| Message ID | UKLdwyvPY-zGlJUe3VfgFr6scXKvmA6hQhpu76J3f54dkFq3PZuSUC_1S2Yxg2hAnHCnS7Jju3U2R1rSOHmmMFjlmaJ9xKzQJujiymuk7tw=@skychase.zone |
|---|---|
| State | New |
| Headers | show |
| Series | rtlhooks: Don't widen a MEM in gen_lowpart_if_possible [PR125728] | expand |
On 11/4/26 8:35 AM, Andrew DeRosier wrote: > r16-160-ge6f89d78c1a752 made a change to simplify-rtx that exposes an > existing latent bug in gen_lowpart_if_possible, which widens a MEM with > byte_lowpart_offset. That offset is negative on big-endian targets, > so on m68k a byte load from foo becomes a longword load from foo-3, > causing an address error. > > Based on how gen_lowpart_for_combine refuses to widen a MEM, this patch > makes gen_lowpart_if_possible return NULL_RTX for a paradoxical lowpart. > > Bootstrapped and regtested on x86_64-pc-linux-gnu. Also tested on > m68k-elf with a Musashi-based 68000 simulator, where it fixes address > errors in 20040709-2.c, 20040709-3.c, pr57875.c and pr79737-1.c at > several optimization levels, with no regressions. > > gcc/ChangeLog: > > PR rtl-optimization/125728 > * rtlhooks.cc (gen_lowpart_if_possible): Return NULL_RTX for a > paradoxical lowpart of a MEM instead of widening the access. > > gcc/testsuite/ChangeLog: > > PR rtl-optimization/125728 > * gcc.target/m68k/pr125728.c: New test. > > Signed-off-by: Andrew DeRosier <andy@skychase.zone> Changes in these routines are always tricky, so before even thinking hard about the patch I threw it into my tester and multiple targets are throwing errors during self-test. So I don't *think* you need a full cross environment. You should be able to configure a cross compiler for one of these targets and try to build. What you'll be looking for is something like this: cc1: internal compiler error: in gen_rtx_SUBREG, at emit-rtl.cc:1062 0x1f4939d internal_error(char const*, ...) ../../..//gcc/gcc/diagnostic-global-context.cc:787 0x892d3d fancy_abort(char const*, int, char const*) ../../..//gcc/gcc/diagnostics/context.cc:1813 0x661e02 gen_rtx_SUBREG(machine_mode, rtx_def*, poly_int<1u, unsigned long>) ../../..//gcc/gcc/emit-rtl.cc:1062 0x661e02 gen_rtx_SUBREG(machine_mode, rtx_def*, poly_int<1u, unsigned long>) ../../..//gcc/gcc/emit-rtl.cc:1060 0xbaa785 test_output_operand_modifier_operand ../../..//gcc/gcc/final.cc:4798 0xbaa785 selftest::final_cc_tests() ../../..//gcc/gcc/final.cc:4820 0x1eaabbb selftest::run_tests() ../../..//gcc/gcc/selftest-run-tests.cc:109 0xff471a toplev::run_self_tests() ../../..//gcc/gcc/toplev.cc:2287 This is from H8 I think, but the others I spot checked had similar behavior (microblaze-linux, moxie-elf, iq2000-elf, h8300-elf). Jeff Jeff
On Tue, Nov 10, 2026 at 8:20 AM Jeff Law <jeffrey.law@oss.qualcomm.com> wrote: > > > > On 11/4/26 8:35 AM, Andrew DeRosier wrote: > > r16-160-ge6f89d78c1a752 made a change to simplify-rtx that exposes an > > existing latent bug in gen_lowpart_if_possible, which widens a MEM with > > byte_lowpart_offset. That offset is negative on big-endian targets, > > so on m68k a byte load from foo becomes a longword load from foo-3, > > causing an address error. > > > > Based on how gen_lowpart_for_combine refuses to widen a MEM, this patch > > makes gen_lowpart_if_possible return NULL_RTX for a paradoxical lowpart. > > > > Bootstrapped and regtested on x86_64-pc-linux-gnu. Also tested on > > m68k-elf with a Musashi-based 68000 simulator, where it fixes address > > errors in 20040709-2.c, 20040709-3.c, pr57875.c and pr79737-1.c at > > several optimization levels, with no regressions. > > > > gcc/ChangeLog: > > > > PR rtl-optimization/125728 > > * rtlhooks.cc (gen_lowpart_if_possible): Return NULL_RTX for a > > paradoxical lowpart of a MEM instead of widening the access. > > > > gcc/testsuite/ChangeLog: > > > > PR rtl-optimization/125728 > > * gcc.target/m68k/pr125728.c: New test. > > > > Signed-off-by: Andrew DeRosier <andy@skychase.zone> > Changes in these routines are always tricky, so before even thinking > hard about the patch I threw it into my tester and multiple targets are > throwing errors during self-test. So I don't *think* you need a full > cross environment. You should be able to configure a cross compiler for > one of these targets and try to build. What you'll be looking for is > something like this: > > > > > > cc1: internal compiler error: in gen_rtx_SUBREG, at emit-rtl.cc:1062 > 0x1f4939d internal_error(char const*, ...) > ../../..//gcc/gcc/diagnostic-global-context.cc:787 > 0x892d3d fancy_abort(char const*, int, char const*) > ../../..//gcc/gcc/diagnostics/context.cc:1813 > 0x661e02 gen_rtx_SUBREG(machine_mode, rtx_def*, poly_int<1u, unsigned long>) > ../../..//gcc/gcc/emit-rtl.cc:1062 > 0x661e02 gen_rtx_SUBREG(machine_mode, rtx_def*, poly_int<1u, unsigned long>) > ../../..//gcc/gcc/emit-rtl.cc:1060 > 0xbaa785 test_output_operand_modifier_operand > ../../..//gcc/gcc/final.cc:4798 > 0xbaa785 selftest::final_cc_tests() > ../../..//gcc/gcc/final.cc:4820 > 0x1eaabbb selftest::run_tests() > ../../..//gcc/gcc/selftest-run-tests.cc:109 > 0xff471a toplev::run_self_tests() > ../../..//gcc/gcc/toplev.cc:2287 > > This is from H8 I think, but the others I spot checked had similar > behavior (microblaze-linux, moxie-elf, iq2000-elf, h8300-elf). Yes I was very much worried about fallout here even. I did try to warn about this fallout in https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125728#c5 but I could have been more verbose there. Thanks, Andrea > > Jeff > > Jeff
On Wednesday, 11 November 2026 at 3:15 AM, Andrea Pinski <andrew.pinski@oss.qualcomm.com> wrote: > > On Tue, Nov 10, 2026 at 8:20 AM Jeff Law <jeffrey.law@oss.qualcomm.com> wrote: > > > > > > > > On 11/4/26 8:35 AM, Andrew DeRosier wrote: > > > r16-160-ge6f89d78c1a752 made a change to simplify-rtx that exposes an > > > existing latent bug in gen_lowpart_if_possible, which widens a MEM with > > > byte_lowpart_offset. That offset is negative on big-endian targets, > > > so on m68k a byte load from foo becomes a longword load from foo-3, > > > causing an address error. > > > > > > Based on how gen_lowpart_for_combine refuses to widen a MEM, this patch > > > makes gen_lowpart_if_possible return NULL_RTX for a paradoxical lowpart. > > > > > > Bootstrapped and regtested on x86_64-pc-linux-gnu. Also tested on > > > m68k-elf with a Musashi-based 68000 simulator, where it fixes address > > > errors in 20040709-2.c, 20040709-3.c, pr57875.c and pr79737-1.c at > > > several optimization levels, with no regressions. > > > > > > gcc/ChangeLog: > > > > > > PR rtl-optimization/125728 > > > * rtlhooks.cc (gen_lowpart_if_possible): Return NULL_RTX for a > > > paradoxical lowpart of a MEM instead of widening the access. > > > > > > gcc/testsuite/ChangeLog: > > > > > > PR rtl-optimization/125728 > > > * gcc.target/m68k/pr125728.c: New test. > > > > > > Signed-off-by: Andrew DeRosier <andy@skychase.zone> > > Changes in these routines are always tricky, so before even thinking > > hard about the patch I threw it into my tester and multiple targets are > > throwing errors during self-test. So I don't *think* you need a full > > cross environment. You should be able to configure a cross compiler for > > one of these targets and try to build. What you'll be looking for is > > something like this: > > > > > > > > > > > > cc1: internal compiler error: in gen_rtx_SUBREG, at emit-rtl.cc:1062 > > 0x1f4939d internal_error(char const*, ...) > > ../../..//gcc/gcc/diagnostic-global-context.cc:787 > > 0x892d3d fancy_abort(char const*, int, char const*) > > ../../..//gcc/gcc/diagnostics/context.cc:1813 > > 0x661e02 gen_rtx_SUBREG(machine_mode, rtx_def*, poly_int<1u, unsigned long>) > > ../../..//gcc/gcc/emit-rtl.cc:1062 > > 0x661e02 gen_rtx_SUBREG(machine_mode, rtx_def*, poly_int<1u, unsigned long>) > > ../../..//gcc/gcc/emit-rtl.cc:1060 > > 0xbaa785 test_output_operand_modifier_operand > > ../../..//gcc/gcc/final.cc:4798 > > 0xbaa785 selftest::final_cc_tests() > > ../../..//gcc/gcc/final.cc:4820 > > 0x1eaabbb selftest::run_tests() > > ../../..//gcc/gcc/selftest-run-tests.cc:109 > > 0xff471a toplev::run_self_tests() > > ../../..//gcc/gcc/toplev.cc:2287 > > > > This is from H8 I think, but the others I spot checked had similar > > behavior (microblaze-linux, moxie-elf, iq2000-elf, h8300-elf). > > Yes I was very much worried about fallout here even. I did try to warn > about this fallout in > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125728#c5 but I could > have been more verbose there. > > Thanks, > Andrea > > > > > Jeff > > > > Jeff > Pulled e26706962, set up an h8300-elf cross compiler... and the self tests passed. It looks like the failing test from that trace (test_output_operand_modifier_operand) is new (f90f14685) and something was fixed about it in 0f8ee5980. After reverting that commit I reproduced the ICE on h8300 and even m68k without this patch. Based on the commit message, I'm thinking the big endian targets were broken in that 4 hour window. https://gcc.gnu.org/cgit/gcc/commit/?id=0f8ee5980094f46eb7ef6398beda7ffe125e7608 Thanks, Andrew
diff --git a/gcc/rtlhooks.cc b/gcc/rtlhooks.cc index a236a6e..d6464f9 100644 --- a/gcc/rtlhooks.cc +++ b/gcc/rtlhooks.cc @@ -115,6 +115,9 @@ gen_lowpart_if_possible (machine_mode mode, rtx x) else if (MEM_P (x)) { /* This is the only other case we handle. */ + if (paradoxical_subreg_p (mode, GET_MODE (x))) + return 0; + poly_int64 offset = byte_lowpart_offset (mode, GET_MODE (x)); rtx new_rtx = adjust_address_nv (x, mode, offset); if (! memory_address_addr_space_p (mode, XEXP (new_rtx, 0), diff --git a/gcc/testsuite/gcc.target/m68k/pr125728.c b/gcc/testsuite/gcc.target/m68k/pr125728.c new file mode 100644 index 0000000..b31d9b8 --- /dev/null +++ b/gcc/testsuite/gcc.target/m68k/pr125728.c @@ -0,0 +1,23 @@ +/* PR rtl-optimization/125728 */ +/* { dg-do compile } */ +/* { dg-options "-O2 -m68000" } */ +/* { dg-final { scan-assembler-not "foo-3" } } */ +/* { dg-final { scan-assembler-not "osc\\+9" } } */ + +extern unsigned char foo; +extern const int table[8]; + +int +index_byte (void) +{ + return table[foo & 7]; +} + +extern unsigned char osc[66] __attribute__ ((aligned (2))); +extern unsigned char lani __attribute__ ((aligned (2))); + +unsigned short +sum_byte (void) +{ + return (unsigned short) lani * 16 + (osc[10] & 15); +}
r16-160-ge6f89d78c1a752 made a change to simplify-rtx that exposes an existing latent bug in gen_lowpart_if_possible, which widens a MEM with byte_lowpart_offset. That offset is negative on big-endian targets, so on m68k a byte load from foo becomes a longword load from foo-3, causing an address error. Based on how gen_lowpart_for_combine refuses to widen a MEM, this patch makes gen_lowpart_if_possible return NULL_RTX for a paradoxical lowpart. Bootstrapped and regtested on x86_64-pc-linux-gnu. Also tested on m68k-elf with a Musashi-based 68000 simulator, where it fixes address errors in 20040709-2.c, 20040709-3.c, pr57875.c and pr79737-1.c at several optimization levels, with no regressions. gcc/ChangeLog: PR rtl-optimization/125728 * rtlhooks.cc (gen_lowpart_if_possible): Return NULL_RTX for a paradoxical lowpart of a MEM instead of widening the access. gcc/testsuite/ChangeLog: PR rtl-optimization/125728 * gcc.target/m68k/pr125728.c: New test. Signed-off-by: Andrew DeRosier <andy@skychase.zone> --- gcc/rtlhooks.cc | 3 +++ gcc/testsuite/gcc.target/m68k/pr125728.c | 23 +++++++++++++++++++++++ 2 files changed, 26 insertions(+) create mode 100644 gcc/testsuite/gcc.target/m68k/pr125728.c