Outline error path

[vchuravy]

Manually outline the error path, otherwise on Julia 1.10 we cause an unecessary GC-frame for every MPI function.

[vchuravy]

On master for Julia 1.10

julia> @code_llvm MPI.Iprobe(MPI.COMM_WORLD)

;  @ /home/vchuravy/src/MPI/src/nonblocking.jl:395 within `Iprobe`
define i8 @julia_Iprobe_610({}* noundef nonnull align 4 dereferenceable(4) %0) #0 {
L18:
  %1 = alloca {}*, align 8
;  @ /home/vchuravy/src/MPI/src/nonblocking.jl:395 within `Iprobe` @ /home/vchuravy/src/MPI/src/nonblocking.jl:395
; ┌ @ /home/vchuravy/src/MPI/src/nonblocking.jl:395 within `#Iprobe#87`
; │┌ @ /home/vchuravy/src/MPI/src/nonblocking.jl:398 within `Iprobe`
; ││┌ @ refpointer.jl:138 within `Ref`
; │││┌ @ refvalue.jl:7 within `RefValue`
      %gcframe7 = alloca [3 x {}*], align 16
      %gcframe7.sub = getelementptr inbounds [3 x {}*], [3 x {}*]* %gcframe7, i64 0, i64 0
      %2 = bitcast [3 x {}*]* %gcframe7 to i8*
      call void @llvm.memset.p0i8.i64(i8* align 16 %2, i8 0, i64 24, i1 true)
      %3 = alloca i32, align 8
      %4 = bitcast i32* %3 to i8*
      %thread_ptr = call i8* asm "movq %fs:0, $0", "=r"() #11
      %tls_ppgcstack = getelementptr i8, i8* %thread_ptr, i64 -8
      %5 = bitcast i8* %tls_ppgcstack to {}****
      %tls_pgcstack = load {}***, {}**** %5, align 8
; └└└└
; ┌ @ refvalue.jl:59 within `getindex`
; │┌ @ Base.jl:37 within `getproperty`
    %6 = bitcast [3 x {}*]* %gcframe7 to i64*
    store i64 4, i64* %6, align 16
    %7 = getelementptr inbounds [3 x {}*], [3 x {}*]* %gcframe7, i64 0, i64 1
    %8 = bitcast {}** %7 to {}***
    %9 = load {}**, {}*** %tls_pgcstack, align 8
    store {}** %9, {}*** %8, align 8
    %10 = bitcast {}*** %tls_pgcstack to {}***
    store {}** %gcframe7.sub, {}*** %10, align 8
    %11 = load i32, i32* inttoptr (i64 140701644651504 to i32*), align 16
    %12 = load i32, i32* inttoptr (i64 140701644651408 to i32*), align 16
    call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %4)
; └└
; ┌ @ /home/vchuravy/src/MPI/src/nonblocking.jl:395 within `#Iprobe#87`
; │┌ @ /home/vchuravy/src/MPI/src/nonblocking.jl:399 within `Iprobe`
; ││┌ @ /home/vchuravy/src/MPI/src/api/generated_api.jl:2124 within `MPI_Iprobe`
; │││┌ @ /home/vchuravy/src/MPI/src/comm.jl:11 within `unsafe_convert`
; ││││┌ @ Base.jl:37 within `getproperty`
       %13 = bitcast {}* %0 to i32*
       %14 = load i32, i32* %13, align 4
; ││└└└
; ││┌ @ refvalue.jl:59 within `getindex`
; │││┌ @ Base.jl:37 within `getproperty`
      %15 = load i64, i64* inttoptr (i64 140701644630800 to i64*), align 16
; ││└└
; ││┌ @ /home/vchuravy/src/MPI/src/api/generated_api.jl:2124 within `MPI_Iprobe`
; │││┌ @ refvalue.jl:53 within `unsafe_convert`
; ││││┌ @ pointer.jl:30 within `convert`
       %bitcast_coercion2 = ptrtoint i32* %3 to i64
; │││└└
     %16 = call i32 inttoptr (i64 140701651319248 to i32 (i32, i32, i32, i64, i64)*)(i32 %11, i32 %12, i32 %14, i64 %bitcast_coercion2, i64 %15)
; │││┌ @ promotion.jl:461 within `==` @ promotion.jl:521
      %.not = icmp eq i32 %16, 0
; │││└
     br i1 %.not, label %L29, label %L26

L26:                                              ; preds = %L18
     call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %4)
     %17 = call nonnull {}* @ijl_box_int32(i32 signext %16)
     %18 = getelementptr inbounds [3 x {}*], [3 x {}*]* %gcframe7, i64 0, i64 2
     store {}* %17, {}** %18, align 16
     store {}* %17, {}** %1, align 8
     %19 = call nonnull {}* @ijl_apply_generic({}* inttoptr (i64 140701645548176 to {}*), {}** nonnull %1, i32 1)
     call void @ijl_throw({}* %19)
     unreachable

L29:                                              ; preds = %L18
; ││└
; ││ @ /home/vchuravy/src/MPI/src/nonblocking.jl:400 within `Iprobe`
; ││┌ @ refvalue.jl:59 within `getindex`
; │││┌ @ Base.jl:37 within `getproperty`
      %20 = load i32, i32* %3, align 8
; ││└└
; ││┌ @ operators.jl:276 within `!=`
; │││┌ @ promotion.jl:461 within `==` @ promotion.jl:521
      %21 = icmp ne i32 %20, 0
; └└└└
  %22 = zext i1 %21 to i8
  %23 = load {}*, {}** %7, align 8
  %24 = bitcast {}*** %tls_pgcstack to {}**
  store {}* %23, {}** %24, align 8
  ret i8 %22
}

On this branch:

julia> @code_llvm MPI.Iprobe(MPI.COMM_WORLD)

;  @ /home/vchuravy/src/MPI/src/nonblocking.jl:395 within `Iprobe`
define i8 @julia_Iprobe_437({}* noundef nonnull align 4 dereferenceable(4) %0) #0 {
L18:
;  @ /home/vchuravy/src/MPI/src/nonblocking.jl:395 within `Iprobe` @ /home/vchuravy/src/MPI/src/nonblocking.jl:395
; ┌ @ /home/vchuravy/src/MPI/src/nonblocking.jl:395 within `#Iprobe#87`
; │┌ @ /home/vchuravy/src/MPI/src/nonblocking.jl:398 within `Iprobe`
; ││┌ @ refpointer.jl:138 within `Ref`
; │││┌ @ refvalue.jl:7 within `RefValue`
      %1 = alloca i32, align 8
      %2 = bitcast i32* %1 to i8*
; └└└└
; ┌ @ refvalue.jl:59 within `getindex`
; │┌ @ Base.jl:37 within `getproperty`
    %3 = load i32, i32* inttoptr (i64 140538370873152 to i32*), align 64
    %4 = load i32, i32* inttoptr (i64 140538370873056 to i32*), align 32
    call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %2)
; └└
; ┌ @ /home/vchuravy/src/MPI/src/nonblocking.jl:395 within `#Iprobe#87`
; │┌ @ /home/vchuravy/src/MPI/src/nonblocking.jl:399 within `Iprobe`
; ││┌ @ /home/vchuravy/src/MPI/src/api/generated_api.jl:2124 within `MPI_Iprobe`
; │││┌ @ /home/vchuravy/src/MPI/src/api/api.jl:143 within `macro expansion`
; ││││┌ @ /home/vchuravy/src/MPI/src/comm.jl:11 within `unsafe_convert`
; │││││┌ @ Base.jl:37 within `getproperty`
        %5 = bitcast {}* %0 to i32*
        %6 = load i32, i32* %5, align 4
; ││└└└└
; ││┌ @ refvalue.jl:59 within `getindex`
; │││┌ @ Base.jl:37 within `getproperty`
      %7 = load i64, i64* inttoptr (i64 140538370852448 to i64*), align 32
; ││└└
; ││┌ @ /home/vchuravy/src/MPI/src/api/generated_api.jl:2124 within `MPI_Iprobe`
; │││┌ @ /home/vchuravy/src/MPI/src/api/api.jl:143 within `macro expansion`
; ││││┌ @ refvalue.jl:53 within `unsafe_convert`
; │││││┌ @ pointer.jl:30 within `convert`
        %bitcast_coercion2 = ptrtoint i32* %1 to i64
; ││││└└
      %8 = call i32 inttoptr (i64 140538377550288 to i32 (i32, i32, i32, i64, i64)*)(i32 %3, i32 %4, i32 %6, i64 %bitcast_coercion2, i64 %7)
; ││││ @ /home/vchuravy/src/MPI/src/api/api.jl:144 within `macro expansion`
; ││││┌ @ operators.jl:276 within `!=`
; │││││┌ @ promotion.jl:461 within `==` @ promotion.jl:521
        %.not = icmp eq i32 %8, 0
; ││││└└
      br i1 %.not, label %L29, label %L26

L26:                                              ; preds = %L18
      call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %2)
; ││││ @ /home/vchuravy/src/MPI/src/api/api.jl:145 within `macro expansion`
      call void @j_mpi_error_439(i32 signext %8) #4
      unreachable

L29:                                              ; preds = %L18
; ││└└
; ││ @ /home/vchuravy/src/MPI/src/nonblocking.jl:400 within `Iprobe`
; ││┌ @ refvalue.jl:59 within `getindex`
; │││┌ @ Base.jl:37 within `getproperty`
      %9 = load i32, i32* %1, align 8
; ││└└
; ││┌ @ operators.jl:276 within `!=`
; │││┌ @ promotion.jl:461 within `==` @ promotion.jl:521
      %10 = icmp ne i32 %9, 0
; └└└└
  %11 = zext i1 %10 to i8
  ret i8 %11
}

[vchuravy]

I have to correct myself, we are not reducing allocations (at least not in the example I show), but the unecessary setup of a GC frame.