[mono][aot] Enable direct call transformation where caller requires no initialization

[kotlarmilos]

Description

Direct call transformation emits a direct call for a patch instead of having a PLT entry. This is possible if a called method is in the same assembly and requires no initialization. In nollvm configuration, when a call is made through a PLT entry, it initializes the got slots used by the method and runs its class cctor if needed. In llvm configuration, methods initialize themselves and they can be called at any time.

The idea is to change nollvm methods to init themselves without a PLT entry. Method entry point should invoke class constructor and initialize all GOT entries used.

Tasks

• [mono][aot] Propose initialization of nollvm methods without a PLT entry #83376
• [mono][aot] Update nollvm methods to init themselves in its prolog #83377
• [mono][aot] Enable direct call transformation for MONO_PATCH_INFO_METHOD #82711
• [mono][aot] Enable direct call transformation for wrappers #83378
• [mono][aot] Enable direct call transformation for inflated methods #83379

[kotlarmilos]

In llvm, there is mono_inited struct that indicates if a method is inited. Here is a llvm method prolog for initialization:

; Function Attrs: noinline uwtable
define hidden monocc void @HelloWorld_Program_Main_string__(i64* %arg_args) #2 gc "coreclr" !dbg !13 {
BB0:
  br label %INIT_BB1, !dbg !17

INIT_BB1:                                         ; preds = %BB0
  %is_inited = load i8, i8* getelementptr inbounds ([1 x i8], [1 x i8]* @mono_inited, i32 0, i32 0), align 1, !dbg !17
  %0 = call i8 @llvm.expect.i8(i8 %is_inited, i8 1), !dbg !17
  %1 = icmp eq i8 %0, 0, !dbg !17
  br i1 %1, label %NOTINITED_BB15, label %INITED_BB2, !dbg !17

...

NOTINITED_BB15:                                   ; preds = %INIT_BB1
  call void @mono_aot_HelloWorld_init_method(i64* bitcast ([11 x i8]* @info_HelloWorld_Program_Main_string__ to i64*)), !dbg !17
  store i8 1, i8* getelementptr inbounds ([1 x i8], [1 x i8]* @mono_inited, i32 0, i32 0), align 1, !dbg !17
  br label %INITED_BB2, !dbg !17
}

NOTINITED_BB15 bblock invokes mono_aot_HelloWorld_init_method that uses mini_llvm_init_method runtime method to invoke .cctr and init GOT slots:

define hidden void @mono_aot_HelloWorld_init_method(i64* %0) #1 {
ENTRY:
  %1 = bitcast i64* %0 to i32*
  %2 = getelementptr i32, i32* %1, i32 0
  %method_index = load i32, i32* %2, align 4
  %3 = getelementptr [0 x i8], [0 x i8]* bitcast ([7 x i8]* @mono_inited to [0 x i8]*), i32 0, i32 %method_index
  %is_inited = load i8, i8* %3, align 1
  %4 = icmp eq i8 %is_inited, 0
  br i1 %4, label %NOT_INITED, label %INITED

INITED:                                           ; preds = %NOT_INITED, %ENTRY
  ret void

NOT_INITED:                                       ; preds = %ENTRY
  %5 = load i64*, i64** @aotconst_aot_module, align 8, !invariant.load !17
  %6 = ptrtoint i64* %5 to i64
  %7 = load void (i64, i64, i64*, i64)*, void (i64, i64, i64*, i64)** @aotconst_jit_icall_mini_llvm_init_method, align 8, !invariant.load !17
  call void %7(i64 ptrtoint (%MonoAotFileInfo* @mono_aot_file_info to i64), i64 %6, i64* %0, i64 0)
  %8 = getelementptr [0 x i8], [0 x i8]* bitcast ([7 x i8]* @mono_inited to [0 x i8]*), i32 0, i32 %method_index
  store i8 1, i8* %8, align 1
  br label %INITED
}

The same approach should be considered for nollvm methods.

/cc: @LeVladIonescu

[vargaz]

I think these are all 9.0 issues.

[kotlarmilos]

@LeVladIonescu @vargaz Can we try to get the implementation merged without the optimizations enabled within .NET 8 time frame?

We can move the following issues to .NET 9:

• [mono][aot] Enable direct call transformation for MONO_PATCH_INFO_METHOD #82711
• [mono][aot] Enable direct call transformation for wrappers #83378
• [mono][aot] Enable direct call transformation for inflated methods #83379

/cc: @SamMonoRT

[kotlarmilos]

The impact is not significant enough / high effort.