Hello dev,
I have a question about unsafe-fp-math function attribute. Currently LLVM merges this attribute for inlining with logical AND. That means if caller has this attribute set but callee has not, LLVM will reset this attribute on caller. But shouldn’t we respect this attribute on caller? If caller wants to perform unsafe fp operation, any code that gets inlined should be allowed to do unsafe fp operation, right? Thanks!
Regards,
Yan
Hello dev,
I have a question about unsafe-fp-math function attribute. Currently LLVM merges this attribute for inlining with logical AND. That means if caller has this attribute set but callee has not, LLVM will reset this attribute on caller. But shouldn’t we respect this attribute on caller? If caller wants to perform unsafe fp operation, any code that gets inlined should be allowed to do unsafe fp operation, right? Thanks!
The answer, by design, is no. At this point, all frontends should be generating per-instruction fast-math flags, and that should also have the property that the relevant semantics are preserved across inlining. The underlying concern is that, under LTO, code that was compiled without -ffast-math can be inlined into code compiled with -ffast-math, and this combined code must continue to work even when the strict semantics are required for correctness for the code compiled without -ffast-math.
-Hal
Regards,
Yan
Hal,
Thanks for the reply. I am trying to understand the underlying concern. That means the inlining of a callee without unsafe-fp-math attribute set will block the unsafe optimization in its caller, right?
Thanks,
Yan
Hal,
Thanks for the reply. I am trying to understand the underlying concern. That means the inlining of a callee without unsafe-fp-math attribute set will block the unsafe optimization in its caller, right?
That may have been true. In recent times, we have per-instruction fast-math attributes, so the unsafe-fp-math setting isn't the controlling factor.
-Hal
Thanks,
Yan
Thanks!