Add config entry to mask out the lower bits in page table pointers.
This is intended to allow users of Dynarmic to pack small integers
inside pointers and update the pair atomically without locks.
These lower bits can be masked out due to the expected alignment in
pointers inside the page table.
For the given usage, using AND on the pointer acts the same way as a
TEST instruction. That said when the mask value is zero, TEST is still
emitted to keep the same behavior.
Here we increase the similarity between the A64 and A32 front-ends in terms of their
page_table handling code. In this commit, we:
* Reserve and use r14 as a register to store the page_table pointer.
* Align the code to be more similar in structure.
* Add a conf member to A32EmitContext.
* Remove scratch argument from EmitVAddrLookup.
* Use pext/pdep where not previously used
* Limit pext/pdep to non-AMD platforms due to slowness on AMD
* Use imul/and as alternatives for AMD and non-BMI2 platforms
The guest program often accesses the NZCV flags directly much less
often than we need to use them for jumps and other such uses.
Therefore, we store our flags in cpsr_nzcv in a x64-friendly format.
This allows for a reduction in conditional jump related code.
Several issues:
1. Several terminal instructions did not stop at the end of a single-step block
2. x64 backend for the A32 frontend sometimes polluted upper_location_descriptor with the single-stepping flag
We also introduce the enable_optimizations parameter to the A32 frontend.
Instead of looking up the page table like:
table[addr >> 12][addr & 0xFFF]
We can use a global offset on the table to query the memory like:
table[addr >> 12][addr]
This saves two instructions on *every* memory access within the recompiler.
Original change by degasus in A64 emitter
Also solves a performance regression initially introduced by b6e8297e369f2dc4758bafe944e51efb8d1a2552,
primarily due to excessively mismatched load/store sizes causing less than optimal load-to-store forwarding.
Removes unnecessary header dependencies that have accumulated over time
as changes have been made. Lessens the amount of files that need to be
rebuilt when the headers change.
Removes a boost header from the public includes in favor of using the
standard-provided std::variant.
The use of boost in public interfaces is often a dealbreaker for some
people. Given we use std::optional in the header already, we can
transition over to std::variant from boost::variant.
With this removal, this makes all of our dependencies internal to the
library itself.
