[X86] Add support for combining FMSUB/FNMADD/FNMSUB ISD nodes with an fneg input.

Previously we could only negate the FMADD opcodes. This used to be mostly ok when we lowered FMA intrinsics during lowering. But with the move to llvm.fma from target specific intrinsics, we can combine (fneg (fma)) to (fmsub) earlier. So if we start with (fneg (fma (fneg))) we would get stuck at (fmsub (fneg)).

This patch fixes that so we can also combine things like (fmsub (fneg)).

llvm-svn: 336304

3 files changed, 14 insertions, 38 deletions

diff --git a/llvm/test/CodeGen/X86/avx2-fma-fneg-combine.ll b/llvm/test/CodeGen/X86/avx2-fma-fneg-combine.ll
index f408f66eef6..9ebc3b8941b 100644
--- a/llvm/test/CodeGen/X86/avx2-fma-fneg-combine.ll
+++ b/llvm/test/CodeGen/X86/avx2-fma-fneg-combine.ll
@@ -65,16 +65,12 @@ declare <4 x float> @llvm.x86.fma.vfnmadd.ss(<4 x float> %a, <4 x float> %b, <4
 define <8 x float> @test4(<8 x float> %a, <8 x float> %b, <8 x float> %c) {
 ; X32-LABEL: test4:
 ; X32:       # %bb.0: # %entry
-; X32-NEXT:    vbroadcastss {{.*#+}} ymm3 = [-0,-0,-0,-0,-0,-0,-0,-0]
-; X32-NEXT:    vxorps %ymm3, %ymm2, %ymm2
-; X32-NEXT:    vfnmsub213ps {{.*#+}} ymm0 = -(ymm1 * ymm0) - ymm2
+; X32-NEXT:    vfnmadd213ps {{.*#+}} ymm0 = -(ymm1 * ymm0) + ymm2
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: test4:
 ; X64:       # %bb.0: # %entry
-; X64-NEXT:    vbroadcastss {{.*#+}} ymm3 = [-0,-0,-0,-0,-0,-0,-0,-0]
-; X64-NEXT:    vxorps %ymm3, %ymm2, %ymm2
-; X64-NEXT:    vfnmsub213ps {{.*#+}} ymm0 = -(ymm1 * ymm0) - ymm2
+; X64-NEXT:    vfnmadd213ps {{.*#+}} ymm0 = -(ymm1 * ymm0) + ymm2
 ; X64-NEXT:    retq
 entry:
   %0 = tail call <8 x float> @llvm.x86.fma.vfmsub.ps.256(<8 x float> %a, <8 x float> %b, <8 x float> %c) #2
diff --git a/llvm/test/CodeGen/X86/avx512-intrinsics-canonical.ll b/llvm/test/CodeGen/X86/avx512-intrinsics-canonical.ll
index c33ae5e6467..b020fdd01ea 100644
--- a/llvm/test/CodeGen/X86/avx512-intrinsics-canonical.ll
+++ b/llvm/test/CodeGen/X86/avx512-intrinsics-canonical.ll
@@ -2507,25 +2507,17 @@ define <4 x float> @test_mm_mask3_fnmsub_round_ss(<4 x float> %__W, <4 x float>
 ; X86-LABEL: test_mm_mask3_fnmsub_round_ss:
 ; X86:       ## %bb.0: ## %entry
 ; X86-NEXT:    movb {{[0-9]+}}(%esp), %al ## encoding: [0x8a,0x44,0x24,0x04]
-; X86-NEXT:    vbroadcastss {{.*#+}} xmm3 = [-0,-0,-0,-0]
-; X86-NEXT:    ## encoding: [0xc4,0xe2,0x79,0x18,0x1d,A,A,A,A]
-; X86-NEXT:    ## fixup A - offset: 5, value: LCPI119_0, kind: FK_Data_4
-; X86-NEXT:    vxorps %xmm3, %xmm1, %xmm1 ## encoding: [0xc5,0xf0,0x57,0xcb]
 ; X86-NEXT:    kmovw %eax, %k1 ## encoding: [0xc5,0xf8,0x92,0xc8]
-; X86-NEXT:    vfmsub231ss %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0xbb,0xd1]
-; X86-NEXT:    ## xmm2 = (xmm0 * xmm1) - xmm2
+; X86-NEXT:    vfnmsub231ss %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0xbf,0xd1]
+; X86-NEXT:    ## xmm2 = -(xmm0 * xmm1) - xmm2
 ; X86-NEXT:    vmovaps %xmm2, %xmm0 ## encoding: [0xc5,0xf8,0x28,0xc2]
 ; X86-NEXT:    retl ## encoding: [0xc3]
 ;
 ; X64-LABEL: test_mm_mask3_fnmsub_round_ss:
 ; X64:       ## %bb.0: ## %entry
-; X64-NEXT:    vbroadcastss {{.*#+}} xmm3 = [-0,-0,-0,-0]
-; X64-NEXT:    ## encoding: [0xc4,0xe2,0x79,0x18,0x1d,A,A,A,A]
-; X64-NEXT:    ## fixup A - offset: 5, value: LCPI119_0-4, kind: reloc_riprel_4byte
-; X64-NEXT:    vxorps %xmm3, %xmm1, %xmm1 ## encoding: [0xc5,0xf0,0x57,0xcb]
 ; X64-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
-; X64-NEXT:    vfmsub231ss %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0xbb,0xd1]
-; X64-NEXT:    ## xmm2 = (xmm0 * xmm1) - xmm2
+; X64-NEXT:    vfnmsub231ss %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0xbf,0xd1]
+; X64-NEXT:    ## xmm2 = -(xmm0 * xmm1) - xmm2
 ; X64-NEXT:    vmovaps %xmm2, %xmm0 ## encoding: [0xc5,0xf8,0x28,0xc2]
 ; X64-NEXT:    retq ## encoding: [0xc3]
 entry:
@@ -3133,21 +3125,17 @@ define <2 x double> @test_mm_mask3_fnmsub_round_sd(<2 x double> %__W, <2 x doubl
 ; X86-LABEL: test_mm_mask3_fnmsub_round_sd:
 ; X86:       ## %bb.0: ## %entry
 ; X86-NEXT:    movb {{[0-9]+}}(%esp), %al ## encoding: [0x8a,0x44,0x24,0x04]
-; X86-NEXT:    vxorpd LCPI143_0, %xmm1, %xmm1 ## encoding: [0xc5,0xf1,0x57,0x0d,A,A,A,A]
-; X86-NEXT:    ## fixup A - offset: 4, value: LCPI143_0, kind: FK_Data_4
 ; X86-NEXT:    kmovw %eax, %k1 ## encoding: [0xc5,0xf8,0x92,0xc8]
-; X86-NEXT:    vfmsub231sd %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf2,0xfd,0x09,0xbb,0xd1]
-; X86-NEXT:    ## xmm2 = (xmm0 * xmm1) - xmm2
+; X86-NEXT:    vfnmsub231sd %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf2,0xfd,0x09,0xbf,0xd1]
+; X86-NEXT:    ## xmm2 = -(xmm0 * xmm1) - xmm2
 ; X86-NEXT:    vmovapd %xmm2, %xmm0 ## encoding: [0xc5,0xf9,0x28,0xc2]
 ; X86-NEXT:    retl ## encoding: [0xc3]
 ;
 ; X64-LABEL: test_mm_mask3_fnmsub_round_sd:
 ; X64:       ## %bb.0: ## %entry
-; X64-NEXT:    vxorpd {{.*}}(%rip), %xmm1, %xmm1 ## encoding: [0xc5,0xf1,0x57,0x0d,A,A,A,A]
-; X64-NEXT:    ## fixup A - offset: 4, value: LCPI143_0-4, kind: reloc_riprel_4byte
 ; X64-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
-; X64-NEXT:    vfmsub231sd %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf2,0xfd,0x09,0xbb,0xd1]
-; X64-NEXT:    ## xmm2 = (xmm0 * xmm1) - xmm2
+; X64-NEXT:    vfnmsub231sd %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf2,0xfd,0x09,0xbf,0xd1]
+; X64-NEXT:    ## xmm2 = -(xmm0 * xmm1) - xmm2
 ; X64-NEXT:    vmovapd %xmm2, %xmm0 ## encoding: [0xc5,0xf9,0x28,0xc2]
 ; X64-NEXT:    retq ## encoding: [0xc3]
 entry:
diff --git a/llvm/test/CodeGen/X86/fma-fneg-combine.ll b/llvm/test/CodeGen/X86/fma-fneg-combine.ll
index 1d698c77438..6d02eaec36f 100644
--- a/llvm/test/CodeGen/X86/fma-fneg-combine.ll
+++ b/llvm/test/CodeGen/X86/fma-fneg-combine.ll
@@ -78,18 +78,10 @@ entry:
 
 
 define <8 x float> @test7(<8 x float> %a, <8 x float> %b, <8 x float> %c) {
-; SKX-LABEL: test7:
-; SKX:       # %bb.0: # %entry
-; SKX-NEXT:    vxorps {{.*}}(%rip){1to8}, %ymm2, %ymm2
-; SKX-NEXT:    vfnmsub213ps {{.*#+}} ymm0 = -(ymm1 * ymm0) - ymm2
-; SKX-NEXT:    retq
-;
-; KNL-LABEL: test7:
-; KNL:       # %bb.0: # %entry
-; KNL-NEXT:    vbroadcastss {{.*#+}} ymm3 = [-0,-0,-0,-0,-0,-0,-0,-0]
-; KNL-NEXT:    vxorps %ymm3, %ymm2, %ymm2
-; KNL-NEXT:    vfnmsub213ps {{.*#+}} ymm0 = -(ymm1 * ymm0) - ymm2
-; KNL-NEXT:    retq
+; CHECK-LABEL: test7:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vfnmadd213ps {{.*#+}} ymm0 = -(ymm1 * ymm0) + ymm2
+; CHECK-NEXT:    retq
 entry:
   %0 = tail call <8 x float> @llvm.x86.fma.vfmsub.ps.256(<8 x float> %a, <8 x float> %b, <8 x float> %c) #2
   %sub.i = fsub <8 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %0