defined tokens for C, M, Ziscr extensions.
separate the base RV32 instructions from the RV64, for potential future
re-use in a RV32-only assembler, from which the RV64-tok can #include
scall, sbreak have been renamed (page 7 of spec),
necessitating some renaming in riscv64-asm.c
riscv-spec-20191213.pdf was used,
in which the "V" extension is not yet ratified.
available under https://riscv.org/technical/specifications/
Tables 16.5–16.7 do not list any "scall"
neither does the privileged spec
3 additional tokens not present in the tables were removed
note that this riscv64-asm.c still contains defects, which will
be addressed in another commit
tccgen.c:
- new function incr_offset(int) to increment a lvalue
- use it in gv/vstore to load/store from/to two-word types
- use it to advance the pointer to struct fields
- use it to load/store structs passed in registers
- structs: always assume that reg-classes of registers are 2^n
- adjust stack space when regsize > sizeof the_struct
x86_64-gen.c:
- return regsize=16 for VT_QLONG/QFLOAT
i386-gen.c:
- pass structs of size(8) as two VT_INT rather than one VT_LLONG
(both should work now)
fixes a82aff3337
fixes fd6d2180c5 (slightly)
In order to detect a "same file" faster, this makes one
restriction to the feature: the basenames are required to
match at least (as in "test.h" and "dir/../test.h")
Also remove obsolete 'pp_once' (cached includes do not persist
across compilations anymore anyway)
This rewrites commits
30fd24abd44e363a1728
3 files changed, 38 insertions(+), 74 deletions(-)
The stack was not aligned when a returned structure was stored on stack.
This resulted in destoying of previous values stored on stack.
See testcase 119 (tst_struct_return_align) where value d is overwritten.
On backends that rely on gfunc_return() to handle structures
returned in registers (like RISC-V), gfunc_return() may generate
invalid loads for structures without VT_LOCAL and VT_LVAL. This
commit fixes it and adds a regression test
(131_return_struct_in_reg)
pragma once can now be used with
test.h
./test.h
and other references to the same file just like gcc/clang.
On linux we can use stat and st_ino to check for the same file.
On windows the st_ino does not work so we calculate a file hash
if the size of the file differs and then compare the hash.
avoid memory leaks with lost CStrings on stack after errors.
tccpp.c:
- use/abuse static Cstring tokcstr where possible
tccgen.c:
- use/abuse static Cstring initstr where possible
tcc.h/libtcc.a:
- add 'stk_data' array to track memory pointer on stack
- add macros stk_push/pop() and cstr_new/free_s()
tccasm.c:
- use that
- use char[16] instead of char* for op.constraint
This avoids 'exit(1)' with errors outside of compilation
(nasty in particular with libtcc usage)
As a sideeffect multiple errors can be seen for linker
errors (such as undefined symbols, relocation errors, ...)
tccgen.c:
- allow cross-compiling 0.0L (cross-compile tcc with tcc)
tccelf.c:
- use alignment of TLS section for PT_TLS
tccpp.c:
- support long double constants on systems that do not support
long doubles (i.e. mark them (VT_DOUBLE | VT_LONG))
tccdefs.h:
#define __has_feature() for android
remove _unaligned (why define a msvc extension under !_WIN32)
Value 0 and values >2GB where printed wrong during tests with yarpgen (v1)
Starting the output with a hashtag make it compatible to TAP
--
Regards ... Detlef
These extensions where used by other compiler
to generate different code on old cpu's (arm and x86 before pentium/pentiumpro)
to prevent processor exceptions / very high delays when accessing an unaligned pointer.
--
Regards ... Detlef
if we include standard headers on glibc-based systems then
we can't use '__attribute__' (those are defined away for unknown
compilers) but must use '__attribute' (or #undef that token).
tccgen.c:
- just track local_stack for small scopes (can't declare variables)
- fix a vla problem with nested scopes
- move debug N_L/RBRAC into curly braced block
Also:
- tccpp.c: move 'label_...' functions to tccgen.c
- tccpp.c: let get_tok_str() say "<no name>" for anonymous symbols
- tcctest.c: let __pa_symbol() work for memory > 2GB
- 119_random_stuff.c: revert strtoll test (no reason to test libc)
- tccdefs.h/tcctest.c: enable bit fncs for _WIN32
- Makefile:
- use i686-linux-gnu instead of i386-linux-gnu for cross-i386
- update 'make help'
- revert umplicit 'make all' with 'make install' (but print warning)
instead in vcheck_cmp(), pass the CODE_OFF_BIT to generators
unless other nocode purposes are set.
This reverts commit cad8739594.
Also in 128_run_atexit.c:
avoid line output disorder which may happen on windows when
tcc itself and runned code are using two different printf
implementations and tcc would print the "[Returns 1]" above
any output from the runned test.
The generated code by yarpgen failed. I traced the problem to
expr_landor. The problem was that nocode_wanted was removing
too much code.
See also testcase 33.
the casted type was lost when a delayed bool was finally converted
to a value. See testcase, in the wrong case the '(unsigned int)' cast
was ignored, and hence the division was signed reulting in -1 instead of
the proper 0x7fffffff.
The following code:
typedef struct {
unsigned int a __attribute__((nodebug));
unsigned int b;
unsigned int : 32;
unsigned int c;
} tst;
Supresses a and also suppresses bitfield padding.
So debugger shows only b and c in above example.
the code in expr_cond save nocode_wanted around some parts of
expression evaluation, but at the wrong spots. If the evaluation
of the condition itself (e.g. in the testcase the first whole ternary
expression) resulted in CODE_OFF, then that was saved, and restored
before return, even if in-between codegen would have CODE_ON'ed already.
Thus the whole CODE_OFF state bled out to outside the expression
evaluation and also disabled the whole if-block. Found by yarpgen v1
(seed 64).
we activate code (CODE_ON) only when the target labels are used,
which doesn't happen during nocode_wanted regions. So, we cannot
just switch off code either during nocode_wanted regions, nothing
would switch it on again (except in the happy coincidences when we
outright save/restore nocode_wanted). See the testcase for one
example, reduced from code generated by yarpgen: in
ext = (xxx || 1) // #1
|| ((xxx && 1) || 1) // #2
code is first suppressed in #1 normally, then (before this commit)
was suppressed via CODE_OFF during #2 (via indirect gjmp),
then the suppression from #1 was undone, but nothing undoes the
suppression from #2 anymore as everything therein was generated
while nocode_wanted was active.
So, we would either need to save/restore nocode_wanted around
some more expressions, activate CODE_ON also with unused labels (breaks
some optimizations we want), or deactivate CODE_OFF only when not
already in nocode_wanted state. This does the latter.
Sym.sym_scope and Sym.f (FuncAttr) share space, so blindly setting
one clobbers the other. Right now this only leads to missing errors
on incompatible typedefs (see testcase), which this commit fixes.
But it points to a larger problem:
Generally we can only manipulate Sym.f for anonymous and field symbols,
not for anything that has a top-level name (basically any proper decl),
because the latter use sym_scope. Luckily the functions type always
contains an anonymous symbol (in sym->type.ref), so we can use that.
But some of the functions attributes actually _do_ apply to the decl,
not the type (e.g. always_inline), so we still have a problem possibly,
when we update an pre-existing type that may already be shared with
another decl.
Would need untangling and perhaps using accessor functions that check
that Sym.f and Sym.sym_scope aren't used for the same symbol.
I have a testfile created by YARPGen (seed=9), which displayed
a negative bss size without this patch.
* tcc needs also a bigger VSTACK_SIZE (i used 512)
--
bye bye ... Detlef
I have a testfile created by YARPGen (seed=9), which displayed
a negative bss size without this patch.
* tcc needs also a bigger VSTACK_SIZE (i used 512)
--
bye bye ... Detlef
The last commit was not a problem with macos.
The problem was that dwarf version < 5 uses different indexes for
files and directories.
They start at 1 for dwarf version < 5 and 0 for dwarf version >= 5.
noneofyourbusiness