[X86][AVX] Fold concat(broadcast(x),broadcast(x)) -> broadcast(x)

Differential Revision: https://reviews.llvm.org/D57514

llvm-svn: 352774

3 files changed, 30 insertions, 82 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index da479eb692c..18240ef0014 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -41651,12 +41651,11 @@ static SDValue combineInsertSubvector(SDNode *N, SelectionDAG &DAG,
         if (SubVec2 == SubVec && ISD::isNormalLoad(Ld) && Vec.hasOneUse())
           return DAG.getNode(X86ISD::SUBV_BROADCAST, dl, OpVT, SubVec);
 
-      // If this is subv_broadcast insert into both halves, use a larger
-      // subv_broadcast.
-      // TODO - handle X86ISD::VBROADCAST as well?
-      if (SubVec.getOpcode() == X86ISD::SUBV_BROADCAST && SubVec == SubVec2)
-        return DAG.getNode(X86ISD::SUBV_BROADCAST, dl, OpVT,
-                           SubVec.getOperand(0));
+      // If this broadcast/subv_broadcast is inserted into both halves, use a
+      // larger broadcast/subv_broadcast.
+      if (SubVec == SubVec2 && (SubVec.getOpcode() == X86ISD::VBROADCAST ||
+                                SubVec.getOpcode() == X86ISD::SUBV_BROADCAST))
+        return DAG.getNode(SubVec.getOpcode(), dl, OpVT, SubVec.getOperand(0));
 
       // If we're inserting all zeros into the upper half, change this to
       // an insert into an all zeros vector. We will match this to a move
diff --git a/llvm/test/CodeGen/X86/subvector-broadcast.ll b/llvm/test/CodeGen/X86/subvector-broadcast.ll
index 066e4dc016a..c0ab9ac47c9 100644
--- a/llvm/test/CodeGen/X86/subvector-broadcast.ll
+++ b/llvm/test/CodeGen/X86/subvector-broadcast.ll
@@ -1582,60 +1582,28 @@ define <4 x i32> @test_2xi32_to_4xi32_mem(<2 x i32>* %vp) {
 }
 
 define <8 x i32> @test_2xi32_to_8xi32_mem(<2 x i32>* %vp) {
-; X32-AVX1-LABEL: test_2xi32_to_8xi32_mem:
-; X32-AVX1:       # %bb.0:
-; X32-AVX1-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-AVX1-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
-; X32-AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; X32-AVX1-NEXT:    retl
-;
-; X32-AVX2-LABEL: test_2xi32_to_8xi32_mem:
-; X32-AVX2:       # %bb.0:
-; X32-AVX2-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; 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:    vbroadcastsd (%eax), %ymm0
-; X32-AVX512-NEXT:    retl
-;
-; X64-AVX1-LABEL: test_2xi32_to_8xi32_mem:
-; X64-AVX1:       # %bb.0:
-; X64-AVX1-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
-; X64-AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; X64-AVX1-NEXT:    retq
-;
-; X64-AVX2-LABEL: test_2xi32_to_8xi32_mem:
-; X64-AVX2:       # %bb.0:
-; X64-AVX2-NEXT:    vbroadcastsd (%rdi), %ymm0
-; X64-AVX2-NEXT:    retq
+; X32-LABEL: test_2xi32_to_8xi32_mem:
+; X32:       # %bb.0:
+; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NEXT:    vbroadcastsd (%eax), %ymm0
+; X32-NEXT:    retl
 ;
-; X64-AVX512-LABEL: test_2xi32_to_8xi32_mem:
-; X64-AVX512:       # %bb.0:
-; X64-AVX512-NEXT:    vbroadcastsd (%rdi), %ymm0
-; X64-AVX512-NEXT:    retq
+; X64-LABEL: test_2xi32_to_8xi32_mem:
+; X64:       # %bb.0:
+; X64-NEXT:    vbroadcastsd (%rdi), %ymm0
+; X64-NEXT:    retq
   %vec = load <2 x i32>, <2 x i32>* %vp
   %res = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
   ret <8 x i32> %res
 }
 
 define <16 x i32> @test_2xi32_to_16xi32_mem(<2 x i32>* %vp) {
-; X32-AVX1-LABEL: test_2xi32_to_16xi32_mem:
-; X32-AVX1:       # %bb.0:
-; X32-AVX1-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-AVX1-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
-; X32-AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; X32-AVX1-NEXT:    vmovaps %ymm0, %ymm1
-; X32-AVX1-NEXT:    retl
-;
-; X32-AVX2-LABEL: test_2xi32_to_16xi32_mem:
-; X32-AVX2:       # %bb.0:
-; X32-AVX2-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-AVX2-NEXT:    vbroadcastsd (%eax), %ymm0
-; X32-AVX2-NEXT:    vmovaps %ymm0, %ymm1
-; X32-AVX2-NEXT:    retl
+; X32-AVX-LABEL: test_2xi32_to_16xi32_mem:
+; X32-AVX:       # %bb.0:
+; X32-AVX-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX-NEXT:    vbroadcastsd (%eax), %ymm0
+; X32-AVX-NEXT:    vmovaps %ymm0, %ymm1
+; X32-AVX-NEXT:    retl
 ;
 ; X32-AVX512-LABEL: test_2xi32_to_16xi32_mem:
 ; X32-AVX512:       # %bb.0:
@@ -1645,18 +1613,11 @@ define <16 x i32> @test_2xi32_to_16xi32_mem(<2 x i32>* %vp) {
 ; X32-AVX512-NEXT:    vpermd %zmm0, %zmm1, %zmm0
 ; X32-AVX512-NEXT:    retl
 ;
-; X64-AVX1-LABEL: test_2xi32_to_16xi32_mem:
-; X64-AVX1:       # %bb.0:
-; X64-AVX1-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
-; X64-AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; X64-AVX1-NEXT:    vmovaps %ymm0, %ymm1
-; X64-AVX1-NEXT:    retq
-;
-; X64-AVX2-LABEL: test_2xi32_to_16xi32_mem:
-; X64-AVX2:       # %bb.0:
-; X64-AVX2-NEXT:    vbroadcastsd (%rdi), %ymm0
-; X64-AVX2-NEXT:    vmovaps %ymm0, %ymm1
-; X64-AVX2-NEXT:    retq
+; X64-AVX-LABEL: test_2xi32_to_16xi32_mem:
+; X64-AVX:       # %bb.0:
+; X64-AVX-NEXT:    vbroadcastsd (%rdi), %ymm0
+; X64-AVX-NEXT:    vmovaps %ymm0, %ymm1
+; X64-AVX-NEXT:    retq
 ;
 ; X64-AVX512-LABEL: test_2xi32_to_16xi32_mem:
 ; X64-AVX512:       # %bb.0:
diff --git a/llvm/test/CodeGen/X86/widened-broadcast.ll b/llvm/test/CodeGen/X86/widened-broadcast.ll
index c192ee6fd28..ecbeb532f27 100644
--- a/llvm/test/CodeGen/X86/widened-broadcast.ll
+++ b/llvm/test/CodeGen/X86/widened-broadcast.ll
@@ -607,21 +607,10 @@ define <8 x i32> @load_splat_8i32_2i32_0101(<2 x i32>* %vp) {
 ; SSE-NEXT:    movdqa %xmm0, %xmm1
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: load_splat_8i32_2i32_0101:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: load_splat_8i32_2i32_0101:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vbroadcastsd (%rdi), %ymm0
-; AVX2-NEXT:    retq
-;
-; AVX512-LABEL: load_splat_8i32_2i32_0101:
-; AVX512:       # %bb.0:
-; AVX512-NEXT:    vbroadcastsd (%rdi), %ymm0
-; AVX512-NEXT:    retq
+; AVX-LABEL: load_splat_8i32_2i32_0101:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vbroadcastsd (%rdi), %ymm0
+; AVX-NEXT:    retq
   %vec = load <2 x i32>, <2 x i32>* %vp
   %res = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
   ret <8 x i32> %res
@@ -639,8 +628,7 @@ define <16 x i32> @load_splat_16i32_2i32_0101(<2 x i32>* %vp) {
 ;
 ; AVX1-LABEL: load_splat_16i32_2i32_0101:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; AVX1-NEXT:    vbroadcastsd (%rdi), %ymm0
 ; AVX1-NEXT:    vmovaps %ymm0, %ymm1
 ; AVX1-NEXT:    retq
 ;