Hello,
This seems to be causing a build regression on the or1k port.
I have not quite figured it out yet but I have bisected to this commit.
The failure is as below, this seems to be caused by the cstoresi4 instruction
produced by
or1k.md. So I think its likely something we are doing funny in the port.
Thanks to Jeff for pointing this out. If you have any idea let me know.
Log:
during RTL pass: ce1
dump file: libgcc2.c.286r.ce1
/home/shorne/work/gnu-toolchain/gcc/libgcc/libgcc2.c: In function
'__muldi3':
/home/shorne/work/gnu-toolchain/gcc/libgcc/libgcc2.c:538:1: internal
compiler error: in emit_move_multi_word, at expr.cc:4492
538 | }
| ^
0x1b094df internal_error(char const*, ...)
/home/shorne/work/gnu-toolchain/gcc/gcc/diagnostic-global-context.cc:517
0x6459c1 fancy_abort(char const*, int, char const*)
/home/shorne/work/gnu-toolchain/gcc/gcc/diagnostic.cc:1815
0x473a2e emit_move_multi_word
/home/shorne/work/gnu-toolchain/gcc/gcc/expr.cc:4492
0x93fd7d emit_move_insn(rtx_def*, rtx_def*)
/home/shorne/work/gnu-toolchain/gcc/gcc/expr.cc:4746
0x9175f1 copy_to_reg(rtx_def*)
/home/shorne/work/gnu-toolchain/gcc/gcc/explow.cc:630
0xd18977 gen_lowpart_general(machine_mode, rtx_def*)
/home/shorne/work/gnu-toolchain/gcc/gcc/rtlhooks.cc:56
0x9414c7 convert_mode_scalar
/home/shorne/work/gnu-toolchain/gcc/gcc/expr.cc:818
0x9421a1 convert_modes(machine_mode, machine_mode, rtx_def*, int)
/home/shorne/work/gnu-toolchain/gcc/gcc/expr.cc:961
0xc0afa0 prepare_operand(insn_code, rtx_def*, int, machine_mode,
machine_mode, int)
/home/shorne/work/gnu-toolchain/gcc/gcc/optabs.cc:4637
0xc0b33a prepare_cmp_insn
/home/shorne/work/gnu-toolchain/gcc/gcc/optabs.cc:4538
0xc0f0dd emit_conditional_move(rtx_def*, rtx_comparison, rtx_def*,
rtx_def*, machine_mode, int)
/home/shorne/work/gnu-toolchain/gcc/gcc/optabs.cc:5098
0x192cc7b noce_emit_cmove
/home/shorne/work/gnu-toolchain/gcc/gcc/ifcvt.cc:1777
0x193365b try_emit_cmove_seq
/home/shorne/work/gnu-toolchain/gcc/gcc/ifcvt.cc:3410
0x193365b noce_convert_multiple_sets_1
/home/shorne/work/gnu-toolchain/gcc/gcc/ifcvt.cc:3705
0x1933c71 noce_convert_multiple_sets
/home/shorne/work/gnu-toolchain/gcc/gcc/ifcvt.cc:3496
0x1938687 noce_process_if_block
/home/shorne/work/gnu-toolchain/gcc/gcc/ifcvt.cc:4045
0x1938687 noce_find_if_block
/home/shorne/work/gnu-toolchain/gcc/gcc/ifcvt.cc:4726
0x1938687 find_if_header
/home/shorne/work/gnu-toolchain/gcc/gcc/ifcvt.cc:4931
0x1938687 if_convert
/home/shorne/work/gnu-toolchain/gcc/gcc/ifcvt.cc:6078
0x1938d01 rest_of_handle_if_conversion
/home/shorne/work/gnu-toolchain/gcc/gcc/ifcvt.cc:6143
-Stafford
On Thu, May 15, 2025 at 10:33:49PM +0300, Dimitar Dimitrov wrote:
> When a paradoxical subreg is detected, validate_subreg exits early, thus
> skipping the important checks later in the function.
>
> Fix by continuing with the checks instead of declaring early that the
> paradoxical subreg is valid.
>
> One of the newly allowed subsequent checks needed to be disabled for
> paradoxical subregs. It turned out that combine attempts to create
> a paradoxical subreg of mem even for strict-alignment targets.
> That is invalid and should eventually be rejected, but is
> temporarily left allowed to prevent regressions for
> armv8l-unknown-linux-gnueabihf.
>
> Tests I did:
> - No regressions were found for C and C++ for the following targets:
> - native x86_64-pc-linux-gnu
> - cross riscv64-unknown-linux-gnu
> - cross riscv32-none-elf
> - Sanity checked armv8l-unknown-linux-gnueabihf by cross-building
> up to including libgcc. I'll monitor Linaro CI bot for the
> full regression test results.
> - Sanity checked powerpc64-unknown-linux-gnu by building native
> toolchain, but could not setup qemu-user for DejaGnu testing.
>
> PR target/119966
>
> gcc/ChangeLog:
>
> * emit-rtl.cc (validate_subreg): Do not exit immediately for
> paradoxical subregs. Filter subsequent tests which are
> not valid for paradoxical subregs.
>
> Co-authored-by: Richard Sandiford <[email protected]>
> Signed-off-by: Dimitar Dimitrov <[email protected]>
> ---
> gcc/emit-rtl.cc | 25 ++++++++++++++++++-------
> 1 file changed, 18 insertions(+), 7 deletions(-)
>
> diff --git a/gcc/emit-rtl.cc b/gcc/emit-rtl.cc
> index 3e2c4309dee..e46b0f9eac4 100644
> --- a/gcc/emit-rtl.cc
> +++ b/gcc/emit-rtl.cc
> @@ -969,10 +969,10 @@ validate_subreg (machine_mode omode, machine_mode imode,
> }
>
> /* Paradoxical subregs must have offset zero. */
> - if (maybe_gt (osize, isize))
> - return known_eq (offset, 0U);
> + if (maybe_gt (osize, isize) && !known_eq (offset, 0U))
> + return false;
>
> - /* This is a normal subreg. Verify that the offset is representable. */
> + /* Verify that the offset is representable. */
>
> /* For hard registers, we already have most of these rules collected in
> subreg_offset_representable_p. */
> @@ -988,9 +988,13 @@ validate_subreg (machine_mode omode, machine_mode imode,
>
> return subreg_offset_representable_p (regno, imode, offset, omode);
> }
> - /* Do not allow SUBREG with stricter alignment than the inner MEM. */
> + /* Do not allow normal SUBREG with stricter alignment than the inner MEM.
> +
> + FIXME: Combine can create paradoxical mem subregs even for
> + strict-alignment targets. Allow it until combine is fixed. */
> else if (reg && MEM_P (reg) && STRICT_ALIGNMENT
> - && MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (omode))
> + && MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (omode)
> + && known_le (osize, isize))
> return false;
>
> /* The outer size must be ordered wrt the register size, otherwise
> @@ -999,7 +1003,7 @@ validate_subreg (machine_mode omode, machine_mode imode,
> if (!ordered_p (osize, regsize))
> return false;
>
> - /* For pseudo registers, we want most of the same checks. Namely:
> + /* For normal pseudo registers, we want most of the same checks. Namely:
>
> Assume that the pseudo register will be allocated to hard registers
> that can hold REGSIZE bytes each. If OSIZE is not a multiple of
> REGSIZE,
> @@ -1008,8 +1012,15 @@ validate_subreg (machine_mode omode, machine_mode
> imode,
> otherwise it is at the lowest offset.
>
> Given that we've already checked the mode and offset alignment,
> - we only have to check subblock subregs here. */
> + we only have to check subblock subregs here.
> +
> + For paradoxical little-endian registers, this check is redundant. The
> + offset has already been validated to be zero.
> +
> + For paradoxical big-endian registers, this check is not valid
> + because the offset is zero. */
> if (maybe_lt (osize, regsize)
> + && known_le (osize, isize)
> && ! (lra_in_progress && (FLOAT_MODE_P (imode) || FLOAT_MODE_P
> (omode))))
> {
> /* It is invalid for the target to pick a register size for a mode
> --
> 2.49.0
>
