[X86][AVX] Match broadcast loads through a bitcast

AVX1 v8i32/v4i64 shuffles are bitcasted to v8f32/v4f64, this patch peeks through any bitcast to check for a load node to allow broadcasts to occur. This is a re-commit of r257055 after r257264 fixed 32-bit broadcast loads of i64 scalars. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@257266 91177308-0d34-0410-b5e6-96231b3b80d8
simoatze · Jan 9, 2016 · 95d397c · 95d397c
1 parent d8eb9f7
commit 95d397c
Show file tree

Hide file tree

Showing 3 changed files with 12 additions and 21 deletions.
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
@@ -8175,6 +8175,11 @@ static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
 
   MVT BroadcastVT = VT;
 
+  // Peek through any bitcast (only useful for loads).
+  SDValue BC = V;
+  while (BC.getOpcode() == ISD::BITCAST)
+    BC = BC.getOperand(0);
+
   // Also check the simpler case, where we can directly reuse the scalar.
   if (V.getOpcode() == ISD::BUILD_VECTOR ||
       (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
@@ -8184,14 +8189,14 @@ static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
     // Only AVX2 has register broadcasts.
     if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
       return SDValue();
-  } else if (MayFoldLoad(V) && !cast<LoadSDNode>(V)->isVolatile()) {
+  } else if (MayFoldLoad(BC) && !cast<LoadSDNode>(BC)->isVolatile()) {
     // 32-bit targets need to load i64 as a f64 and then bitcast the result.
     if (!Subtarget->is64Bit() && VT.getScalarType() == MVT::i64)
       BroadcastVT = MVT::getVectorVT(MVT::f64, VT.getVectorNumElements());
 
     // If we are broadcasting a load that is only used by the shuffle
     // then we can reduce the vector load to the broadcasted scalar load.
-    LoadSDNode *Ld = cast<LoadSDNode>(V);
+    LoadSDNode *Ld = cast<LoadSDNode>(BC);
     SDValue BaseAddr = Ld->getOperand(1);
     EVT AddrVT = BaseAddr.getValueType();
     EVT SVT = BroadcastVT.getScalarType();

diff --git a/test/CodeGen/X86/2012-01-12-extract-sv.ll b/test/CodeGen/X86/2012-01-12-extract-sv.ll
@@ -3,9 +3,7 @@
 define void @endless_loop() {
 ; CHECK-LABEL: endless_loop:
 ; CHECK-NEXT:  # BB#0:
-; CHECK-NEXT:    vmovaps (%eax), %ymm0
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; CHECK-NEXT:    vmovsldup {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; CHECK-NEXT:    vbroadcastss (%eax), %ymm0
 ; CHECK-NEXT:    vmovddup {{.*#+}} xmm1 = xmm0[0,0]
 ; CHECK-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
 ; CHECK-NEXT:    vxorps %xmm2, %xmm2, %xmm2

diff --git a/test/CodeGen/X86/avx-vbroadcast.ll b/test/CodeGen/X86/avx-vbroadcast.ll
@@ -192,18 +192,12 @@ define <8 x i32> @load_splat_8i32_8i32_55555555(<8 x i32>* %ptr) nounwind uwtabl
 ; X32-LABEL: load_splat_8i32_8i32_55555555:
 ; X32:       ## BB#0: ## %entry
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-NEXT:    vmovaps (%eax), %ymm0
-; X32-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; X32-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; X32-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; X32-NEXT:    vbroadcastss 20(%eax), %ymm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: load_splat_8i32_8i32_55555555:
 ; X64:       ## BB#0: ## %entry
-; X64-NEXT:    vmovaps (%rdi), %ymm0
-; X64-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; X64-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; X64-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; X64-NEXT:    vbroadcastss 20(%rdi), %ymm0
 ; X64-NEXT:    retq
 entry:
   %ld = load <8 x i32>, <8 x i32>* %ptr
@@ -304,18 +298,12 @@ define <4 x i64> @load_splat_4i64_4i64_2222(<4 x i64>* %ptr) nounwind uwtable re
 ; X32-LABEL: load_splat_4i64_4i64_2222:
 ; X32:       ## BB#0: ## %entry
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-NEXT:    vmovapd (%eax), %ymm0
-; X32-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; X32-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
-; X32-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; X32-NEXT:    vbroadcastsd 16(%eax), %ymm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: load_splat_4i64_4i64_2222:
 ; X64:       ## BB#0: ## %entry
-; X64-NEXT:    vmovapd (%rdi), %ymm0
-; X64-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; X64-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
-; X64-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; X64-NEXT:    vbroadcastsd 16(%rdi), %ymm0
 ; X64-NEXT:    retq
 entry:
   %ld = load <4 x i64>, <4 x i64>* %ptr