The end goal will be also merging of disasm and cpu decoder to one module and remove the disasm.
Two bug fixes on the way:
TBM: fixed 64-bit TBM instructions with memory access (did 32-bit load instead of 64-bit)
BMI2: fixed operands order for PEXT/PDEP instructions
AVX2: fixed gather instruction decoding bug from decoder alias commit
Bochs instruction emulation handlers won't refer to direct fields of instructions like MODRM.NNN or MODRM.RM anymore.
Use generic source/destination indications like SRC1, SRC2 and DST.
All handlers are modified to support new notation. In addition fetchDecode module was modified to assign sources to instructions properly.
Immediate benefits:
- Removal of several duplicated handlers (FMA3 duplicated with FMA4 is a trivial example)
- Simpler to understand fetch-decode code
Future benefits:
- Integration of disassembler into Bochs CPU module, ability to disasm bx_instruction_c instance (planned)
Huge patch. Almost all source files wre modified.
Averything that required cpu.h include now has it explicitly and there are a lot of files not dependant by CPU at all which will compile a lot faster now ...
1. Review and commit patch
[ 896733 ] Lazy flags, for more instructions, only 1 src op
May be partially, but I hope to get all ideas from patch in
2. Get Bochs speedup after lazy flags optimization
3. Most important for me: improve correctness of emulation by handling several
undocumented EFLAGS modifications. And finally pass
UFLAGS - Undefined Flags Test v 3.0
Copyright (C) Potemkin's Hackers Group (PHG) 1989,1995
The test still fails on > 50% of its checks.
Notes from the author:
Here is another one of my speed up patches. Unlike my previous speedups
this one will help more platforms than just X86. It cleans up the Data
Xfer instructions. Since the Data Xfer instructions are the most often
executed instructions it gives a noticable boost in speed. The basic
optimization technique was to eliminate intermediate variables and pass
a pointer to the final destination or original source to the
read_virtual_whatever and the write_virtual_whatever functions.
there to offer a way to substitute more efficient code
to do the RMW cases. At the moment, they just map to
the normal functions.
Sorry, restored the previous version ...
in cpu.cc out of the main loop, and into the asynchronous
events handling. I went through all the code paths, and
there doesn't seem to be any reason for that code to be
in the hot loop.
Added another accessor for getting instruction data, called
modC0(). A lot of instructions test whether the mod field
of mod-nnn-rm is 0xc0 or not, ie., it's a register operation
and not memory. So I flag this in fetchdecode{,64}.cc.
This added on the order of 1% performance improvement for
a Win95 boot.
Macroized a few leftover calls to Write_RMV_virtual_xyz()
that didn't get modified in the x86-64 merge. Really, they
just call the real function for now, but I want to have them
available to do direct writes with the guest2host TLB pointers.
to bitfields. bxInstruction_c is now 24 bytes, including 4 for
the memory addr resolution function pointer, and 4 for the
execution function pointer (16 + 4 + 4).
Coded more accessors, to abstract access from most code.
with accessors. Had to touch a number of files to update the
access using the new accessors.
Moved rm_addr to the CPU structure, to slim down bxInstruction_c
and to prevent future instruction caching from getting sprayed
with writes to individual rm_addr fields. There only needs to
be one. Though need to deal with instructions which have
static non-modrm addresses, but which are using rm_addr since
that will change.
bxInstruction_c is down to about 40 bytes now. Trying to
get down to 24 bytes.
use accessors. This lets me work on compressing the
size of fetch-decode structure (now called bxInstruction_c).
I've reduced it down to about 76 bytes. We should be able
to do much better soon. I needed the abstraction of the
accessors, so I have a lot of freedom to re-arrange things
without making massive future changes.
Lost a few percent of performance in these mods, but my
main focus was to get the abstraction.
Read-Modify-Write instructions. The first read phase stores
the host pointer in the "pages" field if a direct use pointer
is available. The Write phase first checks if a pointer was
issued and uses it for a direct write if available.
I chose the "pages" field since it needs to be checked by the
write_RMW_virtual variants anyways and thus needs to be
cached anyways.
Mostly the mods where to access.cc, but I did also macro-ize
the calls to write_RMW_virtual...() in files which use it
and cpu.h. Right now, the macro is just a straight pass-through.
I tried expanding it to a quick initial check for the pointer
availability to do the write in-place, with a function call
as a fall-back. That didn't seemed to matter at all.
Booting is not helped by this really. The upper bound of
the gain is 5 or 6%, and that's only if you have a loop that
looks like:
label:
add [eax], ebx ;; mega read-modify-write instruction
jmp label ;; intensive loop.
tries to fix it. The shortcuts to register names such as AX and DL are
#defines in cpu/cpu.h, and they are defined in terms of BX_CPU_THIS_PTR.
When BX_USE_CPU_SMF=1, this works fine. (This is what bochs used for
a long time, and nobody used the SMF=0 mode at all.) To make SMP bochs
work, I had to get SMF=0 mode working for the CPU so that there could
be an array of cpus.
When SMF=0 for the CPU, BX_CPU_THIS_PTR is defined to be "this->" which
only works within methods of BX_CPU_C. Code outside of BX_CPU_C must
reference BX_CPU(num) instead.
- to try to enforce the correct use of AL/AX/DL/etc. shortcuts, they are
now only #defined when "NEED_CPU_REG_SHORTCUTS" is #defined. This is
only done in the cpu/*.cc code.
