SequenceReader performance; elide bounds checks and book-keeping#82765

Open

mgravell wants to merge 23 commits intomainfrom marc/sequencereader

+252-220

Commits on Aug 11, 2023

SequenceReader; elide bounds checks by using ref-oriented approach

mgravell
authored and
jozkee
committed
net6-: don't forget to increment _consumedAtStartOfCurrentSpan

mgravell
authored and
jozkee
committed
- don't use the Unsafe or MemoryMarshal evilness

mgravell
authored and
jozkee
committed
in search, avoid copying out the T value

mgravell
authored and
jozkee
committed
drop unused method

mgravell
authored and
jozkee
committed
reinstate next position; needed to avoid double-fetch in GetNextSpan

mgravell
authored and
jozkee
committed
use better API to get first span in ResetReader

mgravell
authored and
jozkee
committed
simplify Position getter

mgravell
authored and
jozkee
committed
move reset logic to constructor; use constructor in ResetReader

mgravell
authored and
jozkee
committed
always track consumed span when moving to next, even for single segment

mgravell
authored and
jozkee
committed
for single-segment cases, length is trivial; record length at EOF; update position at EOF

mgravell
authored and
jozkee
committed
use named parameter to clarify meaning

mgravell
authored and
jozkee
committed
explicit types (IDE0008)

mgravell
authored and
jozkee
committed
- fix increment bug at end of single-segment buffers

mgravell
authored and
jozkee
committed
simplify advance further

mgravell
authored and
jozkee
committed
remove debug-check from AssertValidPosition() to make the compiler happier

mgravell
authored and
jozkee
committed
TryPeek/TryRead: avoid bounds check by hoisting into locals

mgravell
authored and
jozkee
committed
nit: use local for the addition too (cuts a mov from the JIT output)

mgravell
authored and
jozkee
committed
test fixes

mgravell
authored and
jozkee
committed
Advance: don't short-circuit (we expect both sides to pass)

mgravell
authored and
jozkee
committed
work around JIT regressions

mgravell
authored and
jozkee
committed
nit: TryGetNextBuffer - defer updating result; empty segments are v.rare

mgravell
authored and
jozkee
committed

Commits on Aug 17, 2023

Fix two rebase errors
jozkee
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

SequenceReader performance; elide bounds checks and book-keeping#82765

Commits on Aug 11, 2023

SequenceReader; elide bounds checks by using ref-oriented approach

net6-: don't forget to increment _consumedAtStartOfCurrentSpan

- don't use the Unsafe or MemoryMarshal evilness

in search, avoid copying out the T value

drop unused method

reinstate next position; needed to avoid double-fetch in GetNextSpan

use better API to get first span in ResetReader

simplify Position getter

move reset logic to constructor; use constructor in ResetReader

always track consumed span when moving to next, even for single segment

for single-segment cases, length is trivial; record length at EOF; update position at EOF

use named parameter to clarify meaning

explicit types (IDE0008)

- fix increment bug at end of single-segment buffers

simplify advance further

remove debug-check from AssertValidPosition() to make the compiler happier

TryPeek/TryRead: avoid bounds check by hoisting into locals

nit: use local for the addition too (cuts a `mov` from the JIT output)

test fixes

Advance: don't short-circuit (we expect both sides to pass)

work around JIT regressions

nit: TryGetNextBuffer - defer updating result; empty segments are v.rare

Commits on Aug 17, 2023

Fix two rebase errors