[X86][AVX] Fold broadcast(bitcast(src)) -> bitcast(broadcast(src))

llvm-svn: 352751

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -31938,6 +31938,14 @@ static SDValue combineTargetShuffle(SDValue N, SelectionDAG &DAG,
         return DAG.getNode(X86ISD::VBROADCAST, DL, VT,
                            DAG.getBitcast(SrcVT, Res));
     }
+    // broadcast(bitcast(src)) -> bitcast(broadcast(src))
+    // 32-bit targets have to bitcast i64 to f64, so better to bitcast upward.
+    if (Src.getOpcode() == ISD::BITCAST &&
+        SrcVT.getScalarSizeInBits() == BCVT.getScalarSizeInBits()) {
+      EVT NewVT = EVT::getVectorVT(*DAG.getContext(), BCVT.getScalarType(),
+                                   VT.getVectorNumElements());
+      return DAG.getBitcast(VT, DAG.getNode(X86ISD::VBROADCAST, DL, NewVT, BC));
+    }
     return SDValue();
   }
   case X86ISD::PSHUFD:

llvm/test/CodeGen/X86/avx-vbroadcast.ll

@@ -45,9 +45,9 @@ define <4 x i64> @A2(i64* %ptr, i64* %ptr2) nounwind uwtable readnone ssp {
 ; X64-LABEL: A2:
 ; X64:       ## %bb.0: ## %entry
 ; X64-NEXT:    movq (%rdi), %rax
-; X64-NEXT:    vmovq %rax, %xmm0
 ; X64-NEXT:    movq %rax, (%rsi)
-; X64-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
+; X64-NEXT:    vmovq %rax, %xmm0
+; X64-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
 ; X64-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X64-NEXT:    retq
 entry:
@@ -110,17 +110,17 @@ define <8 x i32> @B3(i32* %ptr, i32* %ptr2) nounwind uwtable readnone ssp {
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %ecx
 ; X32-NEXT:    movl (%ecx), %ecx
-; X32-NEXT:    vmovd %ecx, %xmm0
 ; X32-NEXT:    movl %ecx, (%eax)
+; X32-NEXT:    vmovd %ecx, %xmm0
 ; X32-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
 ; X32-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: B3:
 ; X64:       ## %bb.0: ## %entry
 ; X64-NEXT:    movl (%rdi), %eax
-; X64-NEXT:    vmovd %eax, %xmm0
 ; X64-NEXT:    movl %eax, (%rsi)
+; X64-NEXT:    vmovd %eax, %xmm0
 ; X64-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
 ; X64-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X64-NEXT:    retq

llvm/test/CodeGen/X86/avx2-vbroadcast.ll

@@ -244,8 +244,7 @@ define <16 x i16> @broadcast_mem_v4i16_v16i16(<4 x i16>* %ptr) {
 ; X32-LABEL: broadcast_mem_v4i16_v16i16:
 ; X32:       ## %bb.0:
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; X32-NEXT:    vbroadcastsd %xmm0, %ymm0
+; X32-NEXT:    vbroadcastsd (%eax), %ymm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: broadcast_mem_v4i16_v16i16:

llvm/test/CodeGen/X86/subvector-broadcast.ll

@@ -1592,15 +1592,13 @@ define <8 x i32> @test_2xi32_to_8xi32_mem(<2 x i32>* %vp) {
 ; X32-AVX2-LABEL: test_2xi32_to_8xi32_mem:
 ; X32-AVX2:       # %bb.0:
 ; X32-AVX2-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-AVX2-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; X32-AVX2-NEXT:    vbroadcastsd %xmm0, %ymm0
+; X32-AVX2-NEXT:    vbroadcastsd (%eax), %ymm0
 ; X32-AVX2-NEXT:    retl
 ;
 ; X32-AVX512-LABEL: test_2xi32_to_8xi32_mem:
 ; X32-AVX512:       # %bb.0:
 ; X32-AVX512-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-AVX512-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; X32-AVX512-NEXT:    vbroadcastsd %xmm0, %ymm0
+; X32-AVX512-NEXT:    vbroadcastsd (%eax), %ymm0
 ; X32-AVX512-NEXT:    retl
 ;
 ; X64-AVX1-LABEL: test_2xi32_to_8xi32_mem:
@@ -1635,8 +1633,7 @@ define <16 x i32> @test_2xi32_to_16xi32_mem(<2 x i32>* %vp) {
 ; X32-AVX2-LABEL: test_2xi32_to_16xi32_mem:
 ; X32-AVX2:       # %bb.0:
 ; X32-AVX2-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-AVX2-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; X32-AVX2-NEXT:    vbroadcastsd %xmm0, %ymm0
+; X32-AVX2-NEXT:    vbroadcastsd (%eax), %ymm0
 ; X32-AVX2-NEXT:    vmovaps %ymm0, %ymm1
 ; X32-AVX2-NEXT:    retl
 ;