First cut at AVX machinery support. The only instruction implemented
is VPERMIL2PS, and it's probably buggy. I'm checking this in with the
hope that other people can start helping out with (a) testing this,
and (b) adding instructions.
NDISASM support is not there yet.
For PASS1 warnings, only do them when pass0 == 1. The prior passes
are to be considered training passes. This is a bit awkward if we
then hit an error, but it's better than n repeated warnings.
Don't combine type and size into a single argument; *every* backend
immediately breaks them apart, so it's really just a huge waste of
effort. Additionally, it avoids using short immediates in the
resulting code, which is a bad thing.
OpenWatcom doesn't like 64-bit switch arguments; the change to 64-bit
type arguments caused that to happen in outmacho.c. Hack around it
for now; however, realistically speaking the whole bit stealing thing
is probably a bad idea, especially since virtually all CPUs handle
short immediates better than long ones.
Use a 32-bit limb size ("like a digit, but bigger") for floating-point
conversion. This cuts the number of multiplications per constant by a
factor of four.
This means supporting fractional-limb-sized numbers, so while we're at
it, add support for 8-bit floating point numbers (apparently used in
graphics and in audio compression applications.)
Revamp the address- and prefix-handling code to make more sense in
64-bit mode. We are now a lot closer to where we want to be, but
we're not quite there yet.
ndisasm may very well have problems, or give counterintuitive output.
However, checking it in so we can make forward progress.
Correct the handling of floating-point tokens in the preprocessor.
The preprocessor scanner and the main scanner really are painfully
divergent for no good reason.
Proper use of bool and enum makes code easier to debug. Do more of
it. In particular, we really should stomp out any residual uses of
magic constants that aren't enums or, in some cases, even #defines.
Both C and C++ have "bool", "true" and "false" in lower case; C
requires <stdbool.h> for this, in C++ it is an inherent type built
into the compiler. Use those instead of the old macros; emulate with
a simple typedef enum if unavailable.
Change cloc_t to struct location, and reorder the members so that it
should fit in 16 bytes instead of needing 8 bytes of extra padding on
64-bit machines.
Change loc_t to cloc_t to avoid AIX conflict.
We really shouldn't use _t names at all; they are usually considered
platform types, but worry about that later.
Concentrate compiler dependencies to compiler.h; make sure compiler.h
is included first in every .c file (since some prototypes may depend
on the presence of feature request macros.)
Actually use the conditional inclusion of various functions (totally
broken in previous releases.)
Add special operators to allow the use of floating-point constants in
contexts other than DW/DD/DQ/DT/DO.
As part of this checkin, make MAX_KEYWORD generated by tokhash.pl,
since it knows what all the keywords are so it can tell which one is
the longest.
Implement oword, reso, do, as well as the SO flag to instructions. No
instructions are actually flagged with SO yet, but this allows us to
specify 128-bit sizes in instruction patterns.
This checkin completes what is required to actually generate SSE5
instructions. No support in the disassembler yet.
This checkin covers:
- Support for actually generating DREX prefixes.
- Support for matching operand "operand X must match Y"
Add the SSSE3, SSE4.1 and SSE4.2 instruction sets. Change \332 to be
a literal 0xF2 prefix, by analog with \333 for 0xF3 prefix (the
previous \332 flag changed to \335). This is necessary to get the REX
prefix in the right place for instructions that use it.
We are going to have to go in and change existing instruction patterns
which use these, as well.
Support r/m operands for non-integer operands types, i.e. mmx or xmm
operands. This allows mmx and xmm operands to be written more
compactly, speeding up the assembler.
We have a lot of enumerations; by declaring fields as such, we make it
easier when debugging, since the debugger can display the enumerations
in cleartext. However, make sure exceptional values (like -1) are
included in the enumeration, since the compiler otherwise may not
include it in the valid range of the enumeration.
Implement "REL" and "ABS" modifiers for offsets in 64-bit mode. This
replaces "rip+XXX" type addressing. The infrastructure to set the default
mode is there, but there is nothing to throw the switch just yet.
Get rid of magic open-coded register numbers. We now keep track of
a total of three different kinds of register numbers: the register
enumeration (regs.h), the x86 register value, and the register flags.
That has all the information we need.
Additionally, do massive revamping of the EA generation code and the
REX generation logic.
64-bit support for ndisasm. This is very much an initial attempt, and
there are guaranteed to be bugs in the code. However, some *very*
preliminary testing seems to indicate it's not completely off-base.
