Menu
I am getting odd results from microbenchmarks that can be reproduced with the the example bench_xor_1000_ints from the Rust Book chapter on Benchmark tests.
The section Gotcha: optimizations gives general recommendations to return a value from the Bencher::iter closure and/or to use the black_box function. I ended up with these five variants:
#![feature(test)]
extern crate test;
use test::{Bencher, black_box};
fn xor(x: i32, y: i32) -> i32 { x ^ y }
#[bench]
fn xor_closure_a(b: &mut Bencher) {
b.iter(|| {
let n = black_box(1000);
(0..n).fold(0, |a, b| a ^ b)
});
}
#[bench]
fn xor_closure_b(b: &mut Bencher) {
b.iter(|| {
let n = black_box(1000);
(0..n).fold(0, |a, b| a ^ b);
});
}
#[bench]
fn xor_pointer_a(b: &mut Bencher) {
b.iter(|| {
let n = black_box(1000);
(0..n).fold(0, xor)
});
}
#[bench]
fn xor_pointer_b(b: &mut Bencher) {
b.iter(|| {
let n = black_box(1000);
(0..n).fold(0, xor);
});
}
#[bench] // closure/pointer or semicolon doesn't affect results
fn xor_black_box(b: &mut Bencher) {
b.iter(|| {
black_box((0..1000).fold(0, xor));
});
}
|a, b| a ^ b
xor
b.iter closure returns value (note lack of semicolon)b.iter closure doesn't return value (ends with semicolon)black_box and hope for the besttest tests::xor_black_box ... bench: 69 ns/iter (+/- 1)
test tests::xor_closure_a ... bench: 70 ns/iter (+/- 1)
test tests::xor_closure_b ... bench: 921 ns/iter (+/- 5)
test tests::xor_pointer_a ... bench: 60 ns/iter (+/- 1)
test tests::xor_pointer_b ... bench: 0 ns/iter (+/- 0)
Some observations regarding measurements:
let n = black_box(1000) and inlining it into range (0..1000).fold ... doesn't affect any of the resultsn scales measurements appropriately (except for xor_pointer_b optimized to 0)On the whole, results agree with the general recommendations from Rust Book, except for xor_closure_b benchmark.
More specifically, results with the suffix _a (that return a value) seem to agree with the black-boxed xor_black_box which sounds good to me. And xor_pointer_b optimized to 0 seems legit since its closure doesn't return the value. But the result of xor_closure_b is odd.
Assembly code may clarify this question. How can I prevent the Rust benchmark library from optimizing away my code? gives a good overview on reading assembly from the Rust benchmark tests.
cargo rustc --release -- --emit asm
produces this output:
.text
.file "fpinrust.cgu-0.rs"
.section .rodata.cst16,"aM",@progbits,16
.p2align 4
.LCPI0_0:
.long 0
.long 1
.long 2
.long 3
.LCPI0_1:
.long 4
.long 4
.long 4
.long 4
.LCPI0_2:
.long 8
.long 8
.long 8
.long 8
.LCPI0_3:
.long 12
.long 12
.long 12
.long 12
.LCPI0_4:
.long 16
.long 16
.long 16
.long 16
.LCPI0_5:
.long 20
.long 20
.long 20
.long 20
.LCPI0_6:
.long 24
.long 24
.long 24
.long 24
.LCPI0_7:
.long 28
.long 28
.long 28
.long 28
.LCPI0_8:
.long 32
.long 32
.long 32
.long 32
.section .text._ZN8fpinrust5tests13xor_closure_a17h4df097d1e565a700E,"ax",@progbits
.globl _ZN8fpinrust5tests13xor_closure_a17h4df097d1e565a700E
.p2align 4, 0x90
.type _ZN8fpinrust5tests13xor_closure_a17h4df097d1e565a700E,@function
_ZN8fpinrust5tests13xor_closure_a17h4df097d1e565a700E:
.cfi_startproc
pushq %r14
.Ltmp0:
.cfi_def_cfa_offset 16
pushq %rbx
.Ltmp1:
.cfi_def_cfa_offset 24
subq $40, %rsp
.Ltmp2:
.cfi_def_cfa_offset 64
.Ltmp3:
.cfi_offset %rbx, -24
.Ltmp4:
.cfi_offset %r14, -16
movq %rdi, %r14
leaq 24(%rsp), %rdi
callq _ZN3std4time7Instant3now17h37bccd496c61083dE@PLT
movq (%r14), %r9
testq %r9, %r9
je .LBB0_15
xorl %ecx, %ecx
leaq 8(%rsp), %r8
movdqa .LCPI0_0(%rip), %xmm8
movdqa .LCPI0_1(%rip), %xmm15
movdqa .LCPI0_2(%rip), %xmm2
movdqa .LCPI0_3(%rip), %xmm9
movdqa .LCPI0_4(%rip), %xmm10
movdqa .LCPI0_5(%rip), %xmm11
movdqa .LCPI0_6(%rip), %xmm12
movdqa .LCPI0_7(%rip), %xmm13
movdqa .LCPI0_8(%rip), %xmm14
.p2align 4, 0x90
.LBB0_2:
incq %rcx
movl $1000, 8(%rsp)
#APP
#NO_APP
movl 8(%rsp), %esi
testl %esi, %esi
movl $0, %edx
jle .LBB0_14
xorl %edx, %edx
cmpl $8, %esi
jae .LBB0_5
xorl %edi, %edi
jmp .LBB0_13
.p2align 4, 0x90
.LBB0_5:
movl %esi, %eax
andl $-8, %eax
movl $0, %edi
je .LBB0_13
leal -8(%rax), %edx
movl %edx, %edi
shrl $3, %edi
leal 1(%rdi), %ebx
andl $3, %ebx
pxor %xmm3, %xmm3
cmpl $24, %edx
pxor %xmm4, %xmm4
movdqa %xmm8, %xmm5
jb .LBB0_9
leal -1(%rbx), %edx
subl %edi, %edx
pxor %xmm3, %xmm3
pxor %xmm4, %xmm4
movdqa %xmm8, %xmm5
.p2align 4, 0x90
.LBB0_8:
movdqa %xmm5, %xmm6
paddd %xmm15, %xmm6
movdqa %xmm5, %xmm7
paddd %xmm2, %xmm7
pxor %xmm5, %xmm3
pxor %xmm4, %xmm6
movdqa %xmm5, %xmm4
paddd %xmm9, %xmm4
movdqa %xmm5, %xmm0
paddd %xmm10, %xmm0
pxor %xmm7, %xmm3
movdqa %xmm5, %xmm7
paddd %xmm11, %xmm7
movdqa %xmm5, %xmm1
paddd %xmm12, %xmm1
pxor %xmm4, %xmm7
pxor %xmm6, %xmm7
movdqa %xmm5, %xmm4
paddd %xmm13, %xmm4
paddd %xmm14, %xmm5
pxor %xmm0, %xmm1
pxor %xmm1, %xmm3
pxor %xmm7, %xmm4
addl $4, %edx
jne .LBB0_8
.LBB0_9:
testl %ebx, %ebx
je .LBB0_12
negl %ebx
.p2align 4, 0x90
.LBB0_11:
movdqa %xmm5, %xmm0
pxor %xmm5, %xmm3
paddd %xmm15, %xmm5
paddd %xmm2, %xmm0
pxor %xmm5, %xmm4
incl %ebx
movdqa %xmm0, %xmm5
jne .LBB0_11
.LBB0_12:
pxor %xmm4, %xmm3
pshufd $78, %xmm3, %xmm0
pxor %xmm3, %xmm0
pshufd $229, %xmm0, %xmm1
pxor %xmm0, %xmm1
movd %xmm1, %edx
cmpl %eax, %esi
movl %eax, %edi
je .LBB0_14
.p2align 4, 0x90
.LBB0_13:
xorl %edi, %edx
leal 1(%rdi), %eax
cmpl %eax, %esi
movl %eax, %edi
jne .LBB0_13
.LBB0_14:
movl %edx, 8(%rsp)
#APP
#NO_APP
cmpq %r9, %rcx
jne .LBB0_2
.LBB0_15:
leaq 8(%rsp), %rdi
leaq 24(%rsp), %rsi
callq _ZN3std4time7Instant7elapsed17h0b6076720ddfcc2bE@PLT
movq 8(%rsp), %rax
movl 16(%rsp), %ecx
movq %rax, 8(%r14)
movl %ecx, 16(%r14)
addq $40, %rsp
popq %rbx
popq %r14
retq
.Lfunc_end0:
.size _ZN8fpinrust5tests13xor_closure_a17h4df097d1e565a700E, .Lfunc_end0-_ZN8fpinrust5tests13xor_closure_a17h4df097d1e565a700E
.cfi_endproc
.section .text._ZN8fpinrust5tests13xor_closure_b17h4bdd5e59e5c19a55E,"ax",@progbits
.globl _ZN8fpinrust5tests13xor_closure_b17h4bdd5e59e5c19a55E
.p2align 4, 0x90
.type _ZN8fpinrust5tests13xor_closure_b17h4bdd5e59e5c19a55E,@function
_ZN8fpinrust5tests13xor_closure_b17h4bdd5e59e5c19a55E:
.cfi_startproc
pushq %r14
.Ltmp5:
.cfi_def_cfa_offset 16
pushq %rbx
.Ltmp6:
.cfi_def_cfa_offset 24
subq $40, %rsp
.Ltmp7:
.cfi_def_cfa_offset 64
.Ltmp8:
.cfi_offset %rbx, -24
.Ltmp9:
.cfi_offset %r14, -16
movq %rdi, %r14
leaq 24(%rsp), %rdi
callq _ZN3std4time7Instant3now17h37bccd496c61083dE@PLT
movq (%r14), %rax
testq %rax, %rax
je .LBB1_5
xorl %ecx, %ecx
leaq 8(%rsp), %rdx
.p2align 4, 0x90
.LBB1_2:
movl $1000, 8(%rsp)
#APP
#NO_APP
movl 8(%rsp), %esi
xorl %ebx, %ebx
.p2align 4, 0x90
.LBB1_3:
xorl %edi, %edi
cmpl %esi, %ebx
setl %dil
addl %ebx, %edi
cmpl %esi, %ebx
movl %edi, %ebx
jl .LBB1_3
incq %rcx
#APP
#NO_APP
cmpq %rax, %rcx
jne .LBB1_2
.LBB1_5:
leaq 8(%rsp), %rdi
leaq 24(%rsp), %rsi
callq _ZN3std4time7Instant7elapsed17h0b6076720ddfcc2bE@PLT
movq 8(%rsp), %rax
movl 16(%rsp), %ecx
movq %rax, 8(%r14)
movl %ecx, 16(%r14)
addq $40, %rsp
popq %rbx
popq %r14
retq
.Lfunc_end1:
.size _ZN8fpinrust5tests13xor_closure_b17h4bdd5e59e5c19a55E, .Lfunc_end1-_ZN8fpinrust5tests13xor_closure_b17h4bdd5e59e5c19a55E
.cfi_endproc
.section .rodata.cst16,"aM",@progbits,16
.p2align 4
.LCPI2_0:
.long 0
.long 1
.long 2
.long 3
.LCPI2_1:
.long 4
.long 4
.long 4
.long 4
.LCPI2_2:
.long 8
.long 8
.long 8
.long 8
.LCPI2_3:
.long 12
.long 12
.long 12
.long 12
.LCPI2_4:
.long 16
.long 16
.long 16
.long 16
.LCPI2_5:
.long 20
.long 20
.long 20
.long 20
.LCPI2_6:
.long 24
.long 24
.long 24
.long 24
.LCPI2_7:
.long 28
.long 28
.long 28
.long 28
.LCPI2_8:
.long 32
.long 32
.long 32
.long 32
.section .text._ZN8fpinrust5tests13xor_pointer_a17hca6c00df2b597ce7E,"ax",@progbits
.globl _ZN8fpinrust5tests13xor_pointer_a17hca6c00df2b597ce7E
.p2align 4, 0x90
.type _ZN8fpinrust5tests13xor_pointer_a17hca6c00df2b597ce7E,@function
_ZN8fpinrust5tests13xor_pointer_a17hca6c00df2b597ce7E:
.cfi_startproc
pushq %r14
.Ltmp10:
.cfi_def_cfa_offset 16
pushq %rbx
.Ltmp11:
.cfi_def_cfa_offset 24
subq $40, %rsp
.Ltmp12:
.cfi_def_cfa_offset 64
.Ltmp13:
.cfi_offset %rbx, -24
.Ltmp14:
.cfi_offset %r14, -16
movq %rdi, %r14
leaq 24(%rsp), %rdi
callq _ZN3std4time7Instant3now17h37bccd496c61083dE@PLT
movq (%r14), %r9
testq %r9, %r9
je .LBB2_15
xorl %ecx, %ecx
leaq 8(%rsp), %r8
movdqa .LCPI2_0(%rip), %xmm8
movdqa .LCPI2_1(%rip), %xmm15
movdqa .LCPI2_2(%rip), %xmm2
movdqa .LCPI2_3(%rip), %xmm9
movdqa .LCPI2_4(%rip), %xmm10
movdqa .LCPI2_5(%rip), %xmm11
movdqa .LCPI2_6(%rip), %xmm12
movdqa .LCPI2_7(%rip), %xmm13
movdqa .LCPI2_8(%rip), %xmm14
.p2align 4, 0x90
.LBB2_2:
incq %rcx
movl $1000, 8(%rsp)
#APP
#NO_APP
movl 8(%rsp), %esi
testl %esi, %esi
movl $0, %edx
jle .LBB2_14
xorl %edx, %edx
cmpl $8, %esi
jae .LBB2_5
xorl %edi, %edi
jmp .LBB2_13
.p2align 4, 0x90
.LBB2_5:
movl %esi, %eax
andl $-8, %eax
movl $0, %edi
je .LBB2_13
leal -8(%rax), %edx
movl %edx, %edi
shrl $3, %edi
leal 1(%rdi), %ebx
andl $3, %ebx
pxor %xmm3, %xmm3
cmpl $24, %edx
pxor %xmm4, %xmm4
movdqa %xmm8, %xmm5
jb .LBB2_9
leal -1(%rbx), %edx
subl %edi, %edx
pxor %xmm3, %xmm3
pxor %xmm4, %xmm4
movdqa %xmm8, %xmm5
.p2align 4, 0x90
.LBB2_8:
movdqa %xmm5, %xmm6
paddd %xmm15, %xmm6
movdqa %xmm5, %xmm7
paddd %xmm2, %xmm7
pxor %xmm5, %xmm3
pxor %xmm4, %xmm6
movdqa %xmm5, %xmm4
paddd %xmm9, %xmm4
movdqa %xmm5, %xmm0
paddd %xmm10, %xmm0
pxor %xmm7, %xmm3
movdqa %xmm5, %xmm7
paddd %xmm11, %xmm7
movdqa %xmm5, %xmm1
paddd %xmm12, %xmm1
pxor %xmm4, %xmm7
pxor %xmm6, %xmm7
movdqa %xmm5, %xmm4
paddd %xmm13, %xmm4
paddd %xmm14, %xmm5
pxor %xmm0, %xmm1
pxor %xmm1, %xmm3
pxor %xmm7, %xmm4
addl $4, %edx
jne .LBB2_8
.LBB2_9:
testl %ebx, %ebx
je .LBB2_12
negl %ebx
.p2align 4, 0x90
.LBB2_11:
movdqa %xmm5, %xmm0
pxor %xmm5, %xmm3
paddd %xmm15, %xmm5
paddd %xmm2, %xmm0
pxor %xmm5, %xmm4
incl %ebx
movdqa %xmm0, %xmm5
jne .LBB2_11
.LBB2_12:
pxor %xmm4, %xmm3
pshufd $78, %xmm3, %xmm0
pxor %xmm3, %xmm0
pshufd $229, %xmm0, %xmm1
pxor %xmm0, %xmm1
movd %xmm1, %edx
cmpl %eax, %esi
movl %eax, %edi
je .LBB2_14
.p2align 4, 0x90
.LBB2_13:
xorl %edi, %edx
leal 1(%rdi), %eax
cmpl %eax, %esi
movl %eax, %edi
jne .LBB2_13
.LBB2_14:
movl %edx, 8(%rsp)
#APP
#NO_APP
cmpq %r9, %rcx
jne .LBB2_2
.LBB2_15:
leaq 8(%rsp), %rdi
leaq 24(%rsp), %rsi
callq _ZN3std4time7Instant7elapsed17h0b6076720ddfcc2bE@PLT
movq 8(%rsp), %rax
movl 16(%rsp), %ecx
movq %rax, 8(%r14)
movl %ecx, 16(%r14)
addq $40, %rsp
popq %rbx
popq %r14
retq
.Lfunc_end2:
.size _ZN8fpinrust5tests13xor_pointer_a17hca6c00df2b597ce7E, .Lfunc_end2-_ZN8fpinrust5tests13xor_pointer_a17hca6c00df2b597ce7E
.cfi_endproc
.section .text._ZN8fpinrust5tests13xor_pointer_b17h4ef66678d6655ef4E,"ax",@progbits
.globl _ZN8fpinrust5tests13xor_pointer_b17h4ef66678d6655ef4E
.p2align 4, 0x90
.type _ZN8fpinrust5tests13xor_pointer_b17h4ef66678d6655ef4E,@function
_ZN8fpinrust5tests13xor_pointer_b17h4ef66678d6655ef4E:
.cfi_startproc
pushq %rbx
.Ltmp15:
.cfi_def_cfa_offset 16
subq $32, %rsp
.Ltmp16:
.cfi_def_cfa_offset 48
.Ltmp17:
.cfi_offset %rbx, -16
movq %rdi, %rbx
leaq 16(%rsp), %rdi
callq _ZN3std4time7Instant3now17h37bccd496c61083dE@PLT
movq (%rbx), %rax
testq %rax, %rax
je .LBB3_3
leaq (%rsp), %rcx
.p2align 4, 0x90
.LBB3_2:
movl $1000, (%rsp)
#APP
#NO_APP
#APP
#NO_APP
decq %rax
jne .LBB3_2
.LBB3_3:
leaq (%rsp), %rdi
leaq 16(%rsp), %rsi
callq _ZN3std4time7Instant7elapsed17h0b6076720ddfcc2bE@PLT
movq (%rsp), %rax
movl 8(%rsp), %ecx
movq %rax, 8(%rbx)
movl %ecx, 16(%rbx)
addq $32, %rsp
popq %rbx
retq
.Lfunc_end3:
.size _ZN8fpinrust5tests13xor_pointer_b17h4ef66678d6655ef4E, .Lfunc_end3-_ZN8fpinrust5tests13xor_pointer_b17h4ef66678d6655ef4E
.cfi_endproc
.section .rodata.cst16,"aM",@progbits,16
.p2align 4
.LCPI4_0:
.long 0
.long 1
.long 2
.long 3
.LCPI4_1:
.long 4
.long 4
.long 4
.long 4
.LCPI4_2:
.long 8
.long 8
.long 8
.long 8
.LCPI4_3:
.long 12
.long 12
.long 12
.long 12
.LCPI4_4:
.long 16
.long 16
.long 16
.long 16
.LCPI4_5:
.long 20
.long 20
.long 20
.long 20
.LCPI4_6:
.long 24
.long 24
.long 24
.long 24
.LCPI4_7:
.long 28
.long 28
.long 28
.long 28
.LCPI4_8:
.long 32
.long 32
.long 32
.long 32
.LCPI4_9:
.long 36
.long 36
.long 36
.long 36
.LCPI4_10:
.long 40
.long 40
.long 40
.long 40
.section .text._ZN8fpinrust5tests13xor_black_box17h8af7e6f80fe83dc8E,"ax",@progbits
.globl _ZN8fpinrust5tests13xor_black_box17h8af7e6f80fe83dc8E
.p2align 4, 0x90
.type _ZN8fpinrust5tests13xor_black_box17h8af7e6f80fe83dc8E,@function
_ZN8fpinrust5tests13xor_black_box17h8af7e6f80fe83dc8E:
.cfi_startproc
pushq %rbx
.Ltmp18:
.cfi_def_cfa_offset 16
subq $32, %rsp
.Ltmp19:
.cfi_def_cfa_offset 48
.Ltmp20:
.cfi_offset %rbx, -16
movq %rdi, %rbx
leaq 16(%rsp), %rdi
callq _ZN3std4time7Instant3now17h37bccd496c61083dE@PLT
movq (%rbx), %rax
testq %rax, %rax
je .LBB4_5
xorl %ecx, %ecx
movdqa .LCPI4_1(%rip), %xmm9
movdqa .LCPI4_2(%rip), %xmm10
movdqa .LCPI4_3(%rip), %xmm11
movdqa .LCPI4_4(%rip), %xmm12
movdqa .LCPI4_5(%rip), %xmm13
movdqa .LCPI4_6(%rip), %xmm14
movdqa .LCPI4_7(%rip), %xmm15
movdqa .LCPI4_8(%rip), %xmm0
movdqa .LCPI4_9(%rip), %xmm1
movdqa .LCPI4_10(%rip), %xmm2
leaq (%rsp), %rdx
.p2align 4, 0x90
.LBB4_2:
pxor %xmm3, %xmm3
movl $1000, %esi
pxor %xmm4, %xmm4
movdqa .LCPI4_0(%rip), %xmm5
.p2align 4, 0x90
.LBB4_3:
movdqa %xmm5, %xmm6
paddd %xmm9, %xmm6
movdqa %xmm5, %xmm7
paddd %xmm10, %xmm7
pxor %xmm5, %xmm3
pxor %xmm4, %xmm6
movdqa %xmm5, %xmm4
paddd %xmm11, %xmm4
pxor %xmm7, %xmm3
movdqa %xmm5, %xmm7
paddd %xmm13, %xmm7
pxor %xmm4, %xmm7
movdqa %xmm5, %xmm4
paddd %xmm12, %xmm4
pxor %xmm6, %xmm7
movdqa %xmm5, %xmm6
paddd %xmm14, %xmm6
pxor %xmm4, %xmm6
movdqa %xmm5, %xmm8
paddd %xmm15, %xmm8
pxor %xmm6, %xmm3
movdqa %xmm5, %xmm4
paddd %xmm0, %xmm4
pxor %xmm4, %xmm3
movdqa %xmm5, %xmm4
paddd %xmm1, %xmm4
pxor %xmm8, %xmm4
pxor %xmm7, %xmm4
paddd %xmm2, %xmm5
addl $-40, %esi
jne .LBB4_3
pxor %xmm3, %xmm4
pshufd $78, %xmm4, %xmm3
pxor %xmm4, %xmm3
pshufd $229, %xmm3, %xmm4
pxor %xmm3, %xmm4
incq %rcx
movd %xmm4, (%rsp)
#APP
#NO_APP
#APP
#NO_APP
cmpq %rax, %rcx
jne .LBB4_2
.LBB4_5:
leaq (%rsp), %rdi
leaq 16(%rsp), %rsi
callq _ZN3std4time7Instant7elapsed17h0b6076720ddfcc2bE@PLT
movq (%rsp), %rax
movl 8(%rsp), %ecx
movq %rax, 8(%rbx)
movl %ecx, 16(%rbx)
addq $32, %rsp
popq %rbx
retq
.Lfunc_end4:
.size _ZN8fpinrust5tests13xor_black_box17h8af7e6f80fe83dc8E, .Lfunc_end4-_ZN8fpinrust5tests13xor_black_box17h8af7e6f80fe83dc8E
.cfi_endproc
.section ".note.GNU-stack","",@progbits
Looking at the assembly, xor_closure_a, xor_pointer_a and xor_black_box which all have relatively the same performance, evaluate to similar assembly code. Which is, by the way, a lot more instructions than the amount of ASM for poor xor_closure_b.
This is where my research ends. I would be glad if someone explains, why this blanket of assembly produced by xor_black_box evaluates faster than xor_closure_b. Or put it differently, why the version that should be optimized away by the compiler runs notably slower, and which benchmark variant I should trust?
rustc --version
rustc 1.13.0-nightly (378195665 2016-09-08)
cargo --version
cargo 0.13.0-nightly (afaffa1 2016-09-06)
@Francis Gagné did a fantastic job to investigate the issue in his answer. The steps that took me to reproduce his results are listed below.
Compile benchmarks and emit assembly:
cargo rustc --release -- --test --emit asm
Run produced executable to verify previous behavior:
./target/release/deps/xor --bench
running 2 tests
test tests::xor_closure_b ... bench: 925 ns/iter (+/- 9)
test tests::xor_pointer_b ... bench: 0 ns/iter (+/- 0)
test result: ok. 0 passed; 0 failed; 0 ignored; 2 measured
Compile patched assembly output ./target/release/deps/xor.s
export RUSTLIB=/path/to/lib/rustlib
gcc target/release/deps/xor.s $RUSTLIB/x86_64-unknown-linux-gnu/lib/*.rlib $RUSTLIB/x86_64-unknown-linux-gnu/lib/*.so -pthread -lpthread -lm -ldl
Run updated benchmarks:
./a.out --bench
running 2 tests
test tests::xor_closure_b ... bench: 1 ns/iter (+/- 0)
test tests::xor_pointer_b ... bench: 0 ns/iter (+/- 0)
test result: ok. 0 passed; 0 failed; 0 ignored; 2 measured
694
I'll focus on comparing xor_closure_b and xor_pointer_b, since they should have similar performance (i.e. they should both do nothing).
Update: I made a mistake in my initial analysis, as pointed out by @EOF, so I revised the text below.
First, let's look at the LLVM IR generated for those two functions. (I find LLVM IR easier to read than ASM because it's more structured.)
; Function Attrs: uwtable
define internal void @_ZN3xor13xor_closure_b17hb13913a8d2a27b06E(%"11.test::Bencher"* nocapture dereferenceable(32)) unnamed_addr #0 personality i32 (i32, i32, i64, %"8.unwind::libunwind::_Unwind_Exception"*, %"8.unwind::libunwind::_Unwind_Context"*)* @rust_eh_personality {
entry-block:
%dummy.i.i = alloca {}, align 8
%dummy.i.i.i = alloca i32, align 4
%start1.i = alloca %"1.std::time::Instant", align 8
%tmp_ret2.i = alloca %"1.std::time::Duration", align 8
%1 = bitcast %"1.std::time::Duration"* %tmp_ret2.i to i8*
call void @llvm.lifetime.start(i64 16, i8* %1)
%2 = bitcast %"1.std::time::Instant"* %start1.i to i8*
call void @llvm.lifetime.start(i64 16, i8* %2)
call void @_ZN3std4time7Instant3now17h37bccd496c61083dE(%"1.std::time::Instant"* noalias nocapture nonnull sret dereferenceable(16) %start1.i)
%3 = getelementptr inbounds %"11.test::Bencher", %"11.test::Bencher"* %0, i64 0, i32 0
%4 = load i64, i64* %3, align 8
%5 = icmp eq i64 %4, 0
br i1 %5, label %_ZN4test7Bencher4iter17h0cab611e22e5c5faE.exit, label %bb7.lr.ph.i
bb7.lr.ph.i: ; preds = %entry-block
%6 = bitcast i32* %dummy.i.i.i to i8*
%7 = bitcast {}* %dummy.i.i to i8*
br label %bb7.i
bb7.i: ; preds = %"_ZN3xor13xor_closure_b28_$u7b$$u7b$closure$u7d$$u7d$17hbaf8d82981c57ba0E.exit.i", %bb7.lr.ph.i
%iter.sroa.0.019.i = phi i64 [ 0, %bb7.lr.ph.i ], [ %11, %"_ZN3xor13xor_closure_b28_$u7b$$u7b$closure$u7d$$u7d$17hbaf8d82981c57ba0E.exit.i" ]
call void @llvm.lifetime.start(i64 4, i8* %6) #2
store i32 1000, i32* %dummy.i.i.i, align 4
call void asm "", "r,~{dirflag},~{fpsr},~{flags}"(i32* nonnull %dummy.i.i.i) #2, !srcloc !1
%8 = load i32, i32* %dummy.i.i.i, align 4
call void @llvm.lifetime.end(i64 4, i8* %6) #2
br label %bb7.i.i.i
bb7.i.i.i: ; preds = %bb7.i.i.i, %bb7.i
%iter.sroa.0.0.i.i.i = phi i32 [ 0, %bb7.i ], [ %iter.sroa.0.1.i.i.i, %bb7.i.i.i ]
%9 = icmp slt i32 %iter.sroa.0.0.i.i.i, %8
%10 = zext i1 %9 to i32
%iter.sroa.0.1.i.i.i = add i32 %10, %iter.sroa.0.0.i.i.i
br i1 %9, label %bb7.i.i.i, label %"_ZN3xor13xor_closure_b28_$u7b$$u7b$closure$u7d$$u7d$17hbaf8d82981c57ba0E.exit.i"
"_ZN3xor13xor_closure_b28_$u7b$$u7b$closure$u7d$$u7d$17hbaf8d82981c57ba0E.exit.i": ; preds = %bb7.i.i.i
%11 = add nuw i64 %iter.sroa.0.019.i, 1
call void @llvm.lifetime.start(i64 0, i8* %7)
call void asm "", "r,~{dirflag},~{fpsr},~{flags}"({}* nonnull %dummy.i.i) #2, !srcloc !1
call void @llvm.lifetime.end(i64 0, i8* %7)
%exitcond.i = icmp eq i64 %11, %4
br i1 %exitcond.i, label %_ZN4test7Bencher4iter17h0cab611e22e5c5faE.exit.loopexit, label %bb7.i
_ZN4test7Bencher4iter17h0cab611e22e5c5faE.exit.loopexit: ; preds = %"_ZN3xor13xor_closure_b28_$u7b$$u7b$closure$u7d$$u7d$17hbaf8d82981c57ba0E.exit.i"
br label %_ZN4test7Bencher4iter17h0cab611e22e5c5faE.exit
_ZN4test7Bencher4iter17h0cab611e22e5c5faE.exit: ; preds = %_ZN4test7Bencher4iter17h0cab611e22e5c5faE.exit.loopexit, %entry-block
call void @_ZN3std4time7Instant7elapsed17h0b6076720ddfcc2bE(%"1.std::time::Duration"* noalias nocapture nonnull sret dereferenceable(16) %tmp_ret2.i, %"1.std::time::Instant"* noalias nonnull readonly dereferenceable(16) %start1.i)
%12 = getelementptr inbounds %"1.std::time::Duration", %"1.std::time::Duration"* %tmp_ret2.i, i64 0, i32 0
%13 = getelementptr inbounds %"1.std::time::Duration", %"1.std::time::Duration"* %tmp_ret2.i, i64 0, i32 1
%14 = load i64, i64* %12, align 8
%15 = load i32, i32* %13, align 8
%16 = getelementptr inbounds %"11.test::Bencher", %"11.test::Bencher"* %0, i64 0, i32 1, i32 0
store i64 %14, i64* %16, align 8
%17 = getelementptr inbounds %"11.test::Bencher", %"11.test::Bencher"* %0, i64 0, i32 1, i32 1
store i32 %15, i32* %17, align 4
call void @llvm.lifetime.end(i64 16, i8* %2)
call void @llvm.lifetime.end(i64 16, i8* %1)
ret void
}
; Function Attrs: uwtable
define internal void @_ZN3xor13xor_pointer_b17h7ba0f9760d9fd9f8E(%"11.test::Bencher"* nocapture dereferenceable(32)) unnamed_addr #0 personality i32 (i32, i32, i64, %"8.unwind::libunwind::_Unwind_Exception"*, %"8.unwind::libunwind::_Unwind_Context"*)* @rust_eh_personality {
entry-block:
%dummy.i.i = alloca {}, align 8
%dummy.i.i.i = alloca i32, align 4
%start1.i = alloca %"1.std::time::Instant", align 8
%tmp_ret2.i = alloca %"1.std::time::Duration", align 8
%1 = bitcast %"1.std::time::Duration"* %tmp_ret2.i to i8*
call void @llvm.lifetime.start(i64 16, i8* %1)
%2 = bitcast %"1.std::time::Instant"* %start1.i to i8*
call void @llvm.lifetime.start(i64 16, i8* %2)
call void @_ZN3std4time7Instant3now17h37bccd496c61083dE(%"1.std::time::Instant"* noalias nocapture nonnull sret dereferenceable(16) %start1.i)
%3 = getelementptr inbounds %"11.test::Bencher", %"11.test::Bencher"* %0, i64 0, i32 0
%4 = load i64, i64* %3, align 8
%5 = icmp eq i64 %4, 0
br i1 %5, label %_ZN4test7Bencher4iter17hae343b1316e5897bE.exit, label %bb7.lr.ph.i
bb7.lr.ph.i: ; preds = %entry-block
%6 = bitcast i32* %dummy.i.i.i to i8*
%7 = bitcast {}* %dummy.i.i to i8*
br label %bb7.i
bb7.i: ; preds = %bb7.i, %bb7.lr.ph.i
%iter.sroa.0.019.i = phi i64 [ 0, %bb7.lr.ph.i ], [ %8, %bb7.i ]
%8 = add nuw i64 %iter.sroa.0.019.i, 1
call void @llvm.lifetime.start(i64 4, i8* %6) #2
store i32 1000, i32* %dummy.i.i.i, align 4
call void asm "", "r,~{dirflag},~{fpsr},~{flags}"(i32* nonnull %dummy.i.i.i) #2, !srcloc !1
call void @llvm.lifetime.end(i64 4, i8* %6) #2
call void @llvm.lifetime.start(i64 0, i8* %7)
call void asm "", "r,~{dirflag},~{fpsr},~{flags}"({}* nonnull %dummy.i.i) #2, !srcloc !1
call void @llvm.lifetime.end(i64 0, i8* %7)
%exitcond.i = icmp eq i64 %8, %4
br i1 %exitcond.i, label %_ZN4test7Bencher4iter17hae343b1316e5897bE.exit.loopexit, label %bb7.i
_ZN4test7Bencher4iter17hae343b1316e5897bE.exit.loopexit: ; preds = %bb7.i
br label %_ZN4test7Bencher4iter17hae343b1316e5897bE.exit
_ZN4test7Bencher4iter17hae343b1316e5897bE.exit: ; preds = %_ZN4test7Bencher4iter17hae343b1316e5897bE.exit.loopexit, %entry-block
call void @_ZN3std4time7Instant7elapsed17h0b6076720ddfcc2bE(%"1.std::time::Duration"* noalias nocapture nonnull sret dereferenceable(16) %tmp_ret2.i, %"1.std::time::Instant"* noalias nonnull readonly dereferenceable(16) %start1.i)
%9 = getelementptr inbounds %"1.std::time::Duration", %"1.std::time::Duration"* %tmp_ret2.i, i64 0, i32 0
%10 = getelementptr inbounds %"1.std::time::Duration", %"1.std::time::Duration"* %tmp_ret2.i, i64 0, i32 1
%11 = load i64, i64* %9, align 8
%12 = load i32, i32* %10, align 8
%13 = getelementptr inbounds %"11.test::Bencher", %"11.test::Bencher"* %0, i64 0, i32 1, i32 0
store i64 %11, i64* %13, align 8
%14 = getelementptr inbounds %"11.test::Bencher", %"11.test::Bencher"* %0, i64 0, i32 1, i32 1
store i32 %12, i32* %14, align 4
call void @llvm.lifetime.end(i64 16, i8* %2)
call void @llvm.lifetime.end(i64 16, i8* %1)
ret void
}
If we compare the LLVM IR for xor_closure_b and xor_pointer_b, they look pretty similar. However, one difference stands out: the bb7.i.i.i block has been optimized out in xor_pointer_b, but not in xor_closure_b, for some reason. Here's the block:
bb7.i.i.i: ; preds = %bb7.i.i.i, %bb7.i
%iter.sroa.0.0.i.i.i = phi i32 [ 0, %bb7.i ], [ %iter.sroa.0.1.i.i.i, %bb7.i.i.i ]
%9 = icmp slt i32 %iter.sroa.0.0.i.i.i, %8
%10 = zext i1 %9 to i32
%iter.sroa.0.1.i.i.i = add i32 %10, %iter.sroa.0.0.i.i.i
br i1 %9, label %bb7.i.i.i, label %"_ZN3xor13xor_closure_b28_$u7b$$u7b$closure$u7d$$u7d$17hbaf8d82981c57ba0E.exit.i"
And here it is translated to ASM:
.LBB1_3:
xorl %edi, %edi
cmpl %esi, %ebx
setl %dil
addl %ebx, %edi
cmpl %esi, %ebx
movl %edi, %ebx
jl .LBB1_3
This is a pretty dumb way to loop from 0 to 1000. Changing the code above to this:
.LBB1_3:
cmpl %esi, %ebx
jge .LBB1_3a
incl %ebx
cmpl %esi, %ebx
jl .LBB1_3
.LBB1_3a:
makes the benchmark for xor_closure_b drop from 781 ns/iter (+/- 19) to 270 ns/iter (+/- 7) on my machine.
I cannot say for sure why the code generated by the compiler is so slow, or why it wasn't optimized away in the first place (as it was in xor_pointer_b)... However, it seems that xor_pointer_a and xor_closure_a and faster still because the generated code is vectorized, which has the effect that the loop performs fewer iterations (i.e. the loop is unrolled), by a factor of 32 (e.g. .LBB0_8, the main loop in xor_closure_a, performs 31 iterations, then the rest is handled after the loop).
For reference, I compiled the edited ASM with this command line:
$ gcc target/release/xor-71758a2519026d86.s ~/.multirust/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/x86_64-unknown-linux-gnu/lib/lib{test,term,getopts,rustc_unicode,std,libc,rand,collections,alloc_system,alloc,core,panic_unwind}-411f48d3.rlib -pthread -lpthread -lm -ldl
and I ran it with ./a.out --bench. Also, my CPU is an Intel Core i7-4770K.
173
If you love us? You can donate to us via Paypal or buy me a coffee so we can maintain and grow! Thank you!
Donate Us With
