JSBuiltinReducer::ReduceStringIndexOf: String type unknown if indexOf call is wrapped in function |
|||||||
Issue description
Version: ToT
I came across this while looking at a String.p.indexOf microbenchmark (attached).
JSBuiltinReducer::ReduceStringIndexOf [0] bails out on the
if (search_string_type->Is(Type::String()) &&
position_type->Is(Type::SignedSmall()))
check when the indexOf loop is wrapped in a function, causing at least a 5x increase in runtime (less is better):
out/release/d8 ~/bench-string-p-indexof-in-function.js -- aaaaaaaaaaaaaaaaaaaab a // 1158.757
out/release/d8 ~/bench-string-p-indexof.js -- aaaaaaaaaaaaaaaaaaaab a // 226.996
The difference can become much worse with other args:
out/release/d8 ~/bench-string-p-indexof-in-function.js -- aaaaaaaaaaaaaaaaaaaab ab // 14226.949
out/release/d8 ~/bench-string-p-indexof.js -- aaaaaaaaaaaaaaaaaaaab ab // 213.938
Looks to me like this happens because search_string_type cannot be inferred anymore. Is this expected behavior?
[0] https://cs.chromium.org/chromium/src/v8/src/compiler/js-builtin-reducer.cc?type=cs&q=js-builtin-+package:%5Echromium$&l=1882
,
Apr 18 2017
,
Apr 18 2017
We currently don't have feedback on CallIC inputs. We just discussed a more complex CallIC today, that would be able to collect different feedback, i.e. String feedback on the searchString input.
,
Apr 19 2017
In this example, lowering also results in omission of the (pure) StringIndexOf operator after the first ~10k loop iterations, while the non-lowered version performs 10^8 calls to the StringIndexOf builtin. That explains the huge runtime difference.
,
Dec 11 2017
Once we have the optimization guards we could speculatively assume that the inputs to String.prototype.indexOf are both strings (by inserting CheckString) and then learn via the deoptimization guard. This seems like a reasonable approach. We should also move the lowering to JSCallReducer in that case maybe.
,
Dec 11 2017
Sigurd, I'm assigning this to you, as it's follow-up work from the builtin optimization guards. It's not high priority and should be fairly straight-forward.
,
Dec 18 2017
The following revision refers to this bug: https://chromium.googlesource.com/v8/v8.git/+/8e8d8623a72166ff5c34d5e2a73611fa9421aabd commit 8e8d8623a72166ff5c34d5e2a73611fa9421aabd Author: Sigurd Schneider <sigurds@chromium.org> Date: Mon Dec 18 16:47:28 2017 [turbofan] Add feedback to CheckString node This CL allows deopts from CheckString to disable speculation. Bug: v8:7127 , v8:6270 Change-Id: I029caeb61c509e5eb51b169ac42596d632f7c75a Reviewed-on: https://chromium-review.googlesource.com/831866 Commit-Queue: Sigurd Schneider <sigurds@chromium.org> Reviewed-by: Benedikt Meurer <bmeurer@chromium.org> Cr-Commit-Position: refs/heads/master@{#50172} [modify] https://crrev.com/8e8d8623a72166ff5c34d5e2a73611fa9421aabd/src/compiler/effect-control-linearizer.cc [modify] https://crrev.com/8e8d8623a72166ff5c34d5e2a73611fa9421aabd/src/compiler/js-native-context-specialization.cc [modify] https://crrev.com/8e8d8623a72166ff5c34d5e2a73611fa9421aabd/src/compiler/js-typed-lowering.cc [modify] https://crrev.com/8e8d8623a72166ff5c34d5e2a73611fa9421aabd/src/compiler/property-access-builder.cc [modify] https://crrev.com/8e8d8623a72166ff5c34d5e2a73611fa9421aabd/src/compiler/simplified-operator.cc [modify] https://crrev.com/8e8d8623a72166ff5c34d5e2a73611fa9421aabd/src/compiler/simplified-operator.h
,
Dec 19 2017
The following revision refers to this bug: https://chromium.googlesource.com/v8/v8.git/+/5e18f849535b920da1f2194ea1b56e5ecaa3f3ea commit 5e18f849535b920da1f2194ea1b56e5ecaa3f3ea Author: Sigurd Schneider <sigurds@chromium.org> Date: Tue Dec 19 10:54:37 2017 [turbofan] Add benchmarks for String.indexOf Bug: v8:7127 , v8:6270 Change-Id: Ic35a9b7a5145115736934b0c7de6ace26e9c0e51 Reviewed-on: https://chromium-review.googlesource.com/832966 Commit-Queue: Sigurd Schneider <sigurds@chromium.org> Reviewed-by: Jakob Gruber <jgruber@chromium.org> Reviewed-by: Benedikt Meurer <bmeurer@chromium.org> Cr-Commit-Position: refs/heads/master@{#50191} [modify] https://crrev.com/5e18f849535b920da1f2194ea1b56e5ecaa3f3ea/test/js-perf-test/JSTests.json [modify] https://crrev.com/5e18f849535b920da1f2194ea1b56e5ecaa3f3ea/test/js-perf-test/Strings/run.js [add] https://crrev.com/5e18f849535b920da1f2194ea1b56e5ecaa3f3ea/test/js-perf-test/Strings/string-indexof.js
,
Dec 22 2017
The following revision refers to this bug: https://chromium.googlesource.com/v8/v8.git/+/b44f820559d6f691d68fc07165ceae3201ad0509 commit b44f820559d6f691d68fc07165ceae3201ad0509 Author: Sigurd Schneider <sigurds@chromium.org> Date: Fri Dec 22 08:13:47 2017 [turbofan] Move String.indexOf inlining to call reducer This also adds speculative checks asserting that all arguments are of the right types; each check disables speculation if it fails. Bug: v8:7127 , v8:6270 Change-Id: Ifcb8bc509b86c712f0fab50ef1caee0c3a289e86 Reviewed-on: https://chromium-review.googlesource.com/832449 Commit-Queue: Sigurd Schneider <sigurds@chromium.org> Reviewed-by: Benedikt Meurer <bmeurer@chromium.org> Cr-Commit-Position: refs/heads/master@{#50289} [modify] https://crrev.com/b44f820559d6f691d68fc07165ceae3201ad0509/src/compiler/js-builtin-reducer.cc [modify] https://crrev.com/b44f820559d6f691d68fc07165ceae3201ad0509/src/compiler/js-builtin-reducer.h [modify] https://crrev.com/b44f820559d6f691d68fc07165ceae3201ad0509/src/compiler/js-call-reducer.cc [modify] https://crrev.com/b44f820559d6f691d68fc07165ceae3201ad0509/src/compiler/js-call-reducer.h [modify] https://crrev.com/b44f820559d6f691d68fc07165ceae3201ad0509/src/compiler/typer.cc
,
Jan 10
I just verified, this has been fixed. ToT shows: out.gn/x64.release/d8 bench-string-p-indexof-in-function.js -- aaaaaaaaaaaaaaaaaaaab ab // 122.34799999999998 out.gn/x64.release/d8 bench-string-p-indexof.js -- aaaaaaaaaaaaaaaaaaaab ab // 304.869
,
Jan 10
Very nice! |
|||||||
►
Sign in to add a comment |
|||||||
Comment 1 by jarin@chromium.org
, Apr 18 2017Owner: bmeu...@chromium.org