[Mesa-dev] [PATCH 1/2] i965/vec4: Handle nir_instr_type_deref to silence run-time output

[Ian Romanick]

On 07/03/2018 07:09 AM, Jason Ekstrand wrote:
> Do we do this in fs?  I'm very confused as to how any deref instructions
> can survive this late.  They can for fs but only for images (not
> textures). I would expect them to be killed off by DCE long before we
> get here.

I didn't debug this to great depth.  Once I found a couple shaders that
hit this (apologies for sending the wrong shader to IRC), I ran with a
breakpoint set at the fprintf.  Once that was triggered, I looked at the
instruction type.  I saw that it was always nir_instr_type_deref, so I
looked at fs_visitor::nir_emit_instr to see how it was handled there.  I
then copied that code verbatim.

I think texture accesses in the vertex shader was a red herring.
Looking at the NIR coming out of one of the shaders, I see:


        decl_var  INTERP_MODE_NONE vec4[4] t7
        decl_reg vec4 32 r0[4]

        ...

        vec1 32 ssa_50 = deref_var &t7 (local vec4[4]) 
        vec1 32 ssa_51 = deref_array &(*ssa_50)[0] (local vec4) /* &t7[0] */
        vec2 32 ssa_52 = fmul ssa_5, ssa_49.zw
        vec2 32 ssa_53 = ffma ssa_47, ssa_49.zw, ssa_52
        vec4 32 ssa_54 = txl ssa_53 (coord), ssa_0 (lod), 
        r0[0] = imov ssa_54
        vec1 32 ssa_55 = deref_array &(*ssa_50)[1] (local vec4) /* &t7[1] */
        vec2 32 ssa_56 = fmul ssa_7, ssa_49.zw
        vec2 32 ssa_57 = ffma ssa_47, ssa_49.zw, ssa_56
        vec4 32 ssa_58 = txl ssa_57 (coord), ssa_0 (lod), 
        r0[1] = imov ssa_58
        vec1 32 ssa_59 = deref_array &(*ssa_50)[2] (local vec4) /* &t7[2] */
        vec2 32 ssa_60 = fmul ssa_8, ssa_49.zw
        vec2 32 ssa_61 = ffma ssa_47, ssa_49.zw, ssa_60
        vec4 32 ssa_62 = txl ssa_61 (coord), ssa_0 (lod), 
        r0[2] = imov ssa_62
        vec1 32 ssa_63 = deref_array &(*ssa_50)[3] (local vec4) /* &t7[3] */
        vec2 32 ssa_64 = fmul ssa_9, ssa_49.zw
        vec2 32 ssa_65 = ffma ssa_47, ssa_49.zw, ssa_64
        vec4 32 ssa_66 = txl ssa_65 (coord), ssa_0 (lod), 
        r0[3] = imov ssa_66

Maybe inadequate DCE is the problem?

> On July 3, 2018 00:26:32 "Ian Romanick" <idr at freedesktop.org> wrote:
> 
>> From: Ian Romanick <ian.d.romanick at intel.com>
>>
>> Previously, shader-db runs on Gen5 through Haswell would spew tons of
>> messages like:
>>
>>    VS instruction not yet implemented by NIR->vec4
>>
>> I checked several instance of this, and all of them were
>> nir_instr_type_deref instructions in vertex shaders that had texture
>> accesses (e.g., UnrealEngine4/EffectsCaveDemo/239.shader_test).
>>
>> Solution copied directly from fs_visitor::nir_emit_instr.
>>
>> Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
>> Fixes: 5a02ffb733e nir: Rework lower_locals_to_regs to use deref
>> instructions
>> Cc: Jason Ekstrand <jason at jlekstrand.net>
>> ---
>> src/intel/compiler/brw_vec4_nir.cpp | 4 ++++
>> 1 file changed, 4 insertions(+)
>>
>> diff --git a/src/intel/compiler/brw_vec4_nir.cpp
>> b/src/intel/compiler/brw_vec4_nir.cpp
>> index 7131fa06b4a..5f45d5b0c98 100644
>> --- a/src/intel/compiler/brw_vec4_nir.cpp
>> +++ b/src/intel/compiler/brw_vec4_nir.cpp
>> @@ -168,6 +168,10 @@ vec4_visitor::nir_emit_instr(nir_instr *instr)
>>       nir_emit_undef(nir_instr_as_ssa_undef(instr));
>>       break;
>>
>> +   case nir_instr_type_deref:
>> +      /* Derefs can exist for images but they do nothing */
>> +      break;
>> +
>>    default:
>>       fprintf(stderr, "VS instruction not yet implemented by
>> NIR->vec4\n");
>>       break;
>> -- 
>> 2.14.4