Doing so is not allowed (presumably because compilers are allowed to use a different algorithm for allocating between the two/store extra data such as the length of the array before the pointer).
This bug existed in the original implementation at https://web.archive.org/web/20140708092159/http://www.hydrogenaud.io/forums/index.php?showtopic=52235 and causes Valgrind to emit a warning. Note that this ended up happening even if DSPLLE and the DPL decoder are not enabled in Dolphin.
This fixes a regression from 592ba31. When `a` was a constant 0 and `b`
was a non-constant 0x80000000, the 32-bit negation operation would
overflow, causing an incorrect result. The sign extension needs to happen
before the negation to avoid overflow.
Note that I can't merge the SXTW and NEG into one instruction.
NEG is an alias for SUB with the first operand being set to ZR,
but "SUB (extended register)" treats register 31 as SP instead of ZR.
I've also changed the order for the case where `a` is a constant
0xFFFFFFFF. I don't think the order actually affects correctness here,
but let's use the same order for all the cases since it makes the code
easier to reason about.
This works around an Intel driver bug where, on D3D12 only, dual-source blending behaves incorrectly if the second source is unused on. This bug is visible in skyboxes in Super Mario Sunshine, which first draw clouds and sun flare in greyscale and then draw the sky afterwards with a source factor of 1 and a dest factor of 1-src_color (this results in the clouds being tinted blue). This process is done on an RGB888 framebuffer, so alpha update is disabled. (Color update is enabled; note that if you look at this in Dolphin's fifo analyzer, it won't be enabled because they use the BP mask functionality to only change the blending functions and not alpha/color update, for whatever reason.)
The previous code only updated the PLRU on cache misses, which made it so that the least recently inserted cache block was evicted, instead of the least recently used/hit one.
This regressed in 9d39647f9e (part of #11183, but it was fine in e97d380437), although beforehand it was only implemented for the instruction cache, and the instruction cache hit extremely infrequently when the JIT or cached interpreter is in use, which generally keeps it from behaving correctly (the pure interpreter behaves correctly with it).
I'm not aware of any games that are affected by this, though I did not do extensive testing.
Previously we would only backpatch overflowed address calculations
if the overflow was 0x1000 or less. Now we can handle the full 2 GiB
of overflow in both directions.
I'm also making equivalent changes to JitArm64's code. This isn't because
it needs it – JitArm64 address calculations should never overflow – but
because I wanted to get rid of the 0x100001000 inherited from Jit64 that
makes even less sense for JitArm64 than for Jit64.
