Compilation failure in DXC 1.10.2605.2 about invalid width `i26` Integers

[dj2]

Description
I have a shader which compiles with DXC 1.9.2602 but fails with DXC 1.10.2605.2 (as tested on Godbot). The failures in 1.10.2605.2 are:

error: validation errors
error: Int type 'i26' has an invalid width.
error: Int type 'i26' has an invalid width.
error: Int type 'i26' has an invalid width.
error: Int type 'i26' has an invalid width.
error: Int type 'i26' has an invalid width.
error: Int type 'i26' has an invalid width.
error: Int type 'i26' has an invalid width.
error: Int type 'i26' has an invalid width.
error: Int type 'i26' has an invalid width.

Which, I don't create an 26-bit integers, so I don't know where the i26 comes from.

Steps to Reproduce

Compile the below with DXC with the flags -T ps_6_6 -E fs.

    struct fs_outputs {
      float4 tint_symbol : SV_Target0;
    };
    
    
    Texture2D<float4> tint_resource_table_array[] : register(t51, space43);
    Texture2D tint_resource_table_array_1[] : register(t51, space45);
    SamplerState tint_resource_table_array_2[] : register(s51, space46);
    ByteAddressBuffer tint_resource_table_metadata : register(t52, space42);
    float4 fs_inner() {
      bool v = false;
      if ((2u < tint_resource_table_metadata.Load(0u))) {
        uint3 v_1 = uint3((tint_resource_table_metadata.Load(12u)).xxx);
        v = any((v_1 == uint3(6u, 7u, 34u)));
      } else {
        v = false;
      }
      bool v_2 = v;
      uint v_3 = 0u;
      if (v_2) {
        v_3 = tint_resource_table_metadata.Load(12u);
      } else {
        v_3 = 6u;
      }
      uint texture_kind = v_3;
      uint v_4 = 0u;
      if (v_2) {
        v_4 = 2u;
      } else {
        v_4 = (0u + tint_resource_table_metadata.Load(0u));
      }
      uint v_5 = v_4;
      bool v_6 = false;
      if ((3u < tint_resource_table_metadata.Load(0u))) {
        uint2 v_7 = uint2((tint_resource_table_metadata.Load(16u)).xx);
        v_6 = any((v_7 == uint2(40u, 41u)));
      } else {
        v_6 = false;
      }
      bool v_8 = v_6;
      uint v_9 = 0u;
      if (v_8) {
        v_9 = tint_resource_table_metadata.Load(16u);
      } else {
        v_9 = 41u;
      }
      uint sampler_kind = v_9;
      uint v_10 = 0u;
      if (v_8) {
        v_10 = 3u;
      } else {
        v_10 = (4u + tint_resource_table_metadata.Load(0u));
      }
      uint v_11 = v_10;
      bool v_12 = false;
      if ((sampler_kind == 40u)) {
        bool v_13 = false;
        switch(texture_kind) {
          default:
          {
            v_13 = false;
            break;
          }
          case 1u:
          case 6u:
          case 11u:
          case 16u:
          case 21u:
          case 26u:
          {
            v_13 = true;
            break;
          }
        }
        v_12 = v_13;
      } else {
        v_12 = true;
      }
      float4 v_14 = (0.0f).xxxx;
      if (v_12) {
        v_14 = tint_resource_table_array[v_5].Sample(tint_resource_table_array_2[v_11], (0.0f).xx);
      } else {
        uint v_15 = (4u + tint_resource_table_metadata.Load(0u));
        v_14 = tint_resource_table_array[v_5].Sample(tint_resource_table_array_2[v_15], (0.0f).xx);
      }
      return v_14;
    }
    
    fs_outputs fs() {
      fs_outputs v_16 = {fs_inner()};
      return v_16;
    }

Environment

• DXC version
  • DXC 1.10.2605.2 fails
  • DXC 1.9.2602 compiles
• Host Operating System
  • Godbot
  • Linux

[dj2]

• Bad: https://godbolt.org/z/Tc699E1aj
• Good: https://godbolt.org/z/rMMnG4vYb

[dj2]

Looks like this broke when the validation for integer sizes was enabled: 90ae8d8. It looks like the reason is the compiler is changing the size of the integer used in the switch:

switch(texture_kind) {
    default: {
      v_13 = false;
      break;
    }
    case 1u:
    case 6u:
    case 11u:
    case 16u:
    case 21u:
    case 26u: {
       v_13 = true;
       break;
    }
}

switch.lookup:
  %switch.cast = trunc i32 %v_3.i.02 to i26, !dbg !145
  %switch.downshift = lshr i26 -32472031, %switch.cast, !dbg !145

[dj2]

@damyanp

[damyanp]

@dj2 - sorry for the delay, much of the team was travelling last week.

The "break" here is that the validator is detecting that invalid DXIL was being generated. So, we can work to fix the compiler to not generate invalid DXIL, but I'm interested to know if you've ever seen any runtime issues at all from presenting this shader to various GPUs?

[Cameron187]

I have a website with a WebGPU WGSL shader that was working without runtime issues in Chromium browsers on Windows until Chromium M148 picked up DXC 1.10.2605.2.
Then it started producing the error "GPUPipelineError: DXC compile failed with error: E_FAIL.".
The path in my case is WGSL > Tint > HLSL > DXC > DXIL inside Chromium's D3D12 backend.
The validation error in my case was "error: Int type 'i10' has an invalid width." (a different intermediate width).
I have a shader dump with the validation error that I can provide if that is helpful.

[damyanp]

If it involves a switch statement then it's probably a dupe of this, otherwise please open a new issue. Thank you!

[Cameron187]

Yeah, big switch statement in my fragment shader. thanks.