Codegen regression with llvmcall/inlining

[maleadt]

MWE:

@inline use(i::Int64) =
    Base.llvmcall("""%slot = alloca i64
                     store volatile i64 %0, i64* %slot
                     %value = load volatile i64, i64* %slot
                     ret i64 %value""", Int64, Tuple{Int64}, i)


struct Wrapper{F} <: Core.Function
    f::F
end

function (k::Wrapper)(arg1, arg2)
    (k.f)(arg1, arg2)
    return nothing
end


foobar(unused, used) = use(used)

code_llvm(Wrapper(foobar), Tuple{Type{Int64}, Int64})

Before #27857:

; Function Wrapper
; Location: /home/tbesard/Julia/CUDAnative/wip.jl:13
define void @julia_Wrapper_32376(%jl_value_t addrspace(10)*, i64) {
top:
  %slot.i = alloca i64, align 8
; Function foobar; {
; Location: /home/tbesard/Julia/CUDAnative/wip.jl:18
; Function use; {
; Location: /home/tbesard/Julia/CUDAnative/wip.jl:1
  %slot.i.0.slot.i.0..sroa_cast = bitcast i64* %slot.i to i8*
  call void @llvm.lifetime.start.p0i8(i64 8, i8* %slot.i.0.slot.i.0..sroa_cast)
  store volatile i64 %1, i64* %slot.i, align 8
  %slot.i.0.slot.i.0.slot.0.slot.0.value.i = load volatile i64, i64* %slot.i, align 8
  call void @llvm.lifetime.end.p0i8(i64 8, i8* %slot.i.0.slot.i.0..sroa_cast)
;}}
; Location: /home/tbesard/Julia/CUDAnative/wip.jl:14
  ret void
}

After:

; Function Wrapper
; Location: /home/tbesard/Julia/CUDAnative/wip.jl:13
define void @julia_Wrapper_32797(%jl_value_t addrspace(10)*, i64) {
top:
  %2 = alloca %jl_value_t addrspace(10)*, i32 3
  %gcframe = alloca %jl_value_t addrspace(10)*, i32 3
  %3 = bitcast %jl_value_t addrspace(10)** %gcframe to i8*
  call void @llvm.memset.p0i8.i32(i8* %3, i8 0, i32 24, i32 0, i1 false)
  %thread_ptr = call i8* asm "movq %fs:0, $0", "=r"()
  %ptls_i8 = getelementptr i8, i8* %thread_ptr, i64 -10920
  %ptls = bitcast i8* %ptls_i8 to %jl_value_t***
  %4 = getelementptr %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %gcframe, i32 0
  %5 = bitcast %jl_value_t addrspace(10)** %4 to i64*
  store i64 2, i64* %5
  %6 = getelementptr %jl_value_t**, %jl_value_t*** %ptls, i32 0
  %7 = getelementptr %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %gcframe, i32 1
  %8 = bitcast %jl_value_t addrspace(10)** %7 to %jl_value_t***
  %9 = load %jl_value_t**, %jl_value_t*** %6
  store %jl_value_t** %9, %jl_value_t*** %8
  %10 = bitcast %jl_value_t*** %6 to %jl_value_t addrspace(10)***
  store %jl_value_t addrspace(10)** %gcframe, %jl_value_t addrspace(10)*** %10
  %11 = call %jl_value_t addrspace(10)* @jl_box_int64(i64 signext %1)
  %12 = getelementptr %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %gcframe, i32 2
  store %jl_value_t addrspace(10)* %11, %jl_value_t addrspace(10)** %12
  %13 = getelementptr %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %2, i32 0
  store %jl_value_t addrspace(10)* addrspacecast (%jl_value_t* inttoptr (i64 140506765508720 to %jl_value_t*) to %jl_value_t addrspace(10)*), %jl_value_t addrspace(10)** %13
  %14 = getelementptr %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %2, i32 1
  store %jl_value_t addrspace(10)* addrspacecast (%jl_value_t* inttoptr (i64 140506765445104 to %jl_value_t*) to %jl_value_t addrspace(10)*), %jl_value_t addrspace(10)** %14
  %15 = getelementptr %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %2, i32 2
  store %jl_value_t addrspace(10)* %11, %jl_value_t addrspace(10)** %15
  %16 = call nonnull %jl_value_t addrspace(10)* @jl_invoke(%jl_value_t addrspace(10)* addrspacecast (%jl_value_t* inttoptr (i64 140506779001488 to %jl_value_t*) to %jl_value_t addrspace(10)*), %jl_value_t addrspace(10)** %2, i32 3)
; Location: /home/tbesard/Julia/CUDAnative/wip.jl:14
  %17 = getelementptr %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %gcframe, i32 1
  %18 = load %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %17
  %19 = getelementptr %jl_value_t**, %jl_value_t*** %ptls, i32 0
  %20 = bitcast %jl_value_t*** %19 to %jl_value_t addrspace(10)**
  store %jl_value_t addrspace(10)* %18, %jl_value_t addrspace(10)** %20
  ret void
}

Bisected to 9277d3a.

[KristofferC]

Perhaps 9277d3a#diff-2c25696889a7cf8eed57b80f011be7fbR191?

[JeffBezanson]

Basically a duplicate of #27694; llvmcall needs to be overhauled for the new IR.

[chriselrod]

Here's an example of another regression, this time without using llvmcall:

julia> using StaticArrays, BenchmarkTools
julia> A = @SMatrix randn(8,8);
julia> B = @SMatrix randn(8,8);
julia> @benchmark $A * $B
BenchmarkTools.Trial: 
  memory estimate:  0 bytes
  allocs estimate:  0
  --------------
  minimum time:     16.310 ns (0.00% GC)
  median time:      17.290 ns (0.00% GC)
  mean time:        17.885 ns (0.00% GC)
  maximum time:     63.709 ns (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     998

vs

julia> @benchmark $A * $B
BenchmarkTools.Trial: 
  memory estimate:  0 bytes
  allocs estimate:  0
  --------------
  minimum time:     28.014 ns (0.00% GC)
  median time:      28.112 ns (0.00% GC)
  mean time:        28.664 ns (0.00% GC)
  maximum time:     59.865 ns (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     995

I think it might be copying big tuples when it doesn't inline.
Structs get copied on non-inlined function calls, right? Is that factored into the cost of not inlining?
FWIW, prior to that commit the @code_native shows a bunch of math, and after:

@code_native A * B
	.text
; Function * {
; Location: matrix_multiply.jl:9
	pushq	%r14
	pushq	%rbx
	subq	$520, %rsp              # imm = 0x208
	movq	%rdi, %rbx
; Function _mul; {
; Location: matrix_multiply.jl:75
; Function macro expansion; {
; Location: matrix_multiply.jl:78
	movabsq	$"*;", %rax
	leaq	8(%rsp), %r14
	movq	%r14, %rdi
	callq	*%rax
	movabsq	$__memcpy_avx_unaligned_erms, %rax
;}}
	movl	$512, %edx              # imm = 0x200
	movq	%rbx, %rdi
	movq	%r14, %rsi
	callq	*%rax
	movq	%rbx, %rax
	addq	$520, %rsp              # imm = 0x208
	popq	%rbx
	popq	%r14
	retq
	nopw	%cs:(%rax,%rax)
;}

[JeffBezanson]

Ok, I think something can be done about that.

[JeffBezanson]

That issue captured as #28168. Closing this one as a duplicate of #27694.