[InstCombine] Combine (zext a) mul (zext b) to llvm.umul.with.overflow only if mul has NUW flag

[StephenFan]

Summary

Fixes: https://github.com/llvm/llvm-project/issues/59836

[nikic]

LG

llvm/test/Transforms/InstCombine/overflow-mul.ll

246

nit: Drop nounwind.

280

This commutation test doesn't look useful, because it will be commuted before the transform.

283

Add a variant that actually has the nuw flag? (Not just inferred like in above cases.)

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B),
                                 Result, Overflow)) {
         replaceInstUsesWith(*AddI, Result);
         eraseInstFromFunction(*AddI);
         return replaceInstUsesWith(I, Overflow);
       }
     }
 
     // (zext a) * (zext b)  --> llvm.umul.with.overflow.
-    if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
+    if (match(Op0, m_NUWMul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
       if (Instruction *R = processUMulZExtIdiom(I, Op0, Op1, *this))
         return R;
     }
-    if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
+    if (match(Op1, m_NUWMul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
       if (Instruction *R = processUMulZExtIdiom(I, Op1, Op0, *this))
         return R;
     }
   }
 
   if (Instruction *Res = foldICmpEquality(I))
     return Res;
 

llvm/test/Transforms/InstCombine/overflow-mul.ll

 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %ext = zext i8 %a to i32
   %mul = mul i32 %ext, zext (i8 ptrtoint (ptr @pr21445_data to i8) to i32)
   %and = and i32 %mul, 255
   %cmp = icmp ne i32 %mul, %and
   ret i1 %cmp
 }
+
+; Negative test: mul(zext x, zext y) may overflow.
+define i32 @mul_may_overflow(i32 %x, i32 %y) {
+; CHECK-LABEL: @mul_may_overflow(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[L:%.*]] = zext i32 [[X:%.*]] to i34
+; CHECK-NEXT:    [[R:%.*]] = zext i32 [[Y:%.*]] to i34
+; CHECK-NEXT:    [[MUL34:%.*]] = mul i34 [[L]], [[R]]
+; CHECK-NEXT:    [[OVERFLOW:%.*]] = icmp ult i34 [[MUL34]], 4294967296
+; CHECK-NEXT:    [[RETVAL:%.*]] = zext i1 [[OVERFLOW]] to i32
+; CHECK-NEXT:    ret i32 [[RETVAL]]
+;
+entry:
+  %l = zext i32 %x to i34
+  %r = zext i32 %y to i34
+  %mul34 = mul i34 %l, %r
+  %overflow = icmp ule i34 %mul34, 4294967295
+  %retval = zext i1 %overflow to i32
+  ret i32 %retval
+}
+
+define i32 @mul_known_nuw(i32 %x, i32 %y) {
+; CHECK-LABEL: @mul_known_nuw(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[UMUL:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 [[X:%.*]], i32 [[Y:%.*]])
+; CHECK-NEXT:    [[TMP0:%.*]] = extractvalue { i32, i1 } [[UMUL]], 1
+; CHECK-NEXT:    [[OVERFLOW:%.*]] = xor i1 [[TMP0]], true
+; CHECK-NEXT:    [[RETVAL:%.*]] = zext i1 [[OVERFLOW]] to i32
+; CHECK-NEXT:    ret i32 [[RETVAL]]
+;
+entry:
+  %l = zext i32 %x to i34
+  %r = zext i32 %y to i34
+  %mul34 = mul nuw i34 %l, %r
+  %overflow = icmp ule i34 %mul34, 4294967295
+  %retval = zext i1 %overflow to i32
+  ret i32 %retval
+}