Remove most floating-point/integer intrinsics from the compiler #5724
Conversation
|
Makes sense to me. |
|
Thanks. |
|
I'm really sorry about these silly errors. This attribute does not exist, it has the proper name #[no_mangle] but is not needed. |
I am not convinced that this is a great idea, but I wanted to throw this out here since I noticed it was possible to save a few hundred lines of code with no reduction in functionality. The main ideas of this commit is to remove most of the special implementation for llvm-intrinsics in src/librustc/middle/trans/base.rs and friends, and instead declare them as semi-standard functions instead with a combination of attributes instead. In the end, this means that the compiler doesn't need to know about every intrinsic and we only need to provide definitions for the ones we use in the core/std library and users can provide the rest if they need them.
|
This seems to break the buildbot, if it is related to the modified code, could someone send me the logs? |
|
I thought I heard graydon mention that one of the bots was running out of memory, so it might not be related to this branch. It looks like ther's a build error in the sudoku benchmark http://buildbot.rust-lang.org/builders/auto-mac/builds/764/steps/test/logs/stdio |
|
@brson Oh, that is good to head, because it should work. Especially on OS X, where I did the development. The changed API on the intrinsics might have broken that benchmark. I didn't know that it used the ctlz intrinsic, I didn't grep good enough, mea maxima culpa. That intrinsic is at least partially broken, since it takes a constant parameter. I am working on a patch to provide them as pragmas, then I'll update the PR. If I cannot do that, I'll just revert the CTTZ/CTLZ intrinsics. |
|
Status update: I can now call the CTTZ/CTLZ, but I need to clean up the code and make a proper parser for the new function attribute I added to support constant integer arguments. After it is done, we can see if we want to revert parts of this patch to have the old CTTZ/CTLZ, or the 'new' patch with a small amount of new compiler guck, but which can move all intrinsics into libcore. |
|
Status update: I have a patch that fixes the cttz/ctlz/memmove-issue here auroranockert@f566870 but it needs work, mainly the changes in src/librustc/middle/foreign.rs, and I need some tips on how to best refactor it. |
|
I'm closing this, and instead adding the new code as a PR. |
…anishearth redundant_pattern_matching: avoid non-`const fn` calls in const contexts changelog: Avoid suggesting non-`const fn` calls in const contexts in [`redundant_pattern_matching`] Fixes rust-lang#5697
Rollup of 9 pull requests Successful merges: - rust-lang#5178 (clippy-driver: pass all args to rustc if --rustc is present) - rust-lang#5705 (Downgrade unnested_or_patterns to pedantic) - rust-lang#5709 (Fix ICE in consts::binop) - rust-lang#5710 (typo) - rust-lang#5712 (Remove `bar` from blacklisted names) - rust-lang#5713 (Use lints in Clippy that are enabled in rustc bootstrap) - rust-lang#5716 (Fix typo in wildcard_imports) - rust-lang#5724 (redundant_pattern_matching: avoid non-`const fn` calls in const contexts) - rust-lang#5726 (Fix typo) Failed merges: r? @ghost changelog: rollup
I am not convinced that this is a great idea, but I wanted to throw this out here since I noticed it was possible to save a few hundred lines of code with no reduction in functionality.
The main ideas of this commit is to remove most of the special implementation for llvm-intrinsics in src/librustc/middle/trans/base.rs and friends, and instead declare them as semi-standard functions instead with a combination of attributes instead.
In the end, this means that the compiler doesn't need to know about every intrinsic and we only need to provide definitions for the ones we use in the core/std library and users can provide the rest if they need them.