[gcc(refs/users/aoliva/heads/testme)] [lra] don't spill when disabling fp2sp elimination [PR120424]

[Alexandre Oliva via Gcc-cvs]

https://gcc.gnu.org/g:271a0746cebb60d6ee519a271f28f1fbd2692b4f

commit 271a0746cebb60d6ee519a271f28f1fbd2692b4f
Author: Alexandre Oliva <ol...@adacore.com>
Date:   Sun Jun 22 15:06:53 2025 -0300

    [lra] don't spill when disabling fp2sp elimination [PR120424]
    
    We still have problems with the logic of disabling the fp2sp
    elimination.  An x86_64-linux-gnu bootstrap with
    ix86_frame_pointer_required tweaked to require a frame pointer when
    the frame size is positive miscompiles libiberty/regex.c in stage2,
    and arm-linux-gnueabihf with {BOOT_CFLAGS,TFLAGS}='-O2 -g
    -fnon-call-exceptions -fstack-clash-protection' crashes in gen-fac
    within gmp.
    
    With this simplification, the latter bootstraps successfully (in both
    arm and thumb modes), so that's progress, but spilling *should* work
    here, and libiberty/regex.c o x86_64 still gets miscompiled and causes
    gengtype -r gtype.state to malfunction in stage2, but I can't debug
    this any further, so problems remain.
    
    We have a reliable workaround for arm, namely, disabling the frame
    size-based optimization in arm_frame_pointer_required.
    
    
    for  gcc/ChangeLog
    
            PR rtl-optimization/120424
            * lra-eliminations.cc (lra_update_fp2sp_elimination): Skip
            spilling.

Diff:
---
 gcc/lra-eliminations.cc | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/gcc/lra-eliminations.cc b/gcc/lra-eliminations.cc
index 5713a9680523..931b579e2de8 100644
--- a/gcc/lra-eliminations.cc
+++ b/gcc/lra-eliminations.cc
@@ -1451,9 +1451,21 @@ lra_update_fp2sp_elimination (int *spilled_pseudos)
     /* If lives ranges changed, update the aggregate live ranges in
        slots as well before spilling any further pseudos.  */
     lra_recompute_slots_live_ranges ();
+#if 0
+  /* These spills will take place during update_reg_eliminate anyway,
+     but we probably get improved sharing of slots at this point.  The
+     problem is that it breaks arm-linux-gnueabihf bootstrap with
+     {BOOT_CFLAGS,TFLAGS}='-O2 -g -fnon-call-exceptions
+     -fstack-clash-protection', because gmp's gen-fac crashes in
+     stage2 (arm).  Spilling here *should* work, and some breakage
+     remains: x86_64-linux-gnu with ix86_frame_pointer_required
+     tweaked to return nonzero on positive frame sizes, to exercise
+     this case, miscompiles regex.c in stage2, so gengtype -r
+     gtype.state fails to match filenames.  */
   CLEAR_HARD_REG_SET (set);
   add_to_hard_reg_set (&set, Pmode, HARD_FRAME_POINTER_REGNUM);
   n = spill_pseudos (set, spilled_pseudos);
+#endif
   return n;
 }