https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115043
Bug ID: 115043
Summary: aarch64 locally_streaming function appears to have CFA
note on wrong instruction in prologue
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: matmal01 at gcc dot gnu.org
Target Milestone: ---
Apologies if I'm misunderstanding something here -- but I noticed this RTL
sequence and I believe the `REG_CFA_DEF_CFA` note is on the wrong insn.
I have not observed wrong behaviour coming from this, but figured still worth a
bug report in case it is indeed wrong.
There seem to be a pair of instructions, one doing some special SME operation
and another storing the stack pointer into x11. The instruction doing the
special SME thing has a note saying that it sets the CFA to x11. I would have
expected the note to be on the insn after that records SP into x11.
vshcmd: > cat basic-streaming.c
[[arm::locally_streaming]] void
no_gprs_saved (__SVBool_t x)
{
asm ("");
}
gnu-work [13:19:27] $
vshcmd: > ${install_dir}/aarch64-none-linux-gnu/bin/aarch64-none-linux-gnu-gcc
\
vshcmd: > basic-streaming.c \
vshcmd: > -fdiagnostics-plain-output -march=armv8.2-a+sme+sve
-fno-stack-protector \
vshcmd: > -fdump-rtl-all-all \
vshcmd: > -O -fshrink-wrap -fstack-clash-protection -g -S -o /dev/null
> > > > gnu-work [13:19:36] $
> > > >
> > > >
> > > >
> > > >
vshcmd: > # I'm surprised that the REG_CFA_DEF_CFA note is on the instruction
vshcmd: > # just before we move the stack pointer into x11.
vshcmd: > grep -C 4 REG_CFA_DEF_CFA.*x11
basic-streaming.c.*.late_pro_and_epilogue
(insn/f 15 14 16 2 (set (reg:DI 13 x13)
(const:DI (unspec:DI [
(const_int 288 [0x120])
] UNSPEC_SME_VQ))) "basic-streaming.c":3:1 -1
(expr_list:REG_CFA_DEF_CFA (reg:DI 11 x11)
(nil)))
(insn 16 15 17 2 (set (reg:DI 11 x11)
(reg/f:DI 31 sp)) "basic-streaming.c":3:1 -1
(nil))
gnu-work [13:21:21] $
