This makes codegen faster (by perhaps 10-20% in the case of Jit64,
I didn't measure too closely), which helps speed up NBA Live 2005
a little. But the game still has serious performance issues.
Prefer using eax to isolate the sign bit. This saves a byte when the
destination ends up as r8-15, because those require a REX prefix.
Before:
41 8B C5 mov eax,r13d
41 C1 ED 1F shr r13d,1Fh
44 03 E8 add r13d,eax
41 D1 FD sar r13d,1
After:
41 8B C5 mov eax,r13d
C1 E8 1F shr eax,1Fh
44 03 E8 add r13d,eax
41 D1 FD sar r13d,1
This is done entirely through interpreter fallbacks. It would
probably be possible to implement this using host exception
handlers instead, but I think it would be a lot of complexity
for a rarely used feature, so let's not do it for now.
For performance reasons, there are two settings for this feature:
One setting which does enables just what True Crime: New York City
needs and one setting which enables it all. The latter makes
almost all float instructions fall back to the interpreter.
We were using a "value" register to avoid clobbering physical_addr,
but this isn't actually needed anymore. The only bits we need from
physical_addr after we start clobbering it are bits 5-9, and
those bits are identical in effective_addr and physical_addr,
so we can read them from effective_addr instead.
When making 92d1d60, I checked whether the ~0x1f masking in dcbx
actually was necessary. I came to the conclusion that it wasn't,
so I removed it. However, I hadn't checked the second half of
InvalidateICache closely enough - the masking is actually needed.
This commit re-adds the masking, but this time in C++ code instead
of in jitted code in order to save icache. Though I suppose the
difference doesn't matter all that much, since this is in farcode
and all...
Hopefully fixes https://bugs.dolphin-emu.org/issues/12612.
It was believed that this only mattered when the rounding mode was
set to round to infinity, which games generally don't do, but it
can also affect the sign of the output when the inputs are all zero.
So it turns out you have to pass XMM0 as the clobber register
to HandleNaNs, because HandleNaNs uses BLENDVPD and BLENDVPD
implicitly uses XMM0, and nobody noticed when I broke this in
2c38d64 because nobody plays the one game that needs accurate NaNs.
This reverts commit 66b992cfe4.
A new (additional) correctness issue was revealed in the old
AArch64 code when applying it on top of modern JitArm64:
LSR was being used when LSRV was intended. This commit uses LSRV.
The fast path of using CVTSD2SS/FCVTN rounds the significand if it
can't be exactly represented as a single, whereas the accurate path
instead truncates the significand. So we should only use the fast
path if we know that the lower bits of the significand are not set.
This is not known to affect any games.
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.
These optimizations were already present, but only when d == a. They
also make sense when this condition does not hold.
- imm == 0
Before:
41 BB 00 00 00 00 mov r11d,0
45 2B DF sub r11d,r15d
After:
45 8B DF mov r11d,r15d
41 F7 DB neg r11d
- imm == -1
Before:
41 BD FF FF FF FF mov r13d,0FFFFFFFFh
44 2B EE sub r13d,esi
0F 93 45 68 setae byte ptr [rbp+68h]
After:
44 8B EE mov r13d,esi
41 F7 D5 not r13d
C6 45 68 01 mov byte ptr [rbp+68h],1
fcmpX only updates the FPCC bits, not the C bit.
This was already correctly implemented in the interpreter.
Not known to affect any games, but affects a hardware test.
When I added the software FMA path in 2c38d64 and made us use
it when determinism is enabled, I was assuming that either the
performance impact of software FMA wouldn't be too large or CPUs
that were too old to have FMA instructions were too slow to run
Dolphin well anyway. This was wrong. To give an example, the
netplay performance went from 60 FPS to 30 FPS in one case.
This change makes netplay clients negotiate whether FMA should
be used. If all clients use an x64 CPU that supports FMA, or
AArch64, then FMA is enabled, and otherwise FMA is disabled.
In other words, we sacrifice accuracy if needed to avoid massive
slowdown, but not otherwise. When not using netplay, whether to
enable FMA is simply based on whether the host CPU supports it.
The only remaining case where the software FMA path gets used
under normal circumstances is when an input recording is created
on a CPU with FMA support and then played back on a CPU without.
This is not an especially common scenario (though it can happen),
and TASers are generally less picky about performance and more
picky about accuracy than other users anyway.
With this change, FMA desyncs are avoided between AArch64 and
modern x64 CPUs (unlike before 2c38d64), but we do get FMA
desyncs between AArch64 and old x64 CPUs (like before 2c38d64).
This desync can be avoided by adding a non-FMA path to JitArm64 as
an option, which I will wait with for another pull request so that
we can get the performance regression fixed as quickly as possible.
https://bugs.dolphin-emu.org/issues/12542