Jsut for testing. It works for me (don't break anything)
Small fixes for x86_64-gen.c in "tccpp: fix issues, add tests"
are dropped in flavor of this patch.
Pip Cet:
Okay, here's a first patch that fixes the problem (but I've found
another bug, yet unfixed, in the process), though it's not
particularly pretty code (I tried hard to keep the changes to the
minimum necessary). If we decide to actually get rid of VT_QLONG and
VT_QFLOAT (please, can we?), there are some further simplifications in
tccgen.c that might offset some of the cost of this patch.
The idea is that an integer is no longer enough to describe how an
argument is stored in registers. There are a number of possibilities
(none, integer register, two integer registers, float register, two
float registers, integer register plus float register, float register
plus integer register), and instead of enumerating them I've
introduced a RegArgs type that stores the offsets for each of our
registers (for the other architectures, it's simply an int specifying
the number of registers). If someone strongly prefers an enum, we
could do that instead, but I believe this is a place where keeping
things general is worth it, because this way it should be doable to
add SSE or AVX support.
There is one line in the patch that looks suspicious:
} else {
addr = (addr + align - 1) & -align;
param_addr = addr;
addr += size;
- sse_param_index += reg_count;
}
break;
However, this actually fixes one half of a bug we have when calling a
function with eight double arguments "interrupted" by a two-double
structure after the seventh double argument:
f(double,double,double,double,double,double,double,struct { double
x,y; },double);
In this case, the last argument should be passed in %xmm7. This patch
fixes the problem in gfunc_prolog, but not the corresponding problem
in gfunc_call, which I'll try tackling next.
* fix some macro expansion issues
* add some pp tests in tests/pp
* improved tcc -E output for better diff'ability
* remove -dD feature (quirky code, exotic feature,
didn't work well)
Based partially on ideas / researches from PipCet
Some issues remain with VA_ARGS macros (if used in a
rather tricky way).
Also, to keep it simple, the pp doesn't automtically
add any extra spaces to separate tokens which otherwise
would form wrong tokens if re-read from tcc -E output
(such as '+' '=') GCC does that, other compilers don't.
* cleanups
- #line 01 "file" / # 01 "file" processing
- #pragma comment(lib,"foo")
- tcc -E: forward some pragmas to output (pack, comment(lib))
- fix macro parameter list parsing mess from
a3fc543459a715d7143d
(some coffee might help, next time ;)
- introduce TOK_PPSTR - to have character constants as
written in the file (similar to TOK_PPNUM)
- allow '\' appear in macros
- new functions begin/end_macro to:
- fix switching macro levels during expansion
- allow unget_tok to unget more than one tok
- slight speedup by using bitflags in isidnum_table
Also:
- x86_64.c : fix decl after statements
- i386-gen,c : fix a vstack leak with VLA on windows
- configure/Makefile : build on windows (MSYS) was broken
- tcc_warning: fflush stderr to keep output order (win32)
Author: Philip <pipcet@gmail.com>
Our VLA code can be made a lot simpler (simple enough for
even me to understand it) by giving up on the optimization idea, which
is very tempting. There's a patch to do that attached, feel free to
test and commit it if you like. (It passes all the tests, at least
The old code assumed that if an argument doesn't fit into the available
registers, none of the subsequent arguments do, either. But that's
wrong: passing 7 doubles, then a two-double struct, then another double
should generate code that passes the 9th argument in the 8th register
and the two-double struct on the stack. We now do so.
However, this patch does not yet fix the function calling code to do the
right thing in the same case.
The comment suggests this was meant to detect unions, but in fact it
compared f->c, the union/struct size, against f->next->c, the first
element's offset.
This affected only zero-length structs/unions with a first (zero-length)
element, as in this code:
struct u2 {
};
struct u {
struct u2 u2;
} u;
struct u f(struct u x)
{
return x;
}
However, such structures turned out to be broken anyway, as code like this
was generated for the above f:
0000000000000000 <f>:
0: 55 push %rbp
1: 48 89 e5 mov %rsp,%rbp
4: 48 81 ec 10 00 00 00 sub $0x10,%rsp
b: 66 0f d6 45 f8 movq %xmm0,-0x8(%rbp)
10: 66 0f 6e 45 f8 movd -0x8(%rbp),%xmm0
15: e9 00 00 00 00 jmpq 1a <f+0x1a>
1a: c9 leaveq
1b: c3 retq
I ran into an issue playing with tinycc, and tracked it down to a rather
weird assumption in the function calling code. This breaks only when
varargs and float/double arguments are combined, I think, and only when
calling GCC-generated (or non-TinyCC, at least) code. The problem is we
sometimes generate code like this:
804a468: 4c 89 d9 mov %r11,%rcx
804a46b: b8 01 00 00 00 mov $0x1,%eax
804a470: 48 8b 45 c0 mov -0x40(%rbp),%rax
804a474: 4c 8b 18 mov (%rax),%r11
804a477: 41 ff d3 callq *%r11
for a function call. Note how $eax is first set to the correct value,
then clobbered when we try to load the function pointer into R11. With
the patch, the code generated is:
804a468: 4c 89 d9 mov %r11,%rcx
804a46b: b8 01 00 00 00 mov $0x1,%eax
804a470: 4c 8b 5d c0 mov -0x40(%rbp),%r11
804a474: 4d 8b 1b mov (%r11),%r11
804a477: 41 ff d3 callq *%r11
which is correct.
This becomes an issue when get_reg(RC_INT) is modified not always to
return %rax after a save_regs(0), because then another register (%ecx,
say) is clobbered, and the function passed an invalid argument.
A rather convoluted test case that generates the above code is
included. Please note that the test will not cause a failure because
TinyCC code ignores the %rax argument, but it will cause incorrect
behavior when combined with GCC code, which might wrongly fail to save
XMM registers and cause data corruption.
Verify an immediate value fits into 32 bits before jumping to it/calling
it with a 32-bit immediate operand. Without this fix, code along the
lines of
((int (*)(const char *, ...))140244834372944LL)("hi\n");
will fail mysteriously, even if that decimal constant is the correct
address for printf.
See https://github.com/pipcet/tinycc/tree/bugfix-1
The common code to move a returned structure packed into
registers into memory on the caller side didn't take the
register size into account when allocating local storage,
so sometimes that lead to stack overwrites (e.g. in 73_arm64.c),
on x86_64. This fixes it by generally making gfunc_sret also return
the register size.
Author: Thomas Preud'homme <robotux@celest.fr>
Date: Tue Dec 31 23:51:20 2013 +0800
Move logic for if (int value) to tccgen.c
Move the logic to do a test of an integer value (ex if (0)) out of
arch-specific code to tccgen.c to avoid code duplication. This also
fixes test of long long value which was only testing the bottom half of
such values on 32 bits architectures.
I don't understand why if () in gtst(i) was removed.
This patch allows to compile a linux kernel v.2.4.26
W/o this patch a tcc simply crashes.
- revert to R_X86_64_PC32 for near calls on PE
- revert to s1->section_align set to zero by default
Untested. Compared to release_0_9_26 the pe-image looks back to
normal. There are some differences in dissassembly (r10/r11 usage)
but maybe that's ok.
Refactoring (no logical changes):
- use memcpy in tccgen.c:ieee_finite(double d)
- use union to store attribute flags in Sym
Makefile: "CFLAGS+=-fno-strict-aliasing" basically not necessary
anymore but I left it for now because gcc sometimes behaves
unexpectedly without.
Also:
- configure: back to mode 100755
- tcc.h: remove unused variables tdata/tbss_section
- x86_64-gen.c: adjust gfunc_sret for prototype
The procedure calling standard for ARM architecture mandate the use of
the base standard for variadic function. Therefore, hgen float aggregate
must be returned via stack when greater than 4 bytes and via core
registers else in case of variadic function.
This patch improve gfunc_sret() to take into account whether the
function is variadic or not and make use of gfunc_sret() return value to
determine whether to pass a structure via stack in gfunc_prolog(). It
also take advantage of knowing if a function is variadic or not move
float result value from VFP register to core register in gfunc_epilog().
Move the logic to do a test of an integer value (ex if (0)) out of
arch-specific code to tccgen.c to avoid code duplication. This also
fixes test of long long value which was only testing the bottom half of
such values on 32 bits architectures.
Use comisd / fcompp for float comparison (except TOK_EQ and TOK_NE)
instead of ucomisd / fucompp to detect NaN comparison.
Thanks Vincent Lefèvre for the bug report and for also giving the
solution.
Set *palign for VT_BITFIELD and VT_ARRAY types in classify_x86_64_arg as
else you happen to have in *palign what was already there. This can
cause gfunc_call on !PE systems to consider an array as 16 bytes align
and trigger the assert if the previous argument was 16 bytes aligned.
On ARM with hardfloat calling convention, structure containing 4 fields
or less of the same float type are returned via float registers. This
means that a structure can be returned in up to 4 double registers in a
structure is composed of 4 doubles. This commit adds support for return
of structures in several registers.
VLA storage is now freed when it goes out of scope. This makes it
possible to use a VLA inside a loop without consuming an unlimited
amount of memory.
Combining VLAs with alloca() should work as in GCC - when a VLA is
freed, memory allocated by alloca() after the VLA was created is also
freed. There are some exceptions to this rule when using goto: if a VLA
is in scope at the goto, jumping to a label will reset the stack pointer
to where it was immediately after the last VLA was created prior to the
label, or to what it was before the first VLA was created if the label
is outside the scope of any VLA. This means that in some cases combining
alloca() and VLAs will free alloca() memory where GCC would not.
long double arguments require 16-byte alignment on the stack, which
requires adjustment when the the stack offset is not an evven number of
8-byte words.
I removed the XMM6/7 registers from the register list because they are not used
on Win64 however they are necessary for parameter passing on x86-64. I have now
restored them but not marked them with RC_FLOAT so they will not be used except
for parameter passing.
Also made XMM0-7 available for use as temporary registers, since they
are not used by the ABI. I'd like to do the same with RSI and RDI but
that's trickier since they can be used by gv() as temporary registers
and there isn't a way to disable that.
All tests pass. I think I've caught all the cases assuming only XMM0 is
used. I expect that Win64 is horribly broken by this point though,
because I haven't altered it to cope with XMM1.
I've had to introduce the XMM1 register to get the calling convention
to work properly, unfortunately this has broken a fair bit of code
which assumes that only XMM0 is used.
There are probably still issues on x86-64 I've missed.
I've added a few new tests to abitest, which fail (2x long long and 2x double
in a struct should be passed in registers).
abitest now passes; however test1-3 fail in init_test. All other tests
pass. I need to re-test Win32 and Linux-x86.
I've added a dummy implementation of gfunc_sret to c67-gen.c so it
should now compile, and I think it should behave as before I created
gfunc_sret.
Should fix some warnings wrt. access out of array bounds.
tccelf.c: fix "static function unused" warning
x86_64-gen.c: fix "ctype.ref uninitialzed" warning and cleanup
tcc-win32.txt: remove obsolete limitation notes.
Loads of VT_LLOCAL values (which effectively represent saved
addresses of lvalues) were done in VT_INT type, loosing the upper
32 bits. Needs to be done in VT_PTR type.
When offsetted addresses of global non-static data are computed
multiple times in the same statement the x86_64 backend uses
gen_gotpcrel with offset, which implements an add insn on the
register given. load() uses the R member of the to-be-loaded
value, which doesn't yet have a reg assigned in all cases.
So use the register we're supposed to load the value into as
that register.
The first loop setting up struct arguments must not remove
elements from the vstack (via vtop--), as gen_reg needs them to
potentially evict some argument still held in registers to stack.
Swapping the arg in question to top (and back to its place) also
simplifies the vstore call itself, as not funny save/restore
or some "non-existing" stack elements need to be done.
Generally for a stack a vop-- operation conceptually clobbers
that element, so further references to it aren't allowed anymore.
See also commit 9527c4949f
On x86_64 we need to extend the reg_classes array because load()
is called for (at least) R11 too, which was not part of reg_classes
previously.
- Fix a wrong calculation for size of struct
- Handle cases where struct size isn't multple of 8
- Recover vstack after memcpy for pushing struct
- Add a float parameter for struct_assign_test1 to check SSE alignment
This enables native unwind semantics with longjmp on
win64 by putting an entry into the .pdata section for
each compiled fuction.
Also, the function now use a fixed stack and store arguments
into X(%rsp) rather than using push.
- calling conventions are different:
* only 4 registers
* stack "scratch area" is always reserved
* doubles are mirrored in normal registers
- no GOT or PIC there
- Now we can run tcc -run tcc.c successfully, though there are some bugs.
- Remove jmp_table and got_table and use text_section for got and plt entries.
- Combine buffers in tcc_relocate().
- Use R_X86_64_64 instead of R_X86_64_32 for R_DATA_32 (now the name R_DATA_32 is inappropriate...).
- Add got_table in TCCState. This approach is naive and the distance between executable code and GOT can be longer than 32bit.
- Handle R_X86_64_GOTPCREL properly. We use got_table for TCC_OUTPUT_MEMORY case for now.
- Fix load() and store() so that they access global variables via GOT.
Most change was done in #ifdef TCC_TARGET_X86_64. So, nothing should be broken by this change.
Summary of current status of x86-64 support:
- produces x86-64 object files and executables.
- the x86-64 code generator is based on x86's.
-- for long long integers, we use 64bit registers instead of tcc's generic implementation.
-- for float or double, we use SSE. SSE registers are not utilized well (we only use xmm0 and xmm1).
-- for long double, we use x87 FPU.
- passes make test.
- passes ./libtcc_test.
- can compile tcc.c. The compiled tcc can compile tcc.c, too. (there should be some bugs since the binary size of tcc2 and tcc3 is differ where tcc tcc.c -o tcc2 and tcc2 tcc.c -o tcc3)
- can compile links browser. It seems working.
- not tested well. I tested this work only on my linux box with few programs.
- calling convention of long-double-integer or struct is not exactly the same as GCC's x86-64 ABI.
- implementation of tcc -run is naive (tcc -run tcctest.c works, but tcc -run tcc.c doesn't work). Relocating 64bit addresses seems to be not as simple as 32bit environments.
- shared object support isn't unimplemented
- no bounds checker support
- some builtin functions such as __divdi3 aren't supported
