[Piglit] [PATCH] glsl-1.30: Reproduce a bug in the i965 backend optimizer

[Ian Romanick]

On 09/12/2018 09:17 PM, Ilia Mirkin wrote:
> On Wed, Sep 12, 2018 at 7:59 PM, Ian Romanick <idr at freedesktop.org> wrote:
>> On 09/12/2018 04:33 PM, Ilia Mirkin wrote:
>>> On Wed, Sep 12, 2018 at 7:29 PM, Ian Romanick <idr at freedesktop.org> wrote:
>>>> From: Ian Romanick <ian.d.romanick at intel.com>
>>>>
>>>> The optimizer recently added the ability to replace a compare with a
>>>> subtraction under certain circumstances.  This can fail for integers.
>>>> For inputs a = 0x80000000, b = 4, int(0x80000000) < 4, but
>>>> int(0x80000000) - 4 overflows and results in 0x7ffffffc.  That's not
>>>> less than zero, so the flags get set differently than for (a < b).
>>>>
>>>> This really only affected the signed comparisons because the subtract
>>>> would always have a signed source types, so it wouldn't be seen as a
>>>> match for the compare with unsigned source types.
>>>>
>>>> Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
>>>> Cc: Matt Turner <mattst88 at gmail.com>
>>>> ---
>>>>  tests/spec/CMakeLists.txt                          |  1 +
>>>>  .../fs-absoluteDifference-int.shader_test          | 85 +++++++++++++++++++
>>>>  .../fs-absoluteDifference-uint.shader_test         | 85 +++++++++++++++++++
>>>>  .../vs-absoluteDifference-int.shader_test          | 96 ++++++++++++++++++++++
>>>>  .../vs-absoluteDifference-uint.shader_test         | 96 ++++++++++++++++++++++
>>>>  5 files changed, 363 insertions(+)
>>>>  create mode 100644 tests/spec/glsl-1.30/execution/fs-absoluteDifference-int.shader_test
>>>>  create mode 100644 tests/spec/glsl-1.30/execution/fs-absoluteDifference-uint.shader_test
>>>>  create mode 100644 tests/spec/glsl-1.30/execution/vs-absoluteDifference-int.shader_test
>>>>  create mode 100644 tests/spec/glsl-1.30/execution/vs-absoluteDifference-uint.shader_test
>>>>
>>>> diff --git a/tests/spec/CMakeLists.txt b/tests/spec/CMakeLists.txt
>>>> index 4df9d331d..28abf3634 100644
>>>> --- a/tests/spec/CMakeLists.txt
>>>> +++ b/tests/spec/CMakeLists.txt
>>>> @@ -2,6 +2,7 @@ add_subdirectory (amd_framebuffer_multisample_advanced)
>>>>  add_subdirectory (amd_depth_clamp_separate)
>>>>  add_subdirectory (amd_performance_monitor)
>>>>  add_subdirectory (amd_pinned_memory)
>>>> +add_subdirectory (amd_transform_feedback3_lines_triangles)
>>>>  add_subdirectory (arb_arrays_of_arrays)
>>>>  add_subdirectory (arb_base_instance)
>>>>  add_subdirectory (arb_bindless_texture)
>>>> diff --git a/tests/spec/glsl-1.30/execution/fs-absoluteDifference-int.shader_test b/tests/spec/glsl-1.30/execution/fs-absoluteDifference-int.shader_test
>>>> new file mode 100644
>>>> index 000000000..cdac53cdf
>>>> --- /dev/null
>>>> +++ b/tests/spec/glsl-1.30/execution/fs-absoluteDifference-int.shader_test
>>>> @@ -0,0 +1,85 @@
>>>> +[require]
>>>> +GL >= 3.0
>>>> +GLSL >= 1.30
>>>> +
>>>> +[vertex shader passthrough]
>>>> +
>>>> +[fragment shader]
>>>> +#extension GL_EXT_shader_integer_mix: enable
>>>> +
>>>> +// { A, B, absoluteDifference(A, B) }
>>>> +uniform ivec3 data[40];
>>>> +
>>>> +out vec4 color;
>>>> +
>>>> +uint abs_diff(int a, int b)
>>>> +{
>>>> +    /* This can fail if the compiler replaces the (a < b) with the result of
>>>> +     * one of the subtractions.  For inputs a = 0x80000000, b = 4,
>>>> +     * int(0x80000000) < 4, but int(0x80000000)-4 overflows and results in
>>>> +     * 0x7ffffffc.  That's not less than zero, so the flags get set
>>>> +     * differently than for (a < b).
>>>> +     */
>>>> +#ifndef GL_EXT_shader_integer_mix
>>>> +    return (a < b) ? uint(b - a) : uint(a - b);
>>>> +#else
>>>> +    return mix(uint(a - b), uint(b - a), a < b);
>>>> +#endif
>>>
>>> Why bother with this? It'll end up testing different code depending on
>>> the extensions that the driver supports, which seems like a really bad
>>> idea.
>>
>> We only hit the bug in the mix() case because how how we lower flow
>> control to bcsel.  I thought about having it require
>> GL_EXT_shader_integer_mix.  That means that when a driver enabled that
>> extension the test could transition SKIP->FAIL which seems more likely
>> to be ignored than PASS->FAIL.  I guess now there's a possible
>> FAIL->PASS transition.  Meh.
>>
>> I'm not married to it being either way.  I just want to reproduce the
>> bug that I added to the i965 driver a few months ago. :)
> 
> Hmm. Meh indeed. I dunno though -- skip -> fail when enabling an ext
> seems pretty bad to me. I'd just require EXT_shader_integer_mix. In
> practice, it's enabled everywhere GL 3.0 is (in mesa), so it's all a
> bit academic.

Fair enough.  The bikeshedder in me has difficulty deciding where the
tests should live.  It requires extensions beyond GLSL 1.30, so
spec/glsl-1.30 doesn't seem good.  It's not really a
GL_EXT_shader_integer_mix test, so spec/ext_shader_integer_mix doesn't
seem good either. :)

>   -ilia