[wip] optimize for eventual library size #1087
Conversation
|
looks like this actually breaks windows builds |
hpk42
changed the title
|
now the PR is just adding link-time-optimizations which, if i see it correctly, provide a significant reduction of compile time size. and this seems to work for windows as well. @r10s would be good to know if the new commit on this branch here still leads to a significant reduction of the apk. maybe you can trigger it again sometime ;) This does reduce the information (i guess) that you get with traces/crashes but maybe it's good to offer a full-debug build on get-dc and a size-optimized apk for gplay / get.delta.chat. If we automate apk-building this could all happen automatically ;) |
deltachat-ffi/Cargo.toml
Outdated
| @@ -10,6 +10,11 @@ license = "MIT OR Apache-2.0" | |||
| keywords = ["deltachat", "chat", "openpgp", "email", "encryption"] | |||
| categories = ["cryptography", "std", "email"] | |||
|
|
|||
| #[profile.release] | |||
There was a problem hiding this comment.
Remove these commented out sections.
|
Maybe some link time optimizations during JNI/Python bindings building also can be used, since only at this point the compiler knows which functions are actually used. |
|
seems a if with only link-time-optimizations only half of the size-saving is possible: so, instead of 18mb only 11mb are saved. still not bad, though. |
|
new stats: |
|
On Sat, Dec 21, 2019 at 04:19 -0800, björn petersen wrote:
new stats:
```
-rw-r--r-- 64401673 Dec 20 17:15 deltachat-gplay-release-1.0.2.apk
-rw-r--r-- 46354476 Dec 20 18:12 deltachat-gplay-release-1.0.2-opt-size.apk
-rw-r--r-- 53425757 Dec 21 02:01 deltachat-gplay-release-1.0.2-lto-only.apk
-rw-r--r-- 47303123 Dec 21 12:59 deltachat-gplay-release-1.0.2-lto-opt2.apk
```
k, but it seems it is the "opt-level = 'z'" that breaks windows.
because lto-only did pass the windows tests.
|
|
Our Windows builds are for i686 architecture, so it is likely that |
|
i686 is simply 32bit version on Windows, I think this is more likely to be a compiler bug tbh
…On 21. Dec 2019, 20:21 +0100, Alexander Krotov ***@***.***>, wrote:
Our Windows builds are for i686 architecture, so it is likely that opt-level="z" exposes some memory corruption issue that we don't have on 32-bit architectures. We can try adding i686 builds for Linux to CI and see if they break as well.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub, or unsubscribe.
|
|
@dignifiedquire Sorry, I mean 64-bit architectures, corrected it now. These optimizations are done by LLVM and I expect it to be pretty stable. |
|
we have lto=true already in, moving to resurrection. |
Right now, our release mode optimizes for speed.
As per https://github.com/johnthagen/min-sized-rust this PR tries three-settings to build a size-optimized release. Idea is to check what is the influence on the resulting apk and other artifacts.