[DirectX] Validation error: Access to out-of-bounds memory is disallowed

[farzonl]

Error Test case:

lldb -- dxv validation/GemmSmallGemm_Bias_float_b16c32.dat

Validation error in: hlsl::ValidateFunctionBody

https://github.com/microsoft/DirectXShaderCompiler/blob/adffd31eb02f690d9d1afe86c7fa6f12b4e70aa1/lib/DxilValidation/DxilValidation.cpp#L3426

Erroring instruction:

%63 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* %62, i32 1

This is a little weird index 1 should not be out of bounds.
The only unique difference I see from ValidationTest is we don't do addrspace(3) in the ValidationTest.
https://github.com/microsoft/DirectXShaderCompiler/blob/adffd31eb02f690d9d1afe86c7fa6f12b4e70aa1/tools/clang/unittests/HLSL/ValidationTest.cpp#L1678

[sumitsays]

.

[damyanp]

Immediate action at this point is get more of an understanding of the issue.

[Icohedron]

The seems to be an issue related specifically to groupshared memory: https://godbolt.org/z/9MPc8Wca5

RWBuffer<float> e : register(u0);
groupshared float g[2];
[numthreads(2, 1, 1)]
void CSMain(uint3 DTid : SV_DispatchThreadID) {
  g[DTid.x] = 0;
  GroupMemoryBarrierWithGroupSync();
  e[0] = g[1];
}

The above shader successfully compiles under Clang, but fails validation with many of the same error messages as GemmSmallGemm_Bias_float_b16c32:

error: Module bitcode is invalid.
error: Global variable initializer type does not match global variable type!
i8 addrspace(3)* @g

Function: CSMain: error: Pointer type bitcast must be have same size.
note: at '%2 = bitcast i8 addrspace(3)* @g to [2 x float] addrspace(3)*' in block 'entry' of function 'CSMain'.
Function: CSMain: error: TGSM pointers must originate from an unambiguous TGSM global variable.
note: at '%arrayidx.i = getelementptr inbounds [2 x float], [2 x float] addrspace(3)* %2, i32 0, i32 %1' in block 'entry' of function 'CSMain'.
Function: CSMain: error: I8 can only be used as immediate value for intrinsic or as i8* via bitcast by lifetime intrinsics.
note: at '%3 = getelementptr inbounds i8, i8 addrspace(3)* @g, i32 4' in block 'entry' of function 'CSMain'.
Function: CSMain: error: Access to out-of-bounds memory is disallowed.
note: at '%3 = getelementptr inbounds i8, i8 addrspace(3)* @g, i32 4' in block 'entry' of function 'CSMain'.
Function: CSMain: error: Pointer type bitcast must be have same size.
note: at '%4 = bitcast i8 addrspace(3)* %3 to float addrspace(3)*' in block 'entry' of function 'CSMain'.
Validation failed.

Replacing the groupshared with a UAV compiles and validates successfully.

[Icohedron]

This issue is partially fixed by #142475
The remaining cases are almost certainly all caused by #142836