As noted in the context of d53e6b98a2 ("x86/Intel: correct disassembly
of fsub*/fdiv*") there's no such thing as Intel syntax without Intel
mnemonics. Enforce this on the assembler side, and disentangle command
line option handling on the disassembler side accordingly.
As a result in the opcode table specifying ATTMnemonic|ATTSyntax becomes
redundant with just ATTMnemonic. Drop the now meaningless ATTSyntax and
remove the then no longer accessible templates.
Bring disassembly back in line with what the assembler accepts, thus
also making it self-consistent (with, in particular selector load/store
insns). While there further add D to all affected insns except ARPL
(where S is used, matching LAR/LSL), to also behave correctly in suffix-
always mode.
While there also hook up the Intel variant of the LKGS test.
For whatever reason in c9f5b96bda ("x86: correct handling of LAR and
LSL") I didn't realize that we can easily use Sv instead of going
through mod_table[]. Redo this aspect of that change.
First of all it is entirely unclear why THREE_BYTE_TABLE_PREFIX() was
introduced by bf890a93a7. Nothing uses the .prefix_requirement values
from the two relevant entries.
And then having VEX_Cn_TABLE() and friends take arguments is misleading.
These aren't used (or pointlessly used in the case of VEX_C5_TABLE); the
respective table index is decoded from the insn (or implied in the case
of VEX_C5_TABLE).
Several already use OP_R(), which rejects the memory forms of insns, and
a few others can easily be converted to do so as well. Note that for it
to be able to use BadOp() without forward declaration, OP_Skip_MODRM() is
moved down.
While there add the previously missing PREFIX_OPCODE to legacy opcode
0FD7.
Now that we have OP_R(), use it here as well, while wiring memory-only
operands to OP_M() at the same time. To keep the number of consumed
opcode bytes similar to before, make BadOp() also account for VEX/XOP/
EVEX prefix bytes. To keep that change simple, convert need_vex to an
actual count of prefix bytes (keeping intact all prior boolean uses of
the field).
Note how this improves disassembly of such bad encodings, by at least
leaving a hint towards what a "nearby" instruction is. (For KSHIFT*
change the immediates test testcases use, such that disassembly remains
sufficiently in sync.)
While there also use Ux for VPMOV{B,W,D,Q}2M, where decoding through
mod_table[] was missing in the earlier scheme.
Fold OP_MS() and OP_XS() into OP_R(), paralleling OP_M(). Use operand
names (largely) matching those in the SDM. For 128-bit-only forms use
Uxmm though, marking 256-bit forms as bad. This then allows no longer
going through vex_len_table[] for two of the insns.
Specifically _do not_ continue to mis-use v_mode.
Several already use OP_M(), which rejects the register forms of insns,
and a few others can easily be converted to do so as well. (Note that
FXSAVE_Fixup() wires through to OP_M(). Note further that OP_IndirE(),
which wasn't placed very well anyway, is moved down to avoid the need to
forward-declare BadOp().)
Also adjust formatting of and drop PREFIX_OPCODE from a few adjacent
entries.
Most of them use Mx already for the memory operand, which rejects the
register form of the insn. Use that operand type also for the two EVEX
forms which so far have used EXEvexXNoBcst (and thus failed to reject
the register forms), compensating by flagging broadcast as bad for all
Mx. This way several other insns which don't permit embedded broadcast
either are also covered at the same time.
By changing decode order to do ModR/M.mod last (rather than VEX.L), the
VEX entries (which are already reused by EVEX decoding) can be folded
with their legacy counterparts as well. Note how this change of decode
order also allows removing two auxiliary #define-s, which were
introduced during earlier folding (because of that unhelpful order of
steps).
Introduce macro V to expand to 'v' in the VEX/EVEX case, and replace a
couple of abort()s where legacy code can now legitimately make it. While
there for {,V}LDDQU drop hoing through mod_table[] - OP_M() rejects
register operands quite fine.
Masking is not permitted for certain further insns, not falling in any
of the earlier categories. Introduce the Y macro (not expanding to any
output) to flag such cases.
Note that in a few cases entries already covered otherwise are converted
as well, to continue to allow sharing of the string literals.
Masking is not permitted in this case. See the code comment for how this
is being dealt with.
To avoid excess special casing of modes, have OP_M() call OP_E_memory()
directly.
While only zeroing-masking is possible in this case, this still requires
EVEX.z to be clear. Introduce a "global" flag right here, to be re-used
by checks which need to live in specific operand handlers.
Rather than corrupting disassmbly altogether, flag EVEX.z set as bad
when masking isn't in effect in the first place at the time the
destination operand is actually processed.
In commit 1a3b4f90bc ("x86: convert two pointers to (indexing)
integers") I neglected the fact that compilers may warn about comparing
ptrdiff_t (signed long) with size_t (unsigned long) values. Since just
before we've checked that the value is positive, simply add a cast
(despite my dislike for casts).
This in particular reduces the number of pointers to non-const that we
have (and that could potentially be used for undue modification of
state). As a result, fetch_code()'s 2nd parameter can then also become
pointer-to-const.
The present way of dealing with them - misusing MAX_MNEM_SIZE, which has
nothing to do with insn length - leads to inconsistent results. Since we
allow for up to MAX_CODE_LENGTH - 1 prefix bytes (which then could be
followed by another MAX_CODE_LENGTH "normal" insn bytes until we're done
decoding), size the_buffer[] accordingly.
Move struct dis_private down to be able to use MAX_CODE_LENGTH without
moving its #define. While doing this also alter the order to have the
potentially large array last.
This first of all removes a dependency on bfd_byte and unsigned char
being the same types. It further eliminates the need to mask by 0xff
when fetching values (which wasn't done fully consistently anyway),
improving code legibility.
While there, where possible add const.
Consistently do 64-bit first, VEX.L second, VEX.W third, ModR/M fourth,
and only then prefix, resulting in fewer table entries. Note that in the
course of the re-work
- TILEZERO has a previously missing decode step through rm_table[]
added,
- a wrong M_0 suffix for TILEZERO is also corrected to be M_1 (now an
infix).
Consistently do 64-bit first, ModR/M second, VEX.L third, VEX.W fourth,
and prefix last, resulting in fewer table entries. Note that in the
course of the re-work wrong M_0 suffixes are also corrected to be M_1
(partly infixes now).
Since it ended up confusing while testing the change, also adjust the
test name in x86-64-amx-bad.d (to be distinct from x86-64-amx.d's).
1) i386-dis.c:12055:11: runtime error: left shift of negative value -1
Bit twiddling is best done unsigned, due to UB on overflow of signed
expressions. Fix this by using bfd_vma rather than bfd_signed_vma
everywhere in i386-dis.c except print_displacement.
2) Return get32s and get16 value in a bfd_vma, reducing the need for
temp variables.
3) Introduce get16s and get8s functions to simplify the code.
4) With some optimisation options gcc-13 legitimately complains about
a fall-through in OP_I. Fix that. OP_I also doesn't need to use
"mask" which was wrong for w_mode anyway.
5) Masking with & 0xffffffff is better than casting to unsigned. We
don't know for sure that unsigned int is 32-bit.
6) We also don't know that unsigned char is 8 bits. Mask codep
accesses everywhere. I don't expect binutils will work on anything
other than an 8-bit char host, but if we are masking codep accesses in
some places we might as well be consistent. (Better would be to use
stdint.h types more in binutils.)
opcodes/i386-dis.c: In function ‘print_insn’:
opcodes/i386-dis.c:9865:22: error: storing the address of local
variable ‘priv’ in ‘*info.private_data’ [-Werror=dangling-pointer=]
* i386-dis.c (print_insn): Clear info->private_data before
returning.
For quite come time print_insn() has been storing the address of a local
variable into info->private_data. Since the compiler can't know that the
field won't be accessed again after print_insn() returns, it may kind of
legitimately diagnose this. And recent enough gcc does as of the
introduction of the fetch_error() return paths (replacing setjmp()-based
error handling).
Utilizing that neither prefix_name() nor i386_dis_printf() actually use
info->private_data, zap the pointer in fetch_error(), after having
retrieved it for local use.
A recent change in opcodes/i386-dis.c caused a build failure on my
x86-64 Fedora 36 system, which uses:
$ gcc --version
gcc (GCC) 12.2.1 20221121 (Red Hat 12.2.1-4)
[...]
The error is:
../../binutils-gdb/opcodes/i386-dis.c: In function ‘OP_J’:
../../binutils-gdb/opcodes/i386-dis.c:12705:22: error: ‘val’ may be used uninitialized [-Werror=maybe-uninitialized]
12705 | disp = val & 0x8000 ? val - 0x10000 : val;
| ~~~~^~~~~~~~
This patch fixes the warning.
opcodes/ChangeLog
2023-04-21 Tom Tromey <tromey@adacore.com>
* i386-dis.c (OP_J): Check result of get16.
get64() is unreachable when !BFD64 (due to a check relatively early in
print_insn()). Let's avoid the associated #ifdef-ary (or else we should
extend it to remove more dead code).
Make them return boolean and convert FETCH_DATA() uses to fetch_code().
With this no further users of FETCH_DATA() remain, so the macro and its
backing function are dropped as well.
Leave value types as they were for the helper functions, even if I don't
think that beyond get64() use of bfd_{,signed_}vma is really necessary.
With type change of "disp" in OP_E_memory(), change the 2nd parameter of
print_displacement() to a signed type as well, though (eliminating the
need for a local variable of signed type). This also eliminates the need
for custom printing of '-' in Intel syntax displacement expressions.
While there drop forward declarations which aren't really needed.
