[InstSimplify] Failure to drop check-for-zero after uadd_with_overflow

[LebedevRI]

Bugzilla Link	43246
Version	trunk
OS	Linux
CC	@nikic,@rotateright,@vedantk
Fixed by commit(s)	371349

Extended Description

https://rise4fun.com/Alive/yFB

Name: zz
Pre: C != 0
%1 = add i64 C, %offset
%2 = icmp ne i64 %1, 0
%3 = icmp uge i64 %1, C
%r = and i1 %2, %3
=>
%r = icmp uge i64 %1, C

This comes up in https://reviews.llvm.org/D67122

for the code like

char* test(char& base, unsigned long offset) {
return &base + offset;
}

[nikic]

Do you have an IR example where this fails to fold?

Running

define i1 @​test(i64 %offset) {
%1 = add i64 100, %offset
%2 = icmp ne i64 %1, 0
%3 = icmp uge i64 %1, 100
%r = and i1 %2, %3
ret i1 %r
}

through opt -instcombine gives:

define i1 @​test(i64 %offset) {
%1 = add i64 100, %offset
%2 = icmp uge i64 %1, 100
ret i1 %2
}

instcombine would also combine the add+icmp.

[LebedevRI]

Do you have an IR example where this fails to fold?

https://godbolt.org/z/9FOIww

[LebedevRI]

Do you have an IR example where this fails to fold?

https://godbolt.org/z/9FOIww

in alive i didn't actually mean that C is constant, only that it is non-zero.

[LebedevRI]

Do you have an IR example where this fails to fold?

https://godbolt.org/z/9FOIww

in alive i didn't actually mean that C is constant, only that it is non-zero.

(i'll take a stab at this tomorrow)

[LebedevRI]

https://reviews.llvm.org/D67332
I think that is the right place to patch?

[LebedevRI]

r371349, but not marking as fixed because we should also handle non-canonical case of an actual @​llvm.uadd.with.overflow