Stream: git-wasmtime

Topic: wasmtime / PR #2277 machinst x64: a few small opts

view this post on Zulip Wasmtime GitHub notifications bot (Oct 07 2020 at 16:54):

bnjbvr opened PR #2277 from x64-opt to main.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 07 2020 at 16:55):

bnjbvr edited PR #2277 from x64-opt to main:

(I'd like to test it a bit more before considering review, but: this already enhances the code quality a bit on some hot blocks identified with Valgrind.)

view this post on Zulip Wasmtime GitHub notifications bot (Oct 08 2020 at 16:25):

bnjbvr updated PR #2277 from x64-opt to main:

(I'd like to test it a bit more before considering review, but: this already enhances the code quality a bit on some hot blocks identified with Valgrind.)

view this post on Zulip Wasmtime GitHub notifications bot (Oct 08 2020 at 16:26):

bnjbvr requested cfallin for a review on PR #2277.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 08 2020 at 16:26):

bnjbvr requested cfallin and julian-seward1 for a review on PR #2277.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 08 2020 at 16:26):

bnjbvr has marked PR #2277 as ready for review.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 08 2020 at 17:11):

cfallin submitted PR Review.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 08:51):

julian-seward1 submitted PR Review.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 08:51):

julian-seward1 submitted PR Review.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 08:51):

julian-seward1 created PR Review Comment:

Is this a sufficient test even for the non-const case? For example if v1 = add.i32 xx yy ?

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 08:51):

julian-seward1 created PR Review Comment:

I'm sure this is correct, but .. nevertheless .. it's hard to reason about the correctness of it (in particular, the expression (cst << shift) >> shift, given that there there is no type information here. I'm assuming that cst is a u64 here. Could you please add some type annotation (or something else, that the compiler can check) so as to make that clear to the reader?

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 09:08):

julian-seward1 submitted PR Review.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 09:08):

julian-seward1 submitted PR Review.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 09:08):

julian-seward1 created PR Review Comment:

All 32-bit shifts on Intel (even signed right) will produce bits[63:32] == 0 too. Also, probably a more common case .. a 32-bit load also has this property.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 09:08):

julian-seward1 created PR Review Comment:

Does this require a similar change for loads?

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 09:08):

julian-seward1 created PR Review Comment:

nano-nit: this isn't a peephole optimisation in the 1970s meaning of the phrase. It's just that you're increasing the size/complexity of the pattern that you are matching here. Also, nano-nano-nit, x64, not x86, if we are reserving x86 to mean "32-bit".

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 09:46):

bnjbvr submitted PR Review.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 09:46):

bnjbvr created PR Review Comment:

I don't think so, since a load doesn't "recover" information that would have been lost by the store. Here the store really was losing information about the value of high bits, and it was impossible to infer it back from loads.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 09:49):

bnjbvr submitted PR Review.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 09:49):

bnjbvr created PR Review Comment:

Yes, it is an u64 (so the right shift is logical); i'll add an explicit type annotation for uext_cst and shift!

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 09:57):

bnjbvr submitted PR Review.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 09:57):

bnjbvr created PR Review Comment:

This optimization doesn't hold for non-const values (we'd need information about the high puts of a general i32 input value, which would require range analysis iiuc).

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 09:59):

bnjbvr submitted PR Review.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 09:59):

bnjbvr created PR Review Comment:

Added a TODO.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 10:00):

bnjbvr submitted PR Review.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 10:00):

bnjbvr created PR Review Comment:

Thanks! I should definitely read more 1970s' books :-)

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 10:16):

bnjbvr updated PR #2277 from x64-opt to main:

(I'd like to test it a bit more before considering review, but: this already enhances the code quality a bit on some hot blocks identified with Valgrind.)

view this post on Zulip Wasmtime GitHub notifications bot (Oct 09 2020 at 16:49):

bnjbvr merged PR #2277.

Last updated: Oct 23 2024 at 20:03 UTC