[X86][SSE] Don't replace an existing 32-bit load with its duplicate

If we are already loading a single 32-bit float/integer then just reuse it.

Fix for regression in D16729

llvm-svn: 259991

2 files changed, 47 insertions, 4 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 873d1429352..31be8e3e5f3 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -5677,7 +5677,8 @@ static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
     MVT VecSVT = VT.isFloatingPoint() ? MVT::f32 : MVT::i32;
     MVT VecVT = MVT::getVectorVT(VecSVT, VT.getSizeInBits() / 32);
     if (TLI.isTypeLegal(VecVT)) {
-      SDValue V = CreateLoad(VecSVT, LDBase);
+      SDValue V = LastLoadedElt != 0 ? CreateLoad(VecSVT, LDBase)
+                                     : DAG.getBitcast(VecSVT, EltBase);
       V = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, V);
       V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, VecVT, V);
       return DAG.getBitcast(VT, V);
diff --git a/llvm/test/CodeGen/X86/merge-consecutive-loads-128.ll b/llvm/test/CodeGen/X86/merge-consecutive-loads-128.ll
index 2d991049727..3614d5be0ef 100644
--- a/llvm/test/CodeGen/X86/merge-consecutive-loads-128.ll
+++ b/llvm/test/CodeGen/X86/merge-consecutive-loads-128.ll
@@ -602,6 +602,48 @@ define <16 x i8> @merge_16i8_i8_0123uu67uuuuuzzz(i8* %ptr) nounwind uwtable noin
   %res7 = insertelement <16 x i8> %res6, i8 %val7, i32 7
   %resD = insertelement <16 x i8> %res7, i8     0, i32 13
   %resE = insertelement <16 x i8> %resD, i8     0, i32 14
-  %resF = insertelement <16 x i8> %resE, i8     0, i32 15
-  ret <16 x i8> %resF
-}
+  %resF = insertelement <16 x i8> %resE, i8     0, i32 15

+  ret <16 x i8> %resF

+}

+

+define void @merge_4i32_i32_combine(<4 x i32>* %dst, i32* %src) {

+; SSE-LABEL: merge_4i32_i32_combine:

+; SSE:       # BB#0:

+; SSE-NEXT:    movd {{.*#+}} xmm0 = mem[0],zero,zero,zero

+; SSE-NEXT:    movaps %xmm0, (%rdi)

+; SSE-NEXT:    retq

+;

+; AVX1-LABEL: merge_4i32_i32_combine:

+; AVX1:       # BB#0:

+; AVX1-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero

+; AVX1-NEXT:    vmovaps %xmm0, (%rdi)

+; AVX1-NEXT:    retq

+;

+; AVX2-LABEL: merge_4i32_i32_combine:

+; AVX2:       # BB#0:

+; AVX2-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero

+; AVX2-NEXT:    vmovaps %xmm0, (%rdi)

+; AVX2-NEXT:    retq

+;

+; AVX512F-LABEL: merge_4i32_i32_combine:

+; AVX512F:       # BB#0:

+; AVX512F-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero

+; AVX512F-NEXT:    vmovdqa %xmm0, (%rdi)

+; AVX512F-NEXT:    retq

+;

+; X32-SSE-LABEL: merge_4i32_i32_combine:

+; X32-SSE:       # BB#0:

+; X32-SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax

+; X32-SSE-NEXT:    movl {{[0-9]+}}(%esp), %ecx

+; X32-SSE-NEXT:    movd {{.*#+}} xmm0 = mem[0],zero,zero,zero

+; X32-SSE-NEXT:    movaps %xmm0, (%eax)

+; X32-SSE-NEXT:    retl

+ %1 = getelementptr i32, i32* %src, i32 0

+ %2 = load i32, i32* %1

+ %3 = insertelement <4 x i32> undef, i32 %2, i32 0

+ %4 = shufflevector <4 x i32> %3, <4 x i32> undef, <4 x i32> zeroinitializer

+ %5 = lshr <4 x i32> %4, <i32 0, i32 undef, i32 undef, i32 undef>

+ %6 = and <4 x i32> %5, <i32 -1, i32 0, i32 0, i32 0>

+ store <4 x i32> %6, <4 x i32>* %dst

+ ret void

+}