[Bug 56372] New: musl based exp2() gives very inaccurate results on i686

List overview All Threads

newer

older

[Bug 55883] New: SpeedWave can't...

[Bug 56781] New: srcrrun:...

WineHQ Bugzilla

26 Feb 2024 26 Feb '24

4:20 p.m.

https://bugs.winehq.org/show_bug.cgi?id=56372

Bug ID: 56372 Summary: musl based exp2() gives very inaccurate results on i686 Product: Wine Version: 8.6 Hardware: x86-64 OS: Linux Status: UNCONFIRMED Severity: normal Priority: P2 Component: msvcrt Assignee: wine-bugs@winehq.org Reporter: martin@martin.st CC: piotr@codeweavers.com Regression SHA1: ca0643875961ef62c76e0b9cbd0433002a948d9b Distribution: ---

Since switching the msvcrt implementation of exp2() from internal to musl based, in ca0643875961ef62c76e0b9cbd0433002a948d9b, the accuracy on i686 has become quite bad, to the point of breaking a handful of tests in ffmpeg.

The issue can be reproduced with the following test snippet:

--- #include <stdio.h> #include <math.h>

double input = 5.979471;

int main(int argc, char* argv[]) { printf("exp2(%f) = %f\n", input, exp2(input)); return 0; } ---

Prior to this commit, and with later versions on e.g. x86_64, the output is:

exp2(5.979471) = 63.095753

However, on i686, the output now is:

exp2(5.979471) = 62.968753

This is an error of around 0.2%, which is a couple orders of magnitudes larger than expected/tolerated. It's curious how this issue only appears on i686 though, i.e. it doesn't look like an issue with the musl exp2() implementation algorithm, but some component of it misbehaving on i686.

-- Do not reply to this email, post in Bugzilla using the above URL to reply. You are receiving this mail because: You are watching all bug changes.

Show replies by date

WineHQ Bugzilla

26 Feb 26 Feb

4:22 p.m.

New subject: [Bug 56372] musl based exp2() gives very inaccurate results on i686

https://bugs.winehq.org/show_bug.cgi?id=56372

Martin Storsjö martin@martin.st changed:

What |Removed |Added ---------------------------------------------------------------------------- Hardware|x86-64 |x86

-- Do not reply to this email, post in Bugzilla using the above URL to reply. You are receiving this mail because: You are watching all bug changes.

WineHQ Bugzilla

4:29 p.m.

New subject: [Bug 56372] musl based exp2() gives very inaccurate results on i686

https://bugs.winehq.org/show_bug.cgi?id=56372

--- Comment #1 from Martin Storsjö martin@martin.st --- For extra context; this reproduces with llvm-mingw https://github.com/mstorsjo/llvm-mingw/releases/tag/20231128. With older mingw-w64 releases, calls to exp2() go to mingw-w64-crt's internal implementation, but when targeting UCRT, with the latest few releases, the exp2() calls go to the UCRT exp2() function.

-- Do not reply to this email, post in Bugzilla using the above URL to reply. You are receiving this mail because: You are watching all bug changes.

WineHQ Bugzilla

27 Feb 27 Feb

10:01 a.m.

New subject: [Bug 56372] musl based exp2() gives very inaccurate results on i686

https://bugs.winehq.org/show_bug.cgi?id=56372

Martin Storsjö martin@martin.st changed:

What |Removed |Added ---------------------------------------------------------------------------- CC| |julliard@winehq.org

--- Comment #2 from Martin Storsjö martin@martin.st --- This is a bug with how Wine has integrated the musl sources.

Musl removes excess precision from float values in a number of places, this seems to be critical to get the intended precision, see https://gitlab.winehq.org/wine/wine/-/blob/wine-9.3/libs/musl/src/internal/l....

However, in order for this to work as expected, musl also tries to add specific compiler flags, in order to get the right behaviour. See https://git.musl-libc.org/cgit/musl/tree/configure?h=v1.2.4&id=f5f55d658... - which does this:

tryflag CFLAGS_C99FSE -fexcess-precision=standard \ || { test "$ARCH" = i386 && tryflag CFLAGS_C99FSE -ffloat-store ; }

So wine would need to check for these flags somewhere, and enable them for the musl sources.

-- Do not reply to this email, post in Bugzilla using the above URL to reply. You are receiving this mail because: You are watching all bug changes.

WineHQ Bugzilla

8:27 p.m.

New subject: [Bug 56372] musl based exp2() gives very inaccurate results on i686

https://bugs.winehq.org/show_bug.cgi?id=56372

--- Comment #3 from Martin Storsjö martin@martin.st --- FWIW, the issue reproduces if Wine is built with GCC (either in old full ELF mode, or in PE mode), but if built with Clang (tested in PE mode with llvm-mingw), the issue doesn't seem to appear.

-- Do not reply to this email, post in Bugzilla using the above URL to reply. You are receiving this mail because: You are watching all bug changes.

WineHQ Bugzilla

5 Mar 5 Mar

9:57 a.m.

New subject: [Bug 56372] musl based exp2() gives very inaccurate results on i686

https://bugs.winehq.org/show_bug.cgi?id=56372

blubban@gmail.com changed:

What |Removed |Added ---------------------------------------------------------------------------- CC| |blubban@gmail.com

--- Comment #4 from blubban@gmail.com --- Easier solution: Simply add a volatile or two to those functions. https://godbolt.org/z/sWb58Pb4Y

-- Do not reply to this email, post in Bugzilla using the above URL to reply. You are receiving this mail because: You are watching all bug changes.

WineHQ Bugzilla

13 Apr 13 Apr

5:18 a.m.

New subject: [Bug 56372] musl based exp2() gives very inaccurate results on i686

https://bugs.winehq.org/show_bug.cgi?id=56372

Alex Henrie alexhenrie24@gmail.com changed:

What |Removed |Added ---------------------------------------------------------------------------- CC| |alexhenrie24@gmail.com

--- Comment #5 from Alex Henrie alexhenrie24@gmail.com --- Created attachment 76313 --> https://bugs.winehq.org/attachment.cgi?id=76313 Test program

I'm attaching a compiled copy of Martin's test program in case anyone is interested in the bug but doesn't want to download llvm-mingw in order to reproduce it.

-- Do not reply to this email, post in Bugzilla using the above URL to reply. You are receiving this mail because: You are watching all bug changes.

WineHQ Bugzilla

11:22 p.m.

New subject: [Bug 56372] musl based exp2() gives very inaccurate results on i686

https://bugs.winehq.org/show_bug.cgi?id=56372

Ken Sharp imwellcushtymelike@gmail.com changed:

-- Do not reply to this email, post in Bugzilla using the above URL to reply. You are receiving this mail because: You are watching all bug changes.

WineHQ Bugzilla

15 Oct 15 Oct

11:51 a.m.

New subject: [Bug 56372] musl based exp2() gives very inaccurate results on i686

https://bugs.winehq.org/show_bug.cgi?id=56372

Piotr Caban piotr.caban@gmail.com changed:

--- Comment #6 from Piotr Caban piotr.caban@gmail.com --- Marking as fixed by 3607eae620f12cf42a62c201b0c78e9d8bcfffc3.

-- Do not reply to this email, post in Bugzilla using the above URL to reply. You are receiving this mail because: You are watching all bug changes.

WineHQ Bugzilla

18 Oct 18 Oct

8:17 p.m.

New subject: [Bug 56372] musl based exp2() gives very inaccurate results on i686

https://bugs.winehq.org/show_bug.cgi?id=56372

Alexandre Julliard julliard@winehq.org changed:

What |Removed |Added ---------------------------------------------------------------------------- Status|RESOLVED |CLOSED

--- Comment #7 from Alexandre Julliard julliard@winehq.org --- Closing bugs fixed in 9.20.

-- Do not reply to this email, post in Bugzilla using the above URL to reply. You are receiving this mail because: You are watching all bug changes.

WineHQ Bugzilla

2 Dec 2 Dec

8:58 p.m.

New subject: [Bug 56372] musl based exp2() gives very inaccurate results on i686

https://bugs.winehq.org/show_bug.cgi?id=56372

Michael Stefaniuc mstefani@winehq.org changed:

-- Do not reply to this email, post in Bugzilla using the above URL to reply. You are receiving this mail because: You are watching all bug changes.

WineHQ Bugzilla

16 Apr 16 Apr

6:58 p.m.

New subject: [Bug 56372] musl based exp2() gives very inaccurate results on i686

http://bugs.winehq.org/show_bug.cgi?id=56372

Michael Stefaniuc mstefani@winehq.org changed:

What |Removed |Added ---------------------------------------------------------------------------- Target Milestone|9.0.x |---

--- Comment #8 from Michael Stefaniuc mstefani@winehq.org --- Removing the 9.0.x milestone from bug fixes included in 9.0.1.

-- Do not reply to this email, post in Bugzilla using the above URL to reply. You are receiving this mail because: You are watching all bug changes.

Age (days ago)

427

Last active (days ago)

wine-bugs@winehq.org

11 comments

1 participants

tags (0)

participants (1)

WineHQ Bugzilla