[InstCombine] Optimize overflow check base on uadd.with.overflow result

Fix for https://bugs.llvm.org/show_bug.cgi?id=40846.

This adds a combine for cases where a (a + b) < a style overflow
check is performed, but with a + b being the result of
uadd.with.overflow, so the overflow result is also already available
and we can just use it. Subsequently GVN/CSE will deduplicate the extracts.

We can run into this situation if you have both a uadd.with.overflow
and a manual add + overflow check in the same function (on the same
operands), in which case GVN will rewrite the add to the with.overflow
result and leave you with this pattern.

The implementation is a bit ugly because I'm handling the various
canonicalization edge cases.

This does not yet handle the negated version of this pattern.

Differential Revision: https://reviews.llvm.org/D58644

lib/Transforms/InstCombine/InstCombineCompares.cpp

@@ -5386,6 +5386,36 @@ static Instruction *foldVectorCmp(CmpInst &Cmp,
   return nullptr;
 }
 
+// extract(uadd.with.overflow(A, B), 0) ult A
+//  -> extract(uadd.with.overflow(A, B), 1)
+static Instruction *foldICmpOfUAddOv(ICmpInst &I) {
+  CmpInst::Predicate Pred = I.getPredicate();
+  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
+
+  Value *UAddOv;
+  Value *A, *B;
+  auto UAddOvResultPat = m_ExtractValue<0>(
+      m_Intrinsic<Intrinsic::uadd_with_overflow>(m_Value(A), m_Value(B)));
+  if (match(Op0, UAddOvResultPat) &&
+      ((Pred == ICmpInst::ICMP_ULT && (Op1 == A || Op1 == B)) ||
+       (Pred == ICmpInst::ICMP_EQ && match(Op1, m_ZeroInt()) &&
+        (match(A, m_One()) || match(B, m_One()))) ||
+       (Pred == ICmpInst::ICMP_NE && match(Op1, m_AllOnes()) &&
+        (match(A, m_AllOnes()) || match(B, m_AllOnes())))))
+    // extract(uadd.with.overflow(A, B), 0) < A
+    // extract(uadd.with.overflow(A, 1), 0) == 0
+    // extract(uadd.with.overflow(A, -1), 0) != -1
+    UAddOv = cast<ExtractValueInst>(Op0)->getAggregateOperand();
+  else if (match(Op1, UAddOvResultPat) &&
+           Pred == ICmpInst::ICMP_UGT && (Op0 == A || Op0 == B))
+    // A > extract(uadd.with.overflow(A, B), 0)
+    UAddOv = cast<ExtractValueInst>(Op1)->getAggregateOperand();
+  else
+    return nullptr;
+
+  return ExtractValueInst::Create(UAddOv, 1);
+}
+
 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
   bool Changed = false;
   const SimplifyQuery Q = SQ.getWithInstruction(&I);
@@ -5574,6 +5604,9 @@ Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
   if (Instruction *Res = foldICmpEquality(I))
     return Res;
 
+  if (Instruction *Res = foldICmpOfUAddOv(I))
+    return Res;
+
   // The 'cmpxchg' instruction returns an aggregate containing the old value and
   // an i1 which indicates whether or not we successfully did the swap.
   //

test/Transforms/InstCombine/with_overflow.ll

@@ -356,8 +356,7 @@ define i1 @uadd_res_ult_x(i32 %x, i32 %y, i1* %p) nounwind {
 ; CHECK-NEXT:    [[A:%.*]] = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 [[X:%.*]], i32 [[Y:%.*]])
 ; CHECK-NEXT:    [[B:%.*]] = extractvalue { i32, i1 } [[A]], 1
 ; CHECK-NEXT:    store i1 [[B]], i1* [[P:%.*]], align 1
-; CHECK-NEXT:    [[C:%.*]] = extractvalue { i32, i1 } [[A]], 0
+; CHECK-NEXT:    [[D:%.*]] = extractvalue { i32, i1 } [[A]], 1
-; CHECK-NEXT:    [[D:%.*]] = icmp ult i32 [[C]], [[X]]
 ; CHECK-NEXT:    ret i1 [[D]]
 ;
   %a = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 %x, i32 %y)
@@ -373,8 +372,7 @@ define i1 @uadd_res_ult_y(i32 %x, i32 %y, i1* %p) nounwind {
 ; CHECK-NEXT:    [[A:%.*]] = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 [[X:%.*]], i32 [[Y:%.*]])
 ; CHECK-NEXT:    [[B:%.*]] = extractvalue { i32, i1 } [[A]], 1
 ; CHECK-NEXT:    store i1 [[B]], i1* [[P:%.*]], align 1
-; CHECK-NEXT:    [[C:%.*]] = extractvalue { i32, i1 } [[A]], 0
+; CHECK-NEXT:    [[D:%.*]] = extractvalue { i32, i1 } [[A]], 1
-; CHECK-NEXT:    [[D:%.*]] = icmp ult i32 [[C]], [[Y]]
 ; CHECK-NEXT:    ret i1 [[D]]
 ;
   %a = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 %x, i32 %y)
@@ -391,8 +389,7 @@ define i1 @uadd_res_ugt_x(i32 %xx, i32 %y, i1* %p) nounwind {
 ; CHECK-NEXT:    [[A:%.*]] = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 [[X]], i32 [[Y:%.*]])
 ; CHECK-NEXT:    [[B:%.*]] = extractvalue { i32, i1 } [[A]], 1
 ; CHECK-NEXT:    store i1 [[B]], i1* [[P:%.*]], align 1
-; CHECK-NEXT:    [[C:%.*]] = extractvalue { i32, i1 } [[A]], 0
+; CHECK-NEXT:    [[D:%.*]] = extractvalue { i32, i1 } [[A]], 1
-; CHECK-NEXT:    [[D:%.*]] = icmp ugt i32 [[X]], [[C]]
 ; CHECK-NEXT:    ret i1 [[D]]
 ;
   %x = urem i32 42, %xx ; Thwart complexity-based canonicalization
@@ -410,8 +407,7 @@ define i1 @uadd_res_ugt_y(i32 %x, i32 %yy, i1* %p) nounwind {
 ; CHECK-NEXT:    [[A:%.*]] = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 [[X:%.*]], i32 [[Y]])
 ; CHECK-NEXT:    [[B:%.*]] = extractvalue { i32, i1 } [[A]], 1
 ; CHECK-NEXT:    store i1 [[B]], i1* [[P:%.*]], align 1
-; CHECK-NEXT:    [[C:%.*]] = extractvalue { i32, i1 } [[A]], 0
+; CHECK-NEXT:    [[D:%.*]] = extractvalue { i32, i1 } [[A]], 1
-; CHECK-NEXT:    [[D:%.*]] = icmp ugt i32 [[Y]], [[C]]
 ; CHECK-NEXT:    ret i1 [[D]]
 ;
   %y = urem i32 42, %yy ; Thwart complexity-based canonicalization
@@ -428,8 +424,7 @@ define i1 @uadd_res_ult_const(i32 %x, i1* %p) nounwind {
 ; CHECK-NEXT:    [[A:%.*]] = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 [[X:%.*]], i32 42)
 ; CHECK-NEXT:    [[B:%.*]] = extractvalue { i32, i1 } [[A]], 1
 ; CHECK-NEXT:    store i1 [[B]], i1* [[P:%.*]], align 1
-; CHECK-NEXT:    [[C:%.*]] = extractvalue { i32, i1 } [[A]], 0
+; CHECK-NEXT:    [[D:%.*]] = extractvalue { i32, i1 } [[A]], 1
-; CHECK-NEXT:    [[D:%.*]] = icmp ult i32 [[C]], 42
 ; CHECK-NEXT:    ret i1 [[D]]
 ;
   %a = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 %x, i32 42)
@@ -445,8 +440,7 @@ define i1 @uadd_res_ult_const_one(i32 %x, i1* %p) nounwind {
 ; CHECK-NEXT:    [[A:%.*]] = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 [[X:%.*]], i32 1)
 ; CHECK-NEXT:    [[B:%.*]] = extractvalue { i32, i1 } [[A]], 1
 ; CHECK-NEXT:    store i1 [[B]], i1* [[P:%.*]], align 1
-; CHECK-NEXT:    [[C:%.*]] = extractvalue { i32, i1 } [[A]], 0
+; CHECK-NEXT:    [[D:%.*]] = extractvalue { i32, i1 } [[A]], 1
-; CHECK-NEXT:    [[D:%.*]] = icmp eq i32 [[C]], 0
 ; CHECK-NEXT:    ret i1 [[D]]
 ;
   %a = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 %x, i32 1)
@@ -462,8 +456,7 @@ define i1 @uadd_res_ult_const_minus_one(i32 %x, i1* %p) nounwind {
 ; CHECK-NEXT:    [[A:%.*]] = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 [[X:%.*]], i32 -1)
 ; CHECK-NEXT:    [[B:%.*]] = extractvalue { i32, i1 } [[A]], 1
 ; CHECK-NEXT:    store i1 [[B]], i1* [[P:%.*]], align 1
-; CHECK-NEXT:    [[C:%.*]] = extractvalue { i32, i1 } [[A]], 0
+; CHECK-NEXT:    [[D:%.*]] = extractvalue { i32, i1 } [[A]], 1
-; CHECK-NEXT:    [[D:%.*]] = icmp ne i32 [[C]], -1
 ; CHECK-NEXT:    ret i1 [[D]]
 ;
   %a = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 %x, i32 -1)