tree 11cf3d696b4de0a689bc641b82897baf356f8f4d
parent 13a3a5108df85ebb0d6bcbea0b3d1da2288a249e
author Bob Nystrom <rnystrom@google.com> 1709170396 -0800
committer GitHub <noreply@github.com> 1709170396 -0800
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJl397cCRC1aQ7uu5UhlAAA70oQAHWYESSLeJ6fkkdkz7yELvIF
 uJCG99Qc57VFMIVHmuKLgyhnUSkai9vjZ/AaHljTdcE7NI1AiioeIwNfYqG70La+
 5V2SNoNV9df0jac1LYg1Mn55R0nEnBN4Q39h4R9Ks/qPK4xj4QOMGS7qoPY7WtpA
 hLimkJF0blPTHb5VTtH1PO02roKQWtSc7UXz51oWufrbc2tv7KetRZjpBPIfS2+T
 K4SJ/rB4vDq+cQSTZSQbfnwueCySJxLhlE00+X/9ss1YD9cx2QeYJ127McmeowZn
 CIIGw4q/i25reYKVdTubhPGsA0J/Mmun7P+QkjOA1poNXoHDJ2BlXL8XK5BdyOGv
 O0UcdtXaq2cBZmZfXBX5mn8ubgm6CiYhPnv+tt/4A3MyR1cGBoK5L507fHLkKzHk
 KGAHpLUnfafS2C4Cte5bqHNFGQswhJuZ/lLJOoM32RsZo+TE/gy5aNiSZnnLzhXi
 Ukwv9RPAs7VklhSB0iSEt8UZcHIKD7IcwiAsyusLmBB8a0ncmF+/7EgYfdP65ih/
 4VSRB5JYyCnytNAqC69LKmghSV0amVGhdMJSs1VX11BbxuWXN5P3u+KrM67ligE3
 O4Ux5ES+UwyZQSl8jWgtLoidn+DFCu7BqRlxrlQqyK1OWp5ttBQz1vzrc5fi9UXQ
 DrqWWGZj314PsifVClMP
 =QIi1
 -----END PGP SIGNATURE-----
 

Fix two bugs around pinned states in the Solver and SolutionCache. (#1409)

Fix two bugs around pinned states in the Solver and SolutionCache.

I was working on an unrelated bug fix and I stumbled into a couple of
incorrect formats in some complex examples. After poking around, I found
two bugs:

- In SolutionCache, it always used the given state of the root Piece
  even if that state wasn't explicitly bound in the parent Solution.
  This causes incorrect output if we are separately formatting a child,
  the parent Solution doesn't bind the child's root, and the best state
  for the child is *not* the unsplit state.

  That's a rare enough combination of events that none of the existing
  tests hit it, but I'll have some new tests for the unrelated bug fix
  that do.

- In Solution, when applying constraints between pieces, it doesn't
  take into account conflicts from pinned pieces. This bug was
  introduced by #1407. It wasn't caught because the one test that
  happens to tickle this doesn't tickle it when a certain child piece
  is formatted separately, but does if you disable that optimization.

The fix to SolutionCache also means we no longer use the root piece's
state as part of the cache key. Somewhat surprisingly, it doesn't seem
to be necessary. Taking the state out of the cache key is good because
it means we're able to reuse more cached Solutions in different
contexts. This PR makes the large benchmark about 5% faster.