Optimize register signbit extraction 2 - #8

Merged

dsisco11 merged 21 commits into

masterfrom

optimize-register-signbit-extraction-2

Sep 15, 2026

Commits on Sep 14, 2026

perf: reduce instruction count for movemask_slim
dsisco11
committed
b9dd800
perf: reduce instruction count for 8bit `movemask_slim`
dsisco11
committed
6e06101
perf(RegisterMask): reduce instruction count for `any()` and `none()` methods
dsisco11
committed
9fb3cc3
fix: compile-time constant register construction was absurdly unoptimal
dsisco11
committed
3cbe92d
perf: refactor code that constructs registers from an increasing index sequence to instead use the `make_static_register` helper
dsisco11
committed
1cbe116
perf: improve blend_slow method implementations
dsisco11
committed
05e5761
chore: rename `register_blend_slow` -> `register_blend_constexpr` for clarity
dsisco11
committed
f18bd05
perf: Runtime integer lane insertion
dsisco11
committed
389d2ef
docs: amend optimizatio nreport
dsisco11
committed
12e38ed
perf: improve `make_static_register` codegen for unoptimized builds
dsisco11
committed
ca68534
chore: rename `make_static_register_impl` -> `make_static_register_from_indices`
dsisco11
committed
bfb8e54
perf: slightly improve `register_from_array`
dsisco11
committed
7c5a529
chore: rename `register_from_array` -> `register_from_array_constexpr`
dsisco11
committed
70bab0a
chore: remove dead code
dsisco11
committed
6b04c67
perf: Signed minimum-position reductions
dsisco11
committed
46122af
perf: reduce instruction count for `RegisterMask::all()` method
dsisco11
committed
9539d3b
perf: Runtime 128-bit whole-register bit shifts
dsisco11
committed
4570bcb
fix: build error due to type deduction
dsisco11
committed
dda4bf3
chore: specify SimdImpl return types explicitly
dsisco11
committed
3c26052

Navigation Menu

Optimize register signbit extraction 2 - #8

Optimize register signbit extraction 2#8
dsisco11 merged 21 commits into
masterfrom
optimize-register-signbit-extraction-2

Commits on Sep 14, 2026

perf: reduce instruction count for movemask_slim

perf: reduce instruction count for 8bit `movemask_slim`

perf(RegisterMask): reduce instruction count for `any()` and `none()` methods

fix: compile-time constant register construction was absurdly unoptimal

perf: refactor code that constructs registers from an increasing index sequence to instead use the `make_static_register` helper

perf: improve blend_slow method implementations

chore: rename `register_blend_slow` -> `register_blend_constexpr` for clarity

perf: Runtime integer lane insertion

docs: amend optimizatio nreport

perf: improve `make_static_register` codegen for unoptimized builds

chore: rename `make_static_register_impl` -> `make_static_register_from_indices`

perf: slightly improve `register_from_array`

chore: rename `register_from_array` -> `register_from_array_constexpr`

chore: remove dead code

perf: Signed minimum-position reductions

perf: reduce instruction count for `RegisterMask::all()` method

perf: Runtime 128-bit whole-register bit shifts

fix: build error due to type deduction

chore: specify SimdImpl return types explicitly

Commits on Sep 15, 2026

tests: fix outdated codegen test expectations for register mask operations

chore(SimdImpl): use register alias rather than explicit intrinsic typename