[llvm] d56c647 - [X86][AVX] SimplifyDemandedVectorEltsForTargetNode - reduce width of X86ISD::VPERMILPV

[Simon Pilgrim via llvm-commits]

Author: Simon Pilgrim
Date: 2020-06-27T14:43:03+01:00
New Revision: d56c6475a60aa44b040983fe4fd355399cc4c42d

URL: https://github.com/llvm/llvm-project/commit/d56c6475a60aa44b040983fe4fd355399cc4c42d
DIFF: https://github.com/llvm/llvm-project/commit/d56c6475a60aa44b040983fe4fd355399cc4c42d.diff

LOG: [X86][AVX] SimplifyDemandedVectorEltsForTargetNode - reduce width of X86ISD::VPERMILPV

If we don't need the elements of the upper lanes, reduce the width of the X86ISD::VPERMILPV node.

Added: 
    

Modified: 
    llvm/lib/Target/X86/X86ISelLowering.cpp
    llvm/test/CodeGen/X86/var-permute-256.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 54cbaccadc9e..2e23bcb0ee6e 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -37122,7 +37122,8 @@ bool X86TargetLowering::SimplifyDemandedVectorEltsForTargetNode(
       }
       break;
     }
-      // Target Shuffles.
+      // (Non-Lane Crossing) Target Shuffles.
+    case X86ISD::VPERMILPV:
     case X86ISD::PSHUFB:
     case X86ISD::UNPCKL:
     case X86ISD::UNPCKH:

diff  --git a/llvm/test/CodeGen/X86/var-permute-256.ll b/llvm/test/CodeGen/X86/var-permute-256.ll
index 4c04219b3883..6a9f47d6446f 100644
--- a/llvm/test/CodeGen/X86/var-permute-256.ll
+++ b/llvm/test/CodeGen/X86/var-permute-256.ll
@@ -1114,13 +1114,11 @@ define <4 x i32> @var_shuffle_v4i32_from_v8i32(<8 x i32> %v, <4 x i32> %indices)
 ;
 ; AVX1-LABEL: var_shuffle_v4i32_from_v8i32:
 ; AVX1:       # %bb.0: # %entry
-; AVX1-NEXT:    # kill: def $xmm1 killed $xmm1 def $ymm1
-; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm0[2,3,2,3]
-; AVX1-NEXT:    vpermilps %ymm1, %ymm2, %ymm2
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; AVX1-NEXT:    vpermilps %ymm1, %ymm0, %ymm0
+; AVX1-NEXT:    vpermilps %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3]
+; AVX1-NEXT:    vpermilps %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpgtd {{.*}}(%rip), %xmm1, %xmm1
-; AVX1-NEXT:    vblendvps %xmm1, %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vblendvps %xmm1, %xmm0, %xmm2, %xmm0
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;