[X86] EltsFromConsecutiveLoads - Don't confuse elt count with vector element count...

EltsFromConsecutiveLoads was assuming that the number of input elts was the same as the number of elements in the output vector type when creating a zeroing shuffle, causing an assert when subvectors were being combined instead of just scalars.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@370592 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 5812cc39b2ee0680a567159f31b6a79f31b178c0..c603aa51efcaf109244a144641d9f2323f4bffff 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -7832,17 +7832,22 @@ static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
     // IsConsecutiveLoadWithZeros - we need to create a shuffle of the loaded
     // vector and a zero vector to clear out the zero elements.
     if (!isAfterLegalize && VT.isVector()) {
-      SmallVector<int, 4> ClearMask(NumElems, -1);
-      for (unsigned i = 0; i < NumElems; ++i) {
-        if (ZeroMask[i])
-          ClearMask[i] = i + NumElems;
-        else if (LoadMask[i])
-          ClearMask[i] = i;
-      }
-      SDValue V = CreateLoad(VT, LDBase);
-      SDValue Z = VT.isInteger() ? DAG.getConstant(0, DL, VT)
-                                 : DAG.getConstantFP(0.0, DL, VT);
-      return DAG.getVectorShuffle(VT, DL, V, Z, ClearMask);
+      unsigned NumMaskElts = VT.getVectorNumElements();
+      if ((NumMaskElts % NumElems) == 0) {
+        unsigned Scale = NumMaskElts / NumElems;
+        SmallVector<int, 4> ClearMask(NumMaskElts, -1);
+        for (unsigned i = 0; i < NumElems; ++i) {
+          if (UndefMask[i])
+            continue;
+          int Offset = ZeroMask[i] ? NumMaskElts : 0;
+          for (unsigned j = 0; j != Scale; ++j)
+            ClearMask[(i * Scale) + j] = (i * Scale) + j + Offset;
+        }
+        SDValue V = CreateLoad(VT, LDBase);
+        SDValue Z = VT.isInteger() ? DAG.getConstant(0, DL, VT)
+                                   : DAG.getConstantFP(0.0, DL, VT);
+        return DAG.getVectorShuffle(VT, DL, V, Z, ClearMask);
+      }
     }
   }
 

diff --git a/test/CodeGen/X86/vector-shuffle-avx512.ll b/test/CodeGen/X86/vector-shuffle-avx512.ll
index 45f9306326f4fa17196d1ae3e2aa1bf3ba5db911..c669562dcac2dc679e0ecd318b2734a733b1e2e9 100644 (file)
--- a/test/CodeGen/X86/vector-shuffle-avx512.ll
+++ b/test/CodeGen/X86/vector-shuffle-avx512.ll
@@ -527,3 +527,41 @@ define <16 x float> @test_masked_permps_v16f32(<16 x float>* %vp, <16 x float> %
   %res = select <16 x i1> <i1 1, i1 1, i1 1, i1 0, i1 1, i1 1, i1 0, i1 0, i1 1, i1 1, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0>, <16 x float> %shuf, <16 x float> %vec2
   ret <16 x float> %res
 }
+
+%union1= type { <16 x float> }
+@src1 = external dso_local local_unnamed_addr global %union1, align 64
+
+define void @PR43170(<16 x float>* %a0) {
+; SKX64-LABEL: PR43170:
+; SKX64:       # %bb.0: # %entry
+; SKX64-NEXT:    vmovaps {{.*}}(%rip), %ymm0
+; SKX64-NEXT:    vmovaps %zmm0, (%rdi)
+; SKX64-NEXT:    vzeroupper
+; SKX64-NEXT:    retq
+;
+; KNL64-LABEL: PR43170:
+; KNL64:       # %bb.0: # %entry
+; KNL64-NEXT:    vmovaps {{.*}}(%rip), %ymm0
+; KNL64-NEXT:    vmovaps %zmm0, (%rdi)
+; KNL64-NEXT:    retq
+;
+; SKX32-LABEL: PR43170:
+; SKX32:       # %bb.0: # %entry
+; SKX32-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; SKX32-NEXT:    vmovaps src1, %ymm0
+; SKX32-NEXT:    vmovaps %zmm0, (%eax)
+; SKX32-NEXT:    vzeroupper
+; SKX32-NEXT:    retl
+;
+; KNL32-LABEL: PR43170:
+; KNL32:       # %bb.0: # %entry
+; KNL32-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; KNL32-NEXT:    vmovaps src1, %ymm0
+; KNL32-NEXT:    vmovaps %zmm0, (%eax)
+; KNL32-NEXT:    retl
+entry:
+  %0 = load <8 x float>, <8 x float>* bitcast (%union1* @src1 to <8 x float>*), align 64
+  %1 = shufflevector <8 x float> %0, <8 x float> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  store <16 x float> %1, <16 x float>* %a0, align 64
+  ret void
+}