Dot products deviate from the normal disassembly rules for lane indexed
instruction. Their canonical representation is in the form of:
v0.2s, v0.8b, v0.4b[0] instead of v0.2s, v0.8b, v0.b[0] to try to denote
that these instructions select 4x 1 byte elements instead of a single 1 byte
element.
Previously we were disassembling them following the normal rules, this patch
corrects the disassembly.
gas/
PR gas/22559
* config/tc-aarch64.c (vectype_to_qualifier): Support AARCH64_OPND_QLF_S_4B.
* gas/testsuite/gas/aarch64/dotproduct.d: Update disassembly.
include/
PR gas/22559
* aarch64.h (aarch64_opnd_qualifier): Add AARCH64_OPND_QLF_S_4B.
opcodes/
PR gas/22559
* aarch64-asm.c (aarch64_ins_reglane): Change AARCH64_OPND_QLF_S_B to
AARCH64_OPND_QLF_S_4B
* aarch64-dis.c (aarch64_ext_reglane): Change AARCH64_OPND_QLF_S_B to
AARCH64_OPND_QLF_S_4B
* aarch64-opc.c (aarch64_opnd_qualifiers): Add 4b variant.
* aarch64-tbl.h (QL_V2DOT): Change S_B to S_4B.
Previously parse_vector_type_for_operand was changed to allow the use of 4b
register size for indexed lane instructions. However this had the unintended
side effect of also allowing 4b for normal vector registers.
Because this support was only partial the rest of the tool silently treated
4b as 8b and continued. This patch adds full support for 4b so it can be
properly distinguished from 8b and the correct errors are generated.
With this patch you still can't encode any instruction which actually requires
v<num>.4b but such instructions don't exist so to prevent needing a workaround
in get_vreg_qualifier_from_value this was just omitted.
gas/
PR gas/22529
* config/tc-aarch64.c (vectype_to_qualifier): Support AARCH64_OPND_QLF_V_4B.
* gas/testsuite/gas/aarch64/pr22529.s: New.
* gas/testsuite/gas/aarch64/pr22529.d: New.
* gas/testsuite/gas/aarch64/pr22529.l: New.
include/
PR gas/22529
* opcode/aarch64.h (aarch64_opnd_qualifier): Add AARCH64_OPND_QLF_V_4B.
opcodes/
PR gas/22529
* aarch64-opc.c (aarch64_opnd_qualifiers): Add 4b variant.
Just like for instructions in GPRs, there's no need to have separate
templates for otherwise identical insns acting on XMM or YMM registers
(or memory of the same size).
Use a combination of a single new Reg bit and Byte, Word, Dword, or
Qword instead.
Besides shrinking the number of operand type bits this has the benefit
of making register handling more similar to accumulator handling (a
generic flag is being accompanied by a "size qualifier"). It requires,
however, to split a few insn templates, as it is no longer correct to
have combinations like Reg32|Reg64|Byte. This slight growth in size will
hopefully be outweighed by this change paving the road for folding a
presumably much larger number of templates later on.
Pseudo prefixes must be used on an instruction. Issue an error when
pseudo prefix is used without instruction.
PR gas/22623
* gas/config/tc-i386.c (output_insn): Check pseudo prefix
without instruction.
* testsuite/gas/i386/i386.exp: Run inval-pseudo.
* testsuite/gas/i386/inval-pseudo.l: New file.
* testsuite/gas/i386/inval-pseudo.s: Likewise.
Again these look to be typos: No template currently allows for any two
(or all three) of RegXMM, RegYMM, and RegZMM in a single operand. Quite
clearly ! are missing, after the addition of which the checks for the
first and (if present) second operands also fully match up.
I'm rather certain the missing ! was just a typo, the more with the
similar check in mind that's in the same function a few hundred lines
down (in the body of "if (vex_reg != (unsigned int) ~0)"). Of course
this can't be demonstrated by a test case - internal data structure
consistency is being checked here, and neither form of the check
triggers with any current template.
It is also not really clear to me why operand_type_equal() is being used
in the {X,Y,Z}MM register check here, rather than just testing the
respective bits: Just like Reg32|Reg64 is legal in an operand template,
I don't see why e.g. RegXMM|RegYMM wouldn't be. For example it ought to
be possible to combine
vaddpd, 3, 0x6658, None, 1, CpuAVX, Modrm|Vex|VexOpcode=0|VexVVVV=1|VexW=1|IgnoreSize|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { Xmmword|Unspecified|BaseIndex|Disp8|Disp16|Disp32|Disp32S|RegXMM, RegXMM, RegXMM }
vaddpd, 3, 0x6658, None, 1, CpuAVX, Modrm|Vex=2|VexOpcode=0|VexVVVV=1|VexW=1|IgnoreSize|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { Ymmword|Unspecified|BaseIndex|Disp8|Disp16|Disp32|Disp32S|RegYMM, RegYMM, RegYMM }
into a single template (with setting of VEX.L suitably handled elsewhere
if that's not already happening anyway).
Additionally I don't understand why this uses abort() instead of
gas_assert().
Both of these latter considerations then also apply to the
aforementioned other check in the same function.
opcodes * disassemble.c: Enable disassembler_needs_relocs for PRU.
gas * testsuite/gas/pru/extern.s: New test for print of U16_PMEMM
relocation.
* testsuite/gas/pru/extern.d: New test driver.
Don't use address where symbol gets resolved, as during section
relaxation symbols will slide, instead canonicalize symbols and check
that they are are the same.
This fixes a bug when a relaxed jump goes into the wrong trampoline.
gas/
2017-12-07 Max Filippov <jcmvbkbc@gmail.com>
* config/tc-xtensa.c (xg_order_trampoline_chain): Replace
xg_order_trampoline_chain_entry call with check for
canonicalized symbol equality and offset equality.
While we shouldn't outright reject such (as was wrongly done by commit
4d36230d59 ("x86: Update segment register check in Intel syntax"), as
MASM accepts them even silently, issue (by default) a warning for such
questionable constructs.
gas/
* config/tc-riscv.c (riscv_frag_align_code): New local insn_alignment.
Early return if bytes less than or equal to insn_alignment.
* testsuite/gas/riscv/align-1.l: New.
* testsuite/gas/riscv/align-1.s: New.
* testsuite/gas/riscv/riscv.exp: Use run_dump_tests. Use run_list_test
for align-1.
This should be an obvious fix.
It corrects the register number for IP1 to 17.
gas/
2017-11-29 Renlin Li <renlin.li@arm.com>
* config/tc-aarch64.c (reg_names): Fix IP1 register alias error.
* testsuite/gas/aarch64/register_aliases.s: Add IP0 and IP1 tests.
* testsuite/gas/aarch64/register_aliases.d: Update.
bfd/
* po/Make-in (datadir): Define as @datadir@.
(localedir): Define as @localedir@.
(gnulocaledir, gettextsrcdir): Use @datarootdir@.
binutils/
* po/Make-in (datadir): Define as @datadir@.
(localedir): Define as @localedir@.
(gnulocaledir, gettextsrcdir): Use @datarootdir@.
gas/
* po/Make-in (datadir): Define as @datadir@.
(localedir): Define as @localedir@.
(gnulocaledir, gettextsrcdir): Use @datarootdir@.
gold/
* po/Make-in (datadir): Define as @datadir@.
(localedir): Define as @localedir@.
(gnulocaledir, gettextsrcdir): Use @datarootdir@.
gprof/
* po/Make-in (datadir): Define as @datadir@.
(localedir): Define as @localedir@.
(gnulocaledir, gettextsrcdir): Use @datarootdir@.
ld/
* po/Make-in (datadir): Define as @datadir@.
(localedir): Define as @localedir@.
(gnulocaledir, gettextsrcdir): Use @datarootdir@.
opcodes/
* po/Make-in (datadir): Define as @datadir@.
(localedir): Define as @localedir@.
(gnulocaledir, gettextsrcdir): Use @datarootdir@.
find_trampoline_seg takes noticeable time when assembling source with
many sections. Cache the result of the most recent search and check it
first. No functional changes.
gas/
2017-11-27 Max Filippov <jcmvbkbc@gmail.com>
* config/tc-xtensa.c (find_trampoline_seg): Add static variable
that caches the result of the most recent search.
There is a recurring pattern in assembly files generated by a compiler
where a lot of jumps in a function are going to the same place. When
these jumps are relaxed with trampolines the assembler generates a
separate jump thread from each source.
Create an index of trampoline jump targets for each segment and see if a
jump being relaxed goes to a location from that index, in which case
replace its target with a location of existing trampoline jump that
results in the shortest path to the original target.
gas/
2017-11-27 Max Filippov <jcmvbkbc@gmail.com>
* config/tc-xtensa.c (trampoline_chain_entry, trampoline_chain)
(trampoline_chain_index): New structures.
(trampoline_index): Add chain_index field.
(xg_order_trampoline_chain_entry, xg_sort_trampoline_chain)
(xg_find_chain_entry, xg_get_best_chain_entry)
(xg_order_trampoline_chain, xg_get_trampoline_chain)
(xg_find_best_eq_target, xg_add_location_to_chain)
(xg_create_trampoline_chain, xg_get_single_symbol_slot): New
functions.
(xg_relax_fixups): Call xg_find_best_eq_target to adjust jump
target to point to an existing jump. Call
xg_create_trampoline_chain to create new jump target. Call
xg_add_location_to_chain to add newly created trampoline jump
to the corresponding chain.
(add_jump_to_trampoline): Extract loop searching for a single
slot with a symbol into a separate function, replace that code
with a call to that function.
(relax_frag_immed): Call xg_find_best_eq_target to adjust jump
target to point to an existing jump.
* testsuite/gas/xtensa/all.exp: Add trampoline-2 test.
* testsuite/gas/xtensa/trampoline.d: Adjust absolute addresses
as many duplicate trampoline chains are now coalesced.
* testsuite/gas/xtensa/trampoline.s: Add _nop so that objdump
stays in sync with instruction stream.
* testsuite/gas/xtensa/trampoline-2.l: New test result file.
* testsuite/gas/xtensa/trampoline-2.s: New test source file.
There's almost exact copy of the trampoline placement code in the
search_trampolines function that is used for jumps generated for relaxed
branch instructions. Get rid of the duplication and reuse
xg_find_best_trampoline function for that.
gas/
2017-11-27 Max Filippov <jcmvbkbc@gmail.com>
* config/tc-xtensa.c (search_trampolines, get_best_trampoline):
Remove definitions.
(xg_find_best_trampoline_for_tinsn): New function.
(relax_frag_immed): Replace call to get_best_trampoline with a
call to xg_find_best_trampoline_for_tinsn.
* testsuite/gas/xtensa/trampoline.d: Adjust absolute addresses
as the placement of trampolines for relaxed branches has been
changed.
Replace linked list of trampoline frags with an ordered array, so that
instead of indexing fixups trampolines could be indexed. Keep each array
in the trampoline_seg structure, so there's no need to rebuild it for
every new processed segment. Don't run relaxation for each trampoline
frag, instead run it for each fixup in the current segment that needs
relaxation at the beginning of each relaxation pass. This way the
complexity of this process drops from about O(n^2 * m) to about
O(log n * m), where n is the number of trampoline frags and m is the
number of fixups that need relaxation in the segment.
gas/
2017-11-27 Max Filippov <jcmvbkbc@gmail.com>
* config/tc-xtensa.c (trampoline_index): New structure.
(trampoline_seg): Replace trampoline list with trampoline index.
(xg_find_trampoline, xg_add_trampoline_to_index)
(xg_remove_trampoline_from_index, xg_add_trampoline_to_seg)
(xg_is_trampoline_frag_full, xg_get_fulcrum)
(xg_find_best_trampoline, xg_relax_fixup, xg_relax_fixups)
(xg_is_relaxable_fixup): New functions.
(J_MARGIN): New macro.
(xtensa_create_trampoline_frag): Use xg_add_trampoline_to_seg
instead of open-coded addition to the linked list.
(dump_trampolines): Iterate through the trampoline_seg::index.
(cached_fixupS, cached_fixup, fixup_cacheS, fixup_cache)
(fixup_order, xtensa_make_cached_fixup)
(xtensa_realloc_fixup_cache, xtensa_cache_relaxable_fixups)
(xtensa_find_first_cached_fixup, xtensa_delete_cached_fixup)
(xtensa_add_cached_fixup, check_and_update_trampolines): Remove
definitions.
(xg_relax_trampoline): Extract logic into separate functions,
replace body with a call to xg_relax_fixups.
(search_trampolines): Replace search in linked list with search
in index. Change data type of address-tracking variables from
int to offsetT. Replace abs with labs.
(xg_append_jump): Finish the trampoline frag if it's full.
(add_jump_to_trampoline): Remove trampoline frag from the index
if the frag is full.
* config/tc-xtensa.h (xtensa_frag_type): Remove next_trampoline.
* testsuite/gas/xtensa/trampoline.d: Adjust absolute addresses
as the placement of trampolines has slightly changed.
* testsuite/gas/xtensa/trampoline.s: Add _nop so that objdump
stays in sync with instruction stream.
The split between fragS and trampoline_frag doesn't save much space, but
makes trampolines management much more awkward. Merge trampoline_frag
data into the xtensa_frag_type, which is a part of fragS. No functional
changes.
gas/
2017-11-27 Max Filippov <jcmvbkbc@gmail.com>
* config/tc-xtensa.c (init_trampoline_frag): Replace pointer to
struct trampoline_frag parameter with pointer to fragS.
(xg_append_jump): Remove jump_around parameter.
(struct trampoline_frag): Remove.
(struct trampoline_seg): Change type of trampoline_list from
struct trampoline_frag to fragS.
(xtensa_create_trampoline_frag): Don't allocate struct
trampoline_frag. Initialize new fragS::tc_frag_data fields.
(dump_trampolines, xg_relax_trampoline, search_trampolines)
(get_best_trampoline, init_trampoline_frag)
(add_jump_to_trampoline, relax_frag_immed): Replace pointer to
struct trampoline_frag with a pointer to fragS.
(xg_append_jump): Remove jump_around parameter, use
fragS::tc_frag_data.jump_around_fix instead.
(xg_relax_trampoline, init_trampoline_frag)
(add_jump_to_trampoline): Don't pass jump_around parameter to
xg_append_jump.
* config/tc-xtensa.h (struct xtensa_frag_type): Add new fields:
needs_jump_around, next_trampoline and jump_around_fix.
xtensa_create_trampoline_frag has opencoded fragment equivalent to
find_trampoline_seg. Drop the fragment and use find_trampoline_seg
instead. No functional changes.
gas/
2017-11-27 Max Filippov <jcmvbkbc@gmail.com>
* config/tc-xtensa.c (find_trampoline_seg): Move above the first
use.
(xtensa_create_trampoline_frag): Replace trampoline seg search
code with a call to find_trampoline_seg.
init_trampoline_frag, add_jump_to_trampoline and xg_relax_trampoline add
a jump to the end of a trampoline frag. Extract it into a separate
funciton and use it in all these places. No functional changes.
gas/
2017-11-27 Max Filippov <jcmvbkbc@gmail.com>
* config/tc-xtensa.c (xg_append_jump): New function.
(xg_relax_trampoline, init_trampoline_frag)
(add_jump_to_trampoline): Replace trampoline jump assembling
code with a call to xg_append_jump.
To make measurement and changes easier extract trampoline relaxation
function. No functional changes.
gas/
2017-11-27 Max Filippov <jcmvbkbc@gmail.com>
* config/tc-xtensa.c (xg_relax_trampoline): New function.
(xtensa_relax_frag): Replace trampoline relaxation code with a
call to xg_relax_trampoline.
For one the register type used for masking should be validated. And then
we shouldn't accept input producing encodings which will #UD when
executed, as is the case when EVEX.Z is set while EVEX.AAA is zero.
Despite EVEX encodings not being available in real and VM86 modes,
16-bit addressing still needs to be handled properly for 16-bit
protected mode as well as 16-bit addressing in 32-bit mode. Neither
should displacements be dropped silently by the assembler, nor should
the disassembler fail to correctly scale 8-bit displacements.
Except for %eip-relative addressing, where we don't have a suitable
relocation type silently wrapping at the 4G boundary, consistently
force use of R_X86_64_32 (in ELF terms) instead of its sign-extending
counterpart. This wasn't right in case there was no base register in
the addressing expression.
Make the assembler recognize UD0, supporting only the newer form
expecting a ModR/M byte.
Make assembler and disassembler properly emit / expect a ModR/M byte for
UD1.
For the testsuite, as arch-4 already tests all UDn, avoid producing a
huge delta for other tests using UD2B by making them use UD2 instead.
Tamar Christina
Nick Clifton
Jan Beulich
H.J. Lu
Jim Wilson
Dimitar Dimitrov
Alan Modra
Max Filippov
Peter Bergner
Renlin Li
Stefan Stroe
Igor Tsimbalist