https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120404
Bug ID: 120404
Summary: RISC-V: inline asm FRM write ignored by FRM
save/restore machinery
Product: gcc
Version: 15.0
Status: UNCONFIRMED
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: ---
This is extracted from
gcc/testsuite/gcc.target/riscv/rvv/base/float-point-frm-run-1.c
which passes but I think is wrong.
static void
set_frm (int frm)
{
__asm__ volatile ( "fsrm %0" : :"r"(frm) : "frm");
}
main
set_frm (4);
test_float_point_frm_run_1 (op1, op2, vl);
assert_equal (4, get_frm (), "Outer: frm should be equal"); # prints 4 !=
0
vfloat32m1_t __attribute__ ((noinline))
test_float_point_frm_run_1 (vfloat32m1_t op1, vfloat32m1_t op2, size_t vl)
{
vfloat32m1_t result; # upon entry global frm is 4
set_frm (0); # global frm set to 0
result = __riscv_vfadd_vv_f32m1_rm (op1, result, 1, vl); # local clobber 1
assert_equal (1, get_frm (), "Inner: frm should be equal");
return result; # expected to restore frm 4
}
the save/restore in function ignores the inline asm setting it to 0.
inline asm is the only way global frm state is set in RISC-V, so it being
ignored for the "prev" global seems weird.
I believe this happened because of current implementation (gcc 15 release)
which generates an unconditional early/eager FRM save on function entry which
captured the global state at the start and not when really needed.
My FRM removes the early/eager save and hence fails this test.
Per discussions with Robin this behaviour is
1. Wrong
2. Implementation driven
thus needs to be changed.
