Re: [PATCH] ext-dce: Avoid multi-register paradoxical subregs for reload [PR122438]

On Fri, Jul 31, 2026 at 4:36 AM Karl Mehltretter <[email protected]> wrote:
>
> ext-dce can turn an extension into a paradoxical subreg when the
> upper bits are dead.  With classic reload, this can fail if the modes
> use different numbers of hard registers.  On m68k, SI-to-DI creates an
> invalid overlap and an ICE in postreload.
>
> Skip the rewrite in that case.  This change also fixes PR119953
> on MSP430.  LRA and same-register-count cases are unchanged.
>
> Cross-tested on GCC trunk:
>
> 1) m68k-linux-gnu: the unpatched compiler reproduces PR122438.  The
> gcc.target/m68k suite changes from 128 PASS / 9 FAIL / 1 unsupported
> to 129 PASS / 9 FAIL / 1 unsupported with the new test.
>
> 2) msp430-elf: the new PR119953 test passes.
>
> 3) alpha-linux-gnu and vax-netbsdelf: the tested same-register-count
> extension cases produce identical assembly with and without the patch.
>
> Also tested on GCC 16.1.1 for m68k-linux-gnu.  The unpatched compiler
> reproduces PR122438 while the new test passes with the patch.
>
> Assisted-by: Codex gpt-5.6-terra
>
> gcc/ChangeLog:
>
>         PR rtl-optimization/122438
>         PR target/119953
>         * ext-dce.cc: Include regs.h.
>         (ext_dce_try_optimize_extension): Avoid multi-register
>         paradoxical subregs with classic reload.
>
> gcc/testsuite/ChangeLog:
>
>         PR rtl-optimization/122438
>         PR target/119953
>         * gcc.target/m68k/pr122438.c: New test.
>         * g++.target/msp430/pr119953.C: New test.
>
> Signed-off-by: Karl Mehltretter <[email protected]>
> ---
> OK for trunk?

I am not sure if we want to add hacks still to support reload. It
would be better if those targets are changed over to using LRA.  There
was a plan to remove reload for GCC 16 but it seems like it slipped
until this year at the earliest. Though I have not seen any movement
towards removing it but that is still the plan of record.

Thanks,
Andrea

>
>  gcc/ext-dce.cc                             |  8 +++++
>  gcc/testsuite/g++.target/msp430/pr119953.C | 40 ++++++++++++++++++++++
>  gcc/testsuite/gcc.target/m68k/pr122438.c   | 23 +++++++++++++
>  3 files changed, 71 insertions(+)
>  create mode 100644 gcc/testsuite/g++.target/msp430/pr119953.C
>  create mode 100644 gcc/testsuite/gcc.target/m68k/pr122438.c
>
> diff --git a/gcc/ext-dce.cc b/gcc/ext-dce.cc
> index 9e4b0a429b..8e0c5ab9db 100644
> --- a/gcc/ext-dce.cc
> +++ b/gcc/ext-dce.cc
> @@ -26,6 +26,7 @@ along with GCC; see the file COPYING3.  If not see
>  #include "memmodel.h"
>  #include "insn-config.h"
>  #include "emit-rtl.h"
> +#include "regs.h"
>  #include "expr.h"
>  #include "recog.h"
>  #include "cfganal.h"
> @@ -533,6 +534,13 @@ ext_dce_try_optimize_extension (rtx_insn *insn, rtx set)
>        return;
>      }
>
> +  /* Avoid known classic-reload problems with multi-register paradoxical
> +     subregs (PR122438, PR119953).  */
> +  if (!targetm.lra_p ()
> +      && (hard_regno_nregs (0, GET_MODE (src))
> +         != hard_regno_nregs (0, GET_MODE (inner))))
> +    return;
> +
>    /* Avoid (subreg (mem)) and other constructs which may be valid RTL, but
>       not useful for this optimization.  */
>    if (!(REG_P (inner) || (SUBREG_P (inner) && REG_P (SUBREG_REG (inner)))))
> diff --git a/gcc/testsuite/g++.target/msp430/pr119953.C b/gcc/testsuite/g++.target/msp430/pr119953.C
> new file mode 100644
> index 0000000000..fd2562eff9
> --- /dev/null
> +++ b/gcc/testsuite/g++.target/msp430/pr119953.C
> @@ -0,0 +1,40 @@
> +/* PR target/119953 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2" } */
> +
> +bool equals_tmp;
> +int native ();
> +
> +struct path
> +{
> +  class iterator;
> +  iterator begin ();
> +  iterator end ();
> +} p, hash_path;
> +
> +struct path::iterator
> +{
> +  path operator* () { return p; }
> +  void operator++ () {}
> +  friend bool operator!= (iterator lhs, iterator rhs)
> +  {
> +    return lhs.equals (rhs);
> +  }
> +  bool equals (iterator rhs)
> +  {
> +    if (equals_tmp)
> +      return cur == rhs.cur;
> +    return at_end == rhs.at_end;
> +  }
> +  char cur;
> +  bool at_end;
> +};
> +
> +int
> +hash_value ()
> +{
> +  int seed;
> +  for (auto x : hash_path)
> +    seed ^= native () + 0x9e3779b9 + (seed << 6) + (seed >> 2);
> +  return seed;
> +}
> diff --git a/gcc/testsuite/gcc.target/m68k/pr122438.c b/gcc/testsuite/gcc.target/m68k/pr122438.c
> new file mode 100644
> index 0000000000..353bcc1b8f
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/m68k/pr122438.c
> @@ -0,0 +1,23 @@
> +/* PR rtl-optimization/122438 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -w" } */
> +
> +long describeCell_nLocal;
> +char describeCell_a;
> +char describeCell_zDesc[];
> +int strlen (char *);
> +void describeContent (char *, char *);
> +void localPayload (void);
> +
> +void
> +describeCell (char cType)
> +{
> +  long long nDesc;
> +  if (cType)
> +    nDesc += nDesc;
> +  localPayload ();
> +  if (describeCell_nLocal)
> +    nDesc += strlen (&describeCell_zDesc[nDesc]);
> +  if (cType)
> +    describeContent (&describeCell_a, &describeCell_zDesc[nDesc]);
> +}
> --
> 2.53.0