[InstCombine] complete fold extractvalue (any_mul_with_overflow X, -1)

[Chenbing.Zheng]

Summary

When we do extractvalue (any_mul_with_overflow X, -1) --> (-X and icmp),
which left partly failed to match vector constant with poison element.
This patch try to fix it.

Alive2: https://alive2.llvm.org/ce/z/2rGp_3

[spatel]

Does this confirm with Alive2? (add link to patch summary)

llvm/lib/Transforms/InstCombine/InstructionCombining.cpp

3259–3266

Nit: that's a lot of capital letters. Just leave this named "C"?

3295

We need a test for this transform too? (and confirm that it is safe)

[Chenbing.Zheng]

add alive2 in summary

[Chenbing.Zheng]

llvm/lib/Transforms/InstCombine/InstructionCombining.cpp

3295

It seems that the tests we fix in this patch can also test for this transform, and I add Alive2 conform in summary.

[spatel]

LGTM - see inline for an extra test request.

llvm/lib/Transforms/InstCombine/InstructionCombining.cpp

3295

Ah, sorry - I didn't notice that the test diffs showed the transform was partly enabled (because m_AllOnes accepts undef).

It would still be good to have an explicit test for this path, so please add a test like this with "not-negative-one" constant:

define <4 x i1> @smul_neg1_vec_poison(<4 x i8> %x) {
  %m = call { <4 x i8>, <4 x i1> } @llvm.smul.with.overflow.v4i8(<4 x i8> %x, <4 x i8> <i8 -3, i8 -3, i8 poison, i8 -3>)
  %ov = extractvalue { <4 x i8>, <4 x i1> } %m, 1
  ret <4 x i1> %ov
}

llvm/lib/Transforms/InstCombine/InstructionCombining.cpp

 Instruction *
 InstCombinerImpl::foldExtractOfOverflowIntrinsic(ExtractValueInst &EV) {
   auto *WO = dyn_cast<WithOverflowInst>(EV.getAggregateOperand());
   if (!WO)
     return nullptr;
 
   // extractvalue (any_mul_with_overflow X, -1), 0 --> -X
   Intrinsic::ID OvID = WO->getIntrinsicID();
+  const APInt *C = nullptr;
+  if (match(WO->getRHS(), m_APIntAllowUndef(C))) {
     if (*EV.idx_begin() == 0 &&
         (OvID == Intrinsic::smul_with_overflow ||
          OvID == Intrinsic::umul_with_overflow) &&
-      match(WO->getArgOperand(1), m_AllOnes())) {
+        C->isAllOnes()) {
       return BinaryOperator::CreateNeg(WO->getArgOperand(0));
     }
+  }
 
   // We're extracting from an overflow intrinsic. See if we're the only user.
   // That allows us to simplify multiple result intrinsics to simpler things
   // that just get one value.
   if (!WO->hasOneUse())
     return nullptr;
 
   // Check if we're grabbing only the result of a 'with overflow' intrinsic
InstCombinerImpl::foldExtractOfOverflowIntrinsic(ExtractValueInst &EV) {
 
   // (usub LHS, RHS) overflows when LHS is unsigned-less-than RHS.
   if (OvID == Intrinsic::usub_with_overflow)
     return new ICmpInst(ICmpInst::ICMP_ULT, WO->getLHS(), WO->getRHS());
 
   // If only the overflow result is used, and the right hand side is a
   // constant (or constant splat), we can remove the intrinsic by directly
   // checking for overflow.
-  const APInt *C;
-  if (match(WO->getRHS(), m_APInt(C))) {
+  if (C) {
     // Compute the no-wrap range for LHS given RHS=C, then construct an
     // equivalent icmp, potentially using an offset.
     ConstantRange NWR = ConstantRange::makeExactNoWrapRegion(
         WO->getBinaryOp(), *C, WO->getNoWrapKind());
 
     CmpInst::Predicate Pred;
     APInt NewRHSC, Offset;
     NWR.getEquivalentICmp(Pred, NewRHSC, Offset);

llvm/test/Transforms/InstCombine/with_overflow.ll

 ;
   %m = tail call { <4 x i8>, <4 x i1> } @llvm.smul.with.overflow.v4i8(<4 x i8> %x, <4 x i8> <i8 -1, i8 -1, i8 -1, i8 -1>)
   %r = extractvalue { <4 x i8>, <4 x i1> } %m, 0
   %ov = extractvalue { <4 x i8>, <4 x i1> } %m, 1
   store <4 x i1> %ov, <4 x i1>* %p
   ret <4 x i8> %r
 }
 
-; TODO: partly failed to match vector constant with poison element
-
 define <4 x i8> @smul_neg1_vec_poison(<4 x i8> %x, <4 x i1>* %p) {
 ; CHECK-LABEL: @smul_neg1_vec_poison(
-; CHECK-NEXT:    [[M:%.*]] = tail call { <4 x i8>, <4 x i1> } @llvm.smul.with.overflow.v4i8(<4 x i8> [[X:%.*]], <4 x i8> <i8 -1, i8 -1, i8 poison, i8 -1>)
-; CHECK-NEXT:    [[R:%.*]] = sub <4 x i8> zeroinitializer, [[X]]
-; CHECK-NEXT:    [[OV:%.*]] = extractvalue { <4 x i8>, <4 x i1> } [[M]], 1
+; CHECK-NEXT:    [[R:%.*]] = sub <4 x i8> zeroinitializer, [[X:%.*]]
+; CHECK-NEXT:    [[OV:%.*]] = icmp eq <4 x i8> [[X]], <i8 -128, i8 -128, i8 -128, i8 -128>
 ; CHECK-NEXT:    store <4 x i1> [[OV]], <4 x i1>* [[P:%.*]], align 1
 ; CHECK-NEXT:    ret <4 x i8> [[R]]
 ;
   %m = tail call { <4 x i8>, <4 x i1> } @llvm.smul.with.overflow.v4i8(<4 x i8> %x, <4 x i8> <i8 -1, i8 -1, i8 poison, i8 -1>)
   %r = extractvalue { <4 x i8>, <4 x i1> } %m, 0
   %ov = extractvalue { <4 x i8>, <4 x i1> } %m, 1
   store <4 x i1> %ov, <4 x i1>* %p
   ret <4 x i8> %r

 ;
   %m = tail call { <4 x i8>, <4 x i1> } @llvm.umul.with.overflow.v4i8(<4 x i8> %x, <4 x i8> <i8 -1, i8 -1, i8 -1, i8 -1>)
   %r = extractvalue { <4 x i8>, <4 x i1> } %m, 0
   %ov = extractvalue { <4 x i8>, <4 x i1> } %m, 1
   store <4 x i1> %ov, <4 x i1>* %p
   ret <4 x i8> %r
 }
 
-; TODO: partly failed to match vector constant with poison element
-
 define <4 x i8> @umul_neg1_vec_poison(<4 x i8> %x, <4 x i1>* %p) {
 ; CHECK-LABEL: @umul_neg1_vec_poison(
-; CHECK-NEXT:    [[M:%.*]] = tail call { <4 x i8>, <4 x i1> } @llvm.umul.with.overflow.v4i8(<4 x i8> [[X:%.*]], <4 x i8> <i8 poison, i8 -1, i8 -1, i8 poison>)
-; CHECK-NEXT:    [[R:%.*]] = sub <4 x i8> zeroinitializer, [[X]]
-; CHECK-NEXT:    [[OV:%.*]] = extractvalue { <4 x i8>, <4 x i1> } [[M]], 1
+; CHECK-NEXT:    [[R:%.*]] = sub <4 x i8> zeroinitializer, [[X:%.*]]
+; CHECK-NEXT:    [[OV:%.*]] = icmp ugt <4 x i8> [[X]], <i8 1, i8 1, i8 1, i8 1>
 ; CHECK-NEXT:    store <4 x i1> [[OV]], <4 x i1>* [[P:%.*]], align 1
 ; CHECK-NEXT:    ret <4 x i8> [[R]]
 ;
   %m = tail call { <4 x i8>, <4 x i1> } @llvm.umul.with.overflow.v4i8(<4 x i8> %x, <4 x i8> <i8 poison, i8 -1, i8 -1, i8 poison>)
   %r = extractvalue { <4 x i8>, <4 x i1> } %m, 0
   %ov = extractvalue { <4 x i8>, <4 x i1> } %m, 1
   store <4 x i1> %ov, <4 x i1>* %p
   ret <4 x i8> %r
 }
 
+define <4 x i1> @smul_not_neg1_vec(<4 x i8> %x) {
+; CHECK-LABEL: @smul_not_neg1_vec(
+; CHECK-NEXT:    [[TMP1:%.*]] = add <4 x i8> [[X:%.*]], <i8 -43, i8 -43, i8 -43, i8 -43>
+; CHECK-NEXT:    [[OV:%.*]] = icmp ult <4 x i8> [[TMP1]], <i8 -85, i8 -85, i8 -85, i8 -85>
+; CHECK-NEXT:    ret <4 x i1> [[OV]]
+;
+  %m = call { <4 x i8>, <4 x i1> } @llvm.smul.with.overflow.v4i8(<4 x i8> %x, <4 x i8> <i8 -3, i8 -3, i8 poison, i8 -3>)
+  %ov = extractvalue { <4 x i8>, <4 x i1> } %m, 1
+  ret <4 x i1> %ov
+}
+
 ; TODO: this could be 'shl' and 'icmp'
 
 define i8 @umul_2(i8 %x, i1* %p) {
 ; CHECK-LABEL: @umul_2(
 ; CHECK-NEXT:    [[M:%.*]] = tail call { i8, i1 } @llvm.umul.with.overflow.i8(i8 [[X:%.*]], i8 2)
 ; CHECK-NEXT:    [[R:%.*]] = extractvalue { i8, i1 } [[M]], 0
 ; CHECK-NEXT:    [[OV:%.*]] = extractvalue { i8, i1 } [[M]], 1
 ; CHECK-NEXT:    store i1 [[OV]], i1* [[P:%.*]], align 1