gh-104584: Baby steps towards 8000 generating and executing traces #105924
Conversation
|
I |
Use '#define X 1' instead.
|
Benchmark results say "1.00x slower": https://github.com/faster-cpython/benchmarking/blob/main/results/bm-20230619-3.13.0a0-6b23e0f/bm-20230619-linux-x86_64-gvanrossum-optim_exec-3.13.0a0-6b23e0f-vs-base.png (sorry, access only for Faster CPython team members). |
f05a287 to
0ad18cf
Compare
(Macros inherit flags from components, but shouldn't inherit IS_UOP.)
|
The |
Now test_mmap on Windows passes, and hopefully so does test_threading on UNIX.
Instead of `int oparg` we now have `uint64_t operand`. In the Tier 2 interpreter, we add `int oparg` and set it to a truncated version of `operand` so instructions can reference it.
|
The latest change indeed fixes I had hoped it would also fix I will continue to debug this for a bit, but I'll probably choose to just disable the offending bytecode again. UPDATE: I gave up. I have a hunch it's related to a trace that contains two |
Apparently the fix using SET_LOCALS_FROM_FRAME() isn't enough (it fixes some but not all failures). Possibly the problem is due to having *two* BEFORE_WITH uops in one trace? Unclear what's wrong with that opcode.
|
🤖 New build scheduled with the buildbot fleet by @gvanrossum for commit 4bd7bb6 🤖 If you want to schedule another build, you need to add the 🔨 test-with-refleak-buildbots label again. |
|
Regarding new files, I don't think it will stand in the way of progress if this were merged with the superblock creation code in Python/optimizer.c and the trace interpreter in Python/ceval.c. If at some point in the future that becomes a problem we can refactor it then. I'd like this PR to remain a spike, not become a marathon. Last bits of work:
|
The uops optimizer is enabled in the main interpreter when either: - -Xuops is passed; or - PYTHONUOPS=1 is set. The env var for uops debugging is now called PYTHONUOPSDEBUG. It only has an effect when configured --with-pydebug *and* the uops optimizer is enabled (see above). Also fixed a compiler warning about %ld with an int64 arg.
Those tests now only fail when run with -Xuops. We should probably fix the uop executor instead. Also, test_embed now passes even with -Xuops. This reverts commit 91dd86c.
|
🤖 New build scheduled with the buildbot fleet by @gvanrossum for commit 6606c5d 🤖 If you want to schedule another build, you need to add the 🔨 test-with-refleak-buildbots label again. |
|
@markshannon I have a niggling suspicion that if the very first uop in a trace ends with a failing DEOPT_IF (or ERROR_IF?), it will do the wrong thing, since |
|
The refleaks run failed, but that was an existing issue (#105273). They didn't test there were no refleaks left, because I also pushed the change that requires |
|
This is mostly so we can benchmark how bad the overhead is.