https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120245
Bug ID: 120245
Summary: RISC-V: Avoid FRM read/writes in non FRM related code
paths
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
Assignee: vineetg at gcc dot gnu.org
Reporter: vineetg at gcc dot gnu.org
CC: jeffreyalaw at gmail dot com, pan2.li at intel dot com,
rdapp at gcc dot gnu.org
Target Milestone: ---
For a function which needs FRM save/restore (say due to static rounding mode
setting), RISC-V frm mode switching backend generates FRM read/writes in all
the code paths, even those which don't deal with RM at all.
e.g. gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/math-ceil-0.c
-O3 -march=rv64gcv_zvfh -mabi=lp64d -ffast-math -ftree-vectorize
void test__Float16___builtin_ceilf16 (_Float16 *out, _Float16 *in, unsigned
count)
{
for (unsigned i = 0; i < count; i++)
out[i] = __builtin_ceilf16 (in[i]);
}
test__Float16___builtin_ceilf16:
.LFB2:
frrm a3 # backup RM (backup reg)
beq a2,zero,.L14
bleu a5,a4,.L3
fsrmi 3 # static RM
.L4:
vfcvt.x.f.v v3,v2,v0.t # Uses static RM
vfcvt.f.x.v v1,v3,v0.t # Uses static RM
vfsgnj.vv v1,v1,v2
bne a2,zero,.L4
fsrm a3 # Restore OK
ld ra,24(sp)
ld s0,16(sp)
ld s1,8(sp)
addi sp,sp,32
jr ra
.L3:
.L6:
call ceilf16
frrm a3
bne s0,s2,.L6
fsrm a3 # Restore needless
jr ra
.L14:
fsrm a3 # Restore Needless
ret
the frm state machine has a "sticky" note of static mode even been set and if
so, around function calls, it generates the save/restore even if that code path
doesn't need to.
FWIW The sticky static mode is needed for different reasons (see
gcc/testsuite/gcc.target/riscv/rvv/base/float-point-dynamic-frm-74.c)
Anyhow, if we can delay the intial FRM save to right before actual static RM
(that makes the backup reg live only on needed edges) and at time of restore
check if the edge for insertion has backup reg live - only then insert it.
