Stream: git-wasmtime

Topic: wasmtime / PR #1842 mach backend: allow snapshotting IR g...


view this post on Zulip Wasmtime GitHub notifications bot (Jun 08 2020 at 15:21):

bnjbvr opened PR #1842 from regalloc-snapshotting to master:

Will need to come with https://github.com/bytecodealliance/regalloc.rs/pull/74, and a regalloc.rs bump as well. With this PR in regalloc.rs, this allows the serialization of function's IR graphs into a standalone IR that regalloc.rs can manipulate, allowing replay / perf-testing in regalloc.rs in a standalone fashion.

Right now, it requires the enable-serde feature to be set, an env variable "SERIALIZE_REGALLOC" set to anything, and it'll dump the IR graphs into the /tmp directory. I expect that we don't need a better user experience for this, since it's likely to be used very rarely in general.

view this post on Zulip Wasmtime GitHub notifications bot (Jun 08 2020 at 15:21):

bnjbvr requested cfallin and julian-seward1 for a review on PR #1842.

view this post on Zulip Wasmtime GitHub notifications bot (Jun 08 2020 at 15:21):

bnjbvr requested cfallin and julian-seward1 for a review on PR #1842.

view this post on Zulip Wasmtime GitHub notifications bot (Jun 08 2020 at 16:05):

bjorn3 submitted PR Review.

view this post on Zulip Wasmtime GitHub notifications bot (Jun 08 2020 at 16:05):

bjorn3 created PR Review Comment:

I think the bincode dependency should be behind a separate feature flag. You may want to use enable-serde for clif ir serialization without using snapshotting.

view this post on Zulip Wasmtime GitHub notifications bot (Jun 08 2020 at 19:00):

cfallin submitted PR Review.

view this post on Zulip Wasmtime GitHub notifications bot (Jun 08 2020 at 19:00):

cfallin submitted PR Review.

view this post on Zulip Wasmtime GitHub notifications bot (Jun 08 2020 at 19:00):

cfallin created PR Review Comment:

Perhaps a low-cost UX improvement: can we take a base path from an environment variable, rather than hardcoding /tmp? (Perhaps just the value of SERIALIZE_REGALLOC, or call it SERIALIZE_REGALLOC_PATH.) No need to mkdir or anything fancy. Just would prefer not to have fixed filenames baked into the binary like this.

view this post on Zulip Wasmtime GitHub notifications bot (Jun 08 2020 at 19:00):

cfallin created PR Review Comment:

Agreed; let's have a separate serialize-regalloc feature or something similar.

view this post on Zulip Wasmtime GitHub notifications bot (Jun 10 2020 at 09:18):

julian-seward1 submitted PR Review.

view this post on Zulip Wasmtime GitHub notifications bot (Jun 10 2020 at 13:15):

bnjbvr updated PR #1842 from regalloc-snapshotting to master:

Will need to come with https://github.com/bytecodealliance/regalloc.rs/pull/74, and a regalloc.rs bump as well. With this PR in regalloc.rs, this allows the serialization of function's IR graphs into a standalone IR that regalloc.rs can manipulate, allowing replay / perf-testing in regalloc.rs in a standalone fashion.

Right now, it requires the enable-serde feature to be set, an env variable "SERIALIZE_REGALLOC" set to anything, and it'll dump the IR graphs into the /tmp directory. I expect that we don't need a better user experience for this, since it's likely to be used very rarely in general.

view this post on Zulip Wasmtime GitHub notifications bot (Jun 10 2020 at 15:04):

bnjbvr updated PR #1842 from regalloc-snapshotting to master:

Will need to come with https://github.com/bytecodealliance/regalloc.rs/pull/74, and a regalloc.rs bump as well. With this PR in regalloc.rs, this allows the serialization of function's IR graphs into a standalone IR that regalloc.rs can manipulate, allowing replay / perf-testing in regalloc.rs in a standalone fashion.

Right now, it requires the enable-serde feature to be set, an env variable "SERIALIZE_REGALLOC" set to anything, and it'll dump the IR graphs into the /tmp directory. I expect that we don't need a better user experience for this, since it's likely to be used very rarely in general.

view this post on Zulip Wasmtime GitHub notifications bot (Jun 10 2020 at 15:37):

bnjbvr updated PR #1842 from regalloc-snapshotting to master:

Will need to come with https://github.com/bytecodealliance/regalloc.rs/pull/74, and a regalloc.rs bump as well. With this PR in regalloc.rs, this allows the serialization of function's IR graphs into a standalone IR that regalloc.rs can manipulate, allowing replay / perf-testing in regalloc.rs in a standalone fashion.

Right now, it requires the enable-serde feature to be set, an env variable "SERIALIZE_REGALLOC" set to anything, and it'll dump the IR graphs into the /tmp directory. I expect that we don't need a better user experience for this, since it's likely to be used very rarely in general.

view this post on Zulip Wasmtime GitHub notifications bot (Jun 10 2020 at 16:23):

bnjbvr merged PR #1842.


Last updated: Dec 23 2024 at 12:05 UTC