[DAG] replaceStoreOfInsertLoad - don't fold if the inserted element is implicity truncated

[RKSimon]

Summary

D152276 wasn't handling the case where the inserted element is implicity truncated into the vector - resulting in a i1 element (implictly truncated from i8) overwriting 8 bits instead of 1 bit.

This patch is intended to be merged into 17.x so I've just disallowed any vector element vs inserted element type mismatch - technically we could be more elegant and permit truncated stores (as long as the store is still byte sized), but the use cases for that are so limited I'd prefer to play it safe for now.

Candidate patch for #64655 17.x merge

[jacobly]

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

20514

EltVT is the scalar type being inserted, but operand 0 is the vector being inserted into, so these types can never be equal. Maybe EltVT != Value.getOperand(0).getValueType().getVectorElementType()?

[RKSimon]

fix dumb typo

[RKSimon]

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

20514

Sorry about that :)

[luke]

LGTM, +1 on being conservative for now

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

   SDValue Ptr = ST->getBasePtr();
   SDValue Chain = ST->getChain();
   if (Value.getOpcode() != ISD::INSERT_VECTOR_ELT || !Value.hasOneUse())
     return SDValue();
 
   SDValue Elt = Value.getOperand(1);
   SDValue Idx = Value.getOperand(2);
 
-  // If the element isn't byte sized then we can't compute an offset
+  // If the element isn't byte sized or is implicitly truncated then we can't
+  // compute an offset.
   EVT EltVT = Elt.getValueType();
-  if (!EltVT.isByteSized())
+  if (!EltVT.isByteSized() || EltVT != Value.getOperand(0).getValueType())
     return SDValue();
 
   auto *Ld = dyn_cast<LoadSDNode>(Value.getOperand(0));
   if (!Ld || Ld->getBasePtr() != Ptr ||
       ST->getMemoryVT() != Ld->getMemoryVT() || !ST->isSimple() ||
       !ISD::isNormalStore(ST) ||
       Ld->getAddressSpace() != ST->getAddressSpace() ||
       !Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1)))

llvm/test/CodeGen/X86/pr64655.ll

 ; AVX2-NEXT:    vpsllw $15, %xmm0, %xmm0
 ; AVX2-NEXT:    vpacksswb %xmm0, %xmm0, %xmm0
 ; AVX2-NEXT:    vpmovmskb %xmm0, %eax
 ; AVX2-NEXT:    movb %al, (%rdi)
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: f:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    movb $1, 1(%rdi)
+; AVX512-NEXT:    kmovb (%rdi), %k0
+; AVX512-NEXT:    movb $-3, %al
+; AVX512-NEXT:    kmovd %eax, %k1
+; AVX512-NEXT:    kandb %k1, %k0, %k0
+; AVX512-NEXT:    movb $1, %al
+; AVX512-NEXT:    kmovd %eax, %k1
+; AVX512-NEXT:    kshiftlb $7, %k1, %k1
+; AVX512-NEXT:    kshiftrb $6, %k1, %k1
+; AVX512-NEXT:    korb %k1, %k0, %k0
+; AVX512-NEXT:    kmovb %k0, (%rdi)
 ; AVX512-NEXT:    retq
   %2 = load <8 x i1>, ptr %0
   %3 = insertelement <8 x i1> %2, i1 true, i32 1
   store <8 x i1> %3, ptr %0
   ret void
 }