[llvm] [HLSL][DirectX] Handles GEPs in dxil resource access pass (PR #224754)

[Farzon Lotfi via llvm-commits]

================
@@ -665,9 +666,27 @@ static void createCBufferLoad(IntrinsicInst *II, LoadInst *LI,
       // because arrays and structs are always row aligned, and accesses to
       // vector elements will show up as a load of the vector followed by an
       // extractelement.
+      Value *RowIndex = *LastGEP->idx_begin();
+      uint64_t IndexScale =
+          DL.getTypeAllocSize(LastGEP->getSourceElementType());
+      if (IndexScale < hlsl::CBufferRowSizeInBytes) {
+        assert(hlsl::CBufferRowSizeInBytes % IndexScale == 0 &&
+               "CBuffer GEP index is not row aligned");
+        RowIndex = Builder.CreateExactUDiv(
+            RowIndex,
+            ConstantInt::get(Builder.getInt32Ty(),
+                             hlsl::CBufferRowSizeInBytes / IndexScale));
----------------
farzonl wrote:

Don't do `.getInt32Ty()` the legalizer has a i64 to i32 downgrader and we should just depend on that. But the reason not to do this is GEP indices (RowIndex) are frequently 64-bit integers (i64). If you pass an i32 constant from Builder.getInt32Ty() as the right-hand side to CreateExactUDiv or CreateMul, LLVM will hit an assertion failure for type mismatch (Div operands must have the same type!).
```suggestion
            ConstantInt:::get(IndexTy, hlsl::CBufferRowSizeInBytes / IndexScale));
```

https://github.com/llvm/llvm-project/pull/224754