https://gcc.gnu.org/g:930fc5e2f0a44d4b62890a4eab4efc7eb5537512
commit 930fc5e2f0a44d4b62890a4eab4efc7eb5537512
Author: Alexandre Oliva <ol...@adacore.com>
Date: Wed Jun 18 04:13:19 2025 -0300
[genoutput] mark scratch outputs as eliminable
acats' fdd2a00.read is miscompiled on arm-linux-gnu with -O2
-fstack-clash-protection -march=armv7-a -marm: a clobbered scratch
register in a *iorsi3_compare0_scratch pattern gets initially assigned
to the frame pointer register, but at some point during lra the frame
size grows to nonzero, arm_frame_pointer_required flips to true, and
the fp2sp elimination has to be disabled, so the scratch register gets
spilled to a stack slot.
It needs to get the sfp elimination at that point, because later
rounds of elimination will assume the previous round's offset has
already been applied. But since scratch matches are not regarded as
eliminable by genoutput, we don't attempt elimination in the clobbered
stack slot MEM rtx.
Later on, lra issues a reload for that slot, using a new pseudo
allocated to a hardware register, that gets stored in the stack slot
after the original insn. Elimination in that reload store insn
eventually updates the elimination offset, but it's an incremental
update, assuming that the offset so far has already been applied.
Without applying the initial offset, the store ends up overlapping
with the function's register save area, corrupting a caller's
call-saved register.
AFAICT the old reload's elimination wouldn't be harmed by allowing
elimination in scratch operands, so I'm enabling eliminable for them
regardless. Should it be found to make a difference, we could
presumably set a different bit in eliminable to enable reload and lra
to tell them apart and behave accordingly.
for gcc/ChangeLog
* genoutput.cc (scan_operands): Make MATCH_SCRATCHes eliminable.
Diff:
---
gcc/genoutput.cc | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/gcc/genoutput.cc b/gcc/genoutput.cc
index dd4e7b80c2a9..25d0b8b86467 100644
--- a/gcc/genoutput.cc
+++ b/gcc/genoutput.cc
@@ -478,7 +478,7 @@ scan_operands (class data *d, rtx part, int this_address_p,
d->operand[opno].n_alternatives
= n_occurrences (',', d->operand[opno].constraint) + 1;
d->operand[opno].address_p = 0;
- d->operand[opno].eliminable = 0;
+ d->operand[opno].eliminable = 1;
return;
case MATCH_OPERATOR:
