[Bug target/113933] New: Switch pa to LRA

[Bug target/113933] New: Switch pa to LRA

[sjames at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113933

            Bug ID: 113933
           Summary: Switch pa to LRA
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: sjames at gcc dot gnu.org
                CC: danglin at gcc dot gnu.org, law at gcc dot gnu.org
            Blocks: 113932
  Target Milestone: ---

People are getting eager to require LRA. Please port the PA target to LRA (see
PR113932).


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113932
[Bug 113932] [meta-bug] Targets which should be ported to LRA

[Bug target/113933] Switch pa to LRA

[dave.anglin at bell dot net]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113933

--- Comment #1 from dave.anglin at bell dot net ---
On 2024-02-15 2:01 p.m., sjames at gcc dot gnu.org wrote:
> People are getting eager to require LRA. Please port the PA target to LRA (see
> PR113932).
Having tried this once, I know it's non trivial on PA.

[Bug target/113933] Switch pa to LRA

[segher at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113933

--- Comment #2 from Segher Boessenkool <segher at gcc dot gnu.org> ---
(In reply to dave.anglin from comment #1)
> On 2024-02-15 2:01 p.m., sjames at gcc dot gnu.org wrote:
> > People are getting eager to require LRA. Please port the PA target to LRA (see
> > PR113932).
> Having tried this once, I know it's non trivial on PA.

Then please tell us what problems you are running into?

[Bug target/113933] Switch pa to LRA

[sjames at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113933

Sam James <sjames at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2024-09-19
     Ever confirmed|0                           |1

[Bug target/113933] Switch pa to LRA

[danglin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113933

--- Comment #3 from John David Anglin <danglin at gcc dot gnu.org> ---
This code in pa_emit_move_sequence is reload specific (reg_equiv_mem):

  if (scratch_reg
      && reload_in_progress
      && GET_CODE (operand0) == REG
      && REGNO (operand0) >= FIRST_PSEUDO_REGISTER)
    operand0 = reg_equiv_mem (REGNO (operand0));
  else if (scratch_reg
           && reload_in_progress
           && GET_CODE (operand0) == SUBREG
           && GET_CODE (SUBREG_REG (operand0)) == REG
           && REGNO (SUBREG_REG (operand0)) >= FIRST_PSEUDO_REGISTER)
    {
     /* We must not alter SUBREG_BYTE (operand0) since that would confuse
        the code which tracks sets/uses for delete_output_reload.  */
      rtx temp = gen_rtx_SUBREG (GET_MODE (operand0),
                                 reg_equiv_mem (REGNO (SUBREG_REG (operand0))),
                                 SUBREG_BYTE (operand0));
      operand0 = alter_subreg (&temp, true);
    }

We need something similar for lra.  Otherwise, we hit following error:

dave@mx3210:~/gnu/gcc/objdir/hppa-linux-gnu/libgcc$ sh xx.sh
GNU C17 (GCC) version 15.0.0 20240918 (experimental) [master
r15-3695-gcc62b2c3da1] (hppa-linux-gnu)
        compiled by GNU C version 13.3.0, GMP version 6.3.0, MPFR version
4.2.1, MPC version 1.3.1, isl version isl-0.18-GMP

GGC heuristics: --param ggc-min-expand=100 --param ggc-min-heapsize=131072
Compiler executable checksum: 07d108a61db93a2471fe8e13f11689df
during RTL pass: reload
../../../gcc/libgcc/libgcc2.c: In function ‘__mulvdi3’:
../../../gcc/libgcc/libgcc2.c:376:1: internal compiler error: maximum number of
generated reload insns per insn achieved (90)
  376 | }
      | ^
0x335fd73 internal_error(char const*, ...)
        ../../gcc/gcc/diagnostic-global-context.cc:517
0x12b9e43 lra_constraints(bool)
        ../../gcc/gcc/lra-constraints.cc:5404
0x1291b63 lra(_IO_FILE*, int)
        ../../gcc/gcc/lra.cc:2445
0x11f4d43 do_reload
        ../../gcc/gcc/ira.cc:5976
0x11f5663 execute
        ../../gcc/gcc/ira.cc:6164

[Bug target/113933] Switch pa to LRA

[danglin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113933

--- Comment #4 from John David Anglin <danglin at gcc dot gnu.org> ---
Created attachment 59163
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=59163&action=edit
Initial patch to enable lra

[Bug target/113933] Switch pa to LRA

[danglin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113933

--- Comment #5 from John David Anglin <danglin at gcc dot gnu.org> ---
Created attachment 59164
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=59164&action=edit
Preproccessed source

We have lra crash with this source.

/home/dave/gnu/gcc/objdir/./gcc/cc1 -fpreprocessed _mulvdi3.i -quiet -dumpbase
_mulvdi3.c -dumpbase-ext .c -g -g -g -O2 -O2 -O2 -Wextra -Wall -Wno-narrowing
-Wwrite-strings -Wcast-qual -Wformat=0 -Wstrict-prototypes -Wmissing-prototypes
-Wold-style-definition -version -fno-checking -fbuilding-libgcc
-fno-stack-protector -fPIC -fvisibility=hidden -o _mulvdi3.s

[Bug target/113933] Switch pa to LRA

[danglin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113933

--- Comment #6 from John David Anglin <danglin at gcc dot gnu.org> ---
lra crashes on this insn:

(insn 524 38 523 5 (set (reg:DI 106 [ _16 ])
        (reg:DI 321 [orig:106 _16 ] [106]))
"../../../gcc/libgcc/libgcc2.c":276:4 81 {*pa.md:4260}
     (nil))

[Bug target/113933] Switch pa to LRA

[danglin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113933

John David Anglin <danglin at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #59163|0                           |1
        is obsolete|                            |

--- Comment #7 from John David Anglin <danglin at gcc dot gnu.org> ---
Created attachment 59178
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=59178&action=edit
Patch

This patch gets us to bootstrap on hppa-unknown-linux-gnu and
hppa64-hp-hpux11.11.

For the most part, I fixed the LRA crashes by adjusting pa_secondary_reload
to prevent pa_emit_move_sequence from being called in situations where it
just emitted an unmodified set.  Without reg_equiv_mem, the operands were
not being adjusted and the reload failed.

I was not able to get SECONDARY_MEMORY_NEEDED to work properly on hppa64
target with lra.  It never did copy via memory between general and
floating-point registers.  We ended up with an unrecognized insn if I allowed
reloading
pseudos during LRA.  If I didn't allow them, we got a secondary reload
overflow.  So, I modified 64-bit SImode, SFmode, DImode and DFmode to do
copy using the SP-40 slot in the frame marker.

We have a bunch of regressions.  For example,
FAIL: gcc.dg/enum-alias-3.c (test for excess errors)
FAIL: gcc.dg/non-local-goto-1.c execution test
FAIL: gcc.dg/non-local-goto-2.c execution test
FAIL: gcc.dg/pr49994-1.c (internal compiler error: in add_stores, at
var-tracking.cc:5966)
FAIL: gcc.dg/pr49994-1.c (test for excess errors)

Don't know if the changes to pa_secondary_reload will work with legacy reload.

[Bug target/113933] Switch pa to LRA

[danglin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113933

--- Comment #8 from John David Anglin <danglin at gcc dot gnu.org> ---
Looks like there is a typo in the indirect_goto constraints and that
causes the insn which loads "lab" to be deleted.