The feature isn't universally available on 64-bit CPUs.
Note that in i386-gen.c:isa_dependencies[] I'm only adding it to models
where I'm certain the functionality exists. For Nocona and Core I'm
uncertain in particular.
The attribute really specifies that the sum of register and memory
operands is 4. Express it like that in most places, while using the 2nd
(apart from XOP) CPU feature flags (FMA4) in reversed operand matching
logic.
With the use in build_modrm_byte() gone, part of an assertion there
also becomes meaningless - simplify that at the same time.
With all uses of the opcode modifier field gone, also drop that.
This really isn't a "modifier" and rather ought to live next to the base
opcode anyway. Use the bits we presently have available to fit in the
field, renaming it to opcode_space. As an intended side effect this
helps readability at the use sites, by shortening the references quite a
bit.
In generated code arrange for human readable output, by using the
SPACE_* constants there rather than raw numbers. This may aid debugging
down the road.
All glibc malloc() implementations I've checked have a smallest
allocation size worth of 3 pointers, with an increment worth of 2
pointers. Hence mnemonics with multiple templates can be stored more
efficiently when maintaining the shared "name" field only in the actual
hash entry. (To express the shared nature, also convert "name" to by
pointer-to-const.)
While doing the conversation also pull out common code from the involved
if/else construct in expand_templates().
When generating the mnemonic string table we already set up an
identifier for the following entry in a number of cases. Re-use that on
the next loop iteration rather than re-doing allocation and conversion.
Compact the mnemonic string table such that the tails of longer
mnemonics are re-used for shorter ones, going beyond what compilers
would typically do, but matching what ELF linkers may do when processing
SHF_MERGE|SHF_STRINGS sections. This reduces table size by about 12.5%.
Using full pointers to reference the insn mnemonic strings is not very
efficient. With overall string size presently just slightly over 20k,
even a 16-bit value would suffice. Use "unsigned int" for now, as
there's no good use we could presently make of the otherwise saved 16
bits.
For 64-bit builds this reduces table size by 6.25% (prior to the recent
ISA extension additions it would have been 12.5%), with a similar effect
on cache occupation of table entries accessed. For PIE builds of gas
this also reduces the number of base relocations quite a bit (obviously
independent of bitness).
The newer update-copyright.py fixes file encoding too, removing cr/lf
on binutils/bfdtest2.c and ld/testsuite/ld-cygwin/exe-export.exp, and
embedded cr in binutils/testsuite/binutils-all/ar.exp string match.
TSXLDTRK takes RTM as a prereq. Additionally introduce an umbrella "tsx"
extension option covering both RTM and HLE, paralleling the "abm" one we
already have.
SEV-ES is an extension to SVME. SNP in turn is an extension to SEV-ES,
and yet in turn RMPQUERY is a SNP extension.
Note that cpu_arch[] has no SNP entry, so CPU_ANY_SNP_FLAGS remains
unused (just like CPU_SNP_FLAGS already is).
Like various other features AMX-TILE takes XSAVE as a prereq.
XSAVES, unconditionally using compacted format, in turn effectively
takes XSAVEC as a prereq (an SDM clarification to this effect is in the
works).
Like AVX512-FP16, several other extensions require wider than 16-bit
mask registers. As a result they take AVX512BW as a prereq, not (just)
AVX512F. Which in turn points out wrong expectations in the noavx512-1
testcase.
SSE itself takes FXSR as a prereq. Like AES, PCLMUL, and SHA both GFNI
and KL take SSE2 as a prereq, for operating on packed integers. And
while correcting KL also record it as a prereq to WIDEKL.
Getting both forward and reverse ISA dependencies right / consistent has
been a permanent source of mistakes. Reduce what needs specifying
manually to just the direct forward dependencies. Transitive forward
dependencies as well as reverse ones are now derived and hence cannot go
out of sync anymore (at least in the vast majority of cases; there are a
few special cases to still take care of manually). In the course of this
several CPU_ANY_*_FLAGS disappear, requiring adjustment to the
assembler's cpu_arch[].
Note that to retain the correct reverse dependency of AVX512F wrt
AVX512-VP2INTERSECT, the latter has the previously missing AVX512F
prereq added.
Note further that to avoid adding the following undue prereqs:
* ATHLON, K8, and AMDFAM10 gain CMOV and FXSR,
* IAMCU gains 387,
auxiliary table entries (including a colon-separated modifier) are
introduced in addition to the ones representing from converting the old
table.
To maintain forward-only dependencies between AVX (XOP) and SSE* (SSE4a)
(i.e. "nosse" not disabling AVX), reverse dependency tracking is
artifically suppressed.
As a side effect disabling of SSE or SSE2 will now also disable AES,
PCLMUL, and SHA (respective elements were missing from
CPU_ANY_SSE2_FLAGS).
While originally indeed used for register size checking only, the
attribute has been used for memory operand size checking as well already
for quite a while, with more such uses recently having been added.
Having a "None" field in the vast majority of entries is needlessly
cluttering the overall table. Instead of this being a separate field,
use a representation matching that of Intel SDM and AMD PM for the main
use of the field: Append the value after a / as the separator.
Speed up gas startup by avoiding runtime allocation of the instances of
type "templates". At the same time cut the memory requirement to just
very little over half (not even accounting for any overhead
notes_alloc() may incur) by reusing the "end" slot of a preceding entry
for the "start" slot of the subsequent one.
Now that the table is local to gas, ARRAY_SIZE() can be used to
determine the end of the table. Re-arrange the processing loop in
md_begin() accordingly, at the same time folding the two calls to
notes_alloc() into just one.
Unlike many other architectures, x86 does not share an opcode table
between assembly and disassembly. Any consumer of libopcodes would only
ever access one of the two. Since gas is the only consumer of the
assembly data, move it there. While doing so mark respective entities
"static" in i386-gen (we may want to do away with i386_regtab_size
altogether).
This also shrinks the number of relocations to be processed for
libopcodes.so by about 30%.
With the general use of C99 there's no need anymore to have i386-gen
produce these. For more frequently used ones introduce local #define-s,
while others are simply spelled out directly. While doing this move
some static constants into more narrow scopes.
Note that as a "side effect" this corrects type_names[]'es imm8s entry.
There are just 4 templates using it, which can be easily identified by
other means, as D is set only on a very limited number of FPU templates.
Also move the respective conditional out of the code path taken by all
"reverse match" insns (it probably should have been this way already
before, to avoid the one conditional in the common case).
With this the templates which had FloatR dropped no longer differ from
their AT&T syntax + mnemonic counterparts - the only difference is now
which of the two would be recognized. For this, however, we don't need
two templates - we can simply arrange the condition for setting
Opcode_FloatR accordingly.
Attributes which aren't used together in any single insn template can be
converted from individual booleans to a single enum, as was done for a few
other attributes before. This is more space efficient. Collect together
all attributes which express special operand constraints (and which fit
the criteria for folding).
Earlier tidying still missed an opportunity: There's no need for the
"anyimm" static variable. Instead of using it in the loop to mask
"allowed" (which is necessary to satisfy operand_type_or()'s assertions)
simply use "mask", requiring it to be calculated first. That way the
post-loop masking by "mask" ahead of the operand_type_all_zero() can be
dropped.
