[MLIR][LLVM Dialect][DTLI] Missing DLTI entries

[ghehg]

mlir:llvm
when we do mlir-translate --import-llvm for the simple LLVM IR line like following
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128-Fn32"

to parse datalayout string from LLVM IR, we got following warnings

test.c:0:0: warning: unhandled data layout token: m:e
test.c:0:0: warning: unhandled data layout token: n32:64
test.c:0:0: warning: unhandled data layout token: Fn32

Thats because MLIR DLIT doesn't have representation of Mangling, native int width, as well as function pointer alignment, and thus LLVM dialect cannot import them.
We should add those things to preserve info.

[llvmbot]

@llvm/issue-subscribers-mlir

Author: Guojin (ghehg)

mlir:llvm when we do mlir-translate --import-llvm for the simple LLVM IR line like following `target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128-Fn32"`

to parse datalayout string from LLVM IR, we got following warnings

test.c:0:0: warning: unhandled data layout token: m:e
test.c:0:0: warning: unhandled data layout token: n32:64
test.c:0:0: warning: unhandled data layout token: Fn32

Thats because MLIR DLIT doesn't have representation of Mangling, native int width, as well as function pointer alignment, and thus LLVM dialect cannot import them.
We should add those things to preserve info.

[llvmbot]

@llvm/issue-subscribers-mlir-llvm

Author: Guojin (ghehg)

mlir:llvm when we do mlir-translate --import-llvm for the simple LLVM IR line like following `target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128-Fn32"`

to parse datalayout string from LLVM IR, we got following warnings

test.c:0:0: warning: unhandled data layout token: m:e
test.c:0:0: warning: unhandled data layout token: n32:64
test.c:0:0: warning: unhandled data layout token: Fn32

Thats because MLIR DLIT doesn't have representation of Mangling, native int width, as well as function pointer alignment, and thus LLVM dialect cannot import them.
We should add those things to preserve info.

[rengolin]

@rolfmorel @ftynse we were discussing this on Discord recently.

For LLVM lowering, if we hook TargetMachine to the DLTI dialect, and extract the DL from it, we'll always build a correct string that LLVM can import.

For now, a quick fix is to adjust the string correctly, but this highlights the important of getting the information from the right place.

[rolfmorel]

Just linking #125875 (review) as it is related to obtaining and exposing LLVM target information in a more principled way (also touching on the approach that @rengolin describes up above).

[rolfmorel]

Just linking #125875 (review) as it is related to obtaining and exposing LLVM target information in a more principled way (also touching on the approach that @rengolin describes up above).

Same story for #127416.