Optimization workflow for SPIR-V. Why is spirv-opt only used if you go through GLSLang->SPIR-V?

[mbechard]

Hey,
Looking at the source, it seems like spirv-opt is only used on generated SPIR-V if we use the Slang->glslang->SPIR-V workflow. The optimization passes that are applied are explicitly chosen in Slang though:

slang/source/slang-glslang/slang-glslang.cpp

Line 253 in d90d869

static void glslang_optimizeSPIRV(

rather than using optimizations that GLSLang could already can apply internally https://github.com/KhronosGroup/glslang/blob/f7c910864c67fc8a347184769b6c7bd6ae4c32ad/SPIRV/SpvTools.cpp#L184

Is there a reason why those optimization passes aren't be applied to SPIR-V output directly from Slang code, rather than going through GLSL/GLSLang? It would certainly be faster to compile than introducing a Slang->GLSL stage.

Thanks

[juliusikkala]

Those optimizations are being used in the direct-to-SPIR-V approach too. See:

slang/source/slang/slang-emit.cpp

Line 2782 in 768ae51

SlangResult emitSPIRVForEntryPointsDirectly(

calling

slang/source/slang/slang-emit.cpp

Line 2621 in 768ae51

static SlangResult createArtifactFromIR(

which runs the optimizations using the downstream compiler.

It's just a bit confusing that the same slang-glslang downstream compiler is used both for compiling GLSL->SPIR-V but also for optimizing SPIR-V->SPIR-V.

[mbechard]

Ah ok, thanks for the reply. Yes, I didn't think the glslang portion of the tool would be hit in the Slang->SPIR-V workflow.