https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118732
Bug ID: 118732
Summary: [15 regression] caller-saves optimization can inhibit
shrink wrapping
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: rearnsha at gcc dot gnu.org
CC: jskumari at gcc dot gnu.org, vmakarov at redhat dot com
Target Milestone: ---
Target: arm aarch64
The patch to fix PR111673 introduced a regression on Arm and AArch64 (and maybe
elsewhere) in that we now more aggressively generate caller-save registers, but
can place them before an early branch. This means that the shrink wrap code
cannot operate properly as it cannot save the register before the initial
branch unless the stack frame has been set up.
Consider for example:
void f(int *i)
{
if (!i)
return;
else
{
__builtin_printf("Hi");
*i=0;
}
}
In gcc-14 this would be shrink-wrapped to give
cbz x0, done
...
body
...
done: ret
With gcc-15 there is a store of the parameter i onto the stack, but this is
placed before the conditional branch and this is enough to block
shrink-wrapping
str x0, [sp, offset]
cbz x0, done
...
body
...
It's not clear to me why the store has to be placed before the branch: the
stack slot is dead if we do not enter body, so this is worse code anyway even
if we don't shrink wrap.
