This was causing us to generate invalid code in Darwinia, resulting in a
crash. With assertions enabled this would be picked up in the emitter.
Only implement AddShift optimizations for now because I don't want to do
the remaining optimizations in a bug fix PR.
Fixes Darwinia.
Accidentally we were swapping which sources were the base and which was
the one getting shifted. This wasn't super common so it usually didn't
matter.
Fixes one crash in Darwinia.
use a vec. block indices will be dense in the new IR. This is memory intensive
but seems faster in practice.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Flag DCE needs to do general DCE anyway to converge in one pass. So we can move
the special syscall/atomic logic over to flag DCE and then drop the second DCE
pass altogether. Now local dead code of both is eliminated in a single pass.
Flag DCE is carefully written to converge in a single iteration which makes this
scheme work.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
If both the destination and the flags are dead for an AddWithFlags, we need to
eliminate it in one pass. If we only replace without elimiating, we would need a
second DCE pass to eliminate. We want DCE to finish in one pass, so fix this.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
This has been deadcode since 2020. Drop it so we can focus on what *does* work
and what does matter.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Generally, there are three reasons to track progress:
* Conditional optimizations. E.g. only run DCE if ConstProp succeeds.
* Fixed point optimizations. E.g. keep running the opt loop until convergence.
* Metadata shenianigans.
None of these apply to FEX. We explicitly do not want a nonlinear pass ordering,
instead we want just a few passes that each converge in a single iteration. We
expect them all to make progress when run. As such, tracking progress is a waste
of CPU cycles. Stop doing it.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
New RA does not need it for correctness, and the slight slow down to new RA from
not compacting first is much smaller than the cost of compaction. Overall speeds
up node.js start time by ~6% on top of new RA.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
I recommend viewing the new source file as the diff is quite messy.
---
The old RA commits every "how not to write an RA" sin in the book.
Chaitin spill-one loop? Check.
Potential spilling caused by alignment issues since there's no live range
splitting? Check.
Panic spilling? Check.
Generating an interference graph with linear live ranges, so you get the code
quality of linear scan with the cost of graph colouring? Check.
...
It is wholly unsuitable to any application, and specifically unsuitable for FEX.
---
The new RA exploits a key IR invariant unique to FEX: no values are live across
block boundaries. This is validated.
Because of this invariant, all RA is block local. This lets us use a dead simple
2 pass RA that generates ~optimal code in linear time.
The first pass walks the IR backwards, analyzing the IR. This is a souped up
analogue to liveness analysis.
The second pass walks the IR forward, blasting out registers. If necessary, it
will insert spill and/or shuffle code on the fly. Spilling uses the well-known
furthest-first heuristic, which has excellent results for straight line code.
That's it :-)
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Fails with the new RA when copies are inserted, since RAValidation loses
visibility. Nontrivial to fix, but this particular assert hopefully isn't buying
us a ton. Weaken it for now, we can revisit later.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
To implement GPRPair "properly", we need to be able to shuffle scalars around
the register file. That means we need explicit copy/swap instructions that RA
can generate them. Add some.
Swap is split in a really sketchy way, because of the 1 instruction = 1 dest
requirement. Hopefully that requirement is lifted in the future and then this
goes away.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
After two months of testing I finally have enough confidence that these
default options getting changed is safe enough that most games won't
notice the difference. But the performance differences can be wild.
Might want to come back and reenable this on platforms that support
hardware TSO and LRCPC3 (for the vector feature), but we can care once
those platforms come up to speed.
It doesn't write to its node. Fixes spurious
%7: Arg[0] expects reg0 to contain %4, but it actually contains %16
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
now that we've eliminated cross block liveness, we can do our validation locally
too for a massive simplification.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
All we actually need to validate is that each source has been previously defined
within the block. That checks everything we care about now.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
terser and will aid coalescing, as well as eventual transition to multidest
extracts which is what we'll actually want.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
This avoids a bunch of sharp edges for RA at a small cost when obscure
segment registers are used.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
From a theoretical perspective, we should not need to run RCLSE more than once.
If there are convergence issues with the current implementation, they should be
fixed instead of bandaged around. Fortunately, this has no instcountci changes.
Brings RCLSE cost down from like 12% to 5%.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Ryan Houdek