[Bug target/113796] New: [14] RISC-V rv64gcv vector: Runtime mismatch at -O2

[Bug target/113796] New: [14] RISC-V rv64gcv vector: Runtime mismatch at -O2

[patrick at rivosinc dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113796

            Bug ID: 113796
           Summary: [14] RISC-V rv64gcv vector: Runtime mismatch at -O2
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: patrick at rivosinc dot com
  Target Milestone: ---

Testcase:
signed char a[] = {0x80, 0};
int b;
signed char c;

int main() {
  for (; b < 1; b += 1)
      c = a[b] < 0 ?: a[b] >> 6;

  if (c == 1)
    return 0;
  else
    return 1;
}

Commands:
> /scratch/tc-testing/tc-feb-6/build-rv64gcv/bin/riscv64-unknown-linux-gnu-gcc -march=rv64gcv -O2 red.c -o user-config.out
> QEMU_CPU=rv64,vlen=128,v=true,vext_spec=v1.0,Zve32f=true,Zve64f=true /scratch/tc-testing/tc-feb-6/build-rv64gcv/bin/qemu-riscv64 user-config.out
> echo $?
1

> /scratch/tc-testing/tc-feb-6/build-rv64gcv/bin/riscv64-unknown-linux-gnu-gcc -march=rv64gcv -O1 red.c -o user-config.out
> QEMU_CPU=rv64,vlen=128,v=true,vext_spec=v1.0,Zve32f=true,Zve64f=true /scratch/tc-testing/tc-feb-6/build-rv64gcv/bin/qemu-riscv64 user-config.out
> echo $?
0

Godbolt:
https://godbolt.org/z/c7xrqx7Kz

Tested using r14-8834-gdb5c3f6d952 (not bisected).

Found using fuzzer.

[Bug tree-optimization/113796] [14 Regression] RISC-V rv64gcv vector: Runtime mismatch at -O2

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113796

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=110252
   Target Milestone|---                         |14.0
            Summary|[14] RISC-V rv64gcv vector: |[14 Regression] RISC-V
                   |Runtime mismatch at -O2     |rv64gcv vector: Runtime
                   |                            |mismatch at -O2
                 CC|                            |pinskia at gcc dot gnu.org
          Component|target                      |tree-optimization
   Last reconfirmed|                            |2024-02-07

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed, the issue is in ifcvt.

Before ifcvt we had:
```
  if (_1 >= 0)
    goto <bb 5>; [59.00%]
  else
    goto <bb 6>; [41.00%]

  <bb 5> [local count: 563821836]:
  iftmp.1_9 = _1 >> 6;

  <bb 6> [local count: 955630226]:
  # iftmp.1_5 = PHI <iftmp.1_9(5), 1(4)>
```
Which is fine.

Note the range for _1 is:
```
  # RANGE [irange] signed char [0, 1]
  iftmp.1_9 = _1 >> 6;
```

After ifcvt we get:
```
  iftmp.1_9 = _1 >> 6;
  _25 = _1 < 0;
  _4 = (signed charD.14) _25;
  # RANGE [irange] signed char [0, 1]
  iftmp.1_5 = _4 | iftmp.1_9;
```
Which is incorrect.

ifcvt needs to something similar to what phiopt did in r14-2650-g8c79b49cd4fa74
.

[Bug tree-optimization/113796] [14 Regression] RISC-V rv64gcv vector: Runtime mismatch at -O2

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113796

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|riscv                       |riscv aarch64

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
It can be reproduce on aarch64 with `-O2 -fno-vect-cost-model
-march=armv9-a+sve` too so it is at least a regression there too.

[Bug tree-optimization/113796] [14 Regression] RISC-V rv64gcv vector: Runtime mismatch at -O2

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113796

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Just an FYI :

[apinski@xeond2 upstream-full-cross]$ ./install/bin/aarch64-linux-gnu-gcc -O2
-static t65.c -march=armv9-a+sve2 -fno-vect-cost-model
[apinski@xeond2 upstream-full-cross]$ ./install-qemu/bin/qemu-aarch64 a.out
;echo $?
1
[apinski@xeond2 upstream-full-cross]$ ./install/bin/aarch64-linux-gnu-gcc -O2
-static t65.c -march=armv9-a+sve2
[apinski@xeond2 upstream-full-cross]$ ./install-qemu/bin/qemu-aarch64 a.out
;echo $?
0

[Bug tree-optimization/113796] [14 Regression] RISC-V rv64gcv vector: Runtime mismatch at -O2

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113796

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #1)
> ifcvt needs to something similar to what phiopt did in
> r14-2650-g8c79b49cd4fa74 .

Or ifcvt needs to remove the range info earlier ...
I have not looked into the ifcvt code to see which is the right approach.

But it was introduced by the same commit that introduced PR 110252 (which was
then fixed by the phiopt patch).

[Bug tree-optimization/113796] [14 Regression] ifcvt does not remove range info before folding: Runtime mismatch at -O2

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113796

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|riscv aarch64               |
            Summary|[14 Regression] RISC-V      |[14 Regression] ifcvt does
                   |rv64gcv vector: Runtime     |not remove range info
                   |mismatch at -O2             |before folding: Runtime
                   |                            |mismatch at -O2

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Here is one that fails on x86_64 too:
```
signed char a[] = {0x80, 0x80,0x80,0x80};
int b;
signed char c;

int main() {
  for (; b < sizeof(a); b += 1)
      c = a[b] < 0 ?: a[b] >> 6;

  if (c == 1)
    return 0;
  else
    return 1;
}
```

Compile with `-O2 -fno-vect-cost-model` to see the issue.

[Bug tree-optimization/113796] [14 Regression] ifcvt does not remove range info before folding: Runtime mismatch at -O2

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113796

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
             Status|NEW                         |ASSIGNED

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
Let me take this.

[Bug tree-optimization/113796] [14 Regression] ifcvt does not remove range info before folding: Runtime mismatch at -O2

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113796

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
We're removing flow-sensitive info in combine_blocks, but after inserting
and folding stmts comprising the PHI replacements.  There's possibly
latent issues when building up the predicates themselves since that uses
maybe_fold_or_comparisons without the workaround added for if-combine.

I have a patch resetting flow-sensitive info earlier (and also covering
PHIs).

[Bug tree-optimization/113796] [14 Regression] ifcvt does not remove range info before folding: Runtime mismatch at -O2

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113796

--- Comment #8 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:8636c538b68068cd2a4115fece531dc3e3e3a84a

commit r14-8869-g8636c538b68068cd2a4115fece531dc3e3e3a84a
Author: Richard Biener <rguenther@suse.de>
Date:   Wed Feb 7 13:08:43 2024 +0100

    tree-optimization/113796 - if-conversion and ranges

    The following makes sure to wipe range info before folding the
    COND_EXPRs we insert as part of replacing PHI nodes when combining
    blocks in the if-conversion pass.

            PR tree-optimization/113796
            * tree-if-conv.cc (combine_blocks): Wipe range-info before
            replacing PHIs and inserting predicates.

            * gcc.dg/torture/pr113796.c: New testcase.

[Bug tree-optimization/113796] [14 Regression] ifcvt does not remove range info before folding: Runtime mismatch at -O2

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113796

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |FIXED

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed (but possibly latent on branches of course).