[Bug target/110273] New: i686-w64-mingw32 with -march=znver4 generates AVX instructions without stack alignment

[Bug target/110273] New: i686-w64-mingw32 with -march=znver4 generates AVX instructions without stack alignment

[zfigura at codeweavers dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

            Bug ID: 110273
           Summary: i686-w64-mingw32 with -march=znver4 generates AVX
                    instructions without stack alignment
           Product: gcc
           Version: 13.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: zfigura at codeweavers dot com
                CC: amonakov at gcc dot gnu.org
  Target Milestone: ---

Created attachment 55334
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55334&action=edit
offending source file

Found with Wine.

I've attached a C file that's a minimal-ish reproducer.

I don't currently have a machine with gcc 13.1.0, so I did this using
godbolt.org's "MinGW gcc 13.1.0" target, with "-m32". Full compiler flags are
"-m32 -march=znver4 -O2".

The generated output begins

        pushl   %ebp
        vpxor   %xmm0, %xmm0, %xmm0
        movl    %esp, %ebp
        subl    $424, %esp
        vmovdqa %xmm0, 16(%esp)

which is broken. Using -march=znver3 instead will generate the same vmovdqa
instruction, but align the stack first:

        pushl   %ebp
        vpxor   %xmm0, %xmm0, %xmm0
        movl    %esp, %ebp
        andl    $-16, %esp
        subl    $416, %esp
        leal    36(%esp), %eax
        movl    $380, 8(%esp)
        movl    $0, 4(%esp)
        vmovdqa %xmm0, 16(%esp)

[Bug target/110273] i686-w64-mingw32 with -march=znver4 generates AVX instructions without stack alignment

[zfigura at codeweavers dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

--- Comment #1 from Zebediah Figura <zfigura at codeweavers dot com> ---
Created attachment 55335
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55335&action=edit
gcc -v output, from godbolt

[Bug target/110273] i686-w64-mingw32 with -march=znver4 generates AVX instructions without stack alignment

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
The difference is the inlining of memcpy if I used -minline-all-stringops, the
realignment comes back.

[Bug target/110273] i686-w64-mingw32 with -march=znver4 generates AVX instructions without stack alignment

[amonakov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

--- Comment #3 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
Seems to work fine with explicit '-mincoming-stack-boundary=2' on the command
line, even though it should make no difference for the 32-bit MinGW target.

[Bug target/110273] i686-w64-mingw32 with -march=znver4 generates AVX instructions without stack alignment

[amonakov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

--- Comment #4 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
Further reduced:

void f()
{
    int c[4] = { 0, 0, 0, 0 };
    int cc[8] = { 0 };
    asm("" :: "m"(c), "m"(cc));
}

Also reproducible with -march=skylake-avx512 or even plain -mavx512f,
retitling.

[Bug target/110273] [12/13/14 Regression] i686-w64-mingw32 with -mavx512f generates AVX instructions without stack alignment

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

--- Comment #5 from H.J. Lu <hjl.tools at gmail dot com> ---
GCC doesn't align stack for Windows.  As a workaround, one can pass
-muse-unaligned-vector-move to newer assembler.

[Bug target/110273] [12/13/14 Regression] i686-w64-mingw32 with -mavx512f generates AVX instructions without stack alignment

[amonakov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

--- Comment #6 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
Huh? Just compile the supplied testcases without avx512, you'll see proper
stack realignment.

[Bug target/110273] [12/13/14 Regression] i686-w64-mingw32 with -mavx512f generates AVX instructions without stack alignment

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |12.4

[Bug target/110273] [12/13/14 Regression] i686-w64-mingw32 with -mavx512f generates AVX instructions without stack alignment

[sjames at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

--- Comment #7 from Sam James <sjames at gcc dot gnu.org> ---
We keep getting quite a few reports of this downstream.

[Bug target/110273] [12/13/14 Regression] i686-w64-mingw32 with -mavx512f generates AVX instructions without stack alignment

[amonakov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

--- Comment #8 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
(In reply to Sam James from comment #7)
> We keep getting quite a few reports of this downstream.

Of this mingw32 stack realignment issue specifically, i.e. Wine breakage when
AVX512 is enabled via CFLAGS?

[Bug target/110273] [12/13/14 Regression] i686-w64-mingw32 with -mavx512f generates AVX instructions without stack alignment

[sjames at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

--- Comment #9 from Sam James <sjames at gcc dot gnu.org> ---
Yes - primarily from znver4 users who build with -march=native (or
-march=znver4).

[Bug target/110273] [12/13/14 Regression] i686-w64-mingw32 with -mavx512f generates AVX instructions without stack alignment

[hanno@schwalm-bremen.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

Jens-Hanno Schwalm <hanno@schwalm-bremen.de> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hanno@schwalm-bremen.de

--- Comment #10 from Jens-Hanno Schwalm <hanno@schwalm-bremen.de> ---
Hi, i think we found a very-similar issue in darktable code, you might look at

https://github.com/darktable-org/darktable/pull/15742

In short
- reported with gcc 13.2.1, -march=znver4 or using -fvect-cost-model=dynamic

The code in question was
```
typedef struct dt_draw_curve_t
{
  CurveData c;
  CurveSample csample;
} dt_draw_curve_t;

static inline void dt_draw_curve_set_point(dt_draw_curve_t *c,
                                           const int num,
                                           const float x,
                                           const float y)
{
  c->c.m_anchors[num].x = x;
  c->c.m_anchors[num].y = y;
}


static void _generate_curve_lut( ...)
{
  ...
  dt_iop_rgbcurve_node_t curve_nodes[3][20];
  ...
     for(int k = 0; k < d->params.curve_num_nodes[ch]; k++)
      dt_draw_curve_set_point(d->curve[ch], k,
            curve_nodes[ch][k].x,
            curve_nodes[ch][k].y);
 ...
}
```

The relevant point here was:
even if the curve_nodes are not aligned64 the compiled code looked like

lea (%rcx, %rbx,1), %rdi
vmoaps -0x220(%rbp, %rdi,8), %ymm2
vmoups %ymm2, 0x18(%r15,%rcx,8)

forcing an alignment64 on curve_nodes fixed the problem

reporting this as one of darktable devs - no compiler expert at all - and never
reported here before. Yet this seems a bug and possibly the same are originally
reported here.

[Bug target/110273] [12/13/14 Regression] i686-w64-mingw32 with -mavx512f generates AVX instructions without stack alignment

[sjames at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

--- Comment #11 from Sam James <sjames at gcc dot gnu.org> ---
(In reply to Jens-Hanno Schwalm from comment #10)
> Hi, i think we found a very-similar issue in darktable code, you might look
> at
> 
> https://github.com/darktable-org/darktable/pull/15742
> 

If you're hitting this on another target than i686-w64-mingw32, please file a
new bug. We can always mark it as a dupe if it turns out to be, although I
suspect it isn't here.

[Bug target/110273] [12/13/14 Regression] i686-w64-mingw32 with -mavx512f generates AVX instructions without stack alignment

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

--- Comment #12 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Sam James from comment #11)
> (In reply to Jens-Hanno Schwalm from comment #10)
> > Hi, i think we found a very-similar issue in darktable code, you might look
> > at
> > 
> > https://github.com/darktable-org/darktable/pull/15742
> > 
> 
> If you're hitting this on another target than i686-w64-mingw32, please file
> a new bug. We can always mark it as a dupe if it turns out to be, although I
> suspect it isn't here.

Though it might be PR 110027 ...

[Bug target/110273] [12/13/14 Regression] i686-w64-mingw32 with -mavx512f generates AVX instructions without stack alignment

[zfigura at codeweavers dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

--- Comment #13 from Zeb Figura <zfigura at codeweavers dot com> ---
(In reply to Sam James from comment #11)
> (In reply to Jens-Hanno Schwalm from comment #10)
> > Hi, i think we found a very-similar issue in darktable code, you might look
> > at
> > 
> > https://github.com/darktable-org/darktable/pull/15742
> > 
> 
> If you're hitting this on another target than i686-w64-mingw32, please file
> a new bug. We can always mark it as a dupe if it turns out to be, although I
> suspect it isn't here.

FWIW, I think the relevant part of i686-w64-ming32 is actually just
STACK_REALIGN_DEFAULT. I can reproduce the same lack of alignment with
"-mstackrealign -mavx512 -O2" with i386-linux-gnu, whereas "-mstackrealign
-mavx2 -O2" does align the stack. [-O2 is necessary here otherwise gcc will
just use vmovdqu and not bother aligning the stack. No idea what the more
targeted optimization is.]

I'd assume this is a bug with -mstackrealign, but I also don't understand what
the intended purpose of -mstackrealign is. How does it differ from
-mincoming-stack-boundary=2 or -mpreferred-stack-boundary=2?

[Bug target/110273] [12/13/14 Regression] i686-w64-mingw32 with -mavx512f generates AVX instructions without stack alignment

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

--- Comment #14 from H.J. Lu <hjl.tools at gmail dot com> ---
(In reply to Zeb Figura from comment #13)
> (In reply to Sam James from comment #11)
> > (In reply to Jens-Hanno Schwalm from comment #10)
> > > Hi, i think we found a very-similar issue in darktable code, you might look
> > > at
> > > 
> > > https://github.com/darktable-org/darktable/pull/15742
> > > 
> > 
> > If you're hitting this on another target than i686-w64-mingw32, please file
> > a new bug. We can always mark it as a dupe if it turns out to be, although I
> > suspect it isn't here.
> 
> FWIW, I think the relevant part of i686-w64-ming32 is actually just
> STACK_REALIGN_DEFAULT. I can reproduce the same lack of alignment with
> "-mstackrealign -mavx512 -O2" with i386-linux-gnu, whereas "-mstackrealign
> -mavx2 -O2" does align the stack. [-O2 is necessary here otherwise gcc will
> just use vmovdqu and not bother aligning the stack. No idea what the more
> targeted optimization is.]
> 

For the attached testcase here, GCC 13.2 generates:

[hjl@gnu-cfl-3 tmp]$ gcc -S -mstackrealign -mavx512f -O2 -m32 x.c
[hjl@gnu-cfl-3 tmp]$ head -20 x.s
        .file   "x.c"
        .text
        .p2align 4
        .globl  ddraw7_GetCaps
        .type   ddraw7_GetCaps, @function
ddraw7_GetCaps:
.LFB0:
        .cfi_startproc
        leal    4(%esp), %ecx
        .cfi_def_cfa 1, 0
        andl    $-16, %esp  <<<<< Stack realignment.
        vpxor   %xmm0, %xmm0, %xmm0
        xorl    %eax, %eax
        pushl   -4(%ecx)
        pushl   %ebp
        movl    %esp, %ebp
        .cfi_escape 0x10,0x5,0x2,0x75,0
        pushl   %edi
        pushl   %ecx
        .cfi_escape 0xf,0x3,0x75,0x78,0x6
[hjl@gnu-cfl-3 tmp]$ 

It works for me.  Please file a separate bug if-mstackrealign -mavx512f -O2
doesn't work for you on Linux.

[Bug target/110273] [12/13/14 Regression] i686-w64-mingw32 with -mavx512f generates AVX instructions without stack alignment

[law at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110273

Jeffrey A. Law <law at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P2
                 CC|                            |law at gcc dot gnu.org