https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90427
Bug ID: 90427
Summary: missing "sign flipping" optimization
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: tromey at gcc dot gnu.org
Target Milestone: ---
This test case comes from this blog post:
https://nfrechette.github.io/2019/05/08/sign_flip_optimization/
(which also says that clang 8 performs this optimization).
Consider
float foo_ref(float value) {
value = value * 0.5f; // mulss value, 0.5f
// movss tmp, 1.0f
return 1.0f - value; // subss tmp, value
}
float foo_ref2(float value) {
value = value * -0.5f; // mulss value, -0.5f
return value + 1.0f; // addss value, 1.0f
}
According to the post, these are equivalent. However, gcc
generates different code for them, with the latter being better.
The comments for the first function seem to omit an
instruction that gcc emits, making that version even worse:
movss .LC0(%rip), %xmm1
mulss %xmm0, %xmm1
movss .LC1(%rip), %xmm0
subss %xmm1, %xmm0
However the comments in the second one are correct:
mulss .LC2(%rip), %xmm0
addss .LC1(%rip), %xmm0
Tested with git master from today, using gcc -O2, on x86-64 Fedora 29.
I just made a guess at which component to use.
