The expectation of x86-64-branch-3 for "call" / "jmp" with an obvious
direct destination to translate to an indirect _far_ branch is plain
wrong. The operand size prefix should have no effect at all on the
interpretation of the operand. The main underlying issue here is that
the Intel64 templates of the direct branches don't include Disp16, yet
various assumptions exist that it would always be there when there's
also Disp32/Disp32S, toggled by the operand size prefix (which is
being ignored by direct branches in Intel64 mode).
Along these lines it was also wrong to base the displacement width
decision solely on the operand size prefix: REX.W cancels this effect
and hence needs taking into consideration, too.
A disassembler change is needed here as well: XBEGIN was wrongly treated
the same as direct CALL/JMP, which isn't the case - the operand size
prefix does affect displacement size there, it's merely ignored when it
comes to updating [ER]IP.
In memory operand addressing, which forms of displacement are permitted
besides Disp8 is pretty clearly limited
- outside of 64-bit mode, Disp16 or Disp32 only, depending on address
size (MPX being special in not allowing Disp16),
- in 64-bit mode, Disp32s or Disp64 without address size override, and
solely Disp32 with one.
Adjust assembler and i386-gen to match this, observing that templates
already get adjusted before trying to match them against input depending
on the presence of an address size prefix.
This adjustment logic gets extended to all cases, as certain DispNN
values should also be dropped when there's no such prefix. In fact
behavior of the assembler, perhaps besides the exact diagnostics wording,
should not differ between there being templates applicable to 64-bit and
non-64-bit at the same time, or there being fully separate sets of
templates, with their DispNN settings already reduced accordingly.
This adjustment logic further gets guarded such that there wouldn't be
and Disp<N> conversion based on address size prefix when this prefix
doesn't control the width of the displacement (on branches other than
absolute ones).
These adjustments then also allow folding two MOV templates, which had
been split between 64-bit and non-64-bits variants so far.
Once in this area also
- drop the bogus DispNN from JumpByte templates, leaving just the
correct Disp8 there (compensated by i386_finalize_displacement()
now setting Disp8 on their operands),
- add the missing Disp32S to XBEGIN.
Note that the changes make it necessary to temporarily mark a test as
XFAIL; this will get taken care of by a subsequent patch. The failing
parts are entirely bogus and will get replaced.
* d30v-dis.c (extract_value): Make num param a uint64_t, constify
oper. Use unsigned vars.
(print_insn): Make num var uint64_t. Constify oper and remove now
unnecessary casts on extract_value calls.
(print_insn_d30v): Use unsigned vars. Adjust printf formats.
PR 25281
* sh-dis.c (print_insn_ddt): Properly check validity of MOVX_NOPY
and MOVY_NOPX insns. For invalid cases include 0xf000 in the word
printed. Print .word in more cases.
This patch prevents print_insn_mode recursing into another index mode
byte, which if repeated enough times will overflow private.the_buffer
and scribble over other memory.
* vax-dis.c (print_insn_mode): Stop index mode recursion.
fadd, fmul, and fdiv are now, by ISO/IEC TS 18661-3:2015, defined to
refer to functions from the runtime subsystem.
PR 25277
* microblaze-opcm.h (enum microblaze_instr): Prefix fadd, fmul and
fdiv with "mbi_".
* microblaze-opc.h (opcodes): Adjust to suit.
The arc fix in create_map avoiding signed overflow by casting an
unsigned char to unsigned int before shifting, shows one of the
dangers of blinding doing that. The problem in this case was that the
variable storing the value, newAuxRegister->address, was a long.
Using the unsigned cast meant that the 32-bit value was zero extended
when long is 64 bits. Previously we had a sign extension. Net result
was that comparisons in arcExtMap_auxRegName didn't match. Of course,
I could have cast the 32-bit unsigned value back to signed before
storing in a long, but it's neater to just use an unsigned int for the
address.
opcodes/
* alpha-opc.c (OP): Avoid signed overflow.
* arm-dis.c (print_insn): Likewise.
* mcore-dis.c (print_insn_mcore): Likewise.
* pj-dis.c (get_int): Likewise.
* ppc-opc.c (EBD15, EBD15BI): Likewise.
* score7-dis.c (s7_print_insn): Likewise.
* tic30-dis.c (print_insn_tic30): Likewise.
* v850-opc.c (insert_SELID): Likewise.
* vax-dis.c (print_insn_vax): Likewise.
* arc-ext.c (create_map): Likewise.
(struct ExtAuxRegister): Make "address" field unsigned int.
(arcExtMap_auxRegName): Pass unsigned address.
(dump_ARC_extmap): Adjust.
* arc-ext.h (arcExtMap_auxRegName): Update prototype.
I'm flying blind here, not having an s+core s3 insn set reference,
but this seems reasonably obvious from what is done by the assembler.
s3_do16_rpop does some mixing of imm and reg values to place in the
rpop reg field, but I'm not going to try to fix the disassembly
there.
* score-dis.c (print_insn_score16): Move rpush/rpop imm field
value adjustment so that it doesn't affect reg field too.
The ubsan complaint is fixed by the SBM change, with similar possible
complaints fixed by the EXTRACT change. The rest is just cleanup.
include/
* opcode/crx.h (inst <match>): Make unsigned int.
opcodes/
* crx-dis.c (EXTRACT, SBM): Avoid signed overflow.
(get_number_of_operands, getargtype, getbits, getregname),
(getcopregname, getprocregname, gettrapstring, getcinvstring),
(getregliststring, get_word_at_PC, get_words_at_PC, build_mask),
(powerof2, match_opcode, make_instruction, print_arguments),
(print_arg): Delete forward declarations, moving static to..
(getregname, getcopregname, getregliststring): ..these definitions.
(build_mask): Return unsigned int mask.
(match_opcode): Use unsigned int vars.
Note that using 1u in N32_BIT makes all of N32_BIT, __MASK, __MF, __GF
and __SEXT evaluate as unsigned int (the latter three when when their
v arg is int or smaller). This would be a problem if assigning the
result to a bfd_vma, long, or other type wider than an int since the
__SEXT result would be zero extended to the wider type. Fortunately
nds32 target code doesn't use wider types unnecessarily.
include/
* opcode/nds32.h (N32_BIT): Define using 1u.
(__SEXT): Use __MASK and N32_BIT.
(N32_IMMS): Remove duplicate mask.
opcodes/
* nds32-dis.c (print_insn16, print_insn32): Remove forward decls.
(struct objdump_disasm_info): Delete.
(nds32_parse_audio_ext, nds32_parse_opcode): Cast result of
N32_IMMS to unsigned before shifting left.
Commit 8c9b417187 didn't remove a glaring left shift of a number
that had just been sign extended.
* moxie-dis.c (INST2OFFSET): Don't left shift a signed value.
(print_insn_moxie): Remove unnecessary cast.
This is
*cst = (data[0] << 24) + (data[1] << 16) + (data[2] << 8) + data[3];
data is unsigned char which promotes to int.
* h8300-dis.c (extract_immediate): Avoid signed overflow.
(bfd_h8_disassemble): Likewise.
In the following buf is an unsigned char array, so elements are
promoted to int before arithmetic operations.
if (dis_info.info->endian == BFD_ENDIAN_BIG)
while (n--)
val |= buf[n] << (n*8);
else
for (i = 0; i < n; i++)
val |= buf[i] << (i*8);
* csky-dis.c (csky_chars_to_number): Rewrite. Avoid signed
overflow when collecting bytes of a number.
This was the following in print_with_operands
case 4:
number
= buffer[2] + buffer[3] * 256 + buffer[4] * 65536
+ buffer[5] * 0x1000000;
and buffer[5] * 0x1000000 can indeed overflow. So to fix this we need
to use unsigned arithmetic where overflow semantics are specified.
But number is a long, and the expression is int which will be sign
extended to long. If we make the expression unsigned it will be zero
extended. So make number an int32_t and rearrange a little for some
of the places that need fixing.
* cris-dis.c (print_with_operands): Avoid signed integer
overflow when collecting bytes of a 32-bit integer.
This was:
unsigned long mask = SBM (instruction->match_bits);
with
#define SBM(offs) ((((1 << (32 - offs)) -1) << (offs)))
Well, there are a couple of problems. Firstly, the expression uses
int values (1 rather than 1u or 1ul) resulting in the ubsan error, and
secondly, a zero offs will result in a 32-bit shift which is undefined
if ints are only 32 bits.
* cr16-dis.c (EXTRACT, SBM): Rewrite.
(cr16_match_opcode): Delete duplicate bcond test.
This was the following in fmtconst_val, x is unsigned int.
x = SIGNEXTEND (x, constant_formats[cf].nbits);
Problem is, the SIGNEXTEND macro assumed its arg was a long and sign
extended by shifting left then shifting right, and didn't cast the
arg. So don't do the silly shift thing. It's not guaranteed to work
anyway according to the C standard. ">>" might do a logical shift
even if its args are signed.
* bfin-dis.c (HOST_LONG_WORD_SIZE, XFIELD): Delete.
(SIGNBIT): New.
(MASKBITS, SIGNEXTEND): Rewrite.
(fmtconst): Don't use ? expression now that SIGNEXTEND uses
unsigned arithmetic, instead assign result of SIGNEXTEND back
to x.
(fmtconst_val): Use 1u in shift expression.
