Fix issue #761 stores involving GEP of aggregates of pointers hits in…

[thewilsonator]

…valid BitCast assert

Handle pointer scalar elements in lib/ReplacePointerBitcastsPass.cpp:BuildFromElements so that we don't generate an invalid bitcast from float addrspace(1)* to i32 as the else branch for regular scalar conversions conversion generates bitcasts to integer types.

[alan-baker]

Your test doesn't indicate what the expected code is. Bitcasting from a pointer to an integer is not generally allowed in shader SPIR-V and that appears to be what is happening.

When I checkout this branch I still get assertions that the replace all uses with is being used incorrectly.

[thewilsonator]

Bitcasting from a pointer to an integer is not generally allowed in shader SPIR-V and that appears to be what is happening.

Correct, but this is not present in the source input. The bad cast is generated by the ReplacePointerBitcastsPass because it (currently) doesn't handle pointers in aggregates properly and generates pointer-to-integer bit casts for (aggregate of pointer)-to-pointer bit casts (or the other way around, I don't remember off the top of my head). It should be noted that it does handle pointer-to-pointer bit casts outside of aggregates properly.

Your test doesn't indicate what the expected code is.
When I checkout this branch I still get assertions that the replace all uses with is being used incorrectly.

The test is a simple saxpy kernel, but the intent of the test is to validate that the compiler does not hit assert generating the code for it. Now that CI is enabled I will look into this more, because "It works on my machine".

Thanks!

[thewilsonator]

"It works on my machine".

Seems I accidentally committed the 64-bit test case file, which won't work because compute SPIR-V is effectively 32 bit (the width of gl_InvocationID. Fixed the style and unused variable warnings.

[thewilsonator]

I've managed to work around this by not generating those bitcasts in the first place.