]> gcc.gnu.org Git - gcc.git/commit
aarch64, builtins: Include PR registers in FUNCTION_ARG_REGNO_P etc. [PR109254]
authorJakub Jelinek <jakub@redhat.com>
Sat, 1 Apr 2023 06:55:55 +0000 (08:55 +0200)
committerJakub Jelinek <jakub@redhat.com>
Sat, 1 Apr 2023 06:55:55 +0000 (08:55 +0200)
commitb1f6cb2cc3aad0521ad3181d5107e52be155fd18
tree8685e8eb0e2633073de945b9ae7319ec9ca63d50
parentfc4cde2e6aa4d6ebdf7f70b7b4359fb59a1915ae
aarch64, builtins: Include PR registers in FUNCTION_ARG_REGNO_P etc. [PR109254]

The following testcase is miscompiled on aarch64-linux in the regname pass,
because while the function takes arguments in the p0 register,
FUNCTION_ARG_REGNO_P doesn't reflect that, so DF doesn't know the register is
used in register passing. It sees 2 chains with p1 register and wants to
replace the second one and as DF doesn't know p0 is live at the start of the
function, it will happily use p0 register even when it is used in subsequent
instructions.

The following patch fixes that.  FUNCTION_ARG_REGNO_P returns non-zero
for p0-p3 (unconditionally, seems for the floating/vector registers it
doesn't conditionalize them on TARGET_FLOAT either, but if you want,
I can conditionalize p0-p3 on TARGET_SVE), similarly
targetm.calls.function_value_regno_p returns true for p0-p3 registers
if TARGET_SVE (again for consistency, that function conditionalizes
the float/vector on TARGET_FLOAT).

Now, that change broke bootstrap in libobjc and some
__builtin_apply_args/__builtin_apply/__builtin_return tests.  The
aarch64_get_reg_raw_mode hook already documents that SVE scalable arg/return
passing is fundamentally incompatible with those builtins, but unlike
the floating/vector regs where it forces a fixed vector mode, I think
there is no fixed mode which could be used for p0-p3.  So, I have tweaked
the generic code so that it uses VOIDmode return from that hook to signal
that a register shouldn't be touched by
__builtin_apply_args/__builtin_apply/__builtin_return
despite being mentioned in FUNCTION_ARG_REGNO_P or
targetm.calls.function_value_regno_p.

gcc/
2023-04-01  Jakub Jelinek  <jakub@redhat.com>

PR target/109254
* builtins.cc (apply_args_size): If targetm.calls.get_raw_arg_mode
returns VOIDmode, handle it like if the register isn't used for
passing arguments at all.
(apply_result_size): If targetm.calls.get_raw_result_mode returns
VOIDmode, handle it like if the register isn't used for returning
results at all.
* target.def (get_raw_result_mode, get_raw_arg_mode): Document what it
means to return VOIDmode.
* doc/tm.texi: Regenerated.
* config/aarch64/aarch64.cc (aarch64_function_value_regno_p): Return
TARGET_SVE for P0_REGNUM.
(aarch64_function_arg_regno_p): Also return true for p0-p3.
(aarch64_get_reg_raw_mode): Return VOIDmode for PR_REGNUM_P regs.

gcc/testsuite/
2023-04-01  Jakub Jelinek  <jakub@redhat.com>
    Richard Sandiford  <richard.sandiford@arm.com>

PR target/109254
* gcc.target/aarch64/sve/pr109254.c: New test.
gcc/builtins.cc
gcc/config/aarch64/aarch64.cc
gcc/doc/tm.texi
gcc/target.def
gcc/testsuite/gcc.target/aarch64/sve/pr109254.c [new file with mode: 0644]
This page took 0.060136 seconds and 5 git commands to generate.