MIR 4x slower

[MagaTailor]

Running the benchmark suite from ethcore/parity compiled with a recent nightly, produces mostly equal results, except for one type of benchmark which seems to be hit:

MIR:

test bench_stream_1000_empty_lists   ... bench:     458,104 ns/iter (+/- 21,291)
test bench_stream_nested_empty_lists ... bench:      16,640 ns/iter (+/- 500)
test bench_stream_u256_value         ... bench:      26,468 ns/iter (+/- 353)
test bench_stream_u64_value          ... bench:      13,893 ns/iter (+/- 272)

Ye Olde:

test bench_stream_1000_empty_lists   ... bench:      91,131 ns/iter (+/- 3,154)
test bench_stream_nested_empty_lists ... bench:       4,702 ns/iter (+/- 160)
test bench_stream_u256_value         ... bench:       6,478 ns/iter (+/- 383)
test bench_stream_u64_value          ... bench:       3,047 ns/iter (+/- 132)

Found on ARM Linux but hopefully not exclusive to that platform. Not sure if MIR should be held accountable at this early stage but here it is.

[alexcrichton]

cc @rust-lang/compiler

[nikomatsakis]

@petevine great, thanks! I want to get more organized about MIR benchmarks for runtime in particular.

[nagisa]

x86_64:

Orbit

test bench_stream_1000_empty_lists   ... bench:      26,937 ns/iter (+/- 3,985)
test bench_stream_nested_empty_lists ... bench:         842 ns/iter (+/- 131)
test bench_stream_u256_value         ... bench:       1,405 ns/iter (+/- 54)
test bench_stream_u64_value          ... bench:         582 ns/iter (+/- 19)

time to run ~/.cargo/bin/rustup run nightly cargo bench -p ethcore-util --bench rlp -j1
99% (600.02 real, 13.73 kernel, 581.89 user); 468572k resident

Plain old

test bench_stream_1000_empty_lists   ... bench:      11,147 ns/iter (+/- 362)
test bench_stream_nested_empty_lists ... bench:         278 ns/iter (+/- 247)
test bench_stream_u256_value         ... bench:         609 ns/iter (+/- 40)
test bench_stream_u64_value          ... bench:         154 ns/iter (+/- 13)

time to run ~/.cargo/bin/rustup run nightly cargo bench -p ethcore-util --bench rlp -j1
99% (563.41 real, 14.16 kernel, 544.92 user); 454708k resident

While the difference is not as big it is still very noticeable.

[nagisa]

More recent results:

test bench_stream_1000_empty_lists   ... bench:      23,499 ns/iter (+/- 1,156)
test bench_stream_nested_empty_lists ... bench:         748 ns/iter (+/- 30)
test bench_stream_u256_value         ... bench:       1,294 ns/iter (+/- 45)
test bench_stream_u64_value          ... bench:         563 ns/iter (+/- 16)

99% (527.78 real, 28.33 kernel, 497.98 user); 467064k resident

Thus, a small imporovement, but still a big regression. Waiting for the nightly with static drops to see if and how much that helps.

[eddyb]

@nagisa Suppose I can try with my build of master.

[eddyb]

I get the following results with -Z orbit:

test bench_stream_1000_empty_lists   ... bench:      11,624 ns/iter (+/- 632)
test bench_stream_nested_empty_lists ... bench:         398 ns/iter (+/- 4)
test bench_stream_u256_value         ... bench:         631 ns/iter (+/- 5)
test bench_stream_u64_value          ... bench:         240 ns/iter (+/- 2)

EDIT: With old trans:

test bench_stream_1000_empty_lists   ... bench:      11,544 ns/iter (+/- 88)
test bench_stream_nested_empty_lists ... bench:         486 ns/iter (+/- 4)
test bench_stream_u256_value         ... bench:         731 ns/iter (+/- 22)
test bench_stream_u64_value          ... bench:         305 ns/iter (+/- 3)

[alexbool]

@eddyb It would be also nice to see the results of old trans on your hardware

[eddyb]

@alexbool Yes, I just added those, they hadn't finished yet.

EDIT: Wait, I might not have done that right, let me make sure.
EDIT2: Phew, for a second there I thought I was on #34096. Results are valid.

[alexbool]

@eddyb thanks, looks extremely promising

[nagisa]

These are my results for -Z orbit (the non-orbit are above) on master:

test bench_stream_1000_empty_lists   ... bench:       9,954 ns/iter (+/- 626)
test bench_stream_nested_empty_lists ... bench:         267 ns/iter (+/- 17)
test bench_stream_u256_value         ... bench:         574 ns/iter (+/- 52)
test bench_stream_u64_value          ... bench:         163 ns/iter (+/- 19)

99% (454.66 real, 8.66 kernel, 443.18 user); 437932k resident

thus a considerable speed-up in both run and compile times. Considered resolved, thus closing.

[MagaTailor]

A quick heads-up, Ye Olde strikes back in bigint! (actually MIR trans regresses, for reference see the final part of this comment)

name                     gcc6-llvm3.9-mir ns/iter       gcc6-llvm3.9-old ns/iter                    diff ns/iter   diff %
u128_mul                         1,046,807                     629,104                               -417,703  -39.90%
u256_add                         1,186,208                     628,804                               -557,404  -46.99%
u256_full_mul                    21,829,551                    21,381,247                            -448,304   -2.05%
u256_mul                         1,699,611                     1,198,108                             -501,503  -29.51%
u256_sub                         1,186,208                     628,804                               -557,404  -46.99%
u512_add                         1,015,707                     1,016,607                                  900    0.09%
u512_sub                         1,066,007                     1,066,007                                    0    0.00%