Chris Fallin
f49167e0fe
emit annotations at Info level, for easier selective perf-debugging
2021-05-28 18:40:39 -07:00
Chris Fallin
789651f947
Rework inner allocation-loop code: choose more wisely between splitting and evicting based on costs (and unify the fixed-reg-constraint case)
2021-05-28 17:36:06 -07:00
Chris Fallin
43d7095cbd
Properly split when we hit a fixed conflict
2021-05-28 16:49:32 -07:00
Chris Fallin
7171624750
Don't generate r1->scratch,scratch-r1 sequence for cyclic moves of r1->r1 that are generated to change vreg ownership and keep the checker happy. Seems to eliminate a bit of braindeadness and improve bz2 by ~5-10%.
2021-05-26 21:35:43 -07:00
Chris Fallin
13bde99d7d
bugfix with clean-spill opt: avoid if liverange starts at start of block (this is like a def) or if has starts-at-def flag.
2021-05-26 18:08:41 -07:00
Chris Fallin
e521811b88
Avoid re-spilling to spillslot when still clean: intra-block edition (inter-block needs more analysis and careful thought)
2021-05-26 17:08:14 -07:00
Chris Fallin
dcf6f473ca
inline some things
2021-05-26 00:48:41 -07:00
Chris Fallin
4e0dd1f296
little tweak to avoid a div/mod on every iter of a PReg alloc loop
2021-05-26 00:38:53 -07:00
Chris Fallin
b3dc2b25a5
Alloc spillsets for whole vreg, not just spilled LRs. This is a prerequisite to allowing a "clean" value to remain in spillslot while also in reg and avoiding the re-spill. It should also reduce stack-to-stack moves (though they can still come from progmoves).
2021-05-25 18:19:25 -07:00
Chris Fallin
ca5f24f6b7
Hint the same PReg for both halves of a split
2021-05-24 23:49:47 -07:00
Chris Fallin
7cdcb2031e
Split heuristic: split before entering deeper loop nest
2021-05-24 23:09:05 -07:00
Chris Fallin
8887077b59
small fix: preserve starts-at-def flag when setting liverange weight
2021-05-24 22:45:25 -07:00
Chris Fallin
3382f9a2e8
Split based on first conflict of lowest-weight conflict, not first conflict. Also stop scanning PRegs when max bundle weight in conflict bundle list exceeds current best option.
2021-05-24 22:26:57 -07:00
Chris Fallin
5b47462e0c
Loop depth instead of hot/cold, with fast O(n) loop-depth computation. Use this to compute use weights.
2021-05-24 22:09:41 -07:00
Chris Fallin
5895ae8a2d
Remove precomputed requirements from ranges and bundles; cost of struct size and updates is not worth it. Instead, do a simple conflicting-requirements check on each attempted merge, but minimize cost by (i) only checking after ruling out overlaps, and (ii) only checking if we know one of the sides has at least one non-register constraint.
2021-05-24 21:32:41 -07:00
Chris Fallin
5120681730
Fuzzbug fix for requirement recomputation on minimal bundles with multiple LRs
2021-05-24 17:49:45 -07:00
Chris Fallin
10d926557a
avoid some redundant work by computing initial reqs only once
2021-05-24 15:52:14 -07:00
Chris Fallin
78c009181c
Fuzzbug fix re: new requirements computation and multi-fixed-reg fixup.
2021-05-24 15:47:15 -07:00
Chris Fallin
46feacc654
Fuzzbug fix: don't merge bundles that have conflicting requirements. (Normally splitting would fix this, but let's just not merge in the first place.)
2021-05-24 15:32:05 -07:00
Chris Fallin
59967ff756
TODO-list update: braindump of next ideas to work on.
2021-05-23 20:04:28 -07:00
Chris Fallin
107c09181f
Simple speedup in bundle merge: set bundle while everything is in cache (same pass), and only check non-overlap with debug assertions enabled. Alloc time on clang.wasm: 9.1s old backtracking RA vs. 7.2s with this regalloc2 RA.
2021-05-22 16:36:44 -07:00
Chris Fallin
a6c89b1c01
Avoid O(n^2) in liverange construction: we always build LRs in (reverse) order, so we can just append (prepend) to running list and reverse at end. Likewise for uses.
2021-05-22 15:12:35 -07:00
Chris Fallin
469669155f
Another fuzzbug fix: proper checker-hint ordering when V-R and V-V moves are back-to-back and RReg ownership changes
2021-05-22 14:37:56 -07:00
Chris Fallin
4b46c6388a
Fuzzbug fixes for simpler splitting
2021-05-22 14:26:49 -07:00
Chris Fallin
466ea2cd9a
Simpler / possibly better splitting: split based on conflict position, always, and use a reg hint to put the before-conflict part in the place where we determined it fit before.
2021-05-21 01:34:52 -07:00
Chris Fallin
ec7fdeb8ed
Properly handle RReg-RReg moves in new scheme
2021-05-20 20:38:50 -07:00
Chris Fallin
2a5f571b80
WIP: Handle moves between realregs (pregs) and vregs somewhat specially, by converting into operand constraints
...
Still has a fuzzbug in interaction between R->R and V->R moves. Will
likely rework to make pinned-vreg handling more general but want to save
a checkpoint here; idea for rework:
- set allocs immediately if an Operand is a pinned vreg;
- reserve preg ranges;
- then, in rest of liveness computation / LR construction, convert
pinned-vregs to operands with constraints, but otherwise do not
special-case as we do in this commit.
2021-05-20 19:53:19 -07:00
Chris Fallin
f0b24cf9fa
Remove all-empty-ranges-to-spill-bundle: prioritizing same-alloc for all empty ranges over allowing some to live in registers results in too much spilling
2021-05-20 10:21:22 -07:00
Chris Fallin
ce935c1040
Add all empty LRs to a single "spill bundle", to avoid many small bundles and excessive moves
2021-05-19 22:12:22 -07:00
Chris Fallin
f56676fb8d
Fixed all fuzzer targets (some API changes)
2021-05-19 18:25:34 -07:00
Chris Fallin
f1c6dfe807
Optionally show annotations in final allocation/program dump based on RegallocOptions flag
2021-05-19 16:36:36 -07:00
Chris Fallin
e1f67e860f
Pinned VRegs for use with regalloc.rs shim to support RealRegs.
2021-05-18 22:40:43 -07:00
Chris Fallin
04c8e46787
Only do annotations in debug builds
2021-05-18 18:52:34 -07:00
Chris Fallin
c3513b94b0
Bugfix: don't do a split-at-intermediate-defs split if the first such point is the start of the bundle.
2021-05-18 15:16:19 -07:00
Chris Fallin
4389f16156
debugging log message for liveins
2021-05-18 12:14:59 -07:00
Chris Fallin
8e0d0f1de0
fuzzbug fix
2021-05-18 00:26:38 -07:00
Chris Fallin
328c9004e5
fuzzbug fixes
2021-05-17 23:59:13 -07:00
Chris Fallin
f0fbf3aa0c
Rework data structures: bundles have a SmallVec of ranges, and ranges a SmallVec of uses.
...
Appears to be a small speed improvement on the highly-artificial
fuzz-generator test inputs; Cranelift tests TBD.
2021-05-17 22:44:10 -07:00
Chris Fallin
5b55948feb
Check branch-args for conflicts with edge-move placement.
2021-05-13 17:25:11 -07:00
Chris Fallin
1f9258bea5
Detect undefined liveins.
2021-05-12 01:06:27 -07:00
Chris Fallin
37fa3ec763
Improve prog-move handling: no use/def records, just directly connect the LRs.
...
Also requires some metadata in edit output to properly hook up the
checker in regalloc.rs to track user-moves without seeing the original
insts with operands.
2021-05-11 23:59:12 -07:00
Chris Fallin
6066d02f6f
More annotations
2021-05-11 18:19:40 -07:00
Chris Fallin
b069ae099d
Use hot-code map to augment spill weights of each use
2021-05-11 17:59:10 -07:00
Chris Fallin
e1a37cf0e0
some more stats
2021-05-10 22:53:44 -07:00
Chris Fallin
f7551c68d1
Integrate prog-moves with LR-moves; this should in theory reduce move traffic somewhat
2021-05-10 22:47:57 -07:00
Chris Fallin
0dbf4a790f
Collect full conflict-bundle list, by not ending PhysReg probe on first conflict; this leads to better eviction decisions on bz2
2021-05-09 20:21:57 -07:00
Chris Fallin
b7fd53efc5
Fix checker: after moving edge-moves to prior to last branch of block (for simpler semantics for library user), we can no longer check blockparams; but this is fine because they do not exist in post-regalloc code.
2021-05-09 19:47:37 -07:00
Chris Fallin
4f26b1c78f
Properly handle prog-moves with fixed srcs or dests
2021-05-09 13:35:38 -07:00
Chris Fallin
5c5ea4cb9b
bugfix
2021-05-09 04:11:30 -07:00
Chris Fallin
34421fcc6b
fix to prog-move handling: happens in middle of inst; and insert uses to make later move-insertion happy with this
2021-05-09 03:51:10 -07:00
