git commit 202be274a4 went a little wild in removing trailing spaces
in gas/testsuite/gas/i386/{secidx.d,secrel.d}, causing
x86_64-w64-mingw32 +FAIL: i386 secrel reloc
x86_64-w64-mingw32 +FAIL: i386 secidx reloc
I could have just replaced the trailing space, but let's fix the
objdump output instead. Touches lots of testsuite files.
The following test fails on the armeb-gnu-eabi target:
FAIL: Unwind information for Armv8.1-M.Mainline PACBTI extension
This patch adjusts the expected output for big endian.
The zhinx extension is a sub-extension in zfinx, corresponding to
zfh extension but use GPRs instead of FPRs.
This patch expanded the zfh insn class define, since zfh and zhinx
use the same opcodes, thanks for Nelson's works.
changelog in V2: Add missing classes of 'zfh' and 'zhinx' in
"riscv_multi_subset_supports_ext".
bfd/ChangeLog:
* elfxx-riscv.c (riscv_multi_subset_supports): New extensions.
(riscv_multi_subset_supports_ext): New extensions.
gas/ChangeLog:
* testsuite/gas/riscv/fp-zhinx-insns.d: New test.
* testsuite/gas/riscv/fp-zhinx-insns.s: New test.
include/ChangeLog:
* opcode/riscv.h (enum riscv_insn_class): New INSN classes.
opcodes/ChangeLog:
* riscv-opc.c: Modify INSN_CLASS.
While working on another patch[1] I had need to touch this code in
i386-dis.c:
ins->obufp = ins->mnemonicendp;
for (i = strlen (ins->obuf) + prefix_length; i < 6; i++)
oappend (ins, " ");
oappend (ins, " ");
(*ins->info->fprintf_styled_func)
(ins->info->stream, dis_style_mnemonic, "%s", ins->obuf);
What this code does is add whitespace after the instruction mnemonic
and before the instruction operands.
The problem I ran into when working on this code can be seen by
assembling this input file:
.text
nop
retq
Now, when I disassemble, here's the output. I've replaced trailing
whitespace with '_' so that the issue is clearer:
Disassembly of section .text:
0000000000000000 <.text>:
0: 90 nop
1: c3 retq___
Notice that there's no trailing whitespace after 'nop', but there are
three spaces after 'retq'!
What happens is that instruction mnemonics are emitted into a buffer
instr_info::obuf, then instr_info::mnemonicendp is setup to point to
the '\0' character at the end of the mnemonic.
When we emit the whitespace, this is then added starting at the
mnemonicendp position. Lets consider 'retq', first the buffer is
setup like this:
'r' 'e' 't' 'q' '\0'
Then we add whitespace characters at the '\0', converting the buffer
to this:
'r' 'e' 't' 'q' ' ' ' ' ' ' '\0'
However, 'nop' is actually an alias for 'xchg %rax,%rax', so,
initially, the buffer is setup like this:
'x' 'c' 'h' 'g' '\0'
Then in NOP_Fixup we spot that we have an instruction that is an alias
for 'nop', and adjust the buffer to this:
'n' 'o' 'p' '\0' '\0'
The second '\0' is left over from the original buffer contents.
However, when we rewrite the buffer, we don't afjust mnemonicendp,
which still points at the second '\0' character.
Now, when we insert whitespace we get:
'n' 'o' 'p' '\0' ' ' ' ' ' ' ' ' '\0'
Notice the whitespace is inserted after the first '\0', so, when we
print the buffer, the whitespace is not printed.
The fix for this is pretty easy, I can change NOP_Fixup to adjust
mnemonicendp, but now a bunch of tests start failing, we now produce
whitespace after the 'nop', which the tests don't expect.
So, I could update the tests to expect the whitespace....
...except I'm not a fan of trailing whitespace, so I'd really rather
not.
Turns out, I can pretty easily update the whitespace emitting code to
spot instructions that have zero operands and just not emit any
whitespace in this case. So this is what I've done.
I've left in the fix for NOP_Fixup, I think updating mnemonicendp is
probably a good thing, though this is not really required any more.
I've then updated all the tests that I saw failing to adjust the
expected patterns to account for the change in whitespace.
[1] https://sourceware.org/pipermail/binutils/2022-April/120610.html
bfd_hostptr_t is defined as a type large enough to hold either a long
or a pointer. It mostly appears in the coff backend code in casts.
include/coff/internal.h struct internal_syment and union
internal_auxent have the only uses in data structures, where
comparison with include/coff/external.h and other code reveals that
the type only needs to be large enough for a 32-bit integer or a
pointer. That should mean replacing with uintptr_t is OK.
Requiring C99 means that uses of bfd_uint64_t can be replaced with
uint64_t, and similarly for bfd_int64_t, BFD_HOST_U_64_BIT, and
BFD_HOST_64_BIT. This patch does that, removes #ifdef BFD_HOST_*
and tidies a few places that print 64-bit values.
MASM doesn't support the separate operand form; the modifier belongs
after the instruction instead. Accept this form alongside the original
(now legacy) one. Short of having access to a MASM version to actually
check in how far "after the instruction" is a precise statement in their
documentation, allow both that and the SDM mandated form where the
modifier is on the last register operand (with a possible immediate
operand following).
Sadly the split out function, at least for the time being, needs to cast
away constness at some point, as the two callers disagree in this
regard.
Adjust some, but not all of the testcases.
As a preparatory step to allowing proper non-operand forms of specifying
embedded rounding / SAE, convert the internal representation to non-
operand form. While retaining properties (and in a few cases perhaps
providing more meaningful diagnostics), this means doing away with a few
hundred standalone templates, thus - as a nice side effect - reducing
memory consumption / cache occupancy.
MASM doesn't consider {sae} and alike a separate operand; it is attached
to the last register operand instead, just like spelled out by the SDM.
Make the disassembler follow this first, before also adjusting the
assembler (such that it'll be easy to see that the assembler change
doesn't alter generated code).
MASM doesn't support the {1to<n>} form; DWORD BCST (paralleling
DWORD PTR) and alike are to be used there instead. Accept these forms
alongside the original (now legacy) ones.
Acceptance of the original {1to<n>} operand suffix is retained both for
backwards compatibility and to disambiguate VFPCLASSP{S,D,H} and vector
conversions with shrinking element sizes. I have no insight (yet) into
how MASM expects those to be disambiguated.
Adjust some, but not all of the testcases.
MASM doesn't support the {1to<n>} form; DWORD BCST (paralleling
DWORD PTR) and alike are to be used there instead. Make the disassembler
follow this first, before also adjusting the assembler (such that it'll
be easy to see that the assembler change doesn't alter generated code).
For VFPCLASSP{S,D,H} and vector conversions with shrinking element sizes
the original {1to<n>} operand suffix is retained, to disambiguate
output. I have no insight (yet) into how MASM expects those to be
disambiguated.
Whilst reviewing Luis' proposed change to s_arm_unwind_save_mixed
yesterday I noticed that we were making use of __builting_clzl
directly within the main function, which is not guaranteed to be
portable. Whilst studying the code further, I also realized that it
could be rewritten without using it and also reworked to remove a lot
of unnecessary iterations steps. So this patch does that (and also
removes the source of the warning that Luis was trying to fix).
Finally, with the rewrite we can also simplify the caller of this
routine as the new version can handle all the cases directly.
* config/tc-arm.c (s_arm_unwind_save_mixed): Rewrite without
using __builtin_clzl.
(s_arm_unwind_save): Simplify logic for simple/mixed register saves.
This commit makes RV32 + 'Q' extension (version 2.2 or later) not
conflicting since this combination is no longer prohibited by the
specification.
bfd/ChangeLog:
* elfxx-riscv.c (riscv_parse_check_conflicts): Remove conflict
detection that prohibits RV32Q on 'Q' version 2.2 or later.
gas/ChangeLog:
* testsuite/gas/riscv/march-fail-rv32iq.d: Removed.
* testsuite/gas/riscv/march-fail-rv32iq.l: Likewise.
* testsuite/gas/riscv/march-fail-rv32iq2p0.d: New test
showing RV32IQ fails on 'Q' extension version 2.0.
* testsuite/gas/riscv/march-fail-rv32iq2p0.l: Likewise.
* testsuite/gas/riscv/march-fail-rv32iq2.d: Likewise.
* testsuite/gas/riscv/march-fail-rv32iq-isa-2p2.d: New test
showing RV32IQ fails on ISA specification version 2.2.
* testsuite/gas/riscv/march-ok-rv32iq2p2.d: New test
showing RV32IQ succesds on 'Q' extension version 2.2.
* testsuite/gas/riscv/march-ok-rv32iq-isa-20190608.d: New test
showing RV32IQ succesds on ISA specification 20190608.
Per Power ISA Version 3.1B 3.3.12, isel uses BC field rather than CRB
field present in binutils sources. Also, per 1.6.2, BC has the same
semantics as BA and BB fields, so this should keep the same flags and
mask, only with the different offset.
opcodes/
* ppc-opc.c
(BC): Define new field, with the same definition as CRB field,
but with the PPC_OPERAND_CR_BIT flag present.
gas/
* testsuite/gas/ppc/476.d: Update.
* testsuite/gas/ppc/a2.d: Update.
* testsuite/gas/ppc/e500.d: Update.
* testsuite/gas/ppc/power7.d: Update.
With the upcoming SVP64 extension[0] to PowerPC architecture, it became
evident that PowerPC operand indices no longer fit 8 bits. This patch
switches the underlying type to uint16_t, also introducing a special
typedef so that any future extension goes even smoother.
[0] https://libre-soc.org
include/
* opcode/ppc.h (ppc_opindex_t): New typedef.
(struct powerpc_opcode): Use it.
(PPC_OPINDEX_MAX): Define.
gas/
* write.h (struct fix): Increase size of fx_pcrel_adjust.
Reorganise.
* config/tc-ppc.c (insn_validate): Use ppc_opindex_t for operands.
(md_assemble): Likewise.
(md_apply_fix): Likewise. Mask fx_pcrel_adjust with PPC_OPINDEX_MAX.
(ppc_setup_opcodes): Adjust opcode index assertion.
opcodes/
* ppc-dis.c (skip_optional_operands): Use ppc_opindex_t for
operand pointer.
(lookup_powerpc, lookup_prefix, lookup_vle, lookup_spe2): Likewise.
(print_insn_powerpc): Likewise.
As fmv.x.q and fmv.q.x instructions are RV128-only (not RV64-only),
it should be removed until RV128 support for GNU Binutils is required
again.
gas/ChangeLog:
* testsuite/gas/riscv/fmv.x.q-rv64-fail.d: New failure test.
* testsuite/gas/riscv/fmv.x.q-rv64-fail.l: Likewise.
* testsuite/gas/riscv/fmv.x.q-rv64-fail.s: Likewise.
include/ChangeLog:
* opcode/riscv-opc.h (MATCH_FMV_X_Q, MASK_FMV_X_Q,
MATCH_FMV_Q_X, MASK_FMV_Q_X): Remove RV128-only instructions.
opcodes/ChangeLog:
* riscv-opc.c (riscv_opcodes): Remove RV128-only instructions.
The current assembler accepts system registers FPCXTNS and FPCXTS for Armv8.1-M
Mainline Instructions VSTR, VLDR, VMRS and VMSR.
Assembler should be also allowing FPCXT_NS, fpcxt_ns, fpcxtns, FPCXT_S, fpcxt_s
and fpcxts. This patch fixes the issue.
Like the placeholder types added in 04dfe7aa52 ("Arm64: follow-on to
PR gas/27217 fix"), these are also placeholders which are subsequently
resolved (albeit later, hence this being a separate issue). As for the
resolved types 1 is returned, these pseudo-relocs should also have 1
returned to force retaining of the [eventual] relocations. This is also
spelled out individually for each of them in md_apply_fix().
PR gas/27217
Prior to trying to address PR gas/28888 I noticed anomalies in how
certain insns would / wouldn't be affected in similar ways.
Commit eac4eb8ecb ("Fix a problem assembling AArch64 sources when a
relocation is generated against a symbol that has a defined value") had
two copy-and-paste mistakes, passing the wrong type to
aarch64_force_reloc().
It further failed to add placeholder relocation types to that function's
block of case labels leading to a return of 1. While not of interest for
aarch64_force_relocation() (these placeholders are resolved right in
parse_operands()), calls to aarch64_force_reloc() happen before that
resolution would take place.
* config/tc-arm.c (parse_reg_list): Add handling of mixed register
types.
(reg_names): Enumerate pseudoregister according to mapped physical
register number.
(s_arm_unwind_save_pseudo): Modify function signature.
(s_arm_unwind_save_core): Likewise.
(s_arm_unwind_save_mixed): New function.
(s_arm_unwind_save): Generate register list mask to pass to nested
functions.
* testsuite/gas/arm/unwind-pacbti-m.s: Expand test for mixed
register type lists.
* testsuite/gas/arm/unwind-pacbti-m.d: Likewise.
* testsuite/gas/arm/unwind-pacbti-m-readelf.d: Likewise.
Compilers would put decimal numbers there, so I think we should treat
finding octal numbers the same as finding bignums - ignore them as
actually being comments of some very specific form.
Any construct which to the scrubber looks like a C preprocessor
line/file "directive" is converted to .linefile, but the amount of
checking the scrubber does is minimal (albeit it does let through only
decimal digits for the line part of the contruct). Since the scrubber
conversion is further tied to # being a line comment character, anything
which upon closer inspection turns out not to be a line/file "directive"
is supposed to be treated as a comment, i.e. ignored. Therefore we
cannot use get_absolute_expression(), as this may raise errors. Open-
code the function instead, treating everything not resulting in
O_constant as a comment as well.
Furthermore also bounds-check the parsed value. This bounds check tries
to avoid implementation defined behavior (which may be the raising of an
implementation defined signal), but for now makes the assumption that
int has less than 64 bits. The way bfd_signed_vma (which is what offsetT
aliases) is defined in bfd.h for the BFD64 case I cannot really see a
clean way of avoiding this assumption. Omitting the #ifdef, otoh, would
risk "condition is always false" warnings by compilers.
Convert get_linefile_number() to return bool at this occasion as well.
do_repeat_with_expander() already deals with the "no expander" case
quite fine, so there's really little point having two functions. What it
lacks compared with do_repeat() is a call to sb_build(), which can
simply be moved (and the then redundant sb_new() be avoided). Along with
this moving also flip if the main if()'s condition such that the "no
expander" case is handled first.
When assembling code previously pre-processed by a C compiler, long
enough comments may have been collapsed into "# <line> <file>"
constructs. If we skip these, line numbers (and possibly even file
names) will be off / wrong in both diagnostics and debug info.
First of all convert to switch(), in preparation of adding another
directive here which may not be ignored. While doing so drop dead code:
A string the first two characters of which do not match "if" also wont
match "ifdef" or "ifndef".
s390_insert_operand ()'s val, min and max are encoded PCRel32 values
and need to be left-shifted by 1 before being shown to the user.
Left-shifting negative values is undefined behavior in C, but the
current code does not try to prevent it, causing UBSan to complain.
Fix by casting the values to their unsigned equivalents before
shifting.
The new test failed on s390-linux due to bfd_sprintf_vma trimming
output to 32 bits for 32-bit targets. The test was faulty anyway,
expecting zero as the min end of the range is plainly wrong, but
that's what you get if you cast min to int.
* config/tc-s390.c (s390_insert_operand): Print range error using
PRId64.
* testsuite/gas/s390/zarch-z900-err.l: Correct expected output.
as does not accept instructions like brasl %r0,.-0x100000000, because
of two problems with the generic overflow check:
1. PCRel32 operands are signed, but are treated as unsigned.
2. The allowed range for these operands is [-(1 << 32), (1 << 32) - 1],
and not [-(1 << 31), (1 << 31) - 1].
Fix both by disabling the generic overflow check - it's not needed,
because s390_insert_operand () performs its own.
gas/
* config/tc-s390.c (md_gather_operands): Set fx_no_overflow.
* testsuite/gas/s390/s390.exp: Add zarch-z900-err.
* testsuite/gas/s390/esa-z900.d: New test.
* testsuite/gas/s390/esa-z900.s: New test.
* testsuite/gas/s390/zarch-z900-err.l: New test.
* testsuite/gas/s390/zarch-z900-err.s: New test.
So far an SAE-only specifier was accepted for static-rounding insns,
while SAE-only insns didn't accept static rounding specifiers. If
anything it would make sense the other way around, allowing SAE-only
insns to have the (ignored) rounding mode specified individually rather
than globally via -mevexrcig=. But for now make things match the SDM.
The way SAE templates are constructed was always puzzling me (including
the need for separate templates in the first place), and expressing the
extzra attribute via Imm8 actually has a bad effect: Ordinary immediates
would also be accepted, leading to an extra byte being added after the
instruction (i.e. generating bad code). Before re-working this (in
particular to accept proper Intel syntax there), fix the immediate issue
by adding the so far missing check.
For an unclear reason corresponding AVX512F tests were apparently not
cloned or used as reference here, and instead the bogus legacy forms of
the insns (with the embedded rounding specifier not last) were used.
It's not just REX.W which is ignored with opcode 0x90. The same goes for
REX.R and REX.X as well as empty REX. None of these are forms of
"xchg %eax,%eax" (which would mean zero-extending %eax to %rax), so they
also shouldn't be disassembled this way.
While there simplify things: A single hook function suffices, thus
making it unnecessary to keep two expressions in sync. And checking
ins->address_mode for mode_64bit also is unnecessary, as "rex" can be
non-zero only in that case anyway.
Before the change tc-m68k maintained a list of seen labels.
Alignment check traversed label list to resolve symbol to label.
This caused quadratic slowdown as each symbol was checked against
each label. Worst affected files are the ones built with debugging
enabled as DWARF generates many labels.
The change embeds auxiliary label information right into symbol using
TC_SYMFIELD_TYPE.
Before the change test from PR 29058 did not finish in 10 minutes. After
the change it finishes in 2 seconds.
gas/ChangeLog:
PR 29058
* config/tc-m68k.h (TC_SYMFIELD_TYPE): define as m68k_tc_sy.
* config/tc-m68k.c (m68k_frob_label): Use TC_SYMFIELD_TYPE to
store label information.
With the sole user of the return value gone, convert the return type to
void. This in turn allows simplifying another construct, by moving it
slightly later in the function.
These were used originally to represent "# <line> <file>" constructs
inserted by (typically) compilers when pre-processing. Quite some time
ago they were replaced by .linefile though. Since the original
directives were never documented, we ought to be able to remove support
for them. As a result in a number of case function parameter aren't used
anymore and can hence be dropped.
Commit 7992631e8c ("gas/Dwarf: improve debug info generation from .irp
and alike blocks"), while dealing okay with actual assembly source files
not using .file/.line and alike outside but not inside of .macro, has
undue effects when the logical file/line pair was already overridden:
Line numbers would continuously increment while processing the expanded
macro, while the goal of the PR gas/16908 workaround is to keep the
expansion associated with the line invoking the macro. However, as soon
as enough state was overridden _inside_ the macro to cause as_where() to
no longer fall back top as_where_physical(), honor this by resuming the
bumping of the logical line number.
Note that from_sb_is_expansion's initializer was 1 for an unknown
reason. While renaming the variable and changing its type, also change
the initializer to "expanding_none", which would have been "0" in the
original code. Originally the initializer value itself wasn't ever used
anyway (requiring sb_index != -1), as it necessarily had changed in
input_scrub_include_sb() alongside setting sb_index to other than -1.
Strictly speaking input_scrub_insert_line() perhaps shouldn't use
expanding_none, yet none of the other enumerators fit there either. And
then strictly speaking that function probably shouldn't exist in the
first place. It's used only by tic54x.
Commit 7992631e8c ("gas/Dwarf: improve debug info generation from .irp
and alike blocks"), while dealing okay with actual assembly source files
not using .file/.line and alike outside but not inside of .irp et al,
has undue effects when the logical file/line pair was already
overridden: Line numbers would continuously increment upon every
iteration, thus potentially getting far off. Furthermore it left it to
the user to actually insert .file/.line inside such constructs. Note
though that before aforementioned change things weren't pretty either:
Diagnostics (and debug info) would be associated with the directive
terminating the iteration construct, rather than with the actual lines.
Handle this automatically by simply latching the present line and then
re-instating coordinates first thing on every iteration; note that the
file can't change from what was previously pushed on the scrubber's
state stack, and hence can be taken from there by using a new flavor of
.linefile (which is far better memory-footprint-wise than recording the
full path in the inserted directive). (This then leaves undisturbed any
file/line control occurring in the body of the construct, as these will
only be seen and processed afterwards.)
