AVX-512: Fixed a bug in FMA instruction selection on KNL

The FMA instruction was selected from AVX2 set instead of AVX-512 Differential Revision: http://reviews.llvm.org/D16884 llvm-svn: 259792
llvm · Feb 4, 2016 · 8652827 · 8652827
1 parent 23e44f5
commit 8652827
Show file tree

Hide file tree

Showing 4 changed files with 16 additions and 15 deletions.
diff --git a/llvm/lib/Target/X86/X86InstrFormats.td b/llvm/lib/Target/X86/X86InstrFormats.td
@@ -845,7 +845,7 @@ class AVXPCLMULIi8<bits<8> o, Format F, dag outs, dag ins, string asm,
 class FMA3<bits<8> o, Format F, dag outs, dag ins, string asm,
            list<dag>pattern, InstrItinClass itin = NoItinerary>
       : I<o, F, outs, ins, asm, pattern, itin>, T8PD,
-        VEX_4V, FMASC, Requires<[HasFMA]>;
+        VEX_4V, FMASC, Requires<[HasFMA, NoVLX]>;
 
 // FMA4 Instruction Templates
 class FMA4<bits<8> o, Format F, dag outs, dag ins, string asm,

diff --git a/llvm/test/CodeGen/X86/avx-isa-check.ll b/llvm/test/CodeGen/X86/avx-isa-check.ll
@@ -575,3 +575,9 @@ entry:
   %C = zext <8 x i8> %B to <8 x i16>
   ret <8 x i16> %C
 }
+
+define <4 x float> @test_x86_fmsub_ps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) {
+  %x = fmul <4 x float> %a0, %a1
+  %res = fsub <4 x float> %x, %a2
+  ret <4 x float> %res
+}
diff --git a/llvm/test/CodeGen/X86/avx512-fma.ll b/llvm/test/CodeGen/X86/avx512-fma.ll
@@ -78,16 +78,11 @@ define double @test_x86_fmsub_213(double %a0, double %a1, double %a2) {
 }
 
 define double @test_x86_fmsub_213_m(double %a0, double %a1, double * %a2_ptr) {
-; KNL-LABEL: test_x86_fmsub_213_m:
-; KNL:       ## BB#0:
-; KNL-NEXT:    vfmsub213sd (%rdi), %xmm0, %xmm1
-; KNL-NEXT:    vmovaps %zmm1, %zmm0
-; KNL-NEXT:    retq
-;
-; SKX-LABEL: test_x86_fmsub_213_m:
-; SKX:       ## BB#0:
-; SKX-NEXT:    vfmsub213sd (%rdi), %xmm1, %xmm0
-; SKX-NEXT:    retq
+; ALL-LABEL: test_x86_fmsub_213_m:
+; ALL:       ## BB#0:
+; ALL-NEXT:    vfmsub213sd (%rdi), %xmm0, %xmm1
+; ALL-NEXT:    vmovaps %zmm1, %zmm0
+; ALL-NEXT:    retq
   %a2 = load double , double *%a2_ptr
   %x = fmul double %a0, %a1
   %res = fsub double %x, %a2

diff --git a/llvm/test/CodeGen/X86/avx512bwvl-intrinsics.ll b/llvm/test/CodeGen/X86/avx512bwvl-intrinsics.ll
@@ -1508,15 +1508,15 @@ define <4 x float> @test_mask_vfmadd128_ps_rmka(<4 x float> %a0, <4 x float> %a1
 
 define <4 x float> @test_mask_vfmadd128_ps_rmkz(<4 x float> %a0, <4 x float> %a1, <4 x float>* %ptr_a2) {
   ; CHECK-LABEL: test_mask_vfmadd128_ps_rmkz
-  ; CHECK: vfmadd213ps	(%rdi), %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0x71,0xa8,0x07]
+  ; CHECK: vfmadd213ps	(%rdi), %xmm1, %xmm0 ## encoding: [0x62,0xf2,0x75,0x08,0xa8,0x07]
   %a2 = load <4 x float>, <4 x float>* %ptr_a2
   %res = call <4 x float> @llvm.x86.avx512.mask.vfmadd.ps.128(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2, i8 -1) nounwind
   ret <4 x float> %res
 }
 
 define <4 x float> @test_mask_vfmadd128_ps_rmkza(<4 x float> %a0, <4 x float> %a1, <4 x float>* %ptr_a2) {
   ; CHECK-LABEL: test_mask_vfmadd128_ps_rmkza
-  ; CHECK: vfmadd213ps	(%rdi), %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0x71,0xa8,0x07]
+  ; CHECK: vfmadd213ps	(%rdi), %xmm1, %xmm0 ## encoding: [0x62,0xf2,0x75,0x08,0xa8,0x07]
   %a2 = load <4 x float>, <4 x float>* %ptr_a2, align 4
   %res = call <4 x float> @llvm.x86.avx512.mask.vfmadd.ps.128(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2, i8 -1) nounwind
   ret <4 x float> %res
@@ -1594,7 +1594,7 @@ define <2 x double> @test_mask_vfmadd128_pd_rmk(<2 x double> %a0, <2 x double> %
 
 define <2 x double> @test_mask_vfmadd128_pd_rmkz(<2 x double> %a0, <2 x double> %a1, <2 x double>* %ptr_a2) {
   ; CHECK-LABEL: test_mask_vfmadd128_pd_rmkz
-  ; CHECK: vfmadd213pd	(%rdi), %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0xf1,0xa8,0x07]
+  ; CHECK: vfmadd213pd	(%rdi), %xmm1, %xmm0 ## encoding: [0x62,0xf2,0xf5,0x08,0xa8,0x07]
   %a2 = load <2 x double>, <2 x double>* %ptr_a2
   %res = call <2 x double> @llvm.x86.avx512.mask.vfmadd.pd.128(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2, i8 -1) nounwind
   ret <2 x double> %res
@@ -1624,7 +1624,7 @@ define <4 x double> @test_mask_vfmadd256_pd_rmk(<4 x double> %a0, <4 x double> %
 
 define <4 x double> @test_mask_vfmadd256_pd_rmkz(<4 x double> %a0, <4 x double> %a1, <4 x double>* %ptr_a2) {
   ; CHECK-LABEL: test_mask_vfmadd256_pd_rmkz
-  ; CHECK: vfmadd213pd	(%rdi), %ymm1, %ymm0 ## encoding: [0xc4,0xe2,0xf5,0xa8,0x07]
+  ; CHECK: vfmadd213pd	(%rdi), %ymm1, %ymm0 ## encoding: [0x62,0xf2,0xf5,0x28,0xa8,0x07]
   %a2 = load <4 x double>, <4 x double>* %ptr_a2
   %res = call <4 x double> @llvm.x86.avx512.mask.vfmadd.pd.256(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2, i8 -1) nounwind
   ret <4 x double> %res