https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120624
Bug ID: 120624
Summary: aarch64: Incorrect DCE of a ZA restore in SME code
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Keywords: aarch64-sme, wrong-code
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: rsandifo at gcc dot gnu.org
Target Milestone: ---
Target: aarch64*-*-*
For:
#include <arm_sme.h>
void callee();
__arm_new("za") __arm_locally_streaming int test()
{
svbool_t all = svptrue_b8();
svint8_t expected = svindex_s8(1, 1);
svwrite_hor_za8_m(0, 0, all, expected);
callee();
svint8_t actual = svread_hor_za8_m(svdup_s8(0), all, 0, 0);
return svptest_any(all, svcmpne(all, expected, actual));
}
compiled with -O2 -march=armv9-a+sme, GCC produces:
bl callee()
smstart sm
msr tpidr2_el0, xzr
ptrue p7.b, all
mov w12, 0
index z31.b, #1, #1
mov z30.b, #0
mova z30.b, p7/m, za0h.b[w12, 0]
This is missing the necessary conditional restore of ZA after the call to
callee.
Mode switching, which is responsible for inserting the saves and restores, does
insert a restore into the correct place. However, the restore pattern is
missing an important bit of dataflow information, which means that it can be
incorrectly deleted as dead.
Unfortunately, all the tests for this functionality used calls and asms to
model dataflow, and those don't seem to be affected.
