[flang-commits] [clang] [flang] [llvm] [InstCombine] Canonicalize constant GEPs to i8 source element type (PR #68882)

[Karthikeyan Manivannan via flang-commits]

karthik-man wrote:

I am debugging a Triton issue (https://github.com/triton-lang/triton/issues/4060), where an {i32, i32, i32, i64} struct passed to _vprintf_ is printing the wrong value for the i64. The issue here seems to be that Triton creates a llvm:Module with a default DataLayout. In the default layout, i64 abi alignment is 4(DefaultAlignments in DataLayout.cpp). This causes the optimization in this PR to rewrite the GEP to the i64 as follows: 

`%8 = alloca { i32, i32, i32, i64 }, align 8`
`%12 = getelementptr { i32, i32, i32, i64 }, ptr %8, i32 0, i32 3` 
to
`%6 = alloca { i32, i32, i32, i64 }, align 8`
`%9 = getelementptr inbounds i8, ptr %6, i64 12`  

But _vprintf_ expects the i64 to be at offset 16. 
Is it legal for InstCombine to assume that the DL attached to the Module is the right DL for the target? In other words, is it ok for InstCombine to do optimizations that are target-dependent?

https://github.com/llvm/llvm-project/pull/68882