There's no need for such workarounds anymore now that we use C99
uniformly. This addresses several testsuite failures encountered when
(cross-)building on a 32-bit host.
The aarch64 port handles W registers as aliases of X registers. This is
incorrect because X registers are 64-bit and W registers are 32-bit.
This patch teaches GDB how to handle W registers as pseudo-registers of
32-bit, the bottom half of the X registers.
Testcase included.
When using AArch64 GDB with the QEMU debugging stub (in user mode), we get
additional system registers that GDB doesn't particularly care about, so
it doesn't number those explicitly.
But given the pseudo-register numbers are above the number of real registers,
we need to setup/account for the real registers first before going ahead and
numbering the pseudo-registers. This has to happen at the end of
aarch64_gdbarch_init, after the call to tdesc_use_registers, as that
updates the total number of real registers.
This is in preparation to supporting pointer authentication for bare metal
aarch64 (QEMU).
This location of supervisor instructions is out of place (because many other
privileged instructions are located at the tail but after the supervisor
instructions, we have many unprivileged instructions including bit
manipulation / crypto / vector instructions).
Not only that, this is harmful to implement pseudoinstructions in the latest
'P'-extension proposal (CLROV and RDOV). This commit moves supervisor
instructions after all unprivileged instructions.
opcodes/ChangeLog:
* riscv-opc.c (riscv_opcodes): Adjust indents. Move supervisor
instructions after all unprivileged instructions.
When a class inherits from a typedef'd baseclass, GDB may be unable to
find the baseclass if the user is not using the typedef'd name, as is
tested on gdb.cp/virtbase2.exp; the reason that test case is working
under gcc is that the dwarf generated by gcc links the class to the
original definition of the baseclass, not to the typedef. If the
inheritance is linked to the typedef, such as how clang does it,
gdb.cp/virtbase2.exp starts failing.
This can also be seen in gdb.cp/impl-this.exp, when attempting to print
D::Bint::i, and GDB not being able to find the baseclass Bint.
This happens because searching for baseclasses only uses the macro
TYPE_BASECLASS_NAME, which returns the typedef'd name. However, we can't
switch that macro to checking for typedefs, otherwise we wouldn't be
able to find the typedef'd name anymore. This is fixed by searching for
members or baseclasses by name, we check both the saved name and the
name after checking for typedefs.
This also fixes said long-standing bug in gdb.cp/impl-this.exp when the
compiler adds information about typedefs in the debuginfo.
Although it had csr-dw-regnums.d (for CSRs), it didn't have DWARF register
number test for GPRs/FPRs.
This commit adds dw-regnums.{s,d} to test such registers.
gas/ChangeLog:
* testsuite/gas/riscv/dw-regnums.s: New DWARF register number test
for GPRs/FPRs.
* testsuite/gas/riscv/dw-regnums.d: Likewise.
In next-fork-other-thread.c, there's this loop:
...
do
{
ret = waitpid (pid, &stat, 0);
} while (ret == EINTR);
...
The loop condition tests for "ret == EINTR" but waitpid signals EINTR by
returning -1 and setting errno to EINTR.
Fix this by changing the loop condition to "ret == -1 && errno == EINTR".
Tested on x86_64-linux.
I ran some tests like:
$ make check-gdb TESTS="gdb.base/break.exp"
then, then I went to rerun the tests later, I managed to corrupt the
command line, like this:
$ make check-gdb TESTS="gdb.base/breakff.exp"
the make command did exit with an error, but DejaGnu appeared to
report that every test passed! The tail end of the output looks like
this:
Illegal Argument "no-matching-tests-found"
try "runtest --help" for option list
=== gdb Summary ===
# of expected passes 115
/tmp/build/gdb/gdb version 13.0.50.20220831-git -nw -nx -iex "set height 0" -iex "set width 0" -data-directory /tmp/build/gdb/testsuite/../data-directory
make[3]: *** [Makefile:212: check-single] Error 1
make[3]: Leaving directory '/tmp/build/gdb/testsuite'
make[2]: *** [Makefile:161: check] Error 2
make[2]: Leaving directory '/tmp/build/gdb/testsuite'
make[1]: *** [Makefile:1916: check] Error 2
make[1]: Leaving directory '/tmp/build/gdb'
make: *** [Makefile:13565: check-gdb] Error 2
For a while, I didn't spot that DejaGnu had failed at all, I saw the
115 passes, and thought everything had run correctly - though I was
puzzled that make was reporting an error.
What happens is that in gdb/testsuite/Makefile, in the check-single
rule, we first run DejaGnu, then run the dg-add-core-file-count.sh
script, and finally, we use sed to extract the results from the
gdb.sum file.
In my case, with the invalid test name, DejaGnu fails, but the
following steps are still run, the final result, the 115 passes, is
then extracted from the pre-existing gdb.sum file.
If I use 'make -jN' then the 'check-parallel' rule, rather than the
'check-single' rule is used. In this case the behaviour is slightly
different, the tail end of the output now looks like this:
No matching tests found.
make[4]: Leaving directory '/tmp/build/gdb/testsuite'
find: ‘outputs’: No such file or directory
Usage: ../../../src/gdb/testsuite/../../contrib/dg-extract-results.py [-t tool] [-l variant-list] [-L] log-or-sum-file ...
tool The tool (e.g. g++, libffi) for which to create a
new test summary file. If not specified then output
is created for all tools.
variant-list One or more test variant names. If the list is
not specified then one is constructed from all
variants in the files for <tool>.
sum-file A test summary file with the format of those
created by runtest from DejaGnu.
If -L is used, merge *.log files instead of *.sum. In this
mode the exact order of lines may not be preserved, just different
Running *.exp chunks should be in correct order.
find: ‘outputs’: No such file or directory
Usage: ../../../src/gdb/testsuite/../../contrib/dg-extract-results.py [-t tool] [-l variant-list] [-L] log-or-sum-file ...
tool The tool (e.g. g++, libffi) for which to create a
new test summary file. If not specified then output
is created for all tools.
variant-list One or more test variant names. If the list is
not specified then one is constructed from all
variants in the files for <tool>.
sum-file A test summary file with the format of those
created by runtest from DejaGnu.
If -L is used, merge *.log files instead of *.sum. In this
mode the exact order of lines may not be preserved, just different
Running *.exp chunks should be in correct order.
make[3]: Leaving directory '/tmp/build/gdb/testsuite'
make[2]: Leaving directory '/tmp/build/gdb/testsuite'
make[1]: Leaving directory '/tmp/build/gdb'
Rather than DejaGnu failing, we now get a nice 'No matching tests
found' message, followed by some other noise. This other noise is
first `find` failing, followed by the dg-extract-results.py script
failing.
What happens here is that, in the check-parallel rule, the outputs
directory is deleted before DejaGnu is invoked. Then we try to run
all the tests, and finally we use find and dg-extract-results.py to
combine all the separate .sun and .log files together. However, if
there are no tests run then the outputs/ directory is never created,
so the find command and consequently the dg-extract-results.py script,
fail.
This commit aims to fix the following issues:
(1) For check-single and check-parallel rules, don't run any of the
post-processing steps if DejaGnu failed to run. This will avoid all
the noise after the initial failure of DejaGnu,
(2) For check-single ensure that we don't accidentally report
previous results, this is related to the above, but is worth calling
out as a separate point, and
(3) For check-single, print the 'No matching tests found' message
just like we do for a parallel test run. This makes the parallel and
non-parallel testing behaviour more similar, and I think is clearer
than the current 'Illegal Argument' error message.
Points (1) and (2) will be handled by moving the post processing steps
inside an if block within the recipe. For check-single I propose
deleting the gdb.sum and gdb.log files before running DejaGnu, this is
similar (I think) to how we delete the outputs/ directory in the
check-parallel rule.
For point (3) I plan to split the check-single rule in two, the
existing check-single will be renamed do-check-single, then a new
check-single rule will be added. The new check-single rule can either
depend on the new do-check-single, or will ensure the 'No matching
tests found' message is printed when appropriate.
This commit was inspired by this stackoverflow post:
https://stackoverflow.com/questions/73491793/why-is-there-a-%C2%B1-in-lea-rax-rip-%C2%B1-0xeb3
One of the comments helpfully links to this Python test case:
from pygments import formatters, lexers, highlight
def colorize_disasm(content, gdbarch):
try:
lexer = lexers.get_lexer_by_name("asm")
formatter = formatters.TerminalFormatter()
return highlight(content, lexer, formatter).rstrip().encode()
except:
return None
print(colorize_disasm("lea [rip+0x211] # COMMENT", None).decode())
Run the test case and you should see that the '+' character is
underlined, and could be confused with a combined +/- symbol.
What's happening is that Pygments is failing to parse the input text,
and the '+' is actually being marked in the error style. The error
style is red and underlined.
It is worth noting that the assembly instruction being disassembled
here is an x86-64 instruction in the 'intel' disassembly style, rather
than the default att style. Clearly the Pygments module expects the
att syntax by default.
If we change the test case to this:
from pygments import formatters, lexers, highlight
def colorize_disasm(content, gdbarch):
try:
lexer = lexers.get_lexer_by_name("asm")
lexer.add_filter('raiseonerror')
formatter = formatters.TerminalFormatter()
return highlight(content, lexer, formatter).rstrip().encode()
except:
return None
res = colorize_disasm("lea rax,[rip+0xeb3] # COMMENT", None)
if res:
print(res.decode())
else:
print("No result!")
Here I've added the call: lexer.add_filter('raiseonerror'), and I am
now checking to see if the result is None or not. Running this and
the test now print 'No result!' - instead of styling the '+' in the
error style, we instead give up on the styling attempt.
There are two things we need to fix relating to this disassembly
text. First, Pygments is expecting att style disassembly, not the
intel style that this example uses. Fortunately, Pygments also
supports the intel style, all we need to do is use the 'nasm' lexer
instead of the 'asm' lexer.
However, this leads to the second problem; in our disassembler line we
have '# COMMENT'. The "official" Intel disassembler style uses ';'
for its comment character, however, gas and libopcodes use '#' as the
comment character, as gas uses ';' for an instruction separator.
Unfortunately, Pygments expects ';' as the comment character, and
treats '#' as an error, which means, with the addition of the
'raiseonerror' filter, that any line containing a '#' comment, will
not get styled correctly.
However, as the i386 disassembler never produces a '#' character other
than for comments, we can easily "fix" Pygments parsing of the
disassembly line. This is done by creating a filter. This filter
looks for an Error token with the value '#', we then change this into
a comment token. Every token after this (until the end of the line)
is also converted into a comment.
In this commit I do the following:
1. Check the 'disassembly-flavor' setting and select between the
'asm' and 'nasm' lexers based on the setting. If the setting is not
available then the 'asm' lexer is used by default,
2. Use "add_filter('raiseonerror')" to ensure that the formatted
output will not include any error text, which would be underlined,
and might be confusing,
3. If the 'nasm' lexer is selected, then add an additional filter
that will format '#' and all other text on the line, as a comment,
and
4. If Pygments throws an exception, instead of returning None,
return the original, unmodified content. This will mean that this
one instruction is printed without styling, but GDB will continue to
call into the Python code to style later instructions.
I haven't included a test specifically for the above error case,
though I have manually check that the above case now styles
correctly (with no underline). The existing style tests check that
the disassembler styling still works though, so I know I've not
generally broken things.
One final thought I have after looking at this issue is that I wonder
now if using Pygments for styling disassembly from every architecture
is actually a good idea?
Clearly, the 'asm' lexer is OK with att style x86-64, but not OK with
intel style x86-64, so who knows how well it will handle other random
architectures?
When I first added this feature I tested it against some random
RISC-V, ARM, and X86-64 (att style) code, and it seemed fine, but I
never tried to make an exhaustive check of all instructions, so its
quite possible that there are corner cases where things are styled
incorrectly.
With the above changes I think that things should be a bit better
now. If a particular instruction doesn't parse correctly then our
Pygments based styling code will just not style that one instruction.
This is combined with the fact that many architectures are now moving
to libopcodes based styling, which is much more reliable.
So, I think it is fine to keep using Pygments as a fallback mechanism
for styling all architectures, even if we know it might not be perfect
in all cases.
While working on another issue relating to GDB's use of the Python
Pygments package for disassembly styling I noticed an issue in the
case where the Pygments package raises an exception.
The intention of the current code is that, should the Pygments package
raise an exception, GDB will disable future attempts to call into the
Pygments code. This was intended to prevent repeated errors during
disassembly if, for some reason, the Pygments code isn't working.
Since the Pygments based styling was added, GDB now supports
disassembly styling using libopcodes, but this is only available for
some architectures. For architectures not covered by libopcodes
Pygments is still the only option.
What I observed is that, if I disable the libopcodes styling, then
setup GDB so that the Pygments based styling code will indicate an
error (by returning None), GDB does, as expected, stop using the
Pygments based styling. However, the libopcodes based styling will
instead be used, despite this feature having been disabled.
The problem is that the disassembler output is produced into a
string_file buffer. When we are using Pygments, this buffer is
created without styling support. However, when Pygments fails, we
recreate the buffer with styling support.
The problem is that we should only recreate the buffer with styling
support only if libopcodes styling is enabled. This was an oversight
in commit:
commit 4cbe4ca5da
Date: Mon Feb 14 14:40:52 2022 +0000
gdb: add support for disassembler styling using libopcodes
This commit:
1. Factors out some of the condition checking logic into two new
helper functions use_ext_lang_for_styling and
use_libopcodes_for_styling,
2. Reorders gdb_disassembler::m_buffer and gdb_disassembler::m_dest,
this allows these fields to be initialised m_dest first, which means
that the new condition checking functions can rely on m_dest being
set, even when called from the gdb_disassembler constructor,
3. Make use of the new condition checking functions each time
m_buffer is initialised,
4. Add a new test that forces the Python disassembler styling
function to return None, this will cause GDB to disable use of
Pygments for styling, and
5. When we reinitialise m_buffer, and re-disassemble the
instruction, call reset the in-comment flag. If the instruction
being disassembler ends in a comment then the first disassembly pass
will have set the in-comment flag to true. This shouldn't be a
problem as we will only be using Pygments, and thus performing a
re-disassembly pass, if libopcodes is disabled, so the in-comment
flag will never be checked, even if it is set incorrectly. However,
I think that having the flag set correctly is a good thing, even if
we don't check it (you never know what future uses might come up).
This commit extends the gdb.base/style.exp test to cover disassembler
styling using libopcodes (where available).
The test will try to enable libopcode based styling, if this
works (because such styling is available) then some tests are run to
check that the output is styled, and that the styling can be disabled
using 'set style disassembler enabled off'. If libopcodes styling is
not available on the current architecture then these new tests will be
skipped.
I've moved the Python Pygments module check inside the
test_disable_disassembler_styling function now, so that the test will
be run even when Python Pygments is not available, this allows the new
tests discussed above.
If the Pygments module is not available then the Pygments based tests
will be skipped just as they were before.
After the previous few commit, gdbarch_register_name no longer returns
nullptr. This commit audits all the calls to gdbarch_register_name
and removes any code that checks the result against nullptr.
There should be no visible change after this commit.
Building on the previous commits, this commit goes through the various
gdbarch_register_name methods and removes all the remaining 'return
NULL' cases, I claim that these either couldn't be hit, or should be
returning the empty string.
In all cases the return of NULL was used if the register number being
passed to gdbarch_register_name was "invalid", i.e. negative, or
greater than the total number of declared registers. I don't believe
either of these cases can occur, and the previous commit asserts that
this is the case. As a result we can simplify the code by removing
these checks. In many cases, where the register names are held in an
array, I was able to add a static assert that the array contains the
correct number of strings, after that, a direct access into the array
is fine.
I don't have any means of testing these changes.
Building on the previous commits, in this commit I remove two
instances of 'return NULL' from csky_pseudo_register_name, and replace
them with a return of the empty string.
These two are particularly interesting, and worth pulling into their
own commit, because these returns of NULL appear to be depended on
within other parts of the csky code.
In csky-linux-tdep.c in the register collect/supply code, GDB checks
for the register name being nullptr in order to decide if a target
supports a particular feature or not. I've updated the code to check
for the empty string.
I have no way of testing this change.
This commit adds asserts to gdbarch_register_name that validate the
parameters, and the return value.
The interesting thing here is that gdbarch_register_name is generated
by gdbarch.py, and so, to add these asserts, I need to update the
generation script.
I've added two new arguments for Functions and Methods (as declared in
gdbarch-components.py), these arguments are 'param_checks' and
'result_checks'. Each of these new arguments can be used to list some
expressions that are then used within gdb_assert calls in the
generated code.
The asserts that validate the API as described in the comment I added
to gdbarch_register_name a few commits back; the register number
passed in needs to be a valid cooked register number, and the result
being returned should not be nullptr.
Building on the previous commit, this commit extends the register_name
selftest to check for duplicate register names.
If two registers in the cooked register set (real + pseudo registers)
have the same name, then this will show up as duplicate registers in
the 'info all-registers' output, but the user will only be able to
interact with one copy of the register.
In this commit I extend the selftest that I added in the previous
commit to check for duplicate register names, I didn't include this
functionality in the previous commit because one architecture needed
fixing, and I wanted to keep those fixes separate from the fixes in
the previous commit.
The problematic architecture(s) are powerpc:750 and powerpc:604. In
both of these cases the 'dabr' register appears twice, there's a
definition of dabr in power-oea.xml which is included into both
powerpc-604.xml and powerpc-750.xml. Both of these later two xml
files also define the dabr register.
I'm hopeful that this change shouldn't break anything, but I don't
have the ability to actually test this change, however:
On the gdbserver side, neither powerpc-604.xml nor powerpc-750.xml are
mentioned in gdbserver/configure.srv, which I think means that
gdbserver will never use these descriptions, and,
Within GDB the problematic descriptions are held in the variables
tdesc_powerpc_604 and tdesc_powerpc_750, which are only mentioned in
the variants array in rs6000-tdep.c, this is used when looking up a
description based on the architecture.
For a native Linux target however, this will not be used as
ppc_linux_nat_target::read_description exists, which calls
ppc_linux_match_description, which I don't believe can return either
of the problematic descriptions.
This leaves the other native targets, FreeBSD, AIX, etc. These don't
appear to override the ::read_description method, so will potentially
return the problematic descriptions, but, in each case I think the
::fetch_registers and ::store_registers methods will ignore the dabr
register, which will leave the register as <unavailable>.
So, my proposed solution is to just remove the duplicate register from
each of powerpc-604.xml and powerpc-750.xml, then regenerate the
corresponding C++ source file. With this change made, the selftest
now passes for all architectures.
This commit adds a self-test that checks that gdbarch_register_name
never returns nullptr for any valid register number.
Most architectures already met this requirement, there were just 6
that failed the new selftest, and are updated in this commit.
Beyond the self-tests I don't have any facilities to test that the
architectures I've adjusted still work correctly.
If you review all the various gdbarch_register_name implementations
then you will see that there are far more architectures that seem like
they might return nullptr in some situations, e.g. alpha, avr, bpf,
etc. This commit doesn't attempt to address these cases as non of
them are hit during the selftest. Many of these cases can never be
hit, for example, in alpha_register_name GDB checks for a register
number less than zero, this case can't happen and could be changed
into an assert.
A later commit in this series will have a general cleanup of all the
various register_name methods, and remove all references to NULL from
their code, however, as that commit will be mostly adjusting code that
is never hit, I want to keep those changes separate.
The selftest has been tested on x86-64, but I don't have access to
suitable systems to fully test any of the *-tdep.c code I've changed
in this commit.
After the previous commit, this commit sets out to formalise the API
for gdbarch_register_name. Not every architecture is actually in
compliance with the API I set out here, but I believe that most are.
I think architectures that don't comply with the API laid out here
will fail the gdb.base/completion.exp test.
The claims in the comment are I feel, best demonstrated with the
asserts in this code:
const char *
gdbarch_register_name (struct gdbarch *gdbarch, int regnr)
{
gdb_assert (regnr >= 0);
gdb_assert (regnr < gdbarch_num_cooked_regs (gdbarch));
const char *name = gdbarch->register_name (gdbarch, regnr);
gdb_assert (name != nullptr);
return name;
}
Like I said, I don't believe every architecture follows these rules
right now, which is why I'm not actually adding any asserts. Instead,
this commit adds a comment to gdbarch_register_name, this comment is
where I'd like to get to, rather than where we are right now.
Subsequent commits will fix all targets to be in compliance with this
comment, and will even add the asserts shown above to
gdbarch_register_name.
I noticed a test failure in gdb.base/completion.exp for RISC-V on
a native Linux target, this is the failure:
(gdb) FAIL: gdb.base/completion.exp: complete 'info registers '
The problem is caused by a mismatch in the output of 'maint print
registers' and the completion list for 'info registers'. The 'info
registers' completion list contains less registers than
expected. Additionally, the list of registers extracted from the
'maint print registers' list was wrong too, in some cases the test was
grabbing the register number, rather than a register name,
Both of these problems have the same root cause, riscv_register_name
returns nullptr for some registers when it should return an empty
string.
The gdbarch_register_name API is not clearly documented anywhere, and
at first glance it would appear that the function can return either
nullptr, or an empty string to indicate that a register is not
available on the current target. Indeed, there are plenty of places
in GDB where we compare the output of gdbarch_register_name to both
nullptr and '\0' in order to see if a register is supported or not,
and there are plenty of targets that return empty string in some
cases, and nullptr in others.
However, the 'info registers' completion code (reg_or_group_completer)
clearly depends on user_reg_map_regnum_to_name only returning nullptr
when the passed in regnum is greater than the maximum possible
register number (i.e. after all physical registers, pseudo-registers,
and user-registers), this means that gdbarch_register_name should not
be returning nullptr.
I did consider "fixing" user_reg_map_regnum_to_name, if
gdbarch_register_name returns nullptr, I could convert to an empty
string at this point, but that felt like a real hack, so I discarded
that plan.
The next possibility I considered was "fixing" reg_or_group_completer
to not rely on nullptr to indicate the end marker. Or rather, I could
have reg_or_group_completer use gdbarch_num_cooked_regs, we know that
we should check at least that many register numbers. Then, once we're
passed that limit, we keep checking until we hit a nullptr. This
would absolutely work, and didn't actually feel that bad, but, it
still felt a little weird that gdbarch_register_name could return
nullptr OR the empty string to mean the same thing, so I wondered if
the "right" solution was to have gdbarch_register_name not return
nullptr. With this in mind I tried an experiment:
I added a self-test that, for each architecture, calls
gdbarch_register_name for every register number up to the
gdbarch_num_cooked_regs limit, and checks that the name is not
nullptr.
Only a handful of architectures failed this test, RISC-V being one of
them.
This seems to suggest that most architectures agree that the correct
API for gdbarch_register_name is to return an empty string for
registers that are not supported on the current target, and that
returning nullptr is really a mistake.
In this commit I will update the RISC-V target so that GDB no longer
returns nullptr from riscv_register_name, instead we return the empty
string.
In subsequent commits I will add the selftest that I mention above,
and will fix the targets that fail the selftest.
With this change the gdb.base/completion.exp test now passes.
I noticed a test failure in gdb.base/completion.exp for RISC-V on a
native Linux target. Upon investigation I discovered a couple of
reasons for the failure, this commit addresses one of them. A later
commit will address the other issue.
The completion.exp test makes use of the capture_command_output proc
to collect the output of the 'maint print registers' command. For
RISC-V this command produces a lot of output.
Currently the capture_command_output proc tries to collect the
complete command output in a single expect buffer, and what I see is
an error caused by the expect buffer becoming full.
This commit rewrites capture_command_output to make use of
gdb_test_multiple to collect the command output line at a time, in
this way we avoid overflowing the expect buffer.
The capture_command_output proc has some logic for skipping a prefix
pattern, which is passed in to the proc as an argument. In order to
handle this correctly (only matching the prefix at the start of the
command output), I use two gdb_test_multiple calls, the first spots
and discards the echoed command and the (optional) prefix pattern, the
second gdb_test_multiple call then collects the rest of the command
output line at a time until a prompt is seen.
There is one slight oddity with the current implementation, which I
have changed in my rewrite, this does, slightly, change the behaviour
of the proc.
The current implementation uses this pattern:
-re "[string_to_regexp ${command}]\[\r\n\]+${prefix}(.*)\[\r\n\]+$gdb_prompt $"
Now a typical command output will look like this:
output here\r\n
(gdb)
As the TCL regexp matching is greedy, TCL will try to match as much as
possible in one part of the pattern before moving on to the next.
Thus, when this matches against (.*)[\r\n]+, the (.*) will end up
matching against 'output here\r' and the [\r\n]+ will match '\n' only.
In short the previous implementation would leave the '\r' on the end
of the returned text, but not the final trailing '\n'.
Now clearly I could make a new version of capture_command_output that
maintained this behaviour, but I couldn't see any reason to do this.
So, my new implementation drops the final '\r\n' completely, in our
example above we now return 'output here' with no '\r'.
This change doesn't seem to affect any of the existing tests, but I
thought it was worth mentioning.
Now that the disassembler has two different strategies for laying out
the opcode bytes of an instruction (see /r vs /b for the disassemble
command), I wanted to add support for this to the MI disassemble
command.
Currently the -data-disassemble command takes a single 'mode' value,
which currently has 6 different values (0 -> 5), 3 of these modes
relate to opcode display.
So, clearly I should just add an additional 3 modes to handle the new
opcode format, right?
No, I didn't think that was a great idea either.
So, I wonder, could I adjust the -data-disassemble command in a
backward compatible way, that would allow GDB to move away from using
the mode value altogether?
I think we can.
In this commit, I propose adding two new options to -data-disassemble,
these are:
--opcodes none|bytes|display
--source
Additionally, I will make the mode optional, and default to mode 0 if
no mode value is given. Mode 0 is the simplest, no source code, no
opcodes disassembly mode.
The two new options are only valid for mode 0, if they are used with
any other mode then an error is thrown.
The --opcodes option can add opcodes to the result, with 'bytes' being
equivalent to 'disassemble /b' and 'display' being 'disassemble /r'.
The --source option will enable the /s style source code display, this
is equivalent to modes 4 and 5. There is no way, using the new
command options to get the now deprecated /m style source code
display, that is mode 1 and 3.
My hope is that new users of the MI will not use the mode at all, and
instead will just use the new options to achieve the output they need.
Existing MI users can continue to use the mode, and will not need to
be updated to use the new options.
The argument documentation for -data-disassemble looks like this:
-data-disassemble
[ -s @var{start-addr} -e @var{end-addr} ]
| [ -a @var{addr} ]
| [ -f @var{filename} -l @var{linenum} [ -n @var{lines} ] ]
-- @var{mode}
However, I believe, according to the 'Notation and Terminology'
section, this means that the there are 3 optional location
specification argument groups for the command, followed by a
non-optional '-- mode'.
However, this is not true, one of the location specifications must be
given, i.e. we can't choose to give NO location specification, which
is what the above implies.
I propose that we change this to instead be:
-data-disassemble
( -s @var{start-addr} -e @var{end-addr}
| -a @var{addr}
| -f @var{filename} -l @var{linenum} [ -n @var{lines} ] )
-- @var{mode}
By placing all the location specifications within '( ... )' we
indication that these are a group, from which one of the options,
separated by '|', must be selected.
However, the 'Notation and Terminology' section only describes two
uses for parenthesis: '( GROUP )*' and '( GROUP )+', in the first case
GROUP is repeated zero or more times, and in the second GROUP is
repeated 1 or more times.
Neither of those exactly describe what I want, which is GROUP must
appear exactly once. I propose to extend 'Notation and Terminology'
to include '( GROUP )' which means that GROUP should appear exactly
once.
This change is important because, in a later commit, I want to add
additional optional arguments to the -data-disassemble command, and
things start to get confusing with the original syntax.
In a later commit I want to use operator~ on a gdb_disassembly_flag
flag value. This is currently not possible as gdb_disassembly_flag
is, by default, signed.
This commit just makes this enum unsigned.
There should be no user visible changes after this commit.
This commit changes the format of 'disassemble /r' to match GNU
objdump. Specifically, GDB will now display the instruction bytes in
as 'objdump --wide --disassemble' does.
Here is an example for RISC-V before this patch:
(gdb) disassemble /r 0x0001018e,0x0001019e
Dump of assembler code from 0x1018e to 0x1019e:
0x0001018e <call_me+66>: 03 26 84 fe lw a2,-24(s0)
0x00010192 <call_me+70>: 83 25 c4 fe lw a1,-20(s0)
0x00010196 <call_me+74>: 61 65 lui a0,0x18
0x00010198 <call_me+76>: 13 05 85 6a addi a0,a0,1704
0x0001019c <call_me+80>: f1 22 jal 0x10368 <printf>
End of assembler dump.
And here's an example after this patch:
(gdb) disassemble /r 0x0001018e,0x0001019e
Dump of assembler code from 0x1018e to 0x1019e:
0x0001018e <call_me+66>: fe842603 lw a2,-24(s0)
0x00010192 <call_me+70>: fec42583 lw a1,-20(s0)
0x00010196 <call_me+74>: 6561 lui a0,0x18
0x00010198 <call_me+76>: 6a850513 addi a0,a0,1704
0x0001019c <call_me+80>: 22f1 jal 0x10368 <printf>
End of assembler dump.
There are two differences here. First, the instruction bytes after
the patch are grouped based on the size of the instruction, and are
byte-swapped to little-endian order.
Second, after the patch, GDB now uses the bytes-per-line hint from
libopcodes to add whitespace padding after the opcode bytes, this
means that in most cases the instructions are nicely aligned.
It is still possible for a very long instruction to intrude into the
disassembled text space. The next example is x86-64, before the
patch:
(gdb) disassemble /r main
Dump of assembler code for function main:
0x0000000000401106 <+0>: 55 push %rbp
0x0000000000401107 <+1>: 48 89 e5 mov %rsp,%rbp
0x000000000040110a <+4>: c7 87 d8 00 00 00 01 00 00 00 movl $0x1,0xd8(%rdi)
0x0000000000401114 <+14>: b8 00 00 00 00 mov $0x0,%eax
0x0000000000401119 <+19>: 5d pop %rbp
0x000000000040111a <+20>: c3 ret
End of assembler dump.
And after the patch:
(gdb) disassemble /r main
Dump of assembler code for function main:
0x0000000000401106 <+0>: 55 push %rbp
0x0000000000401107 <+1>: 48 89 e5 mov %rsp,%rbp
0x000000000040110a <+4>: c7 87 d8 00 00 00 01 00 00 00 movl $0x1,0xd8(%rdi)
0x0000000000401114 <+14>: b8 00 00 00 00 mov $0x0,%eax
0x0000000000401119 <+19>: 5d pop %rbp
0x000000000040111a <+20>: c3 ret
End of assembler dump.
Most instructions are aligned, except for the very long instruction.
Notice too that for x86-64 libopcodes doesn't request that GDB group
the instruction bytes. This matches the behaviour of objdump.
In case the user really wants the old behaviour, I have added a new
modifier 'disassemble /b', this displays the instruction byte at a
time. For x86-64, which never groups instruction bytes, /b and /r are
equivalent, but for RISC-V, using /b gets the old layout back (except
that the whitespace for alignment is still present). Consider our
original RISC-V example, this time using /b:
(gdb) disassemble /b 0x0001018e,0x0001019e
Dump of assembler code from 0x1018e to 0x1019e:
0x0001018e <call_me+66>: 03 26 84 fe lw a2,-24(s0)
0x00010192 <call_me+70>: 83 25 c4 fe lw a1,-20(s0)
0x00010196 <call_me+74>: 61 65 lui a0,0x18
0x00010198 <call_me+76>: 13 05 85 6a addi a0,a0,1704
0x0001019c <call_me+80>: f1 22 jal 0x10368 <printf>
End of assembler dump.
Obviously, this patch is a potentially significant change to the
behaviour or /r. I could have added /b with the new behaviour and
left /r alone. However, personally, I feel the new behaviour is
significantly better than the old, hence, I made /r be what I consider
the "better" behaviour.
The reason I prefer the new behaviour is that, when I use /r, I almost
always want to manually decode the instruction for some reason, and
having the bytes displayed in "instruction order" rather than memory
order, just makes this easier.
The 'record instruction-history' command also takes a /r modifier, and
has been modified in the same way as disassemble; /r gets the new
behaviour, and /b has been added to retain the old behaviour.
Finally, the MI command -data-disassemble, is unchanged in behaviour,
this command now requests the raw bytes of the instruction, which is
equivalent to the /b modifier. This means that the MI output will
remain backward compatible.
This commit reduces the number of times we call read_code when
printing the instruction opcode bytes during disassembly.
I've added a new gdb::byte_vector within the
gdb_pretty_print_disassembler class, in line with all the other
buffers that gdb_pretty_print_disassembler needs. This byte_vector is
then resized as needed, and filled with a single read_code call for
each instruction.
There should be no user visible changes after this commit.
Add another test for the output of MI command -data-disassemble. The
new check validates the format of the 'opcodes' field, specifically,
this test checks that the field contains a series of bytes, separated
by a single space.
We also check that the bytes are in the correct order, that is, the
first byte is from the lowest address, and subsequent bytes are from
increasing addresses.
The motivation for this test (besides more tests being generally good)
is that I plan to make changes to how opcode bytes are displayed in
the disassembler output, and I want to ensure that I don't break any
existing MI behaviour.
There should be no user visible changes to GDB after this commit.
This commit relaxes requirements to "fmv.s" instructions from 'F' to ('F'
or 'Zfinx'). The same applies to "fmv.d" and "fmv.q". Note that 'Zhinx'
extension already contains "fmv.h" instruction (as well as 'Zfh').
gas/ChangeLog:
* testsuite/gas/riscv/zfinx.s: Add "fmv.s" instruction.
* testsuite/gas/riscv/zfinx.d: Likewise.
* testsuite/gas/riscv/zdinx.s: Add "fmv.d" instruction.
* testsuite/gas/riscv/zdinx.d: Likewise.
* testsuite/gas/riscv/zqinx.d: Add "fmv.q" instruction.
* testsuite/gas/riscv/zqinx.s: Likewise.
opcodes/ChangeLog:
* riscv-opc.c (riscv_opcodes): Relax requirements to "fmv.[sdq]"
instructions to support those in 'Zfinx'/'Zdinx'/'Zqinx'.
This commit adds certain test cases for 'Zfinx'/'Zdinx'/'Zqinx' extensions
and reorganizes them, fixing coding style while improving coverage.
This is partially based on jiawei's 'Zhinx' testcases.
gas/ChangeLog:
* testsuite/gas/riscv/zfinx.s: Use different registers for
better encode space testing. Make indentation consistent.
Add tests for instruction with rounding mode. Change march
to minimum required extensions. Remove source line.
* testsuite/gas/riscv/zfinx.d: Likewise.
* testsuite/gas/riscv/zdinx.s: Likewise.
* testsuite/gas/riscv/zdinx.d: Likewise.
* testsuite/gas/riscv/zqinx.s: Likewise.
Also use even-numbered registers to use valid register pairs.
* testsuite/gas/riscv/zqinx.d: Likewise.
Signed-off-by: Tsukasa OI <research_trasio@irq.a4lg.com>
Signed-off-by: jiawei <jiawei@iscas.ac.cn>
Pre- and post-increment/decrement are side effects, the behavior of
which is undefined when combined with passing an address of the accessed
variable in the same function invocation. There's no need for the
increments here - simply adding 1 achieves the intended effect without
triggering compiler diagnostics (which are fatal with -Werror).
FENCE.TSO isn't an alias. ZIP and UNZIP in the long run likely are, but
presently they aren't. This fixes disassembly of these insns with
-Mno-aliases.
For disassembly to pick up aliases in favor of underlying insns (helping
readability in the common case), the aliases need to come ahead of the
"base" insns. Slightly more code movement is needed because of insns
with the same name needing to stay next to each other.
Note that the "rorw" alias entry also has the missing INSN_ALIAS added
here.
Clone a few testcases to exercise -Mno-aliases some more, better
covering the differences between the default and that disassembly mode.
With the command in the rule merely being "echo", i386-tbl.h won't be
rebuilt if missing, when at the same time i386-init.h is present and
up-to-date. Use a pattern rule instead to express the multiple targets
correctly (the &: rule separator is supported only by GNU make 4.3 and
newer). Note that now, for the opposite case to work (i386-tbl.h is
up-to-date but i386-init.h is missing), i386-init.h also needs
mentioning as a dependency somewhere: Add a fake dependency for
i386-opc.lo ("fake" because i386-opc.c doesn't include that header).
At the same time use $(AM_V_GEN) in the actual rule, replacing the
earlier (open-coded) "echo". And while there also drop a duplicate
dependency of i386-gen.o on i386-opc.h.
At the example of
extractps $0, %xmm0, %xmm0
insertps $0, %xmm0, %eax
(both having respectively the same mistake of using the wrong kind of
destination register) it is easy to see that current behavior is far
from ideal: The former results in "unsupported instruction" for 32-bit
code simply because the 2nd template we have is a Cpu64 one. Instead we
should aim at emitting the "best" possible error, which will typically
be the one where we passed the largest number of checks. Generalize the
original "specific_error" approach by making it apply to the entire
matching loop, utilizing that line numbers increase as we pass further
checks.
While in some cases deriving an AT&T-style suffix from an Intel syntax
memory operand size specifier is necessary, in many cases this is not
only pointless, but has led to the introduction of various workarounds:
Excessive use of IgnoreSize and NoRex64 as well as the ToDword and
ToQword attributes. Suppress suffix derivation when we can clearly tell
that the memory operand's size isn't going to be needed to infer the
possible need for the low byte/word opcode bit or an operand size prefix
(0x66 or REX.W).
As a result ToDword and ToQword can be dropped entirely, plus a fair
number of IgnoreSize and NoRex64 can also be got rid of. Note that
IgnoreSize needs to remain on legacy encoded SIMD insns with GPR
operand, to avoid emitting an operand size prefix in 16-bit mode. (Since
16-bit code using SIMD insns isn't well tested, clone an existing
testcase just enough to cover a few insns which are potentially
problematic but are being touched here.)
Note that while folding the VCVT{,T}S{S,D}2SI templates, VCVT{,T}SH2SI
isn't included there. This is to fulfill the request of not allowing L
and Q suffixes there, despite the inconsistency with VCVT{,T}S{S,D}2SI.
