Don't build MAIN_MODULE as RELOCATABLE #25522
Conversation
sbc100
force-pushed
the
main_module_static_v2
branch
2 times, most recently
from
cd1b8ce to
e8f95c7
Compare
Also, add an assertion to dynCall. Split out from emscripten-core#25522.
Also, add an assertion to dynCall. Split out from emscripten-core#25522.
sbc100
force-pushed
the
main_module_static_v2
branch
2 times, most recently
from
8c32591 to
d40087a
Compare
|
I still need to get some stats on the benefits of this, but that tests are all now passing so I think this should be good to review now. An open question is: What do we do about |
Also, add an assertion to dynCall. Split out from emscripten-core#25522.
sbc100
force-pushed
the
main_module_static_v2
branch
from
d40087a to
6008112
Compare
|
Yeah, I see what you mean, removing RELOCATABLE first would simplify this a little. I don't have a preference though. I do see some tests still fail here though? |
Also, add an assertion to dynCall. Split out from #25522.
sbc100
force-pushed
the
main_module_static_v2
branch
5 times, most recently
from
637ed91 to
f56f631
Compare
sbc100
force-pushed
the
main_module_static_v2
branch
2 times, most recently
from
0294236 to
d499627
Compare
|
Ok, aside from the problem of what to do with The stats look really promising. See the new PR description for the great improvements measured when building a dylink version of binaryen. |
sbc100
force-pushed
the
main_module_static_v2
branch
from
d499627 to
2c6de78
Compare
| #endif | ||
| for (var [symName, entry] of Object.entries(GOT)) { | ||
| if (entry.value == 0) { | ||
| if (entry.value == {{{ to64(-1) }}}) { |
There was a problem hiding this comment.
Maybe comment what this means? (tombstone?)
| return 0; | ||
| } | ||
| ''', | ||
| 'a: loaded\na: b (prev: (null))\na: c (prev: b)\n', cflags=extra_args) |
sbc100
force-pushed
the
main_module_static_v2
branch
2 times, most recently
from
23d619c to
a2e97ec
Compare
The normal dependency system can handle including these when needed.
The main advantage here is that main module no longer requires relocation entries for symbols defined locally. To show the benefits of this approach I build binaryan_wasm with `-sMAIN_MODULE=1` both before and after this change. Before: Wasm Size: 22.6M __wasm_apply_data_relocs size: 123k After: Wasm Size: 16.6M __wasm_apply_data_relocs size: 0 (no longer exists) Fixes: emscripten-core#12682
sbc100
force-pushed
the
main_module_static_v2
branch
from
a2e97ec to
4425035
Compare
|
I solves the |
3 participants
The main advantage here is that main module no longer requires
relocation entries for symbols defined locally.
To show the benefits of this approach I build binaryan_wasm
with
-sMAIN_MODULE=1both before and after this change.Before:
Wasm Size: 22.6M
__wasm_apply_data_relocs size: 123k
After:
Wasm Size: 16.6M
__wasm_apply_data_relocs size: 0 (no longer exists)
Since this is quite a major change, I've kept the
-sRELOCATABLEoption available so folks that use this to get the old behaviour when
linking their main module. In order to tests this path too now run
all the
@needs_dylinktests in test_core.py with and without-sRELOCTABLE. Hopefully we can remove this configurationcompletely at some point in the future.
Fixes: #12682