This new function is like MOVP2R, except it masks out the lower 12 bits,
returning them instead of writing them to the register. These lower
12 bits can then be used as an offset for LDR/STR. This lets us turn
ADRP+ADD+LDR sequences with a zero offset into ADRP+LDR sequences with
a non-zero offset, saving one instruction.
ARM64 can do perform various types of sign and zero extension on a
register value before using it. The Arm64Emitter already had support for
this, but it was kinda hidden away.
This commit exposes the functionality by making the ExtendSpecifier enum
available everywhere and adding a new ArithOption constructor.
The previous implementation of Force25BitPrecision was essentially a
translation of the x86-64 implementation. It worked, but we can make a
more efficient implementation by using an AArch64 instruction I don't
believe x86-64 has an equivalent of: URSHR. The latency is the same as
before, but the instruction count and register count are both reduced.
Heavily simplify logical immediate encoding.
This is based on the observation that if a valid repeating element
exists, it repeats through `value`. Thus it does not matter which
one you analyse. Thus we skip over the least significent element
if LSB = 1 by masking it out with `inverse_mask_from_trailing_ones`,
to avoid the degenerate case of a stretch of 1 bits going 'round
the end' of the word.
Manually encoding and decoding logical immediates is error-prone.
Using ORRI2R and friends lets us avoid doing the work manually,
but in exchange, there is a runtime performance penalty. It's
probably rather small, but still, it would be nice if we could
let the compiler do the work at compile-time. And that's exactly
what this commit does, so now I have no excuse for trying to
manually write logical immediates anymore.
SPDX standardizes how source code conveys its copyright and licensing
information. See https://spdx.github.io/spdx-spec/1-rationale/ . SPDX
tags are adopted in many large projects, including things like the Linux
kernel.
If we can prove that FCVT will provide a correct conversion,
we can use FCVT. This makes the common case a bit faster
and the less likely cases (unfortunately including zero,
which FCVT actually can convert correctly) a bit slower.
If we know at compile time that the PPC carry flag definitely
has a certain value, we can bake that value into the emitted code
and skip having to read from PPCState.
More or less a complete rewrite of the function which aims
to be equally good or better for each given input, without
relying on special cases like the old implementation did.
In particular, we now have more extensive support for
MOVN, as mentioned in a TODO comment.
I don't really see the use of this. (Maybe in the past it
was used for when we need a constant number of instructions
for backpatching? But we don't use MOVI2R for that now.)
PR 9262 added a bunch of Jit64 optimizations, some of
which were already in JitArm64 and some which weren't.
This change ports the latter ones to JitArm64.
At least on some CPUs (I found out about this from the
Arm Cortex-A76 Software Optimization Guide), using X30
with BLR is one cycle slower than using another register.
Similar in nature to e28d063539 in which
this same change was applied to the x64 emitter.
There's no real requirement to make this const, and this should also
be decided by the calling code, considering we had places that would
simply cast away the const and carry on
CNTVCT_EL0 is force-enabled on all linux plattforms.
Windows is untested, but as this is the best way to get *any* low
overhead performance counters, they likely use it as well.
JosJuice