Bug 105271

bisected to:

40e9f2f13847ddd94e1216088aa00456d7b02d2b is the first bad commit
commit 40e9f2f13847ddd94e1216088aa00456d7b02d2b
Author: Timothy Arceri <timothy.arceri@collabora.com>
Date:   Tue Dec 13 11:37:25 2016 +1100
    i965: disable loop unrolling in GLSL IR
    
    There is a single regression in loop unrolling which is:
    
    loops HURT:   shaders/orbital_explorer.shader_test GS SIMD8:    0 -> 1
    
    However the loop is huge so it seems reasonable not to unroll it. It's
    surprising that GLSL IR does unroll it.
    
    shader-db results BDW:
    
    total instructions in shared programs: 13037455 -> 13036947 (-0.00%)
    instructions in affected programs: 17982 -> 17474 (-2.83%)
    helped: 63
    HURT: 25
    
    total cycles in shared programs: 262217870 -> 262227990 (0.00%)
    cycles in affected programs: 2287046 -> 2297166 (0.44%)
    helped: 969
    HURT: 844
    
    total loops in shared programs: 2951 -> 2952 (0.03%)
    loops in affected programs: 0 -> 1
    helped: 0
    HURT: 1
    
    LOST:   0
    GAINED: 1
    
    Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>

This looks like a buggy shader to me.

Inside one of the loops it contains:

((v687 / (((true ? (0 + ((v687) + 0)) : (- 63176))

Where for the first iteration of the loop v687 == 0 so we end up with:

0 / 0

My guess is GLSL IR just happened to optimise away this expression before constant folding tried to evaluate the constants expression.

Yes, the shader is intentionally buggy, to crash drivers that don't suitably check the input.  Mesa has taken a few patches to that have come out of these fuzz tests in the past.

Ok.

From the GLSL 4.60 spec Section 5.9 (Expressions):

   "Dividing by zero does not cause an exception but does result in an unspecified value."

So it seems to be an existing bug with constant evaluation in NIR. Will see if I can figure out a fix.

This should fix the crash:

https://patchwork.freedesktop.org/patch/207227/

V2 of fix:

https://patchwork.freedesktop.org/patch/207228/

(In reply to Timothy Arceri from comment #6)
> 
> https://patchwork.freedesktop.org/patch/207228/

For some reason, that's not showing up in my e-mail.  R-B none the less.

Should be fixed by:

commit 0c1f37cc2d8555223ade73b244a3ee374be8d9cd
Author: Timothy Arceri <tarceri@itsqueeze.com>
Date:   Wed Feb 28 14:33:55 2018 +1100

    nir: fix interger divide by zero crash during constant folding
    
    From the GLSL 4.60 spec Section 5.9 (Expressions):
    
       "Dividing by zero does not cause an exception but does result in
        an unspecified value."
    
    Fixes: 89285e4d47a6 "nir: add new constant folding infrastructure"
    
    Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
    Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105271

Does anyone have an opinion on the ease and relative utility of covering this case with a piglit test?

Getting a division by zero in constant propagation is trivial.  All we would have to do is write a simple shader_runner test.  Getting it to only trigger in NIR would be a bit trickier, but I think it could be done.  Now that NIR is the only thing doing loop unrolling, we could do something like this:

int j = 0;
for (int i = 0; i < 4; i++)
   j += 42/i;

gl_FragColor.x = 255.0f / float(i);

That should do the trick.

(In reply to Jason Ekstrand from comment #10)
> int j = 0;
> for (int i = 0; i < 4; i++)
>    j += 42/i;
> 
> gl_FragColor.x = 255.0f / float(i);

That should be float(j)

*** Bug 104703 has been marked as a duplicate of this bug. ***