Through MLIR lowering (not LLVM): Missing CIR operations

[gxsoar]

There are two lowering paths from CIR:

1. CIR to LLVM IR dialect
2. CIR to MLIR in-tree dialects

Most effort has been into (1), which the list below doesn't really apply. Here's a the list of missing features in (2):

• cir.alloc_exception
• cir.await
• cir.base_class_addr
• cir.break
• cir.asm
• cir.catch
• cir.catch_param
• cir.ceil
• cir.condition
• cir.continue
• cir.cos
• cir.do
• cir.exp2
• cir.exp
• cir.fabs
• cir.floor
• cir.for
• cir.get_bitfield
• cir.get_global
• cir.get_member
• cir.global
• cir.if
• cir.iterator_begin
• cir.iterator_end
• cir.log10
• cir.libc.memchr
• cir.libc.memcpy
• cir.nearbyint
• cir.objsize
• cir.ptr_diff
• cir.ptr_stride
• cir.resume
• cir.rint
• cir.round
• cir.set_bitfield
• cir.shift
• cir.sin
• cir.sqrt
• cir.stack_restore
• cir.stack_save
• cir.std.find
• cir.throw
• cir.trunc
• cir.try_call
• cir.try
• cir.va.arg
• cir.va.copy
• cir.va.end
• cir.vtable.address_point
• cir.vec.cmp
• cir.vec.create
• cir.vec.extract
• cir.vec.insert
• cir.while
• cir.llvmir.zeroinit

[bcardosolopes]

Btw, thanks for trying to categorize, this is going to be very useful for distributing work into multiple issues.

[gxsoar]

Hello, in the clang/lib/CIR/Lowering/ThroughMLIR/LowerCIRToMLIR.cpp :

void populateCIRToMLIRConversionPatterns(mlir::RewritePatternSet &patterns,
                                         mlir::TypeConverter &converter) {
  patterns.add<CIRReturnLowering, CIRBrOpLowering>(patterns.getContext());

  patterns.add<CIRCmpOpLowering, CIRCallOpLowering, CIRUnaryOpLowering,
               CIRBinOpLowering, CIRLoadOpLowering, CIRConstantOpLowering,
               CIRStoreOpLowering, CIRAllocaOpLowering, CIRFuncOpLowering,
               CIRScopeOpLowering, CIRBrCondOpLowering, CIRTernaryOpLowering,
               CIRYieldOpLowering>(converter, patterns.getContext());
}

I find these operations can be lowered to MLIR, cir.if and cir.global donot have releating passes, which can lower cir.if , cir.global to MLIR, why not lower cir.if to scf.if ? .But clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp , these passes can lower cir operations to llvm ir.

[bcardosolopes]

Oh, my apologies, this is talking about "Through MLIR" path, not our direct lowering, let me re-open! Sorry for the confusion

[bcardosolopes]

I just updated the issue description! Thanks for pushing this

[Lancern]

@bcardosolopes Currently the lowering prepare pass runs only on the LLVMIR lowering path. Should we introduce a similar pass for the MLIR lowering path? Some of the operations listed here can be transformed or canonicalized before lowering.

[GaoXiangYa]

Operation	Category	Difficulty	Progress
cir.cos	Math	Easy	Done
cir.ceil	Math	Easy	Done
cir.exp2	Math	Easy	Done
cir.exp	Math	Easy	Done
cir.fabs	Math	Easy	Done
cir.floor	Math	Easy	Done
cir.log10	Math	Easy	Done
cir.rint	Math	Easy
cir.round	Math	Easy	Done
cir.sin	Math	Easy	Done
cir.sqrt	Math	Easy	Done
cir.shift	Math	Easy	Done
cir.libc.memchr	lib call	Easy
cir.libc.memcpy	lib call	Easy
cir.nearbyint	Lib Call	Easy
cir.get_bitfield	Address	Easy
cir.get_global	Address	Easy	Done
cir.global	Address	Easy	Done
cir.set_bitfield	Address	Easy
cir.std.find	Lib Call	Easy
cir.break	Control Flow	Mid
cir.asm	Control Flow	Mid
cir.condition	Control Flow	Mid	Done
cir.continue	Control Flow	Mid
cir.for	Control Flow	Mid	Done
cir.do	Control Flow	Mid
cir.if	Control Flow	Mid	Done
cir.while	Control Flow	Mid	Done
cir.llvmir.zeroinit	LLVM	Mid
cir.alloc_exception	Exception	Hard
cir.catch	Exception	Hard
cir.catch_param	Exception	Hard
cir.throw	Exception	Hard
cir.trunc	Exception	Hard
cir.try_call	Exception	Hard
cir.resume	Exception	Hard
cir.try	Exception	Hard
cir.va.arg	Paramenter List	Hard
cir.va.copy	Paramenter List	Hard
cir.va.end	Paramenter List	Hard
cir.vec.cmp	SIMD vector	Hard
cir.vec.create	SIMD vector	Hard
cir.vec.extract	SIMD vector	Hard
cir.vec.insert	SIMD vector	Hard
cir.iterator_begin	Iterator	Hard
cir.iterator_end	Iterator	Hard
cir.ptr_diff	Pointer	Hard
cir.ptr_stride	Pointer	Hard	Done
cir.stack_restore	Stack Operation	Hard
cir.stack_save	Stack Operation	Hard
cir.base_class_addr	Class	Hard
cir.objsize	Normal	Hard
cir.get_member	Class	Hard
cir.vtable.address_point	Vtable	Hard
cir.await	Asynchronous programming	Hard

Hello, I have conducted research on these operations, classified them, and made a simple assessment of the difficulty of implementing these operations. I want to solve this issue. The current idea is to solve it one by one from easy to difficult, starting with math-related operations.

[bcardosolopes]

Sounds good

[bcardosolopes]

@bcardosolopes Currently the lowering prepare pass runs only on the LLVMIR lowering path. Should we introduce a similar pass for the MLIR lowering path? Some of the operations listed here can be transformed or canonicalized before lowering.

Sure, we should just add LoweringPrepare for that pipeline. There shouldn't be anything LLVM specific there, though somethings naturally go that direction (e.g. lowering globals). If we find that we need to different decisions, we can think about how to refactor that once we get into those.

[bcardosolopes]

@orbiri this table could be interesting for you too!

[Lancern]

Sure, we should just add LoweringPrepare for that pipeline. There shouldn't be anything LLVM specific there, though somethings naturally go that direction (e.g. lowering globals). If we find that we need to different decisions, we can think about how to refactor that once we get into those.

I checked this again. The current LoweringPrepare pass is included in the C++-to-MLIR pipeline (i.e. the pipeline built by clang frontend). But it is not included in the pipelines built by cir-opt or cir-translate. Is it intentional? Should we add the pass in those pipelines?

[bcardosolopes]

But it is not included in the pipelines built by cir-opt or cir-translate. Is it intentional?

LoweringPrepare needs the AST, and we don't have that (yet) by standalone invocations of cir-opt or cir-translate on top of CIR files. :(

Should we add the pass in those pipelines?

We should probably add it to clang driven "through MLIR" codegen, yes.

[GaoXiangYa]

Hello , I find that some libc functions lowering to mlir can use func.call methods, however, I think it is not necessary to do this. And like some vtable operations, it is difficult to lowering them to mlir. @bcardosolopes

[bcardosolopes]

Hello , I find that some libc functions lowering to mlir can use func.call methods, however, I think it is not necessary to do this. And like some vtable operations, it is difficult to lowering them to mlir.

Sorry, I'm not sure I understand the concern, can you elaborate?

[GaoXiangYa]

What I mean is how to translate libc function calls into MLIR , and whether operations related to some vtable can be translated into MLIR? @bcardosolopes

[bcardosolopes]

I understand your question, but I don't understand the specifics:

I find that some libc functions lowering to mlir can use func.call methods, however, I think it is not necessary to do this

If you don't want to use func.call methods, what do you want to use instead?

some vtable operations, it is difficult to lowering them to mlir

Why? What did you tried that didn't work?

Also, on a tangent, what is it your goal on working on MLIR support? Are you just trying to cover all CIR lowering possible for MLIR?

[GaoXiangYa]

We are currently attempting to lower C code to CIR and then further to MLIR for performance analysis. However, for library function calls like libc, we haven't found a good solution to lower them to MLIR yet, and MLIR currently does not support C struct type. @bcardosolopes

[bcardosolopes]

for library function calls like libc, we haven't found a good solution to lower them to MLIR yet

Can you elaborate on why func.call isn't good enough? Are you looking into more higher level representations for libc functions in MLIR? I'm not aware of any, indeed. Btw, I'm asking more questions case I'm curious to understand :)

We started mapping some libc functions in CIR, see memchr for instance, but approaches like that would only be useful if you were doing your perf analysis on top of CIR, not MLIR dialects.