Recommit r352255 "[SelectionDAG][X86] Don't use SEXTLOAD for promoting masked loads in the type legalizer"

This did not cause the buildbot failure it was previously reverted for.

Original commit message:

I'm not sure why we were using SEXTLOAD. EXTLOAD seems more appropriate since we don't care about the upper bits.

This patch changes this and then modifies the X86 post legalization combine to emit a extending shuffle instead of a sign_extend_vector_inreg. Could maybe use an any_extend_vector_inre

On AVX512 targets I think we might be able to use a masked vpmovzx and not have to expand this at all.

llvm-svn: 352433

3 files changed, 20 insertions, 8 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
index 62fadc1e623..e94ccc08d99 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
@@ -554,7 +554,7 @@ SDValue DAGTypeLegalizer::PromoteIntRes_MLOAD(MaskedLoadSDNode *N) {
   SDLoc dl(N);
   SDValue Res = DAG.getMaskedLoad(NVT, dl, N->getChain(), N->getBasePtr(),
                                   N->getMask(), ExtPassThru, N->getMemoryVT(),
-                                  N->getMemOperand(), ISD::SEXTLOAD);
+                                  N->getMemOperand(), ISD::EXTLOAD);
   // Legalize the chain result - switch anything that used the old chain to
   // use the new one.
   ReplaceValueWith(SDValue(N, 1), Res.getValue(1));
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index d98a0d17a71..92001c77097 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -37694,7 +37694,7 @@ static SDValue combineMaskedLoad(SDNode *N, SelectionDAG &DAG,
         return Blend;
   }
 
-  if (Mld->getExtensionType() != ISD::SEXTLOAD)
+  if (Mld->getExtensionType() != ISD::EXTLOAD)
     return SDValue();
 
   // Resolve extending loads.
@@ -37764,8 +37764,20 @@ static SDValue combineMaskedLoad(SDNode *N, SelectionDAG &DAG,
                                      Mld->getBasePtr(), NewMask, WidePassThru,
                                      Mld->getMemoryVT(), Mld->getMemOperand(),
                                      ISD::NON_EXTLOAD);
-  SDValue NewVec = getExtendInVec(/*Signed*/true, dl, VT, WideLd, DAG);
-  return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
+
+  SDValue SlicedVec = DAG.getBitcast(WideVecVT, WideLd);
+  SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
+  for (unsigned i = 0; i != NumElems; ++i)
+    ShuffleVec[i * SizeRatio] = i;
+
+  // Can't shuffle using an illegal type.
+  assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
+         "WideVecVT should be legal");
+  SlicedVec = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
+                                   DAG.getUNDEF(WideVecVT), ShuffleVec);
+  SlicedVec = DAG.getBitcast(VT, SlicedVec);
+
+  return DCI.CombineTo(N, SlicedVec, WideLd.getValue(1), true);
 }
 
 /// If exactly one element of the mask is set for a non-truncating masked store,
diff --git a/llvm/test/CodeGen/X86/masked_load.ll b/llvm/test/CodeGen/X86/masked_load.ll
index 5a456bd2d71..8d28f45d988 100644
--- a/llvm/test/CodeGen/X86/masked_load.ll
+++ b/llvm/test/CodeGen/X86/masked_load.ll
@@ -1388,7 +1388,7 @@ define <2 x i32> @load_v2i32_v2i32(<2 x i32> %trigger, <2 x i32>* %addr, <2 x i3
 ; AVX1-NEXT:    vmaskmovps (%rdi), %xmm0, %xmm2
 ; AVX1-NEXT:    vpermilps {{.*#+}} xmm1 = xmm1[0,2,2,3]
 ; AVX1-NEXT:    vblendvps %xmm0, %xmm2, %xmm1, %xmm0
-; AVX1-NEXT:    vpmovsxdq %xmm0, %xmm0
+; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: load_v2i32_v2i32:
@@ -1400,7 +1400,7 @@ define <2 x i32> @load_v2i32_v2i32(<2 x i32> %trigger, <2 x i32>* %addr, <2 x i3
 ; AVX2-NEXT:    vpmaskmovd (%rdi), %xmm0, %xmm2
 ; AVX2-NEXT:    vpermilps {{.*#+}} xmm1 = xmm1[0,2,2,3]
 ; AVX2-NEXT:    vblendvps %xmm0, %xmm2, %xmm1, %xmm0
-; AVX2-NEXT:    vpmovsxdq %xmm0, %xmm0
+; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
 ; AVX2-NEXT:    retq
 ;
 ; AVX512F-LABEL: load_v2i32_v2i32:
@@ -1412,7 +1412,7 @@ define <2 x i32> @load_v2i32_v2i32(<2 x i32> %trigger, <2 x i32>* %addr, <2 x i3
 ; AVX512F-NEXT:    kshiftlw $14, %k0, %k0
 ; AVX512F-NEXT:    kshiftrw $14, %k0, %k1
 ; AVX512F-NEXT:    vmovdqu32 (%rdi), %zmm0 {%k1}
-; AVX512F-NEXT:    vpmovsxdq %xmm0, %xmm0
+; AVX512F-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
@@ -1423,7 +1423,7 @@ define <2 x i32> @load_v2i32_v2i32(<2 x i32> %trigger, <2 x i32>* %addr, <2 x i3
 ; AVX512VLBW-NEXT:    vptestnmq %xmm0, %xmm0, %k1
 ; AVX512VLBW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
 ; AVX512VLBW-NEXT:    vmovdqu32 (%rdi), %xmm0 {%k1}
-; AVX512VLBW-NEXT:    vpmovsxdq %xmm0, %xmm0
+; AVX512VLBW-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
 ; AVX512VLBW-NEXT:    retq
   %mask = icmp eq <2 x i32> %trigger, zeroinitializer
   %res = call <2 x i32> @llvm.masked.load.v2i32.p0v2i32(<2 x i32>* %addr, i32 4, <2 x i1>%mask, <2 x i32>%dst)