unreachable!() hit when running strange code emitted by spirv-opt

[khyperia]

When talking with @jimblandy about naga's expression/statement representation, I wondered how naga handles SPIR-V control flow that does something like a multi-level break.

I remembered that spirv-opt emits a strange code pattern when inlining a function with a return statement nested inside control flow: it wraps the inlined function in a switch (0) { default: { the function's code } }, and any return statement of the inlined function branches to the end of the switch statement. SPIR-V allows skipping over merge blocks (pretty insane if you ask me, but that's how it is), kinda like a multi-level break, so this is valid.

So, I wrote up a GLSL example with a return statement in some control flow in an inlined function, threw it through glslang and spirv-opt, took the spir-v binary and pushed it through naga with my silly wgpu test app, and naga panicked.

thread 'main' panicked at 'internal error: entered unreachable code: Expression [15] is already cached!', C:\Users\khype\.cargo\registry\src\github.com-1ecc6299db9ec823\naga-0.4.2\src\back\spv\writer.rs:2380:9

I'm on Windows, if that's relevant. The GLSL and SPIR-V shader binary can be played with on shader-playground, although I've included both below for convenience. The SPIR-V binary passes spirv-val. (The shader-playground link does have a "download binary" button, though, so you don't have to go through the trouble of assembling it)

#version 460

layout (location = 1) in float thingy;
layout (location = 0) out float fragColor;

float f() {
    if (thingy > 1) {
        if (thingy > 2) {
            return 1;
        }
    }
    return 2;
}

void main()
{
	fragColor = f();
}

Disassembled SPIR-V, click to expand

; SPIR-V
; Version: 1.0
; Generator: Khronos Glslang Reference Front End; 10
; Bound: 56
; Schema: 0
               OpCapability Shader
          %1 = OpExtInstImport "GLSL.std.450"
               OpMemoryModel Logical GLSL450
               OpEntryPoint Fragment %main "main" %thingy %fragColor
               OpExecutionMode %main OriginUpperLeft
               OpSource GLSL 460
               OpName %main "main"
               OpName %thingy "thingy"
               OpName %fragColor "fragColor"
               OpDecorate %thingy Location 1
               OpDecorate %fragColor Location 0
       %void = OpTypeVoid
          %3 = OpTypeFunction %void
      %float = OpTypeFloat 32
%_ptr_Input_float = OpTypePointer Input %float
     %thingy = OpVariable %_ptr_Input_float Input
    %float_1 = OpConstant %float 1
       %bool = OpTypeBool
    %float_2 = OpConstant %float 2
%_ptr_Output_float = OpTypePointer Output %float
  %fragColor = OpVariable %_ptr_Output_float Output
       %uint = OpTypeInt 32 0
     %uint_0 = OpConstant %uint 0
       %main = OpFunction %void None %3
          %5 = OpLabel
               OpSelectionMerge %53 None
               OpSwitch %uint_0 %44
         %44 = OpLabel
         %45 = OpLoad %float %thingy
         %46 = OpFOrdGreaterThan %bool %45 %float_1
               OpSelectionMerge %52 None
               OpBranchConditional %46 %47 %52
         %47 = OpLabel
         %49 = OpFOrdGreaterThan %bool %45 %float_2
               OpSelectionMerge %51 None
               OpBranchConditional %49 %50 %51
         %50 = OpLabel
               OpBranch %53
         %51 = OpLabel
               OpBranch %52
         %52 = OpLabel
               OpBranch %53
         %53 = OpLabel
         %55 = OpPhi %float %float_1 %50 %float_2 %52
               OpStore %fragColor %55
               OpReturn
               OpFunctionEnd

[khyperia]

shoot, my bad! I was testing on wgpu's most recent release, but it looks like this was fixed in naga's master.