https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120424
Bug ID: 120424
Summary: [arm] -fnon-call-exceptions -fstack-clash-protection
triggers lra-eliminations bug
Product: gcc
Version: 14.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: aoliva at gcc dot gnu.org
Reporter: aoliva at gcc dot gnu.org
Target Milestone: ---
Created attachment 61514
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=61514&action=edit
testcase regresses on arm-eabi, arm-linux-gnu, arm-vx7r2, ... starting at gcc
14
The attached C++ testcase triggers the problem. It's rewritten in C++ from
libgnat (where -fnon-call-exceptions is enabled by default, and
-fstack-clash-protection is enabled explicitly). I don't know yet whether C++
is essential, but the throw appears to be needed to trip the issue.
g() has an empty frame until reload spills a register that needs to survive a
function call. That spilling flips arm_frame_pointer_required from false to
true, so lra_update_fp2sp_elimination proceeds to disable that elimination
possibility, and ultimately we fail to adjust the spill stack slot's negative
offset, so it remains below the stack pointer.
The workaround is to disable the size == 0 frame pointer optimization. I'm
still investigating towards a proper fix.
