss (%rdi), %xmm0, %xmm0 # sched: [9:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_minss: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    minss %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    minss (%rdi), %xmm0 # sched: [9:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_minss:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vminss %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    vminss (%rdi), %xmm0, %xmm0 # sched: [9:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_minss: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    minss %xmm1, %xmm0 # sched: [2:1.00] +; BTVER2-SSE-NEXT:    minss (%rdi), %xmm0 # sched: [7:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_minss:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vminss %xmm1, %xmm0, %xmm0 # sched: [2:1.00]  ; BTVER2-NEXT:    vminss (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_minss: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    minss %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    minss (%rdi), %xmm0 # sched: [10:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_minss:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vminss %xmm1, %xmm0, %xmm0 # sched: [3:1.00] @@ -1550,6 +2504,13 @@ define void @test_movaps(<4 x float> *%a0, <4 x float> *%a1) {  ; SLM-NEXT:    movaps %xmm0, (%rsi) # sched: [1:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_movaps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    movaps (%rdi), %xmm0 # sched: [6:0.50] +; SANDY-SSE-NEXT:    addps %xmm0, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    movaps %xmm0, (%rsi) # sched: [5:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_movaps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vmovaps (%rdi), %xmm0 # sched: [6:0.50] @@ -1557,6 +2518,13 @@ define void @test_movaps(<4 x float> *%a0, <4 x float> *%a1) {  ; SANDY-NEXT:    vmovaps %xmm0, (%rsi) # sched: [5:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_movaps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    movaps (%rdi), %xmm0 # sched: [6:0.50] +; HASWELL-SSE-NEXT:    addps %xmm0, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    movaps %xmm0, (%rsi) # sched: [1:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_movaps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vmovaps (%rdi), %xmm0 # sched: [6:0.50] @@ -1564,6 +2532,13 @@ define void @test_movaps(<4 x float> *%a0, <4 x float> *%a1) {  ; HASWELL-NEXT:    vmovaps %xmm0, (%rsi) # sched: [1:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_movaps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    movaps (%rdi), %xmm0 # sched: [5:0.50] +; BROADWELL-SSE-NEXT:    addps %xmm0, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    movaps %xmm0, (%rsi) # sched: [1:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_movaps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vmovaps (%rdi), %xmm0 # sched: [5:0.50] @@ -1571,6 +2546,13 @@ define void @test_movaps(<4 x float> *%a0, <4 x float> *%a1) {  ; BROADWELL-NEXT:    vmovaps %xmm0, (%rsi) # sched: [1:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_movaps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    movaps (%rdi), %xmm0 # sched: [6:0.50] +; SKYLAKE-SSE-NEXT:    addps %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    movaps %xmm0, (%rsi) # sched: [1:1.00] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_movaps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vmovaps (%rdi), %xmm0 # sched: [6:0.50] @@ -1578,6 +2560,13 @@ define void @test_movaps(<4 x float> *%a0, <4 x float> *%a1) {  ; SKYLAKE-NEXT:    vmovaps %xmm0, (%rsi) # sched: [1:1.00]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_movaps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    movaps (%rdi), %xmm0 # sched: [6:0.50] +; SKX-SSE-NEXT:    addps %xmm0, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    movaps %xmm0, (%rsi) # sched: [1:1.00] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_movaps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vmovaps (%rdi), %xmm0 # sched: [6:0.50] @@ -1585,6 +2574,13 @@ define void @test_movaps(<4 x float> *%a0, <4 x float> *%a1) {  ; SKX-NEXT:    vmovaps %xmm0, (%rsi) # sched: [1:1.00]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_movaps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    movaps (%rdi), %xmm0 # sched: [5:1.00] +; BTVER2-SSE-NEXT:    addps %xmm0, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    movaps %xmm0, (%rsi) # sched: [1:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_movaps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmovaps (%rdi), %xmm0 # sched: [5:1.00] @@ -1592,6 +2588,13 @@ define void @test_movaps(<4 x float> *%a0, <4 x float> *%a1) {  ; BTVER2-NEXT:    vmovaps %xmm0, (%rsi) # sched: [1:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_movaps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    movaps (%rdi), %xmm0 # sched: [8:0.50] +; ZNVER1-SSE-NEXT:    addps %xmm0, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    movaps %xmm0, (%rsi) # sched: [1:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_movaps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmovaps (%rdi), %xmm0 # sched: [8:0.50] @@ -1628,36 +2631,71 @@ define <4 x float> @test_movhlps(<4 x float> %a0, <4 x float> %a1) {  ; SLM-NEXT:    movhlps {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_movhlps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_movhlps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_movhlps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_movhlps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_movhlps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_movhlps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_movhlps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_movhlps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_movhlps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_movhlps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_movhlps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:0.50] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_movhlps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:0.50]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_movhlps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_movhlps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:0.50] @@ -1689,9 +2727,18 @@ define void @test_movhps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {  ; SLM:       # %bb.0:  ; SLM-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [4:1.00]  ; SLM-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00] -; SLM-NEXT:    pextrq $1, %xmm1, (%rdi) # sched: [4:2.00] +; SLM-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1] sched: [1:1.00] +; SLM-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_movhps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [7:1.00] +; SANDY-SSE-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00] +; SANDY-SSE-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1] sched: [1:1.00] +; SANDY-SSE-NEXT:    movlps %xmm1, (%rdi) # sched: [5:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_movhps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [7:1.00] @@ -1699,6 +2746,14 @@ define void @test_movhps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {  ; SANDY-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [5:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_movhps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00] +; HASWELL-SSE-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1] sched: [1:1.00] +; HASWELL-SSE-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_movhps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00] @@ -1706,6 +2761,14 @@ define void @test_movhps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {  ; HASWELL-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [2:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_movhps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00] +; BROADWELL-SSE-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1] sched: [1:1.00] +; BROADWELL-SSE-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_movhps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00] @@ -1713,6 +2776,14 @@ define void @test_movhps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {  ; BROADWELL-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [2:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_movhps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00] +; SKYLAKE-SSE-NEXT:    addps %xmm0, %xmm1 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1] sched: [1:1.00] +; SKYLAKE-SSE-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_movhps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00] @@ -1720,6 +2791,14 @@ define void @test_movhps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {  ; SKYLAKE-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [2:1.00]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_movhps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00] +; SKX-SSE-NEXT:    addps %xmm0, %xmm1 # sched: [4:0.33] +; SKX-SSE-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1] sched: [1:1.00] +; SKX-SSE-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_movhps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00] @@ -1727,6 +2806,14 @@ define void @test_movhps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {  ; SKX-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [2:1.00]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_movhps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00] +; BTVER2-SSE-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1] sched: [1:0.50] +; BTVER2-SSE-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_movhps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00] @@ -1734,6 +2821,14 @@ define void @test_movhps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {  ; BTVER2-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [6:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_movhps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [8:0.50] +; ZNVER1-SSE-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1] sched: [1:0.50] +; ZNVER1-SSE-NEXT:    movlps %xmm1, (%rdi) # sched: [1:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_movhps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [8:0.50] @@ -1771,42 +2866,84 @@ define <4 x float> @test_movlhps(<4 x float> %a0, <4 x float> %a1) {  ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_movlhps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00] +; SANDY-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_movlhps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00]  ; SANDY-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_movlhps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00] +; HASWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_movlhps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00]  ; HASWELL-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_movlhps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00] +; BROADWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_movlhps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00]  ; BROADWELL-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_movlhps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00] +; SKYLAKE-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; 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: [4:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_movlhps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00] +; SKX-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_movlhps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00]  ; SKX-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_movlhps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:0.50] +; BTVER2-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_movlhps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:0.50]  ; BTVER2-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_movlhps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:0.50] +; ZNVER1-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_movlhps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:0.50] @@ -1839,6 +2976,13 @@ define void @test_movlps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {  ; SLM-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_movlps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [7:1.00] +; SANDY-SSE-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00] +; SANDY-SSE-NEXT:    movlps %xmm1, (%rdi) # sched: [5:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_movlps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vmovlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [7:1.00] @@ -1846,6 +2990,13 @@ define void @test_movlps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {  ; SANDY-NEXT:    vmovlps %xmm0, (%rdi) # sched: [5:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_movlps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [6:1.00] +; HASWELL-SSE-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_movlps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vmovlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [6:1.00] @@ -1853,6 +3004,13 @@ define void @test_movlps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {  ; HASWELL-NEXT:    vmovlps %xmm0, (%rdi) # sched: [1:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_movlps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [6:1.00] +; BROADWELL-SSE-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_movlps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vmovlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [6:1.00] @@ -1860,6 +3018,13 @@ define void @test_movlps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {  ; BROADWELL-NEXT:    vmovlps %xmm0, (%rdi) # sched: [1:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_movlps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [6:1.00] +; SKYLAKE-SSE-NEXT:    addps %xmm0, %xmm1 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_movlps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vmovlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [6:1.00] @@ -1867,6 +3032,13 @@ define void @test_movlps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {  ; SKYLAKE-NEXT:    vmovlps %xmm0, (%rdi) # sched: [1:1.00]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_movlps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [6:1.00] +; SKX-SSE-NEXT:    addps %xmm0, %xmm1 # sched: [4:0.33] +; SKX-SSE-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_movlps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vmovlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [6:1.00] @@ -1874,6 +3046,13 @@ define void @test_movlps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {  ; SKX-NEXT:    vmovlps %xmm0, (%rdi) # sched: [1:1.00]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_movlps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [6:1.00] +; BTVER2-SSE-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_movlps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmovlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [6:1.00] @@ -1881,6 +3060,13 @@ define void @test_movlps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {  ; BTVER2-NEXT:    vmovlps %xmm0, (%rdi) # sched: [1:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_movlps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [8:0.50] +; ZNVER1-SSE-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    movlps %xmm1, (%rdi) # sched: [1:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_movlps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmovlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [8:0.50] @@ -1915,36 +3101,71 @@ define i32 @test_movmskps(<4 x float> %a0) {  ; SLM-NEXT:    movmskps %xmm0, %eax # sched: [1:0.50]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_movmskps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    movmskps %xmm0, %eax # sched: [2:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_movmskps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vmovmskps %xmm0, %eax # sched: [2:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_movmskps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    movmskps %xmm0, %eax # sched: [3:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_movmskps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vmovmskps %xmm0, %eax # sched: [3:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_movmskps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    movmskps %xmm0, %eax # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_movmskps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vmovmskps %xmm0, %eax # sched: [3:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_movmskps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    movmskps %xmm0, %eax # sched: [2:1.00] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_movmskps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vmovmskps %xmm0, %eax # sched: [2:1.00]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_movmskps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    movmskps %xmm0, %eax # sched: [2:1.00] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_movmskps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vmovmskps %xmm0, %eax # sched: [2:1.00]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_movmskps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    movmskps %xmm0, %eax # sched: [3:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_movmskps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmovmskps %xmm0, %eax # sched: [3:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_movmskps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    movmskps %xmm0, %eax # sched: [1:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_movmskps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmovmskps %xmm0, %eax # sched: [1:1.00] @@ -1976,36 +3197,71 @@ define void @test_movntps(<4 x float> %a0, <4 x float> *%a1) {  ; SLM-NEXT:    movntps %xmm0, (%rdi) # sched: [1:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_movntps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    movntps %xmm0, (%rdi) # sched: [5:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_movntps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vmovntps %xmm0, (%rdi) # sched: [5:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_movntps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    movntps %xmm0, (%rdi) # sched: [1:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_movntps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vmovntps %xmm0, (%rdi) # sched: [1:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_movntps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    movntps %xmm0, (%rdi) # sched: [1:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_movntps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vmovntps %xmm0, (%rdi) # sched: [1:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_movntps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    movntps %xmm0, (%rdi) # sched: [1:1.00] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_movntps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vmovntps %xmm0, (%rdi) # sched: [1:1.00]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_movntps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    movntps %xmm0, (%rdi) # sched: [1:1.00] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_movntps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vmovntps %xmm0, (%rdi) # sched: [1:1.00]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_movntps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    movntps %xmm0, (%rdi) # sched: [3:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_movntps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmovntps %xmm0, (%rdi) # sched: [3:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_movntps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    movntps %xmm0, (%rdi) # sched: [1:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_movntps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmovntps %xmm0, (%rdi) # sched: [1:0.50] @@ -2036,6 +3292,13 @@ define void @test_movss_mem(float* %a0, float* %a1) {  ; SLM-NEXT:    movss %xmm0, (%rsi) # sched: [1:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_movss_mem: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [6:0.50] +; SANDY-SSE-NEXT:    addss %xmm0, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    movss %xmm0, (%rsi) # sched: [5:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_movss_mem:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [6:0.50] @@ -2043,6 +3306,13 @@ define void @test_movss_mem(float* %a0, float* %a1) {  ; SANDY-NEXT:    vmovss %xmm0, (%rsi) # sched: [5:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_movss_mem: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:0.50] +; HASWELL-SSE-NEXT:    addss %xmm0, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    movss %xmm0, (%rsi) # sched: [1:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_movss_mem:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:0.50] @@ -2050,6 +3320,13 @@ define void @test_movss_mem(float* %a0, float* %a1) {  ; HASWELL-NEXT:    vmovss %xmm0, (%rsi) # sched: [1:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_movss_mem: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:0.50] +; BROADWELL-SSE-NEXT:    addss %xmm0, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    movss %xmm0, (%rsi) # sched: [1:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_movss_mem:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:0.50] @@ -2057,6 +3334,13 @@ define void @test_movss_mem(float* %a0, float* %a1) {  ; BROADWELL-NEXT:    vmovss %xmm0, (%rsi) # sched: [1:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_movss_mem: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:0.50] +; SKYLAKE-SSE-NEXT:    addss %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    movss %xmm0, (%rsi) # sched: [1:1.00] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_movss_mem:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:0.50] @@ -2064,6 +3348,13 @@ define void @test_movss_mem(float* %a0, float* %a1) {  ; SKYLAKE-NEXT:    vmovss %xmm0, (%rsi) # sched: [1:1.00]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_movss_mem: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:0.50] +; SKX-SSE-NEXT:    addss %xmm0, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    movss %xmm0, (%rsi) # sched: [1:1.00] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_movss_mem:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:0.50] @@ -2071,6 +3362,13 @@ define void @test_movss_mem(float* %a0, float* %a1) {  ; SKX-NEXT:    vmovss %xmm0, (%rsi) # sched: [1:1.00]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_movss_mem: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:1.00] +; BTVER2-SSE-NEXT:    addss %xmm0, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    movss %xmm0, (%rsi) # sched: [1:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_movss_mem:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:1.00] @@ -2078,6 +3376,13 @@ define void @test_movss_mem(float* %a0, float* %a1) {  ; BTVER2-NEXT:    vmovss %xmm0, (%rsi) # sched: [1:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_movss_mem: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [8:0.50] +; ZNVER1-SSE-NEXT:    addss %xmm0, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    movss %xmm0, (%rsi) # sched: [1:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_movss_mem:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [8:0.50] @@ -2109,39 +3414,74 @@ define <4 x float> @test_movss_reg(<4 x float> %a0, <4 x float> %a1) {  ;  ; SLM-LABEL: test_movss_reg:  ; SLM:       # %bb.0: -; SLM-NEXT:    blendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:1.00] +; SLM-NEXT:    movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_movss_reg: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_movss_reg:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.50]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_movss_reg: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_movss_reg:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.33]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_movss_reg: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_movss_reg:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.33]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_movss_reg: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.33] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_movss_reg:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.33]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_movss_reg: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_movss_reg:  ; SKX:       # %bb.0: -; SKX-NEXT:    vmovss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.33] +; SKX-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_movss_reg: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.50] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_movss_reg:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.50]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_movss_reg: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_movss_reg:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.50] @@ -2172,6 +3512,13 @@ define void @test_movups(<4 x float> *%a0, <4 x float> *%a1) {  ; SLM-NEXT:    movups %xmm0, (%rsi) # sched: [1:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_movups: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    movups (%rdi), %xmm0 # sched: [6:0.50] +; SANDY-SSE-NEXT:    addps %xmm0, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    movups %xmm0, (%rsi) # sched: [5:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_movups:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vmovups (%rdi), %xmm0 # sched: [6:0.50] @@ -2179,6 +3526,13 @@ define void @test_movups(<4 x float> *%a0, <4 x float> *%a1) {  ; SANDY-NEXT:    vmovups %xmm0, (%rsi) # sched: [5:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_movups: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    movups (%rdi), %xmm0 # sched: [6:0.50] +; HASWELL-SSE-NEXT:    addps %xmm0, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    movups %xmm0, (%rsi) # sched: [1:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_movups:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vmovups (%rdi), %xmm0 # sched: [6:0.50] @@ -2186,6 +3540,13 @@ define void @test_movups(<4 x float> *%a0, <4 x float> *%a1) {  ; HASWELL-NEXT:    vmovups %xmm0, (%rsi) # sched: [1:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_movups: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    movups (%rdi), %xmm0 # sched: [5:0.50] +; BROADWELL-SSE-NEXT:    addps %xmm0, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    movups %xmm0, (%rsi) # sched: [1:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_movups:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vmovups (%rdi), %xmm0 # sched: [5:0.50] @@ -2193,6 +3554,13 @@ define void @test_movups(<4 x float> *%a0, <4 x float> *%a1) {  ; BROADWELL-NEXT:    vmovups %xmm0, (%rsi) # sched: [1:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_movups: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    movups (%rdi), %xmm0 # sched: [6:0.50] +; SKYLAKE-SSE-NEXT:    addps %xmm0, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    movups %xmm0, (%rsi) # sched: [1:1.00] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_movups:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vmovups (%rdi), %xmm0 # sched: [6:0.50] @@ -2200,6 +3568,13 @@ define void @test_movups(<4 x float> *%a0, <4 x float> *%a1) {  ; SKYLAKE-NEXT:    vmovups %xmm0, (%rsi) # sched: [1:1.00]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_movups: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    movups (%rdi), %xmm0 # sched: [6:0.50] +; SKX-SSE-NEXT:    addps %xmm0, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    movups %xmm0, (%rsi) # sched: [1:1.00] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_movups:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vmovups (%rdi), %xmm0 # sched: [6:0.50] @@ -2207,6 +3582,13 @@ define void @test_movups(<4 x float> *%a0, <4 x float> *%a1) {  ; SKX-NEXT:    vmovups %xmm0, (%rsi) # sched: [1:1.00]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_movups: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    movups (%rdi), %xmm0 # sched: [5:1.00] +; BTVER2-SSE-NEXT:    addps %xmm0, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    movups %xmm0, (%rsi) # sched: [1:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_movups:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmovups (%rdi), %xmm0 # sched: [5:1.00] @@ -2214,6 +3596,13 @@ define void @test_movups(<4 x float> *%a0, <4 x float> *%a1) {  ; BTVER2-NEXT:    vmovups %xmm0, (%rsi) # sched: [1:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_movups: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    movups (%rdi), %xmm0 # sched: [8:0.50] +; ZNVER1-SSE-NEXT:    addps %xmm0, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    movups %xmm0, (%rsi) # sched: [1:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_movups:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmovups (%rdi), %xmm0 # sched: [8:0.50] @@ -2245,42 +3634,84 @@ define <4 x float> @test_mulps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; SLM-NEXT:    mulps (%rdi), %xmm0 # sched: [8:2.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_mulps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    mulps %xmm1, %xmm0 # sched: [5:1.00] +; SANDY-SSE-NEXT:    mulps (%rdi), %xmm0 # sched: [11:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_mulps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vmulps %xmm1, %xmm0, %xmm0 # sched: [5:1.00]  ; SANDY-NEXT:    vmulps (%rdi), %xmm0, %xmm0 # sched: [11:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_mulps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    mulps %xmm1, %xmm0 # sched: [5:0.50] +; HASWELL-SSE-NEXT:    mulps (%rdi), %xmm0 # sched: [11:0.50] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_mulps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vmulps %xmm1, %xmm0, %xmm0 # sched: [5:0.50]  ; HASWELL-NEXT:    vmulps (%rdi), %xmm0, %xmm0 # sched: [11:0.50]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_mulps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    mulps %xmm1, %xmm0 # sched: [3:0.50] +; BROADWELL-SSE-NEXT:    mulps (%rdi), %xmm0 # sched: [8:0.50] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_mulps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vmulps %xmm1, %xmm0, %xmm0 # sched: [3:0.50]  ; BROADWELL-NEXT:    vmulps (%rdi), %xmm0, %xmm0 # sched: [8:0.50]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_mulps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    mulps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    mulps (%rdi), %xmm0 # sched: [10:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_mulps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vmulps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    vmulps (%rdi), %xmm0, %xmm0 # sched: [10:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_mulps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    mulps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    mulps (%rdi), %xmm0 # sched: [10:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_mulps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vmulps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    vmulps (%rdi), %xmm0, %xmm0 # sched: [10:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_mulps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    mulps %xmm1, %xmm0 # sched: [2:1.00] +; BTVER2-SSE-NEXT:    mulps (%rdi), %xmm0 # sched: [7:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_mulps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmulps %xmm1, %xmm0, %xmm0 # sched: [2:1.00]  ; BTVER2-NEXT:    vmulps (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_mulps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    mulps %xmm1, %xmm0 # sched: [3:0.50] +; ZNVER1-SSE-NEXT:    mulps (%rdi), %xmm0 # sched: [10:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_mulps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmulps %xmm1, %xmm0, %xmm0 # sched: [3:0.50] @@ -2311,42 +3742,84 @@ define float @test_mulss(float %a0, float %a1, float *%a2) {  ; SLM-NEXT:    mulss (%rdi), %xmm0 # sched: [8:2.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_mulss: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    mulss %xmm1, %xmm0 # sched: [5:1.00] +; SANDY-SSE-NEXT:    mulss (%rdi), %xmm0 # sched: [11:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_mulss:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vmulss %xmm1, %xmm0, %xmm0 # sched: [5:1.00]  ; SANDY-NEXT:    vmulss (%rdi), %xmm0, %xmm0 # sched: [11:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_mulss: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    mulss %xmm1, %xmm0 # sched: [5:0.50] +; HASWELL-SSE-NEXT:    mulss (%rdi), %xmm0 # sched: [10:0.50] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_mulss:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vmulss %xmm1, %xmm0, %xmm0 # sched: [5:0.50]  ; HASWELL-NEXT:    vmulss (%rdi), %xmm0, %xmm0 # sched: [10:0.50]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_mulss: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    mulss %xmm1, %xmm0 # sched: [3:0.50] +; BROADWELL-SSE-NEXT:    mulss (%rdi), %xmm0 # sched: [8:0.50] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_mulss:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vmulss %xmm1, %xmm0, %xmm0 # sched: [3:0.50]  ; BROADWELL-NEXT:    vmulss (%rdi), %xmm0, %xmm0 # sched: [8:0.50]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_mulss: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    mulss %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    mulss (%rdi), %xmm0 # sched: [9:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_mulss:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vmulss %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    vmulss (%rdi), %xmm0, %xmm0 # sched: [9:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_mulss: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    mulss %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    mulss (%rdi), %xmm0 # sched: [9:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_mulss:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vmulss %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    vmulss (%rdi), %xmm0, %xmm0 # sched: [9:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_mulss: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    mulss %xmm1, %xmm0 # sched: [2:1.00] +; BTVER2-SSE-NEXT:    mulss (%rdi), %xmm0 # sched: [7:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_mulss:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmulss %xmm1, %xmm0, %xmm0 # sched: [2:1.00]  ; BTVER2-NEXT:    vmulss (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_mulss: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    mulss %xmm1, %xmm0 # sched: [3:0.50] +; ZNVER1-SSE-NEXT:    mulss (%rdi), %xmm0 # sched: [10:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_mulss:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmulss %xmm1, %xmm0, %xmm0 # sched: [3:0.50] @@ -2381,42 +3854,84 @@ define <4 x float> @test_orps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2  ; SLM-NEXT:    orps (%rdi), %xmm0 # sched: [4:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_orps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    orps %xmm1, %xmm0 # sched: [1:1.00] +; SANDY-SSE-NEXT:    orps (%rdi), %xmm0 # sched: [7:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_orps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vorps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]  ; SANDY-NEXT:    vorps (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_orps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    orps %xmm1, %xmm0 # sched: [1:1.00] +; HASWELL-SSE-NEXT:    orps (%rdi), %xmm0 # sched: [7:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_orps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vorps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]  ; HASWELL-NEXT:    vorps (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_orps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    orps %xmm1, %xmm0 # sched: [1:1.00] +; BROADWELL-SSE-NEXT:    orps (%rdi), %xmm0 # sched: [6:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_orps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vorps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]  ; BROADWELL-NEXT:    vorps (%rdi), %xmm0, %xmm0 # sched: [6:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_orps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    orps %xmm1, %xmm0 # sched: [1:0.33] +; SKYLAKE-SSE-NEXT:    orps (%rdi), %xmm0 # sched: [7:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_orps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vorps %xmm1, %xmm0, %xmm0 # sched: [1:0.33]  ; SKYLAKE-NEXT:    vorps (%rdi), %xmm0, %xmm0 # sched: [7:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_orps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    orps %xmm1, %xmm0 # sched: [1:0.33] +; SKX-SSE-NEXT:    orps (%rdi), %xmm0 # sched: [7:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_orps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vorps %xmm1, %xmm0, %xmm0 # sched: [1:0.33]  ; SKX-NEXT:    vorps (%rdi), %xmm0, %xmm0 # sched: [7:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_orps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    orps %xmm1, %xmm0 # sched: [1:0.50] +; BTVER2-SSE-NEXT:    orps (%rdi), %xmm0 # sched: [6:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_orps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vorps %xmm1, %xmm0, %xmm0 # sched: [1:0.50]  ; BTVER2-NEXT:    vorps (%rdi), %xmm0, %xmm0 # sched: [6:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_orps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    orps %xmm1, %xmm0 # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    orps (%rdi), %xmm0 # sched: [8:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_orps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vorps %xmm1, %xmm0, %xmm0 # sched: [1:0.25] @@ -2463,6 +3978,16 @@ define void @test_prefetch(i8* %a0) optsize {  ; SLM-NEXT:    #NO_APP  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_prefetch: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    #APP +; SANDY-SSE-NEXT:    prefetchnta (%rdi) # sched: [5:0.50] +; SANDY-SSE-NEXT:    prefetcht0 (%rdi) # sched: [5:0.50] +; SANDY-SSE-NEXT:    prefetcht1 (%rdi) # sched: [5:0.50] +; SANDY-SSE-NEXT:    prefetcht2 (%rdi) # sched: [5:0.50] +; SANDY-SSE-NEXT:    #NO_APP +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_prefetch:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    #APP @@ -2473,6 +3998,16 @@ define void @test_prefetch(i8* %a0) optsize {  ; SANDY-NEXT:    #NO_APP  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_prefetch: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    #APP +; HASWELL-SSE-NEXT:    prefetchnta (%rdi) # sched: [5:0.50] +; HASWELL-SSE-NEXT:    prefetcht0 (%rdi) # sched: [5:0.50] +; HASWELL-SSE-NEXT:    prefetcht1 (%rdi) # sched: [5:0.50] +; HASWELL-SSE-NEXT:    prefetcht2 (%rdi) # sched: [5:0.50] +; HASWELL-SSE-NEXT:    #NO_APP +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_prefetch:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    #APP @@ -2483,6 +4018,16 @@ define void @test_prefetch(i8* %a0) optsize {  ; HASWELL-NEXT:    #NO_APP  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_prefetch: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    #APP +; BROADWELL-SSE-NEXT:    prefetchnta (%rdi) # sched: [5:0.50] +; BROADWELL-SSE-NEXT:    prefetcht0 (%rdi) # sched: [5:0.50] +; BROADWELL-SSE-NEXT:    prefetcht1 (%rdi) # sched: [5:0.50] +; BROADWELL-SSE-NEXT:    prefetcht2 (%rdi) # sched: [5:0.50] +; BROADWELL-SSE-NEXT:    #NO_APP +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_prefetch:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    #APP @@ -2493,6 +4038,16 @@ define void @test_prefetch(i8* %a0) optsize {  ; BROADWELL-NEXT:    #NO_APP  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_prefetch: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    #APP +; SKYLAKE-SSE-NEXT:    prefetchnta (%rdi) # sched: [5:0.50] +; SKYLAKE-SSE-NEXT:    prefetcht0 (%rdi) # sched: [5:0.50] +; SKYLAKE-SSE-NEXT:    prefetcht1 (%rdi) # sched: [5:0.50] +; SKYLAKE-SSE-NEXT:    prefetcht2 (%rdi) # sched: [5:0.50] +; SKYLAKE-SSE-NEXT:    #NO_APP +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_prefetch:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    #APP @@ -2503,6 +4058,16 @@ define void @test_prefetch(i8* %a0) optsize {  ; SKYLAKE-NEXT:    #NO_APP  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_prefetch: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    #APP +; SKX-SSE-NEXT:    prefetchnta (%rdi) # sched: [5:0.50] +; SKX-SSE-NEXT:    prefetcht0 (%rdi) # sched: [5:0.50] +; SKX-SSE-NEXT:    prefetcht1 (%rdi) # sched: [5:0.50] +; SKX-SSE-NEXT:    prefetcht2 (%rdi) # sched: [5:0.50] +; SKX-SSE-NEXT:    #NO_APP +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_prefetch:  ; SKX:       # %bb.0:  ; SKX-NEXT:    #APP @@ -2513,6 +4078,16 @@ define void @test_prefetch(i8* %a0) optsize {  ; SKX-NEXT:    #NO_APP  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_prefetch: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    #APP +; BTVER2-SSE-NEXT:    prefetchnta (%rdi) # sched: [5:1.00] +; BTVER2-SSE-NEXT:    prefetcht0 (%rdi) # sched: [5:1.00] +; BTVER2-SSE-NEXT:    prefetcht1 (%rdi) # sched: [5:1.00] +; BTVER2-SSE-NEXT:    prefetcht2 (%rdi) # sched: [5:1.00] +; BTVER2-SSE-NEXT:    #NO_APP +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_prefetch:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    #APP @@ -2523,6 +4098,16 @@ define void @test_prefetch(i8* %a0) optsize {  ; BTVER2-NEXT:    #NO_APP  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_prefetch: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    #APP +; ZNVER1-SSE-NEXT:    prefetchnta (%rdi) # sched: [8:0.50] +; ZNVER1-SSE-NEXT:    prefetcht0 (%rdi) # sched: [8:0.50] +; ZNVER1-SSE-NEXT:    prefetcht1 (%rdi) # sched: [8:0.50] +; ZNVER1-SSE-NEXT:    prefetcht2 (%rdi) # sched: [8:0.50] +; ZNVER1-SSE-NEXT:    #NO_APP +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_prefetch:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    #APP @@ -2560,6 +4145,13 @@ define <4 x float> @test_rcpps(<4 x float> %a0, <4 x float> *%a1) {  ; SLM-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.50]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_rcpps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    rcpps %xmm0, %xmm1 # sched: [5:1.00] +; SANDY-SSE-NEXT:    rcpps (%rdi), %xmm0 # sched: [11:1.00] +; SANDY-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_rcpps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vrcpps %xmm0, %xmm0 # sched: [5:1.00] @@ -2567,6 +4159,13 @@ define <4 x float> @test_rcpps(<4 x float> %a0, <4 x float> *%a1) {  ; SANDY-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_rcpps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    rcpps %xmm0, %xmm1 # sched: [5:1.00] +; HASWELL-SSE-NEXT:    rcpps (%rdi), %xmm0 # sched: [11:1.00] +; HASWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_rcpps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vrcpps %xmm0, %xmm0 # sched: [5:1.00] @@ -2574,6 +4173,13 @@ define <4 x float> @test_rcpps(<4 x float> %a0, <4 x float> *%a1) {  ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_rcpps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    rcpps %xmm0, %xmm1 # sched: [5:1.00] +; BROADWELL-SSE-NEXT:    rcpps (%rdi), %xmm0 # sched: [10:1.00] +; BROADWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_rcpps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vrcpps %xmm0, %xmm0 # sched: [5:1.00] @@ -2581,6 +4187,13 @@ define <4 x float> @test_rcpps(<4 x float> %a0, <4 x float> *%a1) {  ; BROADWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_rcpps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    rcpps %xmm0, %xmm1 # sched: [4:1.00] +; SKYLAKE-SSE-NEXT:    rcpps (%rdi), %xmm0 # sched: [10:1.00] +; SKYLAKE-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_rcpps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vrcpps %xmm0, %xmm0 # sched: [4:1.00] @@ -2588,6 +4201,13 @@ define <4 x float> @test_rcpps(<4 x float> %a0, <4 x float> *%a1) {  ; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_rcpps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    rcpps %xmm0, %xmm1 # sched: [4:1.00] +; SKX-SSE-NEXT:    rcpps (%rdi), %xmm0 # sched: [10:1.00] +; SKX-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_rcpps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vrcpps %xmm0, %xmm0 # sched: [4:1.00] @@ -2595,6 +4215,13 @@ define <4 x float> @test_rcpps(<4 x float> %a0, <4 x float> *%a1) {  ; SKX-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_rcpps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    rcpps %xmm0, %xmm1 # sched: [2:1.00] +; BTVER2-SSE-NEXT:    rcpps (%rdi), %xmm0 # sched: [7:1.00] +; BTVER2-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_rcpps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vrcpps (%rdi), %xmm1 # sched: [7:1.00] @@ -2602,6 +4229,13 @@ define <4 x float> @test_rcpps(<4 x float> %a0, <4 x float> *%a1) {  ; BTVER2-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_rcpps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    rcpps %xmm0, %xmm1 # sched: [5:0.50] +; ZNVER1-SSE-NEXT:    rcpps (%rdi), %xmm0 # sched: [12:0.50] +; ZNVER1-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_rcpps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vrcpps (%rdi), %xmm1 # sched: [12:0.50] @@ -2643,6 +4277,14 @@ define <4 x float> @test_rcpss(float %a0, float *%a1) {  ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_rcpss: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    rcpss %xmm0, %xmm0 # sched: [5:1.00] +; SANDY-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [6:0.50] +; SANDY-SSE-NEXT:    rcpss %xmm1, %xmm1 # sched: [5:1.00] +; SANDY-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_rcpss:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vrcpss %xmm0, %xmm0, %xmm0 # sched: [5:1.00] @@ -2651,6 +4293,14 @@ define <4 x float> @test_rcpss(float %a0, float *%a1) {  ; SANDY-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_rcpss: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    rcpss %xmm0, %xmm0 # sched: [5:1.00] +; HASWELL-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:0.50] +; HASWELL-SSE-NEXT:    rcpss %xmm1, %xmm1 # sched: [5:1.00] +; HASWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_rcpss:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vrcpss %xmm0, %xmm0, %xmm0 # sched: [5:1.00] @@ -2659,6 +4309,14 @@ define <4 x float> @test_rcpss(float %a0, float *%a1) {  ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_rcpss: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    rcpss %xmm0, %xmm0 # sched: [5:1.00] +; BROADWELL-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:0.50] +; BROADWELL-SSE-NEXT:    rcpss %xmm1, %xmm1 # sched: [5:1.00] +; BROADWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_rcpss:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vrcpss %xmm0, %xmm0, %xmm0 # sched: [5:1.00] @@ -2667,6 +4325,14 @@ define <4 x float> @test_rcpss(float %a0, float *%a1) {  ; BROADWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_rcpss: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    rcpss %xmm0, %xmm0 # sched: [4:1.00] +; SKYLAKE-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:0.50] +; SKYLAKE-SSE-NEXT:    rcpss %xmm1, %xmm1 # sched: [4:1.00] +; SKYLAKE-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_rcpss:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vrcpss %xmm0, %xmm0, %xmm0 # sched: [4:1.00] @@ -2675,6 +4341,14 @@ define <4 x float> @test_rcpss(float %a0, float *%a1) {  ; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_rcpss: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    rcpss %xmm0, %xmm0 # sched: [4:1.00] +; SKX-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:0.50] +; SKX-SSE-NEXT:    rcpss %xmm1, %xmm1 # sched: [4:1.00] +; SKX-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_rcpss:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vrcpss %xmm0, %xmm0, %xmm0 # sched: [4:1.00] @@ -2683,6 +4357,14 @@ define <4 x float> @test_rcpss(float %a0, float *%a1) {  ; SKX-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_rcpss: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:1.00] +; BTVER2-SSE-NEXT:    rcpss %xmm0, %xmm0 # sched: [7:1.00] +; BTVER2-SSE-NEXT:    rcpss %xmm1, %xmm1 # sched: [7:1.00] +; BTVER2-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_rcpss:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:1.00] @@ -2691,6 +4373,14 @@ define <4 x float> @test_rcpss(float %a0, float *%a1) {  ; BTVER2-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_rcpss: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [8:0.50] +; ZNVER1-SSE-NEXT:    rcpss %xmm0, %xmm0 # sched: [12:0.50] +; ZNVER1-SSE-NEXT:    rcpss %xmm1, %xmm1 # sched: [12:0.50] +; ZNVER1-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_rcpss:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [8:0.50] @@ -2732,6 +4422,13 @@ define <4 x float> @test_rsqrtps(<4 x float> %a0, <4 x float> *%a1) {  ; SLM-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.50]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_rsqrtps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    rsqrtps %xmm0, %xmm1 # sched: [5:1.00] +; SANDY-SSE-NEXT:    rsqrtps (%rdi), %xmm0 # sched: [11:1.00] +; SANDY-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_rsqrtps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vrsqrtps %xmm0, %xmm0 # sched: [5:1.00] @@ -2739,6 +4436,13 @@ define <4 x float> @test_rsqrtps(<4 x float> %a0, <4 x float> *%a1) {  ; SANDY-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_rsqrtps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    rsqrtps %xmm0, %xmm1 # sched: [5:1.00] +; HASWELL-SSE-NEXT:    rsqrtps (%rdi), %xmm0 # sched: [11:1.00] +; HASWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_rsqrtps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vrsqrtps %xmm0, %xmm0 # sched: [5:1.00] @@ -2746,6 +4450,13 @@ define <4 x float> @test_rsqrtps(<4 x float> %a0, <4 x float> *%a1) {  ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_rsqrtps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    rsqrtps %xmm0, %xmm1 # sched: [5:1.00] +; BROADWELL-SSE-NEXT:    rsqrtps (%rdi), %xmm0 # sched: [10:1.00] +; BROADWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_rsqrtps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vrsqrtps %xmm0, %xmm0 # sched: [5:1.00] @@ -2753,6 +4464,13 @@ define <4 x float> @test_rsqrtps(<4 x float> %a0, <4 x float> *%a1) {  ; BROADWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_rsqrtps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    rsqrtps %xmm0, %xmm1 # sched: [4:1.00] +; SKYLAKE-SSE-NEXT:    rsqrtps (%rdi), %xmm0 # sched: [10:1.00] +; SKYLAKE-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_rsqrtps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vrsqrtps %xmm0, %xmm0 # sched: [4:1.00] @@ -2760,6 +4478,13 @@ define <4 x float> @test_rsqrtps(<4 x float> %a0, <4 x float> *%a1) {  ; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_rsqrtps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    rsqrtps %xmm0, %xmm1 # sched: [4:1.00] +; SKX-SSE-NEXT:    rsqrtps (%rdi), %xmm0 # sched: [10:1.00] +; SKX-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_rsqrtps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vrsqrtps %xmm0, %xmm0 # sched: [4:1.00] @@ -2767,6 +4492,13 @@ define <4 x float> @test_rsqrtps(<4 x float> %a0, <4 x float> *%a1) {  ; SKX-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_rsqrtps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    rsqrtps %xmm0, %xmm1 # sched: [2:1.00] +; BTVER2-SSE-NEXT:    rsqrtps (%rdi), %xmm0 # sched: [7:1.00] +; BTVER2-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_rsqrtps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vrsqrtps (%rdi), %xmm1 # sched: [7:1.00] @@ -2774,6 +4506,13 @@ define <4 x float> @test_rsqrtps(<4 x float> %a0, <4 x float> *%a1) {  ; BTVER2-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_rsqrtps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    rsqrtps %xmm0, %xmm1 # sched: [5:0.50] +; ZNVER1-SSE-NEXT:    rsqrtps (%rdi), %xmm0 # sched: [12:0.50] +; ZNVER1-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_rsqrtps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vrsqrtps (%rdi), %xmm1 # sched: [12:0.50] @@ -2815,6 +4554,14 @@ define <4 x float> @test_rsqrtss(float %a0, float *%a1) {  ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_rsqrtss: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    rsqrtss %xmm0, %xmm0 # sched: [5:1.00] +; SANDY-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [6:0.50] +; SANDY-SSE-NEXT:    rsqrtss %xmm1, %xmm1 # sched: [5:1.00] +; SANDY-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_rsqrtss:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vrsqrtss %xmm0, %xmm0, %xmm0 # sched: [5:1.00] @@ -2823,6 +4570,14 @@ define <4 x float> @test_rsqrtss(float %a0, float *%a1) {  ; SANDY-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_rsqrtss: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    rsqrtss %xmm0, %xmm0 # sched: [5:1.00] +; HASWELL-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:0.50] +; HASWELL-SSE-NEXT:    rsqrtss %xmm1, %xmm1 # sched: [5:1.00] +; HASWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_rsqrtss:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vrsqrtss %xmm0, %xmm0, %xmm0 # sched: [5:1.00] @@ -2831,6 +4586,14 @@ define <4 x float> @test_rsqrtss(float %a0, float *%a1) {  ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_rsqrtss: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    rsqrtss %xmm0, %xmm0 # sched: [5:1.00] +; BROADWELL-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:0.50] +; BROADWELL-SSE-NEXT:    rsqrtss %xmm1, %xmm1 # sched: [5:1.00] +; BROADWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_rsqrtss:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vrsqrtss %xmm0, %xmm0, %xmm0 # sched: [5:1.00] @@ -2839,6 +4602,14 @@ define <4 x float> @test_rsqrtss(float %a0, float *%a1) {  ; BROADWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_rsqrtss: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    rsqrtss %xmm0, %xmm0 # sched: [4:1.00] +; SKYLAKE-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:0.50] +; SKYLAKE-SSE-NEXT:    rsqrtss %xmm1, %xmm1 # sched: [4:1.00] +; SKYLAKE-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_rsqrtss:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vrsqrtss %xmm0, %xmm0, %xmm0 # sched: [4:1.00] @@ -2847,6 +4618,14 @@ define <4 x float> @test_rsqrtss(float %a0, float *%a1) {  ; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_rsqrtss: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    rsqrtss %xmm0, %xmm0 # sched: [4:1.00] +; SKX-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:0.50] +; SKX-SSE-NEXT:    rsqrtss %xmm1, %xmm1 # sched: [4:1.00] +; SKX-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_rsqrtss:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vrsqrtss %xmm0, %xmm0, %xmm0 # sched: [4:1.00] @@ -2855,6 +4634,14 @@ define <4 x float> @test_rsqrtss(float %a0, float *%a1) {  ; SKX-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_rsqrtss: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:1.00] +; BTVER2-SSE-NEXT:    rsqrtss %xmm0, %xmm0 # sched: [7:1.00] +; BTVER2-SSE-NEXT:    rsqrtss %xmm1, %xmm1 # sched: [7:1.00] +; BTVER2-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_rsqrtss:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:1.00] @@ -2863,6 +4650,14 @@ define <4 x float> @test_rsqrtss(float %a0, float *%a1) {  ; BTVER2-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_rsqrtss: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [8:0.50] +; ZNVER1-SSE-NEXT:    rsqrtss %xmm0, %xmm0 # sched: [5:0.50] +; ZNVER1-SSE-NEXT:    rsqrtss %xmm1, %xmm1 # sched: [5:0.50] +; ZNVER1-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_rsqrtss:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [8:0.50] @@ -2902,36 +4697,71 @@ define void @test_sfence() {  ; SLM-NEXT:    sfence # sched: [1:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_sfence: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    sfence # sched: [1:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_sfence:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    sfence # sched: [1:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_sfence: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    sfence # sched: [2:0.33] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_sfence:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    sfence # sched: [2:0.33]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_sfence: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    sfence # sched: [2:0.33] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_sfence:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    sfence # sched: [2:0.33]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_sfence: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    sfence # sched: [2:0.33] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_sfence:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    sfence # sched: [2:0.33]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_sfence: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    sfence # sched: [2:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_sfence:  ; SKX:       # %bb.0:  ; SKX-NEXT:    sfence # sched: [2:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_sfence: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    sfence # sched: [1:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_sfence:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    sfence # sched: [1:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_sfence: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    sfence # sched: [1:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_sfence:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    sfence # sched: [1:0.50] @@ -2963,6 +4793,13 @@ define <4 x float> @test_shufps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%  ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_shufps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:1.00] +; SANDY-SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,3],mem[0,0] sched: [7:1.00] +; SANDY-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_shufps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:1.00] @@ -2970,6 +4807,13 @@ define <4 x float> @test_shufps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%  ; SANDY-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_shufps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:1.00] +; HASWELL-SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,3],mem[0,0] sched: [7:1.00] +; HASWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_shufps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:1.00] @@ -2977,6 +4821,13 @@ define <4 x float> @test_shufps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%  ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_shufps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:1.00] +; BROADWELL-SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,3],mem[0,0] sched: [6:1.00] +; BROADWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_shufps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:1.00] @@ -2984,6 +4835,13 @@ define <4 x float> @test_shufps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%  ; BROADWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_shufps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:1.00] +; SKYLAKE-SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,3],mem[0,0] sched: [7:1.00] +; SKYLAKE-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_shufps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:1.00] @@ -2991,6 +4849,13 @@ define <4 x float> @test_shufps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%  ; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_shufps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:1.00] +; SKX-SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,3],mem[0,0] sched: [7:1.00] +; SKX-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_shufps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:1.00] @@ -2998,6 +4863,13 @@ define <4 x float> @test_shufps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%  ; SKX-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_shufps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:0.50] +; BTVER2-SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,3],mem[0,0] sched: [6:1.00] +; BTVER2-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_shufps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:0.50] @@ -3005,6 +4877,13 @@ define <4 x float> @test_shufps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%  ; BTVER2-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_shufps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:0.50] +; ZNVER1-SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,3],mem[0,0] sched: [8:0.50] +; ZNVER1-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_shufps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:0.50] @@ -3041,6 +4920,13 @@ define <4 x float> @test_sqrtps(<4 x float> %a0, <4 x float> *%a1) {  ; SLM-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.50]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_sqrtps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    sqrtps %xmm0, %xmm1 # sched: [14:1.00] +; SANDY-SSE-NEXT:    sqrtps (%rdi), %xmm0 # sched: [20:1.00] +; SANDY-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_sqrtps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vsqrtps %xmm0, %xmm0 # sched: [14:1.00] @@ -3048,6 +4934,13 @@ define <4 x float> @test_sqrtps(<4 x float> %a0, <4 x float> *%a1) {  ; SANDY-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_sqrtps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    sqrtps %xmm0, %xmm1 # sched: [13:1.00] +; HASWELL-SSE-NEXT:    sqrtps (%rdi), %xmm0 # sched: [19:1.00] +; HASWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_sqrtps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vsqrtps %xmm0, %xmm0 # sched: [14:1.00] @@ -3055,6 +4948,13 @@ define <4 x float> @test_sqrtps(<4 x float> %a0, <4 x float> *%a1) {  ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_sqrtps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    sqrtps %xmm0, %xmm1 # sched: [13:1.00] +; BROADWELL-SSE-NEXT:    sqrtps (%rdi), %xmm0 # sched: [18:1.00] +; BROADWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_sqrtps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vsqrtps %xmm0, %xmm0 # sched: [14:1.00] @@ -3062,6 +4962,13 @@ define <4 x float> @test_sqrtps(<4 x float> %a0, <4 x float> *%a1) {  ; BROADWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_sqrtps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    sqrtps %xmm0, %xmm1 # sched: [13:1.00] +; SKYLAKE-SSE-NEXT:    sqrtps (%rdi), %xmm0 # sched: [19:1.00] +; SKYLAKE-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_sqrtps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vsqrtps %xmm0, %xmm0 # sched: [12:1.00] @@ -3069,6 +4976,13 @@ define <4 x float> @test_sqrtps(<4 x float> %a0, <4 x float> *%a1) {  ; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_sqrtps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    sqrtps %xmm0, %xmm1 # sched: [12:1.00] +; SKX-SSE-NEXT:    sqrtps (%rdi), %xmm0 # sched: [18:1.00] +; SKX-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_sqrtps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vsqrtps %xmm0, %xmm0 # sched: [12:1.00] @@ -3076,6 +4990,13 @@ define <4 x float> @test_sqrtps(<4 x float> %a0, <4 x float> *%a1) {  ; SKX-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_sqrtps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    sqrtps %xmm0, %xmm1 # sched: [21:21.00] +; BTVER2-SSE-NEXT:    sqrtps (%rdi), %xmm0 # sched: [26:21.00] +; BTVER2-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_sqrtps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vsqrtps (%rdi), %xmm1 # sched: [26:21.00] @@ -3083,6 +5004,13 @@ define <4 x float> @test_sqrtps(<4 x float> %a0, <4 x float> *%a1) {  ; BTVER2-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_sqrtps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    sqrtps %xmm0, %xmm1 # sched: [20:1.00] +; ZNVER1-SSE-NEXT:    sqrtps (%rdi), %xmm0 # sched: [27:1.00] +; ZNVER1-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_sqrtps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vsqrtps (%rdi), %xmm1 # sched: [27:1.00] @@ -3124,6 +5052,14 @@ define <4 x float> @test_sqrtss(<4 x float> %a0, <4 x float> *%a1) {  ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_sqrtss: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    sqrtss %xmm0, %xmm0 # sched: [14:1.00] +; SANDY-SSE-NEXT:    movaps (%rdi), %xmm1 # sched: [6:0.50] +; SANDY-SSE-NEXT:    sqrtss %xmm1, %xmm1 # sched: [14:1.00] +; SANDY-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_sqrtss:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vsqrtss %xmm0, %xmm0, %xmm0 # sched: [114:1.00] @@ -3132,6 +5068,14 @@ define <4 x float> @test_sqrtss(<4 x float> %a0, <4 x float> *%a1) {  ; SANDY-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_sqrtss: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    sqrtss %xmm0, %xmm0 # sched: [13:1.00] +; HASWELL-SSE-NEXT:    movaps (%rdi), %xmm1 # sched: [6:0.50] +; HASWELL-SSE-NEXT:    sqrtss %xmm1, %xmm1 # sched: [13:1.00] +; HASWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_sqrtss:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vsqrtss %xmm0, %xmm0, %xmm0 # sched: [14:1.00] @@ -3140,6 +5084,14 @@ define <4 x float> @test_sqrtss(<4 x float> %a0, <4 x float> *%a1) {  ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_sqrtss: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    sqrtss %xmm0, %xmm0 # sched: [13:1.00] +; BROADWELL-SSE-NEXT:    movaps (%rdi), %xmm1 # sched: [5:0.50] +; BROADWELL-SSE-NEXT:    sqrtss %xmm1, %xmm1 # sched: [13:1.00] +; BROADWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_sqrtss:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vsqrtss %xmm0, %xmm0, %xmm0 # sched: [14:1.00] @@ -3148,6 +5100,14 @@ define <4 x float> @test_sqrtss(<4 x float> %a0, <4 x float> *%a1) {  ; BROADWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_sqrtss: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    sqrtss %xmm0, %xmm0 # sched: [13:1.00] +; SKYLAKE-SSE-NEXT:    movaps (%rdi), %xmm1 # sched: [6:0.50] +; SKYLAKE-SSE-NEXT:    sqrtss %xmm1, %xmm1 # sched: [13:1.00] +; SKYLAKE-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_sqrtss:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vsqrtss %xmm0, %xmm0, %xmm0 # sched: [12:1.00] @@ -3156,6 +5116,14 @@ define <4 x float> @test_sqrtss(<4 x float> %a0, <4 x float> *%a1) {  ; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_sqrtss: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    sqrtss %xmm0, %xmm0 # sched: [12:1.00] +; SKX-SSE-NEXT:    movaps (%rdi), %xmm1 # sched: [6:0.50] +; SKX-SSE-NEXT:    sqrtss %xmm1, %xmm1 # sched: [12:1.00] +; SKX-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_sqrtss:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vsqrtss %xmm0, %xmm0, %xmm0 # sched: [12:1.00] @@ -3164,6 +5132,14 @@ define <4 x float> @test_sqrtss(<4 x float> %a0, <4 x float> *%a1) {  ; SKX-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_sqrtss: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    movaps (%rdi), %xmm1 # sched: [5:1.00] +; BTVER2-SSE-NEXT:    sqrtss %xmm0, %xmm0 # sched: [26:21.00] +; BTVER2-SSE-NEXT:    sqrtss %xmm1, %xmm1 # sched: [26:21.00] +; BTVER2-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_sqrtss:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vmovaps (%rdi), %xmm1 # sched: [5:1.00] @@ -3172,6 +5148,14 @@ define <4 x float> @test_sqrtss(<4 x float> %a0, <4 x float> *%a1) {  ; BTVER2-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_sqrtss: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    movaps (%rdi), %xmm1 # sched: [8:0.50] +; ZNVER1-SSE-NEXT:    sqrtss %xmm0, %xmm0 # sched: [27:1.00] +; ZNVER1-SSE-NEXT:    sqrtss %xmm1, %xmm1 # sched: [27:1.00] +; ZNVER1-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_sqrtss:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vmovaps (%rdi), %xmm1 # sched: [8:0.50] @@ -3206,42 +5190,84 @@ define i32 @test_stmxcsr() {  ; SLM-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [3:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_stmxcsr: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    stmxcsr -{{[0-9]+}}(%rsp) # sched: [5:1.00] +; SANDY-SSE-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [5:0.50] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_stmxcsr:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vstmxcsr -{{[0-9]+}}(%rsp) # sched: [5:1.00]  ; SANDY-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [5:0.50]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_stmxcsr: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    stmxcsr -{{[0-9]+}}(%rsp) # sched: [2:1.00] +; HASWELL-SSE-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [5:0.50] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_stmxcsr:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vstmxcsr -{{[0-9]+}}(%rsp) # sched: [2:1.00]  ; HASWELL-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [5:0.50]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_stmxcsr: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    stmxcsr -{{[0-9]+}}(%rsp) # sched: [2:1.00] +; BROADWELL-SSE-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [5:0.50] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_stmxcsr:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vstmxcsr -{{[0-9]+}}(%rsp) # sched: [2:1.00]  ; BROADWELL-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [5:0.50]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_stmxcsr: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    stmxcsr -{{[0-9]+}}(%rsp) # sched: [2:1.00] +; SKYLAKE-SSE-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [5:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_stmxcsr:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vstmxcsr -{{[0-9]+}}(%rsp) # sched: [2:1.00]  ; SKYLAKE-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [5:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_stmxcsr: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    stmxcsr -{{[0-9]+}}(%rsp) # sched: [2:1.00] +; SKX-SSE-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [5:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_stmxcsr:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vstmxcsr -{{[0-9]+}}(%rsp) # sched: [2:1.00]  ; SKX-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [5:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_stmxcsr: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    stmxcsr -{{[0-9]+}}(%rsp) # sched: [1:1.00] +; BTVER2-SSE-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [5:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_stmxcsr:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vstmxcsr -{{[0-9]+}}(%rsp) # sched: [1:1.00]  ; BTVER2-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [5:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_stmxcsr: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    stmxcsr -{{[0-9]+}}(%rsp) # sched: [100:?] +; ZNVER1-SSE-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [8:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_stmxcsr:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vstmxcsr -{{[0-9]+}}(%rsp) # sched: [100:?] @@ -3274,42 +5300,84 @@ define <4 x float> @test_subps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; SLM-NEXT:    subps (%rdi), %xmm0 # sched: [6:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_subps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    subps %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    subps (%rdi), %xmm0 # sched: [9:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_subps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vsubps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    vsubps (%rdi), %xmm0, %xmm0 # sched: [9:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_subps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    subps %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    subps (%rdi), %xmm0 # sched: [9:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_subps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vsubps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    vsubps (%rdi), %xmm0, %xmm0 # sched: [9:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_subps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    subps %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    subps (%rdi), %xmm0 # sched: [8:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_subps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vsubps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    vsubps (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_subps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    subps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    subps (%rdi), %xmm0 # sched: [10:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_subps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vsubps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    vsubps (%rdi), %xmm0, %xmm0 # sched: [10:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_subps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    subps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    subps (%rdi), %xmm0 # sched: [10:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_subps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vsubps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    vsubps (%rdi), %xmm0, %xmm0 # sched: [10:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_subps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    subps %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    subps (%rdi), %xmm0 # sched: [8:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_subps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vsubps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    vsubps (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_subps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    subps %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    subps (%rdi), %xmm0 # sched: [10:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_subps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vsubps %xmm1, %xmm0, %xmm0 # sched: [3:1.00] @@ -3340,42 +5408,84 @@ define float @test_subss(float %a0, float %a1, float *%a2) {  ; SLM-NEXT:    subss (%rdi), %xmm0 # sched: [6:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_subss: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    subss %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    subss (%rdi), %xmm0 # sched: [9:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_subss:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vsubss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    vsubss (%rdi), %xmm0, %xmm0 # sched: [9:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_subss: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    subss %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    subss (%rdi), %xmm0 # sched: [8:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_subss:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vsubss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    vsubss (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_subss: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    subss %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    subss (%rdi), %xmm0 # sched: [8:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_subss:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vsubss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    vsubss (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_subss: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    subss %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    subss (%rdi), %xmm0 # sched: [9:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_subss:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vsubss %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    vsubss (%rdi), %xmm0, %xmm0 # sched: [9:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_subss: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    subss %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    subss (%rdi), %xmm0 # sched: [9:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_subss:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vsubss %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    vsubss (%rdi), %xmm0, %xmm0 # sched: [9:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_subss: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    subss %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    subss (%rdi), %xmm0 # sched: [8:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_subss:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vsubss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    vsubss (%rdi), %xmm0, %xmm0 # sched: [8:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_subss: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    subss %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    subss (%rdi), %xmm0 # sched: [10:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_subss:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vsubss %xmm1, %xmm0, %xmm0 # sched: [3:1.00] @@ -3430,6 +5540,20 @@ define i32 @test_ucomiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; SLM-NEXT:    movzbl %dl, %eax # sched: [1:0.50]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_ucomiss: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    ucomiss %xmm1, %xmm0 # sched: [2:1.00] +; SANDY-SSE-NEXT:    setnp %al # sched: [1:0.50] +; SANDY-SSE-NEXT:    sete %cl # sched: [1:0.50] +; SANDY-SSE-NEXT:    andb %al, %cl # sched: [1:0.33] +; SANDY-SSE-NEXT:    ucomiss (%rdi), %xmm0 # sched: [8:1.00] +; SANDY-SSE-NEXT:    setnp %al # sched: [1:0.50] +; SANDY-SSE-NEXT:    sete %dl # sched: [1:0.50] +; SANDY-SSE-NEXT:    andb %al, %dl # sched: [1:0.33] +; SANDY-SSE-NEXT:    orb %cl, %dl # sched: [1:0.33] +; SANDY-SSE-NEXT:    movzbl %dl, %eax # sched: [1:0.33] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_ucomiss:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vucomiss %xmm1, %xmm0 # sched: [2:1.00] @@ -3444,6 +5568,20 @@ define i32 @test_ucomiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; SANDY-NEXT:    movzbl %dl, %eax # sched: [1:0.33]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_ucomiss: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    ucomiss %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    setnp %al # sched: [1:0.50] +; HASWELL-SSE-NEXT:    sete %cl # sched: [1:0.50] +; HASWELL-SSE-NEXT:    andb %al, %cl # sched: [1:0.25] +; HASWELL-SSE-NEXT:    ucomiss (%rdi), %xmm0 # sched: [8:1.00] +; HASWELL-SSE-NEXT:    setnp %al # sched: [1:0.50] +; HASWELL-SSE-NEXT:    sete %dl # sched: [1:0.50] +; HASWELL-SSE-NEXT:    andb %al, %dl # sched: [1:0.25] +; HASWELL-SSE-NEXT:    orb %cl, %dl # sched: [1:0.25] +; HASWELL-SSE-NEXT:    movzbl %dl, %eax # sched: [1:0.25] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_ucomiss:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vucomiss %xmm1, %xmm0 # sched: [3:1.00] @@ -3458,6 +5596,20 @@ define i32 @test_ucomiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; HASWELL-NEXT:    movzbl %dl, %eax # sched: [1:0.25]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_ucomiss: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    ucomiss %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    setnp %al # sched: [1:0.50] +; BROADWELL-SSE-NEXT:    sete %cl # sched: [1:0.50] +; BROADWELL-SSE-NEXT:    andb %al, %cl # sched: [1:0.25] +; BROADWELL-SSE-NEXT:    ucomiss (%rdi), %xmm0 # sched: [8:1.00] +; BROADWELL-SSE-NEXT:    setnp %al # sched: [1:0.50] +; BROADWELL-SSE-NEXT:    sete %dl # sched: [1:0.50] +; BROADWELL-SSE-NEXT:    andb %al, %dl # sched: [1:0.25] +; BROADWELL-SSE-NEXT:    orb %cl, %dl # sched: [1:0.25] +; BROADWELL-SSE-NEXT:    movzbl %dl, %eax # sched: [1:0.25] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_ucomiss:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vucomiss %xmm1, %xmm0 # sched: [3:1.00] @@ -3472,6 +5624,20 @@ define i32 @test_ucomiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; BROADWELL-NEXT:    movzbl %dl, %eax # sched: [1:0.25]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_ucomiss: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    ucomiss %xmm1, %xmm0 # sched: [2:1.00] +; SKYLAKE-SSE-NEXT:    setnp %al # sched: [1:0.50] +; SKYLAKE-SSE-NEXT:    sete %cl # sched: [1:0.50] +; SKYLAKE-SSE-NEXT:    andb %al, %cl # sched: [1:0.25] +; SKYLAKE-SSE-NEXT:    ucomiss (%rdi), %xmm0 # sched: [7:1.00] +; SKYLAKE-SSE-NEXT:    setnp %al # sched: [1:0.50] +; SKYLAKE-SSE-NEXT:    sete %dl # sched: [1:0.50] +; SKYLAKE-SSE-NEXT:    andb %al, %dl # sched: [1:0.25] +; SKYLAKE-SSE-NEXT:    orb %cl, %dl # sched: [1:0.25] +; SKYLAKE-SSE-NEXT:    movzbl %dl, %eax # sched: [1:0.25] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_ucomiss:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vucomiss %xmm1, %xmm0 # sched: [2:1.00] @@ -3486,6 +5652,20 @@ define i32 @test_ucomiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; SKYLAKE-NEXT:    movzbl %dl, %eax # sched: [1:0.25]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_ucomiss: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    ucomiss %xmm1, %xmm0 # sched: [2:1.00] +; SKX-SSE-NEXT:    setnp %al # sched: [1:0.50] +; SKX-SSE-NEXT:    sete %cl # sched: [1:0.50] +; SKX-SSE-NEXT:    andb %al, %cl # sched: [1:0.25] +; SKX-SSE-NEXT:    ucomiss (%rdi), %xmm0 # sched: [7:1.00] +; SKX-SSE-NEXT:    setnp %al # sched: [1:0.50] +; SKX-SSE-NEXT:    sete %dl # sched: [1:0.50] +; SKX-SSE-NEXT:    andb %al, %dl # sched: [1:0.25] +; SKX-SSE-NEXT:    orb %cl, %dl # sched: [1:0.25] +; SKX-SSE-NEXT:    movzbl %dl, %eax # sched: [1:0.25] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_ucomiss:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vucomiss %xmm1, %xmm0 # sched: [2:1.00] @@ -3500,6 +5680,20 @@ define i32 @test_ucomiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; SKX-NEXT:    movzbl %dl, %eax # sched: [1:0.25]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_ucomiss: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    ucomiss %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    setnp %al # sched: [1:0.50] +; BTVER2-SSE-NEXT:    sete %cl # sched: [1:0.50] +; BTVER2-SSE-NEXT:    andb %al, %cl # sched: [1:0.50] +; BTVER2-SSE-NEXT:    ucomiss (%rdi), %xmm0 # sched: [8:1.00] +; BTVER2-SSE-NEXT:    setnp %al # sched: [1:0.50] +; BTVER2-SSE-NEXT:    sete %dl # sched: [1:0.50] +; BTVER2-SSE-NEXT:    andb %al, %dl # sched: [1:0.50] +; BTVER2-SSE-NEXT:    orb %cl, %dl # sched: [1:0.50] +; BTVER2-SSE-NEXT:    movzbl %dl, %eax # sched: [1:0.50] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_ucomiss:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vucomiss %xmm1, %xmm0 # sched: [3:1.00] @@ -3514,6 +5708,20 @@ define i32 @test_ucomiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {  ; BTVER2-NEXT:    movzbl %dl, %eax # sched: [1:0.50]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_ucomiss: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    ucomiss %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    setnp %al # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    sete %cl # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    andb %al, %cl # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    ucomiss (%rdi), %xmm0 # sched: [10:1.00] +; ZNVER1-SSE-NEXT:    setnp %al # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    sete %dl # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    andb %al, %dl # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    orb %cl, %dl # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    movzbl %dl, %eax # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_ucomiss:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vucomiss %xmm1, %xmm0 # sched: [3:1.00] @@ -3557,6 +5765,13 @@ define <4 x float> @test_unpckhps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_unpckhps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00] +; SANDY-SSE-NEXT:    unpckhps {{.*#+}} xmm1 = xmm1[2],mem[2],xmm1[3],mem[3] sched: [7:1.00] +; SANDY-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_unpckhps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00] @@ -3564,6 +5779,13 @@ define <4 x float> @test_unpckhps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; SANDY-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_unpckhps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00] +; HASWELL-SSE-NEXT:    unpckhps {{.*#+}} xmm1 = xmm1[2],mem[2],xmm1[3],mem[3] sched: [7:1.00] +; HASWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_unpckhps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00] @@ -3571,6 +5793,13 @@ define <4 x float> @test_unpckhps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_unpckhps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00] +; BROADWELL-SSE-NEXT:    unpckhps {{.*#+}} xmm1 = xmm1[2],mem[2],xmm1[3],mem[3] sched: [6:1.00] +; BROADWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_unpckhps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00] @@ -3578,6 +5807,13 @@ define <4 x float> @test_unpckhps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; BROADWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_unpckhps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00] +; SKYLAKE-SSE-NEXT:    unpckhps {{.*#+}} xmm1 = xmm1[2],mem[2],xmm1[3],mem[3] sched: [7:1.00] +; SKYLAKE-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_unpckhps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00] @@ -3585,6 +5821,13 @@ define <4 x float> @test_unpckhps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_unpckhps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00] +; SKX-SSE-NEXT:    unpckhps {{.*#+}} xmm1 = xmm1[2],mem[2],xmm1[3],mem[3] sched: [7:1.00] +; SKX-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_unpckhps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00] @@ -3592,6 +5835,13 @@ define <4 x float> @test_unpckhps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; SKX-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_unpckhps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:0.50] +; BTVER2-SSE-NEXT:    unpckhps {{.*#+}} xmm1 = xmm1[2],mem[2],xmm1[3],mem[3] sched: [6:1.00] +; BTVER2-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_unpckhps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:0.50] @@ -3599,6 +5849,13 @@ define <4 x float> @test_unpckhps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; BTVER2-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_unpckhps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:0.50] +; ZNVER1-SSE-NEXT:    unpckhps {{.*#+}} xmm1 = xmm1[2],mem[2],xmm1[3],mem[3] sched: [8:0.50] +; ZNVER1-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_unpckhps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:0.50] @@ -3634,6 +5891,13 @@ define <4 x float> @test_unpcklps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_unpcklps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00] +; SANDY-SSE-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] sched: [7:1.00] +; SANDY-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_unpcklps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00] @@ -3641,6 +5905,13 @@ define <4 x float> @test_unpcklps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; SANDY-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_unpcklps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00] +; HASWELL-SSE-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] sched: [7:1.00] +; HASWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_unpcklps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00] @@ -3648,6 +5919,13 @@ define <4 x float> @test_unpcklps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_unpcklps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00] +; BROADWELL-SSE-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] sched: [6:1.00] +; BROADWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_unpcklps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00] @@ -3655,6 +5933,13 @@ define <4 x float> @test_unpcklps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; BROADWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_unpcklps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00] +; SKYLAKE-SSE-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] sched: [7:1.00] +; SKYLAKE-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_unpcklps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00] @@ -3662,6 +5947,13 @@ define <4 x float> @test_unpcklps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_unpcklps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00] +; SKX-SSE-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] sched: [7:1.00] +; SKX-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_unpcklps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00] @@ -3669,6 +5961,13 @@ define <4 x float> @test_unpcklps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; SKX-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_unpcklps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:0.50] +; BTVER2-SSE-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] sched: [6:1.00] +; BTVER2-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_unpcklps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:0.50] @@ -3676,6 +5975,13 @@ define <4 x float> @test_unpcklps(<4 x float> %a0, <4 x float> %a1, <4 x float>  ; BTVER2-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_unpcklps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:0.50] +; ZNVER1-SSE-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] sched: [8:0.50] +; ZNVER1-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_unpcklps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:0.50] @@ -3712,42 +6018,84 @@ define <4 x float> @test_xorps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a  ; SLM-NEXT:    xorps (%rdi), %xmm0 # sched: [4:1.00]  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_xorps: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    xorps %xmm1, %xmm0 # sched: [1:1.00] +; SANDY-SSE-NEXT:    xorps (%rdi), %xmm0 # sched: [7:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_xorps:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    vxorps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]  ; SANDY-NEXT:    vxorps (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_xorps: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    xorps %xmm1, %xmm0 # sched: [1:1.00] +; HASWELL-SSE-NEXT:    xorps (%rdi), %xmm0 # sched: [7:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_xorps:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    vxorps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]  ; HASWELL-NEXT:    vxorps (%rdi), %xmm0, %xmm0 # sched: [7:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_xorps: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    xorps %xmm1, %xmm0 # sched: [1:1.00] +; BROADWELL-SSE-NEXT:    xorps (%rdi), %xmm0 # sched: [6:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_xorps:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    vxorps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]  ; BROADWELL-NEXT:    vxorps (%rdi), %xmm0, %xmm0 # sched: [6:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_xorps: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    xorps %xmm1, %xmm0 # sched: [1:0.33] +; SKYLAKE-SSE-NEXT:    xorps (%rdi), %xmm0 # sched: [7:0.50] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_xorps:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    vxorps %xmm1, %xmm0, %xmm0 # sched: [1:0.33]  ; SKYLAKE-NEXT:    vxorps (%rdi), %xmm0, %xmm0 # sched: [7:0.50]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_xorps: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    xorps %xmm1, %xmm0 # sched: [1:0.33] +; SKX-SSE-NEXT:    xorps (%rdi), %xmm0 # sched: [7:0.50] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_xorps:  ; SKX:       # %bb.0:  ; SKX-NEXT:    vxorps %xmm1, %xmm0, %xmm0 # sched: [1:0.33]  ; SKX-NEXT:    vxorps (%rdi), %xmm0, %xmm0 # sched: [7:0.50]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_xorps: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    xorps %xmm1, %xmm0 # sched: [1:0.50] +; BTVER2-SSE-NEXT:    xorps (%rdi), %xmm0 # sched: [6:1.00] +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_xorps:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vxorps %xmm1, %xmm0, %xmm0 # sched: [1:0.50]  ; BTVER2-NEXT:    vxorps (%rdi), %xmm0, %xmm0 # sched: [6:1.00]  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_xorps: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    xorps %xmm1, %xmm0 # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    xorps (%rdi), %xmm0 # sched: [8:0.50] +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_xorps:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vxorps %xmm1, %xmm0, %xmm0 # sched: [1:0.25] @@ -3796,6 +6144,14 @@ define <4 x float> @test_fnop() nounwind {  ; SLM-NEXT:    #NO_APP  ; SLM-NEXT:    retq # sched: [4:1.00]  ; +; SANDY-SSE-LABEL: test_fnop: +; SANDY-SSE:       # %bb.0: +; SANDY-SSE-NEXT:    #APP +; SANDY-SSE-NEXT:    nop # sched: [1:?] +; SANDY-SSE-NEXT:    #NO_APP +; SANDY-SSE-NEXT:    xorps %xmm0, %xmm0 # sched: [1:1.00] +; SANDY-SSE-NEXT:    retq # sched: [1:1.00] +;  ; SANDY-LABEL: test_fnop:  ; SANDY:       # %bb.0:  ; SANDY-NEXT:    #APP @@ -3804,6 +6160,14 @@ define <4 x float> @test_fnop() nounwind {  ; SANDY-NEXT:    vxorps %xmm0, %xmm0, %xmm0 # sched: [1:1.00]  ; SANDY-NEXT:    retq # sched: [1:1.00]  ; +; HASWELL-SSE-LABEL: test_fnop: +; HASWELL-SSE:       # %bb.0: +; HASWELL-SSE-NEXT:    #APP +; HASWELL-SSE-NEXT:    nop # sched: [1:0.25] +; HASWELL-SSE-NEXT:    #NO_APP +; HASWELL-SSE-NEXT:    xorps %xmm0, %xmm0 # sched: [1:1.00] +; HASWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; HASWELL-LABEL: test_fnop:  ; HASWELL:       # %bb.0:  ; HASWELL-NEXT:    #APP @@ -3812,6 +6176,14 @@ define <4 x float> @test_fnop() nounwind {  ; HASWELL-NEXT:    vxorps %xmm0, %xmm0, %xmm0 # sched: [1:1.00]  ; HASWELL-NEXT:    retq # sched: [7:1.00]  ; +; BROADWELL-SSE-LABEL: test_fnop: +; BROADWELL-SSE:       # %bb.0: +; BROADWELL-SSE-NEXT:    #APP +; BROADWELL-SSE-NEXT:    nop # sched: [1:0.25] +; BROADWELL-SSE-NEXT:    #NO_APP +; BROADWELL-SSE-NEXT:    xorps %xmm0, %xmm0 # sched: [1:1.00] +; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00] +;  ; BROADWELL-LABEL: test_fnop:  ; BROADWELL:       # %bb.0:  ; BROADWELL-NEXT:    #APP @@ -3820,6 +6192,14 @@ define <4 x float> @test_fnop() nounwind {  ; BROADWELL-NEXT:    vxorps %xmm0, %xmm0, %xmm0 # sched: [1:1.00]  ; BROADWELL-NEXT:    retq # sched: [7:1.00]  ; +; SKYLAKE-SSE-LABEL: test_fnop: +; SKYLAKE-SSE:       # %bb.0: +; SKYLAKE-SSE-NEXT:    #APP +; SKYLAKE-SSE-NEXT:    nop # sched: [1:0.25] +; SKYLAKE-SSE-NEXT:    #NO_APP +; SKYLAKE-SSE-NEXT:    xorps %xmm0, %xmm0 # sched: [1:0.33] +; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKYLAKE-LABEL: test_fnop:  ; SKYLAKE:       # %bb.0:  ; SKYLAKE-NEXT:    #APP @@ -3828,6 +6208,14 @@ define <4 x float> @test_fnop() nounwind {  ; SKYLAKE-NEXT:    vxorps %xmm0, %xmm0, %xmm0 # sched: [1:0.33]  ; SKYLAKE-NEXT:    retq # sched: [7:1.00]  ; +; SKX-SSE-LABEL: test_fnop: +; SKX-SSE:       # %bb.0: +; SKX-SSE-NEXT:    #APP +; SKX-SSE-NEXT:    nop # sched: [1:0.25] +; SKX-SSE-NEXT:    #NO_APP +; SKX-SSE-NEXT:    xorps %xmm0, %xmm0 # sched: [1:0.33] +; SKX-SSE-NEXT:    retq # sched: [7:1.00] +;  ; SKX-LABEL: test_fnop:  ; SKX:       # %bb.0:  ; SKX-NEXT:    #APP @@ -3836,6 +6224,14 @@ define <4 x float> @test_fnop() nounwind {  ; SKX-NEXT:    vxorps %xmm0, %xmm0, %xmm0 # sched: [1:0.33]  ; SKX-NEXT:    retq # sched: [7:1.00]  ; +; BTVER2-SSE-LABEL: test_fnop: +; BTVER2-SSE:       # %bb.0: +; BTVER2-SSE-NEXT:    xorps %xmm0, %xmm0 # sched: [1:0.50] +; BTVER2-SSE-NEXT:    #APP +; BTVER2-SSE-NEXT:    nop # sched: [1:0.50] +; BTVER2-SSE-NEXT:    #NO_APP +; BTVER2-SSE-NEXT:    retq # sched: [4:1.00] +;  ; BTVER2-LABEL: test_fnop:  ; BTVER2:       # %bb.0:  ; BTVER2-NEXT:    vxorps %xmm0, %xmm0, %xmm0 # sched: [1:0.50] @@ -3844,6 +6240,14 @@ define <4 x float> @test_fnop() nounwind {  ; BTVER2-NEXT:    #NO_APP  ; BTVER2-NEXT:    retq # sched: [4:1.00]  ; +; ZNVER1-SSE-LABEL: test_fnop: +; ZNVER1-SSE:       # %bb.0: +; ZNVER1-SSE-NEXT:    xorps %xmm0, %xmm0 # sched: [1:0.25] +; ZNVER1-SSE-NEXT:    #APP +; ZNVER1-SSE-NEXT:    nop # sched: [1:?] +; ZNVER1-SSE-NEXT:    #NO_APP +; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50] +;  ; ZNVER1-LABEL: test_fnop:  ; ZNVER1:       # %bb.0:  ; ZNVER1-NEXT:    vxorps %xmm0, %xmm0, %xmm0 # sched: [1:0.25] | 

