Reapply "[X86] Limit X86InterleavedAccessGroup to handle the same type case only"

The current implementation assumes the destination type of shuffle is the same as the decomposed ones. Add the check to avoid crush when the condition is not satisfied.

This fixes PR37616.

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D102751

Author: phoebewang

llvm/lib/Target/X86/X86InterleavedAccess.cpp

@@ -724,30 +724,34 @@ bool X86InterleavedAccessGroup::lowerIntoOptimizedSequence() {
   auto *ShuffleTy = cast<FixedVectorType>(Shuffles[0]->getType());
 
   if (isa<LoadInst>(Inst)) {
-    // Try to generate target-sized register(/instruction).
-    decompose(Inst, Factor, ShuffleTy, DecomposedVectors);
-
     auto *ShuffleEltTy = cast<FixedVectorType>(Inst->getType());
     unsigned NumSubVecElems = ShuffleEltTy->getNumElements() / Factor;
-    // Perform matrix-transposition in order to compute interleaved
-    // results by generating some sort of (optimized) target-specific
-    // instructions.
-
     switch (NumSubVecElems) {
     default:
       return false;
     case 4:
-      transpose_4x4(DecomposedVectors, TransposedVectors);
-      break;
     case 8:
     case 16:
     case 32:
     case 64:
-      deinterleave8bitStride3(DecomposedVectors, TransposedVectors,
-                              NumSubVecElems);
+      if (ShuffleTy->getNumElements() != NumSubVecElems)
+        return false;
       break;
     }
 
+    // Try to generate target-sized register(/instruction).
+    decompose(Inst, Factor, ShuffleTy, DecomposedVectors);
+
+    // Perform matrix-transposition in order to compute interleaved
+    // results by generating some sort of (optimized) target-specific
+    // instructions.
+
+    if (NumSubVecElems == 4)
+      transpose_4x4(DecomposedVectors, TransposedVectors);
+    else
+      deinterleave8bitStride3(DecomposedVectors, TransposedVectors,
+                              NumSubVecElems);
+
     // Now replace the unoptimized-interleaved-vectors with the
     // transposed-interleaved vectors.
     for (unsigned i = 0, e = Shuffles.size(); i < e; ++i)

llvm/test/CodeGen/X86/x86-interleaved-access.ll

@@ -1930,3 +1930,22 @@ define void @splat4_v4i64_load_store(<4 x i64>* %s, <16 x i64>* %d) {
   store <16 x i64> %r, <16 x i64>* %d, align 8
   ret void
 }
+
+define <2 x i64> @PR37616(<16 x i64>* %a0) {
+; AVX1-LABEL: PR37616:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vmovaps 16(%rdi), %xmm0
+; AVX1-NEXT:    vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
+; AVX1-NEXT:    retq
+;
+; AVX2OR512-LABEL: PR37616:
+; AVX2OR512:       # %bb.0:
+; AVX2OR512-NEXT:    vmovaps (%rdi), %ymm0
+; AVX2OR512-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],mem[0],ymm0[2],mem[2]
+; AVX2OR512-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX2OR512-NEXT:    vzeroupper
+; AVX2OR512-NEXT:    retq
+  %load = load <16 x i64>, <16 x i64>* %a0, align 128
+  %shuffle = shufflevector <16 x i64> %load, <16 x i64> undef, <2 x i32> <i32 2, i32 6>
+  ret <2 x i64> %shuffle
+}