Bug 118691 - [15 Regression] gcc_r in SPECCPU 2017 miscompare on train dataset
Summary: [15 Regression] gcc_r in SPECCPU 2017 miscompare on train dataset
Status: RESOLVED FIXED
Alias: None
Product: gcc
Classification: Unclassified
Component: middle-end (show other bugs)
Version: 15.0
: P3 normal
Target Milestone: 15.0
Assignee: Victor Do Nascimento
URL:
Keywords: wrong-code
Depends on:
Blocks: spec
  Show dependency treegraph
 
Reported: 2025-01-29 12:35 UTC by Tamar Christina
Modified: 2025-02-18 07:56 UTC (History)
2 users (show)

See Also:
Host:
Target: aarch64*
Build:
Known to work:
Known to fail:
Last reconfirmed: 2025-02-12 00:00:00


Attachments

Note You need to log in before you can comment on or make changes to this bug.
Description Tamar Christina 2025-01-29 12:35:33 UTC
During the training part the run fails with:

200.c: In function 'htls_.clone.35':
200.c:26853:7: benchmark internal error: in ?, at fold-const.c:2677

during fdo1 with -mcpu=neoverse-v1 -Ofast -flto=auto

I'll bisect and reduce but this is a preliminary submission

This reappeared somewhere between 6e0b048fb5bc0809048ef8f487830ad26f4b87cf..9a4bb95a4e68b6f90a16f337b0b4cdb9af957ab1

But could have just been temporarily hidden.  Before that it was also failing between a7ae0c31245a7db7abf2e80d0016510afe9c8ad0..979ca3ba366da7177f427e049f67673ec3e35442
Comment 1 Richard Biener 2025-01-29 12:56:19 UTC
Please add -fno-strict-aliasing and try again.
Comment 2 Tamar Christina 2025-01-29 13:00:23 UTC
(In reply to Richard Biener from comment #1)
> Please add -fno-strict-aliasing and try again.

Already on.  Full options are:

-fprofile-generate -mcpu=neoverse-v1 -Ofast -fomit-frame-pointer -flto=auto -g3 -fno-strict-aliasing -fgnu89-inline -std=gnu17
Comment 3 Andrew Pinski 2025-01-29 13:55:08 UTC
Isn't this a dup of bug 115450 ?
Comment 4 Tamar Christina 2025-01-29 14:14:35 UTC
(In reply to Andrew Pinski from comment #3)
> Isn't this a dup of bug 115450 ?

Don't believe so. This is only showing up with PGO for me, but it's only during training, so I suspect -fprofile-generate is doing something as it doesn't show up on the job without it.
Comment 5 Tamar Christina 2025-01-29 17:59:21 UTC
Ok, I just realized why it fails for PGO,

that's because PGO uses the train dataset.

And indeed I can reproduce it now on `train` without LTO.

ref is still fine which is likely why the other CIs didn't see it.
Comment 6 Richard Biener 2025-02-03 14:53:21 UTC
Works for me on x86_64-linux with -Ofast -march=znver4
Comment 7 Tamar Christina 2025-02-03 15:02:19 UTC
(In reply to Richard Biener from comment #6)
> Works for me on x86_64-linux with -Ofast -march=znver4

Yeah still failing here. I'll track down the change in code this week. It's on my list for the week.
Comment 8 Tamar Christina 2025-02-18 07:56:19 UTC
Fixed by g:589d79e6268b055422a7b6c11cd0a8a4f2531a8c