This completes the transition started in Knight scripts during their M2libc
migration at f2094a9e5b
from : to
--BaseAddress : --base-address
--BigEndian : --big-endian
--LittleEndian : --little-endian
The plan for mescc-tools is to deprecate the old (pre-1.0.0) flags.
Note that test0101 compiles "hex2 0.3" and verifies that it works by
linking an example. This 0.3 version requires the old-style flags, so the
calls to test/results/test0101-*-binary are left unchanged.
hex2 in mescc-tools defaults to executable output file since 1.0.0.
See 0a553f720d39dae4231611671e4163a2d2422e0a there.
This completes the removal started in Knight scripts during their M2libc
migration at f2094a9e5b
Note that test0101 compiles "hex2 0.3" and verifies that it works by
linking an example. Calls to test/results/test0101-*-binary never used the
flag because we don't call the linked example (we just check its SHA-256).
We were checking an empty proof because the test writes to stderr and we
were redirecting stdout to the file. With the correct redirection the hash
is different, so proof.answer is also adjusted.
The x86 test was calling blood-elf with the 64bits flag. Removed to ask for
32bits behaviour.
For consistency, --BaseAddress is now using the usual value for the arch.
Also, unused input file is removed here.
Every AArch64 test script is affected because AArch64 -> aarch64. test0101
needs i386 -> x86 because it verifies the resulting AArch64 binary (hex2)
creating some x86 proof.
Also, a couple of x86 tests (test0000 and test0025) that are M2libc aware,
again because i386 -> x86 (hex2 ELF parts).
For any test there's a directory with a script for each architecture.
These scripts generate intermediate files in the dir during compilation,
with the same names no matter the arch. So if one tries to execute the
scripts in parallel they all fight for the files.
To allow for parallel testing, now we create a temporal dir named after
the arch to isolate the intermediate files. Scripts are patched
accordingly.
This also allows easier clean-up: now, removing the temporal dirs is
almost all that we need (and the per-test .gitignore files are replaced
with a couple of lines in test/.gitignore). A common script replaces
the per-test cleanup.sh. It gets the number of the test as an argument
and takes care of some minor special cases. The makefile is adapted to
use this new test/cleanup_test.sh script.
Those special files are created during the last stage of each test, but
only if the host arch matches the test. They're _not_ isolated here
because it's harder (due to how we do the checksum) and they cannot
collide unless emulation is used and several calls to make are made at
the same time.
- disables --bootstrap-mode where it was enabled, except for test1000
which needs to build in bootstrap-mode to check that M2-Planet can be
compiled by cc_*
- the only hold-out is test0021 which needs getcwd which is missing in
M2libc
- note that test0106 needs a fix to fgetc or it will never notice it has
reached the end of stdin
- no longer generate CONSTANT lines for #defined constants
- "#define FOO 1" turns "return FOO;" into "return 1;"
- Macros expanding into a macro is not supported:
#define FOO 1
#define BAR FOO
will expand BAR to FOO, not to 1
- Expansion inside #if conditions is not supported.
- Function-like macros are not supported.
- // and # lines always appear in the result of read_all_tokens
- in bootstrap mode, // tokens (but not the subsequent lines) and #
lines are stripped. no preprocessing happens.
- in non-bootstrap mode, // lines are stripped and # lines stay for the
preprocessing phase
- updates tests to pass --bootstrap-mode when necessary
- support for #if, #elif, #else, #endif, but no macro expansion just
yet. Any macro is treated as undefined and expands to 0.
for example:
./bin/M2-Planet -E test.c
#if 1 == FOO
foo
#else
bar
#endif
results in
bar
- change the lexer to not remove preprocessor directives
- preprocess() removes any lines starting with a macro directive that it
doesn't understand (which is currently all of them)
- needed to determine start and end of macro directives
- the main parser doesn't need them, so strip out the newline tokens
before parsing to avoid changing it
Note that this is a backwards incompatible change: Unlike POSIX, exit()
will not flush all open streams, so you are responsible for flushing or
closing all your streams before successful exit.
Also, file.c will now also require calloc.c (and malloc.c).
Updated the test cases accordingly.
Only implemented and tested for x86.
The close syscall was missing an indirection and therefore closed random
file descriptors (very visible in strace output).
Test 0104 did not properly null terminate the envp, resulting in -EFAULT
on execve syscall in case the next value in memory does not happen to be
0 (which it seems to be right now).
Paraphrasing OriansJ's IRC message while working on kaem:
GCC needs fflush(stdout) to get matching behavior, as M2-Planet doesn't
buffer. I guess we will need to make a fflush function (it'll do nothing
but return 0).
Until now, each element on the compiled program stack was 128 bits long,
half of them unused because only one 64 bits value was stored.
Now it's 64 bits long, so we don't waste all that memory.
We workaround the architectural alignment restriction of the SP register by
using a free regular register. X18 is for platform use so it seems a good
candidate for this task.
At _start we copy the value of SP into X18. SP is not used anymore. When a
definition refers to SP it doesn't mean it literaly now, because here we
redefine (without renaming) the involved M1 macro definitions to operate
on X18 (easier transition; abstraction). INIT_SP is introduced.
The function arguments are passed via stack, so the offset from the Base
Pointer ("depth") to each of them is different now. Changes both to
compiler code (cc_core.c) and libc reflect that the arguments are 8 bytes
(instead of 16 bytes) apart now. Note that SUB_X0_32 and SUB_X0_48 are
removed, because we only need 8, 16 and 24 bytes subtraction to reach the
arguments from asm libc functions. SUB_X0_8 and SUB_X0_24 are introduced.
Now, aarch64 is valid for --architecture on the command-line. Accordingly,
the global variable Architecture is set to the new AARCH64 constant.
Manual page updated.
Register size is configured to 8 bytes.
Because, as OriansJ found, we don't want to read the first bytes of the
value as a pointer. See global_load() for programs like:
char* s = "xyzzy\n";
int main(void){ file_print(s, 1); return 0; }
There's an alternative: "fix the GLOBAL behavior to be type sensitive" but
this indirect store was choosen because it "preserve[s] the ability to
assign new strings to a global char*".
deesix
Jeremiah Orians
Sanne Wouda
Michael Schierl
nimaje
fosslinux
Jan Nieuwenhuizen
Andrius Štikonas
deesix