Failure to optimize __builtin_mul_overflow pattern

[GabrielRavier]

Bugzilla Link	48769
Resolution	FIXED
Resolved on	Sep 07, 2021 11:04
Version	trunk
OS	Linux
Depends On	#25679
CC	@LebedevRI,@RKSimon,@nickdesaulniers,@nikic,@rotateright
Fixed by commit(s)	13ec913, 35fa7b8

Extended Description

int f1(unsigned x, unsigned y)
{
unsigned int r = x * y;
return x && ((int)r / (int)x) != (int)y;
}

This can be optimized to :

int f2(unsigned x, unsigned y)
{
int res;
return __builtin_mul_overflow((int)x, (int)y, &res);
}

This optimization is done by GCC, but not by LLVM.

See also comparison here : https://godbolt.org/z/f9Yv9c

[GabrielRavier]

assigned to @LebedevRI

[RKSimon]

Alive2: https://alive2.llvm.org/ce/z/gi2rzw

---

define i1 @​src(i8 %0, i8 %1) {
%2:
%3 = mul i8 %1, %0
%4 = sdiv i8 %3, %0
%5 = icmp ne i8 %4, %1
ret i1 %5
}
=>
define i1 @​tgt(i8 %0, i8 %1) {
%2:
%3 = smul_overflow {i8, i1} %0, %1
%4 = extractvalue {i8, i1} %3, 1
ret i1 %4
}
Transformation seems to be correct!

---

define i1 @​src(i8 %0, i8 %1) {
%2:
%3 = mul i8 %1, %0
%4 = udiv i8 %3, %0
%5 = icmp ne i8 %4, %1
ret i1 %5
}
=>
define i1 @​tgt(i8 %0, i8 %1) {
%2:
%3 = umul_overflow {i8, i1} %0, %1
%4 = extractvalue {i8, i1} %3, 1
ret i1 %4
}
Transformation seems to be correct!

[LebedevRI]

Fixed in 13ec913, thanks.

[RKSimon]

Fixed in 13ec913, thanks.

Reverted at 91f7a4f

[nickdesaulniers]

Does https://reviews.llvm.org/D108928 help this get re-landed?

[LebedevRI]

Relanded in 35fa7b8.