x64: reduce work in non-stack using leaf functions #5352

abrown · 2022-11-30T01:03:08Z

Cranelift has had the ability for some time to identify leaf functions;
by Cranelift's definition, a leaf function is one that knows of no other
call signatures. #1148 noted how it would be a good idea to avoid extra
frame setup work in leaf functions and #2960 implemented this for
aarch64 and s390x. This improvement was not made for x64 due to some
test failures. This change avoids any frame setup for non-stack-using
leaf functions in x64.

Because this updates the generated code, there are multiple sets of tests
to update -- each of these is separated here into its own commit. I will follow
up in the comments with benchmark results.

abrown · 2022-11-30T01:23:35Z

Here is the raw output of the benchmarking I did on my current system. It is hard to interpret the results (is it too noisy? not enough runs?) but here they are!

with the default cycles measurement:

$ taskset --cpu-list 4-5 target/release/sightglass-cli benchmark -e /tmp/main/libengine.so -e /tmp/x64-leaf-functions/libengine.so instantiation :: cycles :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [147078 176131.13 996137] main/libengine.so [145930 167314.60 252683] x64-leaf-functions/libengine.so instantiation :: cycles :: benchmarks/bz2/benchmark.wasm No difference in performance. [112360 132696.79 208371] main/libengine.so [112495 136652.85 230477] x64-leaf-functions/libengine.so instantiation :: cycles :: benchmarks/spidermonkey/benchmark.wasm No difference in performance. [368797 405530.78 502881] main/libengine.so [371210 400482.08 545122] x64-leaf-functions/libengine.so compilation :: cycles :: benchmarks/bz2/benchmark.wasm No difference in performance. [209965278 214840438.75 230091486] main/libengine.so [208568512 213673819.47 231217610] x64-leaf-functions/libengine.so execution :: cycles :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [5163256 5316183.14 5716048] main/libengine.so [5176338 5307186.55 5609513] x64-leaf-functions/libengine.so execution :: cycles :: benchmarks/spidermonkey/benchmark.wasm No difference in performance. [520827395 523585282.78 529830174] main/libengine.so [520812995 524419123.61 533595170] x64-leaf-functions/libengine.so compilation :: cycles :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [394379561 403469591.05 421661750] main/libengine.so [395049806 403016360.00 416050865] x64-leaf-functions/libengine.so execution :: cycles :: benchmarks/bz2/benchmark.wasm No difference in performance. [72041670 73242723.27 75147038] main/libengine.so [72171609 73209119.80 74873711] x64-leaf-functions/libengine.so compilation :: cycles :: benchmarks/spidermonkey/benchmark.wasm No difference in performance.

[9360042865 9429678348.56 9517907602] main/libengine.so [9375595405 9427505767.96 9512712326] x64-leaf-functions/libengine.so

specifying the `perf-counters` measurement:

$ taskset --cpu-list 4-5 target/release/sightglass-cli benchmark -e /tmp/main/libengine.so -e /tmp/x64-leaf-functions/libengine.so --measure perf-counters instantiation :: instructions-retired :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [34327 37575.33 71666] main/libengine.so [34070 36209.87 61232] x64-leaf-functions/libengine.so execution :: cache-misses :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [737 1490.90 3397] main/libengine.so [749 1533.36 3791] x64-leaf-functions/libengine.so instantiation :: cpu-cycles :: benchmarks/bz2/benchmark.wasm No difference in performance. [87330 108012.70 140941] main/libengine.so [84135 110894.47 243127] x64-leaf-functions/libengine.so instantiation :: cache-accesses :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [2719 3153.40 6610] main/libengine.so [2754 3082.12 4528] x64-leaf-functions/libengine.so instantiation :: cache-accesses :: benchmarks/bz2/benchmark.wasm No difference in performance. [3122 4161.20 7037] main/libengine.so [2782 4070.35 5860] x64-leaf-functions/libengine.so compilation :: cache-misses :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [43175 63266.12 88724] main/libengine.so [45842 64507.86 80471] x64-leaf-functions/libengine.so instantiation :: cpu-cycles :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [78651 91367.71 218899] main/libengine.so [79408 89770.91 211968] x64-leaf-functions/libengine.so instantiation :: cache-misses :: benchmarks/bz2/benchmark.wasm No difference in performance. [1407 1752.31 2235] main/libengine.so [1436 1730.65 2168] x64-leaf-functions/libengine.so execution :: cache-misses :: benchmarks/bz2/benchmark.wasm No difference in performance. [3077 5807.68 10172] main/libengine.so [3144 5877.77 13198] x64-leaf-functions/libengine.so compilation :: cpu-cycles :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [6155268 6616904.93 7457577] main/libengine.so [6126225 6690847.15 7678328] x64-leaf-functions/libengine.so execution :: cache-accesses :: benchmarks/bz2/benchmark.wasm No difference in performance. [81061 102216.36 143392] main/libengine.so [77609 103227.71 147886] x64-leaf-functions/libengine.so compilation :: cache-misses :: benchmarks/spidermonkey/benchmark.wasm No difference in performance. [1651915 1872738.93 2029320] main/libengine.so [1664784 1889497.05 2135265] x64-leaf-functions/libengine.so execution :: cache-accesses :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [7158 9385.02 13399] main/libengine.so [7644 9313.02 12838] x64-leaf-functions/libengine.so compilation :: cpu-cycles :: benchmarks/bz2/benchmark.wasm No difference in performance. [2128438 2336464.61 3559953] main/libengine.so [2140962 2353988.06 3179803] x64-leaf-functions/libengine.so compilation :: cache-accesses :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [212241 232225.91 244981] main/libengine.so [212462 230856.73 244761] x64-leaf-functions/libengine.so execution :: cache-misses :: benchmarks/spidermonkey/benchmark.wasm No difference in performance. [60862 77279.98 122802] main/libengine.so [58255 77696.47 123312] x64-leaf-functions/libengine.so instantiation :: cpu-cycles :: benchmarks/spidermonkey/benchmark.wasm No difference in performance. [104568 116986.62 131490] main/libengine.so [99399 117542.93 247601] x64-leaf-functions/libengine.so instantiation :: cache-accesses :: benchmarks/spidermonkey/benchmark.wasm No difference in performance. [3808 4391.63 4937] main/libengine.so [3876 4372.42 4783] x64-leaf-functions/libengine.so compilation :: cache-misses :: benchmarks/bz2/benchmark.wasm No difference in performance. [21638 26831.51 35754] main/libengine.so [21900 26730.18 36534] x64-leaf-functions/libengine.so instantiation :: cache-misses :: benchmarks/spidermonkey/benchmark.wasm No difference in performance. [1936 2156.04 2496] main/libengine.so [1836 2149.08 2508] x64-leaf-functions/libengine.so compilation :: cpu-cycles :: benchmarks/spidermonkey/benchmark.wasm No difference in performance. [165986856 171495577.67 179161139] main/libengine.so [166456173 172035488.63 181114155] x64-leaf-functions/libengine.so compilation :: cache-accesses :: benchmarks/spidermonkey/benchmark.wasm No difference in performance. [4855667 5045583.38 5187008] main/libengine.so [4835172 5061308.56 5188155] x64-leaf-functions/libengine.so instantiation :: cache-misses :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [1376 1507.87 1900] main/libengine.so [1394 1510.24 1704] x64-leaf-functions/libengine.so instantiation :: instructions-retired :: benchmarks/bz2/benchmark.wasm No difference in performance. [43563 54983.42 71508] main/libengine.so [36901 55061.22 65926] x64-leaf-functions/libengine.so compilation :: cache-accesses :: benchmarks/bz2/benchmark.wasm No difference in performance. [72326 76693.08 84404] main/libengine.so [72462 76612.53 82601] x64-leaf-functions/libengine.so execution :: cache-accesses :: benchmarks/spidermonkey/benchmark.wasm No difference in performance. [531146 553039.20 577617] main/libengine.so [528802 552658.15 590599] x64-leaf-functions/libengine.so execution :: cpu-cycles :: benchmarks/spidermonkey/benchmark.wasm No difference in performance. [824144011 827131703.61 841908153] main/libengine.so [824077161 826662188.48 838574889] x64-leaf-functions/libengine.so execution :: cpu-cycles :: benchmarks/bz2/benchmark.wasm No difference in performance. [112027942 112932466.82 115093344] main/libengine.so [112063326 112871592.82 114011536] x64-leaf-functions/libengine.so execution :: cpu-cycles :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [7555372 7635027.45 7825127] main/libengine.so [7547343 7631211.45 7825383] x64-leaf-functions/libengine.so compilation :: instructions-retired :: benchmarks/spidermonkey/benchmark.wasm No difference in performance. [207935281 208620688.17 209390775] main/libengine.so [208127260 208673878.55 209613081] x64-leaf-functions/libengine.so compilation :: instructions-retired :: benchmarks/bz2/benchmark.wasm No difference in performance. [2479620 2550178.19 3062957] main/libengine.so [2475971 2550707.19 3058800] x64-leaf-functions/libengine.so compilation :: instructions-retired :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [8860156 8959184.54 9418626] main/libengine.so [8857013 8960071.13 9438139] x64-leaf-functions/libengine.so instantiation :: instructions-retired :: benchmarks/spidermonkey/benchmark.wasm No difference in performance. [43827 44870.44 45809] main/libengine.so [44017 44867.48 46051] x64-leaf-functions/libengine.so execution :: instructions-retired :: benchmarks/spidermonkey/benchmark.wasm No difference in performance. [3203656920 3203780565.63 3204071799] main/libengine.so [3203655116 3203783546.14 3204082241] x64-leaf-functions/libengine.so execution :: instructions-retired :: benchmarks/pulldown-cmark/benchmark.wasm No difference in performance. [26394462 26394717.19 26395376] main/libengine.so [26394464 26394724.88 26395349] x64-leaf-functions/libengine.so execution :: instructions-retired :: benchmarks/bz2/benchmark.wasm No difference in performance.

[319667712 319667714.29 319667716] main/libengine.so [319667712 319667714.26 319667716] x64-leaf-functions/libengine.so

cfallin

The changes to enable this are pleasantly simple (I guess because we already have the generic infra for it) -- thanks! Seeing 1.8k lines of prologues/epilogues disappear in the tests is nice.

Happy to r+ once the test failures are resolved (unfortunately the failure on CI does not seem to be very informative; just a segfault in filetests)...

cranelift/codegen/src/isa/x64/mod.rs

Cranelift has had the ability for some time to identify leaf functions; by Cranelift's definition, a leaf function is one that knows of no other call signatures. bytecodealliance#1148 noted how it would be a good idea to avoid extra frame setup work in leaf functions and bytecodealliance#2960 implemented this for aarch64 and s390x. This improvement was not made for x64 due to some test failures. This change avoids any frame setup for non-stack-using leaf functions in x64.

bjorn3 · 2022-12-02T11:17:07Z

cranelift/codegen/src/isa/x64/abi.rs

+            || num_clobbered_callee_saves > 0
+            || frame_storage_size > 0)
+            // TODO
+            && !cfg!(target_os = "macos")


Maybe my suggestion at #4469 (comment) would work? "non-leaf" could be used by default and "always" could be used on macOS.

github-actions bot added cranelift Issues related to the Cranelift code generator cranelift:area:x64 Issues related to x64 codegen labels Nov 30, 2022

cfallin reviewed Nov 30, 2022

View reviewed changes

cranelift/codegen/src/isa/x64/mod.rs Outdated Show resolved Hide resolved

abrown force-pushed the x64-leaf-functions branch 2 times, most recently from cf67ec5 to 1cc2b42 Compare November 30, 2022 23:06

abrown added 4 commits December 1, 2022 10:40

test: update golden output tests

094c665

test: update systemv unwinding test output

bd011d8

test: update CLIF precise output

b2d3e98

abrown force-pushed the x64-leaf-functions branch from 1cc2b42 to 6e15742 Compare December 1, 2022 18:41

fix: disable frame setup elision on macos

d9945d4

abrown force-pushed the x64-leaf-functions branch from 6e15742 to d9945d4 Compare December 1, 2022 19:49

bjorn3 reviewed Dec 2, 2022

View reviewed changes

tertsdiepraam mentioned this pull request May 1, 2024

WIP: cranelift: omit function prologue and epilogue on x64 if possible #8516

Draft

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

x64: reduce work in non-stack using leaf functions #5352

x64: reduce work in non-stack using leaf functions #5352

abrown commented Nov 30, 2022

abrown commented Nov 30, 2022

cfallin left a comment

bjorn3 Dec 2, 2022

x64: reduce work in non-stack using leaf functions #5352

Are you sure you want to change the base?

x64: reduce work in non-stack using leaf functions #5352

Conversation

abrown commented Nov 30, 2022

abrown commented Nov 30, 2022

cfallin left a comment

Choose a reason for hiding this comment

bjorn3 Dec 2, 2022

Choose a reason for hiding this comment