Code

Initial implementation of regalloc2.#1

Code

Merged

cfallin merged 155 commits intobytecodealliance:mainbytecodealliance/regalloc2:mainfrom cfallin:initial-regalloccfallin/regalloc2:initial-regallocCopy head branch name to clipboardSep 1, 2021

+12,133

Commits

Commits on Apr 14, 2021

• Initial public commit of regalloc2.

  cfallin

  committed
  8e923b0

  Copy full SHA for 8e923b0

• Heuristic improvement: reg-scan offset by inst location.

  Show description for 33ac6cb
  cfallin

  committed
  33ac6cb

  Copy full SHA for 33ac6cb

Commits on Apr 18, 2021

• Add support for reftypes/stackmaps and Stack constraints, and misc API changes.

  Show description for a08b012
  cfallin

  committed
  a08b012

  Copy full SHA for a08b012

• Changes from review comments.

  cfallin

  committed
  940c1b7

  Copy full SHA for 940c1b7

• Add GitHub CI config.

  cfallin

  committed
  34ab744

  Copy full SHA for 34ab744

• Factored out test program and fuzzing features; core crate now only depends on smallvec and log.

  cfallin

  committed
  414f3f8

  Copy full SHA for 414f3f8

Commits on May 1, 2021

• Misc usability and functionality enhancements:

  Show description for 49c54b6
  cfallin

  committed
  49c54b6

  Copy full SHA for 49c54b6

Commits on May 4, 2021

• Derive Ord/hash on OperandOrAllocation.

  cfallin

  committed
  9e7021c

  Copy full SHA for 9e7021c

Commits on May 6, 2021

• Allow multiple defs per vreg (i.e., accept non-SSA code).

  Show description for 15ed2d6
  cfallin

  committed
  15ed2d6

  Copy full SHA for 15ed2d6

• BitVec::get() takes immutable self

  cfallin

  committed
  48fbc23

  Copy full SHA for 48fbc23

• MachineEnv fields are public

  cfallin

  committed
  ab828b6

  Copy full SHA for ab828b6

• Properly handle multiple same-fixed-reg constraints to the same vreg in one inst.

  cfallin

  committed
  80cdd0c

  Copy full SHA for 80cdd0c

• Some performance tweaks -- try to reduce register probe count with soem more hints. Also fix spill-weight caching.

  cfallin

  committed
  1a7a0c5

  Copy full SHA for 1a7a0c5

• Some micro-optimizations in BitVec.

  cfallin

  committed
  747c56c

  Copy full SHA for 747c56c

Commits on May 7, 2021

• Handle moves specially with move-insertion logic rather than ordinary operand/inst handling

  cfallin

  committed
  e2beb47

  Copy full SHA for e2beb47

• BitVec perf: use adaptive hybrid chunked small-array + FxHashMap.

  cfallin

  committed
  07a5a88

  Copy full SHA for 07a5a88

• Some memory-size/bitpacking optimizations

  cfallin

  committed
  02b6516

  Copy full SHA for 02b6516

• Parameterize adaptive-map size in BitVec.

  cfallin

  committed
  a148dcc

  Copy full SHA for a148dcc

• Some perf opts in liveness computation and set impl:

  Show description for 2ff02b5
  cfallin

  committed
  2ff02b5

  Copy full SHA for 2ff02b5

• Fuzzbugfix: actually do need eager liveness computation; must uphold invariant that all earlier-in-postorder blocks have full livein sets.

  cfallin

  committed
  2ba5185

  Copy full SHA for 2ba5185

• Some stats for loop effects on liveins: 487k loop set-unions (441 loops) in one func in bz2 -- fix TBD

  cfallin

  committed
  42582e0

  Copy full SHA for 42582e0

• Replace approximate liveness with true iterative liveness; turns out it is better to improve accuracy so that later stages of the allocator have less wasted work/interference

  cfallin

  committed
  3713d61

  Copy full SHA for 3713d61

• Inlining on btree commitment map comparators for a 10% win

  cfallin

  committed
  0f3454b

  Copy full SHA for 0f3454b

Commits on May 8, 2021

• Optimizations: (i) range-summary array; (ii) early exit from btree probe loop (one conflict bundle is enough, empirically)

  cfallin

  committed
  3ddcf05

  Copy full SHA for 3ddcf05

• Pinned-VReg mechanism.

  cfallin

  committed
  4f63467

  Copy full SHA for 4f63467

• Inline all the things (ProgPoint edition)

  cfallin

  committed
  df59b5e

  Copy full SHA for df59b5e

• sort_unstable (quicksort) everywhere

  cfallin

  committed
  a453501

  Copy full SHA for a453501

• Some preallocation and removal of one u32 from LiveRange struct

  cfallin

  committed
  bfe1c63

  Copy full SHA for bfe1c63

• Some structure packing: Use now fits in three u32s.

  cfallin

  committed
  040c3c8

  Copy full SHA for 040c3c8

• More efficient queue_bundles (saves 18% on clang.wasm)

  cfallin

  committed
  d2cc4f1

  Copy full SHA for d2cc4f1

• Support mod (modify) operands, for better efficiency with regalloc.rs/Cranelift shim.

  cfallin

  committed
  a6e3128

  Copy full SHA for a6e3128

• More efficient live-range creation re: uses

  cfallin

  committed
  4185eab

  Copy full SHA for 4185eab

• Struct-of-array transform: pull LiveRangeHot out of LiveRange with just range and next-in-bundle link

  cfallin

  committed
  e41b010

  Copy full SHA for e41b010

• Some minor opts: inlining, and smallvec reuse

  cfallin

  committed
  ed339ab

  Copy full SHA for ed339ab

• Bugfix: Mod with dead def spans both Before and After positions

  cfallin

  committed
  179ef0e

  Copy full SHA for 179ef0e

• Bugfix: process parallel moves separately for Int and Float classes

  cfallin

  committed
  3d0d760

  Copy full SHA for 3d0d760

Commits on May 9, 2021

• Handle moves by joining LRs at inst boundary, not middle of move inst

  cfallin

  committed
  00c64f6

  Copy full SHA for 00c64f6

• Fix related to move handling

  cfallin

  committed
  f1fc9a8

  Copy full SHA for f1fc9a8

• Error checking: properly signal a crit-edge requirement failure (used for regalloc.rs fuzzer)

  cfallin

  committed
  b9e8988

  Copy full SHA for b9e8988

• Fix crit-edge detection logic in CFGInfo analysis

  cfallin

  committed
  095a883

  Copy full SHA for 095a883

• Remove sanity-check logic in range summary construction -- zero-length

  Show description for 509c5dc
  cfallin

  committed
  509c5dc

  Copy full SHA for 509c5dc

• fuzzbug fix in range-summary iter

  cfallin

  committed
  9fdc69e

  Copy full SHA for 9fdc69e

• assert fix: RegClass doesn't need to match for spillslots (can be reused across classes)

  cfallin

  committed
  c380b0d

  Copy full SHA for c380b0d

• Edge moves always before jumps, never after; semantics are too subtle otherwise (client needs to handle specially)

  cfallin

  committed
  8d7530d

  Copy full SHA for 8d7530d

• fix to prog-move handling: happens in middle of inst; and insert uses to make later move-insertion happy with this

  cfallin

  committed
  34421fc

  Copy full SHA for 34421fc

• bugfix

  cfallin

  committed
  5c5ea4c

  Copy full SHA for 5c5ea4c

• Properly handle prog-moves with fixed srcs or dests

  cfallin

  committed
  4f26b1c

  Copy full SHA for 4f26b1c

Commits on May 10, 2021

• 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 pos…

  Show description for b7fd53e
  cfallin

  committed
  b7fd53e

  Copy full SHA for b7fd53e

• Collect full conflict-bundle list, by not ending PhysReg probe on first conflict; this leads to better eviction decisions on bz2

  cfallin

  committed
  0dbf4a7

  Copy full SHA for 0dbf4a7

Commits on May 11, 2021

• Integrate prog-moves with LR-moves; this should in theory reduce move traffic somewhat

  cfallin

  committed
  f7551c6

  Copy full SHA for f7551c6

• some more stats

  cfallin

  committed
  e1a37cf

  Copy full SHA for e1a37cf

Commits on May 12, 2021

• Use hot-code map to augment spill weights of each use

  cfallin

  committed
  b069ae0

  Copy full SHA for b069ae0

• More annotations

  cfallin

  committed
  6066d02

  Copy full SHA for 6066d02

• Improve prog-move handling: no use/def records, just directly connect the LRs.

  Show description for 37fa3ec
  cfallin

  committed
  37fa3ec

  Copy full SHA for 37fa3ec

• Detect undefined liveins.

  cfallin

  committed
  1f9258b

  Copy full SHA for 1f9258b

Commits on May 14, 2021

• Check branch-args for conflicts with edge-move placement.

  cfallin

  committed
  5b55948

  Copy full SHA for 5b55948

Commits on May 18, 2021

• Rework data structures: bundles have a SmallVec of ranges, and ranges a SmallVec of uses.

  Show description for f0fbf3a
  cfallin

  committed
  f0fbf3a

  Copy full SHA for f0fbf3a

• fuzzbug fixes

  cfallin

  committed
  328c900

  Copy full SHA for 328c900

• fuzzbug fix

  cfallin

  committed
  8e0d0f1

  Copy full SHA for 8e0d0f1

• debugging log message for liveins

  cfallin

  committed
  4389f16

  Copy full SHA for 4389f16

• Bugfix: don't do a split-at-intermediate-defs split if the first such point is the start of the bundle.

  cfallin

  committed
  c3513b9

  Copy full SHA for c3513b9

Commits on May 19, 2021

• Only do annotations in debug builds

  cfallin

  committed
  04c8e46

  Copy full SHA for 04c8e46

• Pinned VRegs for use with regalloc.rs shim to support RealRegs.

  cfallin

  committed
  e1f67e8

  Copy full SHA for e1f67e8

• Optionally show annotations in final allocation/program dump based on RegallocOptions flag

  cfallin

  committed
  f1c6dfe

  Copy full SHA for f1c6dfe

Commits on May 20, 2021

• Fixed all fuzzer targets (some API changes)

  cfallin

  committed
  f56676f

  Copy full SHA for f56676f

• Add all empty LRs to a single "spill bundle", to avoid many small bundles and excessive moves

  cfallin

  committed
  ce935c1

  Copy full SHA for ce935c1

• 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

  cfallin

  committed
  f0b24cf

  Copy full SHA for f0b24cf

Commits on May 21, 2021

• WIP: Handle moves between realregs (pregs) and vregs somewhat specially, by converting into operand constraints

  Show description for 2a5f571
  cfallin

  committed
  2a5f571

  Copy full SHA for 2a5f571

• Properly handle RReg-RReg moves in new scheme

  cfallin

  committed
  ec7fdeb

  Copy full SHA for ec7fdeb

• 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.

  cfallin

  committed
  466ea2c

  Copy full SHA for 466ea2c

Commits on May 22, 2021

• Fuzzbug fixes for simpler splitting

  cfallin

  committed
  4b46c63

  Copy full SHA for 4b46c63

• Another fuzzbug fix: proper checker-hint ordering when V-R and V-V moves are back-to-back and RReg ownership changes

  cfallin

  committed
  4696691

  Copy full SHA for 4696691

• 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.

  cfallin

  committed
  a6c89b1

  Copy full SHA for a6c89b1

• 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.…

  Show description for 107c091
  cfallin

  committed
  107c091

  Copy full SHA for 107c091

Commits on May 24, 2021

• TODO-list update: braindump of next ideas to work on.

  cfallin

  committed
  59967ff

  Copy full SHA for 59967ff

• 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.)

  cfallin

  committed
  46feacc

  Copy full SHA for 46feacc

• Fuzzbug fix re: new requirements computation and multi-fixed-reg fixup.

  cfallin

  committed
  78c0091

  Copy full SHA for 78c0091

• avoid some redundant work by computing initial reqs only once

  cfallin

  committed
  10d9265

  Copy full SHA for 10d9265

Commits on May 25, 2021

• Fuzzbug fix for requirement recomputation on minimal bundles with multiple LRs

  cfallin

  committed
  5120681

  Copy full SHA for 5120681

• 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 c…

  Show description for 5895ae8
  cfallin

  committed
  5895ae8

  Copy full SHA for 5895ae8

• Loop depth instead of hot/cold, with fast O(n) loop-depth computation. Use this to compute use weights.

  cfallin

  committed
  5b47462

  Copy full SHA for 5b47462

• 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.

  cfallin

  committed
  3382f9a

  Copy full SHA for 3382f9a

• small fix: preserve starts-at-def flag when setting liverange weight

  cfallin

  committed
  8887077

  Copy full SHA for 8887077

• Split heuristic: split before entering deeper loop nest

  cfallin

  committed
  7cdcb20

  Copy full SHA for 7cdcb20

• Hint the same PReg for both halves of a split

  cfallin

  committed
  ca5f24f

  Copy full SHA for ca5f24f

Commits on May 26, 2021

• 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 st…

  Show description for b3dc2b2
  cfallin

  committed
  b3dc2b2

  Copy full SHA for b3dc2b2

• little tweak to avoid a div/mod on every iter of a PReg alloc loop

  cfallin

  committed
  4e0dd1f

  Copy full SHA for 4e0dd1f

• inline some things

  cfallin

  committed
  dcf6f47

  Copy full SHA for dcf6f47

Commits on May 27, 2021

• Avoid re-spilling to spillslot when still clean: intra-block edition (inter-block needs more analysis and careful thought)

  cfallin

  committed
  e521811

  Copy full SHA for e521811

• 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.

  cfallin

  committed
  13bde99

  Copy full SHA for 13bde99

• 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 impro…

  Show description for 7171624
  cfallin

  committed
  7171624

  Copy full SHA for 7171624

Commits on May 28, 2021

• Properly split when we hit a fixed conflict

  cfallin

  committed
  43d7095

  Copy full SHA for 43d7095

Commits on May 29, 2021

• Rework inner allocation-loop code: choose more wisely between splitting and evicting based on costs (and unify the fixed-reg-constraint case)

  cfallin

  committed
  789651f

  Copy full SHA for 789651f

• emit annotations at Info level, for easier selective perf-debugging

  cfallin

  committed
  f49167e

  Copy full SHA for f49167e

Commits on Jun 1, 2021

• Fuzzbug fix: properly check for conflicting reqs before merging bundles (cached values are not computed yet)

  cfallin

  committed
  44ca189

  Copy full SHA for 44ca189

• Fuzzbug fix: fix some weirdness with BTree iteration inner loop

  cfallin

  committed
  e49727d

  Copy full SHA for e49727d

• fuzzbug fix: restore clean error exit required by regalloc.rs fuzzer on too-many-live-regs error

  cfallin

  committed
  2614eac

  Copy full SHA for 2614eac

Commits on Jun 2, 2021

• Fuzzbug fix

  cfallin

  committed
  a2a770e

  Copy full SHA for a2a770e

• BTreeMap probe fix (fuzzbug): BTree does not interact nicely with LiveRangeKey definition of equality; need to probe with one-less-than start to get proper range

  cfallin

  committed
  2fe276c

  Copy full SHA for 2fe276c

• Add a perf idea to TODO list

  cfallin

  committed
  dc2b0d1

  Copy full SHA for dc2b0d1

Commits on Jun 3, 2021

• Implement spill-bundle: move all empty ranges, and empty leading/trailing pieces surrounding split points, to a single spill bundle, in an attempt to avoid excessive movement

  cfallin

  committed
  6a0739b

  Copy full SHA for 6a0739b

Commits on Jun 4, 2021

• merge bundles much faster by just concatenating range-lists and unstable-sorting, rather than a merge-sort-like traversal. Rust stdlib sort is very optimized. clang.wasm 9.1s -> 6.8s now.

  cfallin

  committed
  00e4240

  Copy full SHA for 00e4240

• Adaptive commitment-map scanning: re-probe from root if we skip too many entries in linear BTree scan

  cfallin

  committed
  5560499

  Copy full SHA for 5560499

• Fix fuzzbug: properly detect too-many-live-regs condition on fuzzing input. Must be careful in how we probe the BTree when we have multiple "equal" (overlapping) keys.

  cfallin

  committed
  30f42a8

  Copy full SHA for 30f42a8

Commits on Jun 5, 2021

• Fix to checker: analyze all blocks, even if out-state of entry block is empty

  cfallin

  committed
  0eaa0fd

  Copy full SHA for 0eaa0fd

Commits on Jun 7, 2021

• Split-at-first-conflict: first conflict is first of (start of our range), (start of conflict range), not just the latter; otherwise we have a too-early split sometimes

  cfallin

  committed
  2be7bdb

  Copy full SHA for 2be7bdb

Commits on Jun 8, 2021

• WIP: redundant-move elimination.

  cfallin

  committed
  c6bcd3c

  Copy full SHA for c6bcd3c

• Redundant move eliminator.

  cfallin

  committed
  940bc40

  Copy full SHA for 940bc40

• WIP.

  cfallin

  committed
  0f270e5

  Copy full SHA for 0f270e5

• do not remove redundant move if we don't have local (within-basic-block) vreg tags, as this might throw off the checker

  cfallin

  committed
  e33790d

  Copy full SHA for e33790d

Commits on Jun 10, 2021

• Some fuzzbug fixes

  cfallin

  committed
  f898b8d

  Copy full SHA for f898b8d

• Working redundant-move elimination

  cfallin

  committed
  2851ac8

  Copy full SHA for 2851ac8

• Improve redundant-move elimination: don't skip the case where we don't hvae vreg metadata within a BB. Instead, keep the checker happy by feeding more metadata to it on every move.

  cfallin

  committed
  4ba7b2f

  Copy full SHA for 4ba7b2f

• Use hashset to avoid linear scan in conflict-bundle-set deduplication

  cfallin

  committed
  fcbf384

  Copy full SHA for fcbf384

Commits on Jun 11, 2021

• TODO list update

  cfallin

  committed
  09b2dd4

  Copy full SHA for 09b2dd4

• Avoid stack-to-stack moves by allocating an extra spillslot and re-using the scratch reg instead.

  cfallin

  committed
  1bd1248

  Copy full SHA for 1bd1248

• Update TODO list

  cfallin

  committed
  ea81422

  Copy full SHA for ea81422

• Always recompute prio when recomputing bundle properties; otherwise is zero on new bundles and causes spill weight to be zeroed

  cfallin

  committed
  a686d5a

  Copy full SHA for a686d5a

Commits on Jun 18, 2021

• Add design document.

  cfallin

  committed
  6ec6207

  Copy full SHA for 6ec6207

• Fix typo (thanks @bjorn3).

  Show description for 6944bc4
  

  cfallin

  and

  bjorn3

  authored
  6944bc4

  Copy full SHA for 6944bc4

• Cleanup: split allocator implemntation into 11 files of more reasonable size.

  cfallin

  committed
  b36a563

  Copy full SHA for b36a563

Commits on Jun 19, 2021

• Add doc-comment note on Edit that stack-to-stack moves are never generated.

  cfallin

  committed
  36975b8

  Copy full SHA for 36975b8

• Remove unused pred_pos

  cfallin

  committed
  caf7274

  Copy full SHA for caf7274

• Keep internal modules private, but re-export under fuzzing feature flag

  cfallin

  committed
  50eb6fc

  Copy full SHA for 50eb6fc

• Add fixed-non-allocatable operand support.

  cfallin

  committed
  736f636

  Copy full SHA for 736f636

• Make bitvec public; it is used by regalloc.rs shim too.

  cfallin

  committed
  22eed0a

  Copy full SHA for 22eed0a

• TODO update: make note on idea for large-input support

  cfallin

  committed
  a58d36f

  Copy full SHA for a58d36f

• reduce nesting level in DESIGN.md to make pandoc-to-pdf happy

  cfallin

  committed
  21fb233

  Copy full SHA for 21fb233

Commits on Jun 21, 2021

• Revert "Add fixed-non-allocatable operand support."

  Show description for 245c212
  cfallin

  committed
  245c212

  Copy full SHA for 245c212

Commits on Jun 22, 2021

• Remove infinite-loop check: it is not a high enough bound in some pathological cases (e.g., gc::many_live_refs test in wasmtime), and it has served its purpose in testing. We can rely on more det…

  Show description for f27abc9
  cfallin

  committed
  f27abc9

  Copy full SHA for f27abc9

• Fix perf issue with many safepoints.

  Show description for 66d6821
  cfallin

  committed
  66d6821

  Copy full SHA for 66d6821

Commits on Jul 28, 2021

• Fix heuristic-cost function overflow with high loop depth (found by @Amanieu).

  cfallin

  committed
  4c193a9

  Copy full SHA for 4c193a9

Commits on Aug 12, 2021

• Update TODO with new items from private feedback from @julian-seward1.

  cfallin

  committed
  c9e8a87

  Copy full SHA for c9e8a87

• Rename OperandPolicy to OperandConstraint as per feedback from @julian-seward1.

  cfallin

  committed
  84285c2

  Copy full SHA for 84285c2

• Use "vector" or "sequence" in lieu of "list" in DESIGN.md, except when referring to linked lists.

  Show description for a591535
  cfallin

  committed
  a591535

  Copy full SHA for a591535

• Add some more detail about spill bundles to design doc.

  cfallin

  committed
  0c79584

  Copy full SHA for 0c79584

• Some more design-doc and TODO updates from @julian-seward1's feedback.

  cfallin

  committed
  38323e0

  Copy full SHA for 38323e0

• Convert all log::debug to log::trace.

  cfallin

  committed
  3e1e0f3

  Copy full SHA for 3e1e0f3

• Improve domtree as per @Amanieu's feedback.

  cfallin

  committed
  1f30958

  Copy full SHA for 1f30958

• Fix comment in postorder.rs.

  cfallin

  committed
  b76b774

  Copy full SHA for b76b774

• Review feedback.

  cfallin

  committed
  2f85643

  Copy full SHA for 2f85643

• Review feedback.

  cfallin

  committed
  7652b4b

  Copy full SHA for 7652b4b

• Review feedback: bitvec: struct-like enum variants, and factor out one-item cache.

  cfallin

  committed
  82b7e6b

  Copy full SHA for 82b7e6b

• BitVec: remove zero words to avoid expanding when unnecessary.

  cfallin

  committed
  eaf8647

  Copy full SHA for eaf8647

• Derive PartialOrd/Ord/Hash for Operand.

  cfallin

  committed
  c071e44

  Copy full SHA for c071e44

• Debug output for Operands: omit default/most common positions.

  cfallin

  committed
  ffc06b2

  Copy full SHA for ffc06b2

• Add malloc/free optimization to TODO

  cfallin

  committed
  f1a989f

  Copy full SHA for f1a989f

• Fix BitVec::get_or_insert to scan only once.

  cfallin

  committed
  8ed83e3

  Copy full SHA for 8ed83e3

Commits on Aug 13, 2021

• Replace remaining instances of use of debug feature with debug_assertions.

  Show description for 69ad31f
  cfallin

  committed
  69ad31f

  Copy full SHA for 69ad31f

Commits on Aug 14, 2021

• Fix bug in refactored BitVec (found by @Amanieu).

  cfallin

  committed
  e10bffb

  Copy full SHA for e10bffb

Commits on Aug 31, 2021

• Address review comments: more doc comments and some minor refactorings.

  cfallin

  committed
  6d313f2

  Copy full SHA for 6d313f2

• Addressed more review comments.

  cfallin

  committed
  3a18564

  Copy full SHA for 3a18564

Commits on Sep 1, 2021

• Rename operand positions to Early and Late, and make weights f16/f32 values.

  cfallin

  committed
  b19fa48

  Copy full SHA for b19fa48

• Address review comments.

  cfallin

  committed
  6389071

  Copy full SHA for 6389071

• Address review comments.

  cfallin

  committed
  6f0893d

  Copy full SHA for 6f0893d