[SPIR-V] Fix unsupported downcast in LegalizePointerCast pass

[Keenuts]

The following pattern is generated from some offload test suite:

; RUN: llc -O0 -mtriple=spirv1.6-unknown-vulkan1.3-compute -o - %s

target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-n8:16:32:64-G10"
target triple = "spirv1.6-unknown-vulkan1.3-compute"

@.str = private unnamed_addr constant [3 x i8] c"In\00", align 1

; Function Attrs: mustprogress nofree noinline norecurse nosync nounwind willreturn memory(readwrite, inaccessiblemem: none)
define void @main() local_unnamed_addr #0 {
entry:
  %4 = tail call target("spirv.VulkanBuffer", [0 x <4 x half>], 12, 0) @llvm.spv.resource.handlefrombinding.tspirv.VulkanBuffer_a0v4f16_12_0t(i32 0, i32 0, i32 1, i32 0, i1 false, ptr nonnull @.str)
  call void @llvm.spv.assign.type.tspirv.VulkanBuffer_a0v4f16_12_0t(target("spirv.VulkanBuffer", [0 x <4 x half>], 12, 0) %4, metadata target("spirv.VulkanBuffer", [0 x <4 x half>], 12, 0) poison)
  %8 = tail call noundef align 8 dereferenceable(8) ptr addrspace(11) @llvm.spv.resource.getpointer.p11.tspirv.VulkanBuffer_a0v4f16_12_0t(target("spirv.VulkanBuffer", [0 x <4 x half>], 12, 0) %4, i32 1)
  call void @llvm.spv.assign.ptr.type.p11(ptr addrspace(11) %8, metadata <4 x half> poison, i32 11)
  %9 = load <4 x half>, ptr addrspace(11) %8, align 8

  %15 = call ptr addrspace(11) @llvm.spv.ptrcast.p11.p11(ptr addrspace(11) %8, metadata i8 poison, i32 11)
  call void @llvm.spv.assign.ptr.type.p11(ptr addrspace(11) %8, metadata i8 poison, i32 11)

  %16 = call ptr addrspace(11) (i1, ptr addrspace(11), ...) @llvm.spv.gep.p11.p11(i1 true, ptr addrspace(11) %8, i64 2)
  call void @llvm.spv.assign.ptr.type.p11(ptr addrspace(11) %16, metadata i8 poison, i32 11)

  %17 = call ptr addrspace(11) @llvm.spv.ptrcast.p11.p11(ptr addrspace(11) %16, metadata half poison, i32 11)
  call void @llvm.spv.assign.ptr.type.p11(ptr addrspace(11) %17, metadata half poison, i32 11)

  %18 = load half, ptr addrspace(11) %17, align 2
  ret void
}

declare target("spirv.VulkanBuffer", [0 x <4 x half>], 12, 0) @llvm.spv.resource.handlefrombinding.tspirv.VulkanBuffer_a0v4f16_12_0t(i32 %0, i32 %1, i32 %2, i32 %3, i1 %4, ptr %5) #1
declare ptr addrspace(11) @llvm.spv.resource.getpointer.p11.tspirv.VulkanBuffer_a0v4f16_12_0t(target("spirv.VulkanBuffer", [0 x <4 x half>], 12, 0) %0, i32 %1) #1

declare void @llvm.spv.assign.type.tspirv.VulkanBuffer_a0v4f16_12_0t(target("spirv.VulkanBuffer", [0 x <4 x half>], 12, 0) %0, metadata %1) nounwind
declare void @llvm.spv.assign.ptr.type.p11(ptr addrspace(11) %0, metadata %1, i32 immarg %2) nounwind
declare ptr addrspace(11) @llvm.spv.ptrcast.p11.p11(ptr addrspace(11) %0, metadata %1, i32 immarg %2) nounwind
declare ptr addrspace(11) @llvm.spv.gep.p11.p11(i1 immarg %0, ptr addrspace(11) %1, ...) nounwind

attributes #0 = { mustprogress nofree noinline norecurse nosync nounwind willreturn memory(readwrite, inaccessiblemem: none) "approx-func-fp-math"="true" "frame-pointer"="all" "hlsl.numthreads"="1,1,1" "hlsl.shader"="compute" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" }
attributes #1 = { mustprogress nocallback nofree nosync nounwind willreturn memory(none) }

Removing the intermediate ptrcast solves the issue. For some reason, the assign.ptr.type intrinsics are generated on already refined types. In this sample, the first load helps the backend correctly determine the pointer to be a <4 x half>*. A subsequent load also determines the type to be a half*, and this reconciliation is correctly handled by the pass. But in the middle, a i8* type is added because of the GEP.
Maybe the type scavenger should be fixed so the GEP type is correctly found?

@s-perron

[Keenuts]

Similar issue can also be reproduces from this HLSL sample:

// RUN: %clang --driver-mode=dxc %s -T ps_6_8 -E main -O3 -spirv
RWBuffer<float4> buffer;

[shader("pixel")]
void main() {
  float4 a = buffer[0];
  float b = buffer[0].x;

  buffer[0].x = a.y + b;
}

[s-perron]

The structuredbuffer problem is different than the RWBuffset problem. The structured buffer problem is because of optimizations by llvm-opt. If there is a GEP on a pointer where the offset becomes constant, the optimizer will translate that into a GEP on an i8* type with the number of bytes for the offset. That is what is causing the pointer cast. This is complicated because the indices on the GEP need to be replaced. I don't think the type scavenger is the best place for that. I think doing that in legalize pointer cast is the rigth place. I'm about 30% done fixing that up.

The problem with the RWBuffer is an assumption that we can take the first using of the spv_resource_getpointer and assume the type of the use is the correct type for the pointer. Your example shows that is plainly wrong. We should update SPIRVEmitIntrinsics.cpp for that case first. Then try to figure out what to do in legalize-pointer-casts.