* Add `.llvmbc`/`.llvmir` to `ObjectFormat` for emitting unoptimized IR
* Replace `--verbose-air`, `--verbose-intern-pool`,
`--verbose-generic-instances`, `--verbose-llvm-ir`, and
`--verbose-llvm-bc` with various `--debug-log` scopes
* `--debug-log illegal-air`: pre-legalized Air
* `--debug-log legal-air`: post-legalized Air
* `--debug-log air`: final Air with liveness
* `--debug-log ip-stats`: InternPool statistics
* `--debug-log ip-values`: InternPool values
* `--debug-log ip-generic-instances`: InternPool generic instances
* `--debug-log llvm-ir`: pre-optimized LLVM IR
* `--debug-log opt-llvm-ir`: post-optimized LLVM IR
* Delete `--verbose-cimport` which does nothing
* Delete `--verbose-llvm-cpu-features` which is equivalent to looking at
the outputs of `--show-builtin` and `--debug-log llvm-ir`
Add four assembly files for half-float to integer conversions using
QFloat (qf16) on Hexagon V79.
Each function provides IEEE round-to-nearest-even or fast
round-half-away-from-zero conversion for signed and unsigned 8-bit and
16-bit integer destinations.
V79 lacks direct half-float to integer conversion. V75 and below have
native IEEE vcvt instructions,
while V81 and newer have dedicated hardware support.
Co-authored-by: Sumanth Gundapaneni <sgundapa@quicinc.com>
Both on models and architecture levels this time.
As part of this, I removed neon from the baseline model because that isn't
actually part of the ARMv7 baseline on any platform I'm aware of, and it made
our baseline binaries not work on some real boards, e.g. Tegra 2. We keep
vfp3d16 since virtually every ARMv7 board has VFP.
To not regress PlayStation Vita performance, I also added a kermit model with
neon enabled. This was necessary since vortex_a9 lost all NEON/VFP features by
default.
It was really confusing that feature_overrides was pulling double duty for both
features and models. Now there's a clean split between the two. The new
cpu_overrides also subsumes omit_cpus. This in turn uncovered a bug where some
x86 models were not getting the smep/smap features.
dependencies inside build.zig.zon now support `paths` and `version`
fields which can be used to affect the filtered files of the package as
well as the hash.
On macOS with the stage2 C backend, we use the software AES
implementation because the C backend can't pass vectors to
inline assembly.
But because the hardware still supports hardware AES, we use to
set `has_hardware_support` to true.
Fix that and improve readability by the way.
Use doubling and Montgomery squaring to speed up R2 construction
for larger moduli.
For small (1 or 2 limbs) moduli, the shift-based method remains
faster, so keep it.