Force Method Inlining in Burst
Burst 1.0.1 is a patch-level update to the original 1.0.0 release, but it actually contains a useful new feature: we’re now able to force methods to be inlined. Read on to see how!
Update: A Russian translation of this article is available.
First, let’s create a Burst-compiled job that has a big method:
[BurstCompile] struct BigMethodJob : IJob { [ReadOnly] public int Val1; [ReadOnly] public int Val2; [WriteOnly] public NativeArray<float> Result; public void Execute() { float f = PointlessAddition(Val1); f += PointlessAddition(Val2); Result[0] = f; } float PointlessAddition(int val) { float f = 0; f += val+1; f += val+2; f += val+3; f += val+4; f += val+5; f += val+6; f += val+7; f += val+8; f += val+9; f += val+10; f += val+11; f += val+12; f += val+13; f += val+14; f += val+15; f += val+16; f += val+17; f += val+18; f += val+19; f += val+20; f += val+21; f += val+22; f += val+23; f += val+24; f += val+25; f += val+26; f += val+27; f += val+28; f += val+29; f += val+30; f += val+31; f += val+32; f += val+33; f += val+34; f += val+35; f += val+36; f += val+37; f += val+38; f += val+39; f += val+40; return f; } }
The job is pretty pointless, but has enough code to prevent Burst from inlining the body of PointlessAddition into the two call sites in Execute. Let’s prove that by looking at Burst Inspector:
; Execute push r14 push rbx push rax mov rbx, rdi mov edi, dword ptr [rbx] movabs r14, offset ".LBigMethodJob.PointlessAddition(BigMethodJob* this, int val)_4CCC87A9ADCAE58C" call r14 movss dword ptr [rsp + 4], xmm0 mov edi, dword ptr [rbx + 4] call r14 addss xmm0, dword ptr [rsp + 4] mov rax, qword ptr [rbx + 8] movss dword ptr [rax], xmm0 add rsp, 8 pop rbx pop r14 ret ; PointlessAddition movd xmm0, edi pshufd xmm0, xmm0, 0 movabs rax, offset .LCPI1_0 movdqa xmm8, xmmword ptr [rax] paddd xmm8, xmm0 movabs rax, offset .LCPI1_1 movdqa xmm2, xmmword ptr [rax] paddd xmm2, xmm0 movabs rax, offset .LCPI1_2 movdqa xmm3, xmmword ptr [rax] paddd xmm3, xmm0 movabs rax, offset .LCPI1_3 movdqa xmm4, xmmword ptr [rax] paddd xmm4, xmm0 movabs rax, offset .LCPI1_4 movdqa xmm5, xmmword ptr [rax] paddd xmm5, xmm0 movabs rax, offset .LCPI1_5 movdqa xmm6, xmmword ptr [rax] paddd xmm6, xmm0 movabs rax, offset .LCPI1_6 movdqa xmm7, xmmword ptr [rax] paddd xmm7, xmm0 movabs rax, offset .LCPI1_7 movdqa xmm1, xmmword ptr [rax] paddd xmm1, xmm0 cvtdq2ps xmm1, xmm1 cvtdq2ps xmm7, xmm7 cvtdq2ps xmm6, xmm6 cvtdq2ps xmm5, xmm5 addps xmm5, xmm6 addps xmm5, xmm1 addps xmm5, xmm7 cvtdq2ps xmm1, xmm4 cvtdq2ps xmm3, xmm3 cvtdq2ps xmm2, xmm2 cvtdq2ps xmm4, xmm8 addps xmm4, xmm2 addps xmm4, xmm1 addps xmm4, xmm5 addps xmm4, xmm3 movabs rax, offset .LCPI1_8 movdqa xmm1, xmmword ptr [rax] paddd xmm1, xmm0 movabs rax, offset .LCPI1_9 paddd xmm0, xmmword ptr [rax] cvtdq2ps xmm0, xmm0 cvtdq2ps xmm1, xmm1 addps xmm1, xmm0 movaps xmm2, xmm4 movhlps xmm2, xmm2 addps xmm2, xmm4 haddps xmm2, xmm2 movaps xmm0, xmm1 movhlps xmm0, xmm0 addps xmm0, xmm1 haddps xmm0, xmm0 addss xmm0, xmm2 ret
The meaning of most of this assembly doesn’t really matter. The important part is right at the top in Execute where we see two call instructions. These call the PointlessAddition method twice. Then we see the rest of the instructions are for PointlessAddition. This shows that the method is being called rather than having its instructions directly pasted into Execute.
Now let’s try a new feature from Burst 1.0.1 and enable inlining of the method with [MethodImpl]:
[MethodImpl(MethodImplOptions.AggressiveInlining)] float PointlessAddition(int val)
Now when we look in Burst Inspector we see no call instructions and a lot more total instructions:
mov ecx, dword ptr [rdi] mov eax, dword ptr [rdi + 4] lea edx, [rcx + 1] cvtsi2ss xmm0, edx lea edx, [rcx + 2] cvtsi2ss xmm1, edx lea edx, [rcx + 3] cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 4] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 5] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 6] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 7] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 8] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 9] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 10] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 11] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 12] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 13] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 14] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 15] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 16] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 17] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 18] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 19] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 20] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 21] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 22] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 23] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 24] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 25] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 26] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 27] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 28] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 29] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 30] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 31] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 32] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 33] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 34] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 35] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 36] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 37] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 lea edx, [rcx + 38] xorps xmm1, xmm1 cvtsi2ss xmm1, edx addss xmm1, xmm2 lea edx, [rcx + 39] xorps xmm2, xmm2 cvtsi2ss xmm2, edx addss xmm2, xmm1 add ecx, 40 xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 1] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx lea ecx, [rax + 2] cvtsi2ss xmm3, ecx addss xmm3, xmm1 addss xmm3, xmm2 lea ecx, [rax + 3] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm3 lea ecx, [rax + 4] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 5] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 6] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 7] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 8] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 9] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 10] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 11] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 12] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 13] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 14] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 15] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 16] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 17] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 18] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 19] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 20] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 21] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 22] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 23] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 24] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 25] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 26] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 27] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 28] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 29] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 30] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 31] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 32] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 33] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 34] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 35] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 36] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 37] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 lea ecx, [rax + 38] xorps xmm2, xmm2 cvtsi2ss xmm2, ecx addss xmm2, xmm1 lea ecx, [rax + 39] xorps xmm1, xmm1 cvtsi2ss xmm1, ecx addss xmm1, xmm2 add eax, 40 xorps xmm2, xmm2 cvtsi2ss xmm2, eax addss xmm2, xmm1 addss xmm2, xmm0 mov rax, qword ptr [rdi + 8] movss dword ptr [rax], xmm2 ret
This has effectively removed the function calls and duplicated the function within Execute.
Now that we’ve seen method inlining work, the question becomes whether we should use it or not and under what circumstances. As an anecdote, I needed to add a lot of code to the method in order for Burst to not automatically inline it. This means that it’s already aggressively inlining methods, even without the attribute. So [MethodImpl] should only be necessary when either Burst isn’t inlining a short method that it really should have inlined. Large methods generally don’t need to be inlined because the cost of calling them is relatively low compared to the cost of executing their body. This leads to a rule of thumb- force method inlining sparingly.
#1 by sschoener on July 15th, 2019 ·
Thanks for writing this! About the last part: The main benefit of inlining is often not that you are not paying the cost for the call anymore (though that is certainly relevant for very small procedures) but rather that you enable the compiler to perform more optimizations: It now has all the context from the callsite. See e.g. https://en.m.wikipedia.org/wiki/Inline_expansion#Effect_on_performance – that does however not change the conclusion that inlining shoild be used sparingly :)
#2 by David Wu on July 18th, 2019 ·
I am not sure if I understand why inlining of burst code should be used sparingly.
I would think that for most job type functions the cost of inlining (more code and cache pressure) is small compared to the data being processed. Code caches are very large these days.
In most games, a very small percent of the code is performance-critical and aggressively inlining it won’t significantly increase the code footprint as a whole.
#3 by jackson on July 21st, 2019 ·
Inlining should be used liberally, but Burst already does that. Forcing inlining with
[MethodImpl]should be used sparingly, as in the small percent of performance-critical code you mention.