[mypyc] Remove unused labels and gotos #15244
Conversation
This PR removes labels which are never jumped to, as well as removing gotos that jump directly to the next label (and removes the now redundant label). I had a small issue with the `GetAttr` op code as it would insert C code after the end of the branch which could not be detected at the IR level, as seen in mypyc/mypyc#600 (comment) . Here are some basic stats on the file sizes before and after this PR: Building with `MYPY_OPT_LEVEL=0`, `MYPY_DEBUG_LEVEL=1`: | Branch | `.so` size | `.c` size | `.c` lines | |----------|-----------------|-----------------|-----------------| | `master` | 43.9 MB | 79.7 MB | 2290675 | | This PR | 43.6 MB (-0.6%) | 78.6 MB (-1.4%) | 2179967 (-4.8%) | Building with `MYPY_OPT_LEVEL=3`, `MYPY_DEBUG_LEVEL=0`: | Branch | `.so` size | `.c` size | `.c` lines | |----------|------------|-----------|------------| | `master` | 26.5 MB | 79.7 MB | 2291316 | | This PR | 26.5 MB | 78.7 MB | 2179701 | (I don't know why the number of lines changes between optimization levels. It's only a couple hundred lines, perhaps Mypyc is emitting different code depending on optimization level?) Unoptimized builds seem to benefit the most, while optimized builds look prety much the same. I didn't check to see if they are *exactly* the same binary, but they are probably close if not the same. The biggest win is the drop in the number of lines C code being generated. Closes mypyc/mypyc#600
|
(Note the 32-bit test started failing on master and is unrelated to your PR) |
There was a problem hiding this comment.
Hey thank you for the PR, your perseverance is appreciated :)
I'm not finished reviewing — I'd like to double check some more potentially problematic ops — but here are my initial comments. Overall, it looks good. It's impressive that this reduces the self-compile LOC by ~5%.
As usual, it's worth noting I'm a new reviewer (and not a committer!) so I may get some things wrong. Please speak up if I do!
There was a problem hiding this comment.
LGTM*. I wasn't able to break your PR with the Cast op :)
Although... Cast ops still can merge two blocks (just like GetAttr) so it could have a possibility of breaking the control flow (as the branch gotos might not be at the end). I couldn't get it to break as a failed cast seemingly never jumps to the next block (unlike GetAttr). Honestly, I'm still pretty hesitant with the correctness / robustness of this PR.
Anyway, let's wait for Jukka's review now. Thanks!
| # Find blocks that are never jumped to or are only jumped to from the | ||
| # block directly above it. This allows for more labels and gotos to be | ||
| # eliminated during code generation. |
There was a problem hiding this comment.
Only labels are eliminated here, no?
There was a problem hiding this comment.
This code only marks the basic blocks as "referenced" if they meet the conditions above, nothing actually gets eliminated here. And yes, during codegen, labels won't be emitted for these "unreferenced" blocks, same for gotos (assuming they are jumping to the next block).
This PR removes labels which are never jumped to, as well as removing gotos that jump directly to the next label (and removes the now redundant label). I had a small issue with the
GetAttrop code as it would insert C code after the end of the branch which could not be detected at the IR level, as seen in mypyc/mypyc#600 (comment) .Here are some basic stats on the file sizes before and after this PR:
Building with
MYPY_OPT_LEVEL=0,MYPY_DEBUG_LEVEL=1:.sosize.csize.clinesmasterBuilding with
MYPY_OPT_LEVEL=3,MYPY_DEBUG_LEVEL=0:.sosize.csize.clinesmaster(I don't know why the number of lines changes between optimization levels. It's only a couple hundred lines, perhaps Mypyc is emitting different code depending on optimization level?)
Unoptimized builds seem to benefit the most, while optimized builds look pretty much the same. I didn't check to see if they are exactly the same binary, but they are probably close if not the same. The biggest win is the drop in the number of lines C code being generated.
Closes mypyc/mypyc#600