mirror of git://git.psyced.org/git/psyclpc
82 lines
3.7 KiB
Plaintext
82 lines
3.7 KiB
Plaintext
gcc version 2.9-aix51-020209:
|
|
make_func.y::read_func_spec() a too large fprintf() statement
|
|
causes the compiler to generated faulty code on the assignment which
|
|
causes the mkfunc program to crash. Splitting the print statement
|
|
into two solved the problem.
|
|
|
|
Visual Age C 6.0, AIX 5.2:
|
|
make_func.y::read_func_spec() the particularly sized fprintf() statement
|
|
causes the compiler to generated faulty code on the following assignment,
|
|
which in turn causes the mkfunc program to generate faulty data. Splitting
|
|
the print statement into two solved the problem.
|
|
|
|
gcc 2.9, 3.2 - Cygwin
|
|
Highest optimization creates buggy code when the total size of local
|
|
variables gets too big (>= 8KBytes).
|
|
|
|
gcc 3.0
|
|
Optimization option -funroll-all-loops seems to create buggy code.
|
|
|
|
gcc 2.91.66 19990314/Linux (egcs-1.1.2) on Linux/x86:
|
|
Computing MIN_LONG % -1 causes a floating point exception.
|
|
|
|
xlc on AIX 3.4:
|
|
Optimization level -O4 -qipa crashes in the code generation if a void
|
|
returning function is casted to a value returning one.
|
|
|
|
egcs 2.91.66 on Linux 2.2.15 on Sun SparcStation4:
|
|
The optimizer neglects to reload CPU registers after a setjmp() in
|
|
complex functions (namely eval_instruction() in interpret.c). The
|
|
workaround is to put the setjmp()-recover branch into a separate function
|
|
and making sure that it can't be inlined (e.g. by calling it indirectly).
|
|
|
|
gcc 2.8.1 on Solaris 5.7; egcs 2.91.16 on SunOS 4.1.3_U1, and probably
|
|
others:
|
|
Compiling with optimization causes the compiler (not the assembler!) to
|
|
fall into catatonia for the configuration test for a float/branch
|
|
incompatibility. Workaround is to make sure that configure compiles this
|
|
test without optimization.
|
|
|
|
gcc 2.9-beos-980929 on BeOS/x86:
|
|
Compiling without optimization causes the compiler to crash on simulate.c
|
|
|
|
gcc 2.7.2.1-i486-linux on Linux 2.0.33:
|
|
Optimizing compiles may pessimize the expression
|
|
(svalue.type == x && svalue.x.closure_type == y)
|
|
to 'comparison is always zero' if both fields share the same longword
|
|
in memory (closure.c is one victim).
|
|
Makeshift solution: Assigning one of the fields to a local variable
|
|
causes the expression to produce the correct result.
|
|
|
|
gcc 1.36 on Next / NextStep 2.1 with flags '-g' / '-g -fomit_frame_pointer' :
|
|
The calculated offset to a6 for parameter access in free_svalue() is wrong.
|
|
Symptom: mud crashes early.
|
|
Makeshift solution: use '-O -fomit-frame-pointer'
|
|
|
|
gcc 2.2.2 , when told to optimize interpret.c , usually dies with a segv
|
|
or currupted internal data structures in cc1.
|
|
|
|
gcc 2.3.3 , when told to optimize lex.c , will garble the code for #include.
|
|
Symptom: load_object will fail when #include is used, saying it can't
|
|
#include the file.
|
|
|
|
gcc 2.x on 680x0 hardware: the large interpret.c switch table is built with
|
|
unsigned short words, but is processed like signed short words.
|
|
The special thing about 680x0 is that the switch in interpret.c is
|
|
shorter than 64 KB.
|
|
Symptoms: when LPC instructions are to be executed the code to which
|
|
lies beyond the first 32 KB of the switch(), there is weird behaviour,
|
|
a C-level illegal intruction or a crash with undetermined, although most
|
|
times quite short, latency.
|
|
Makeshift solution: fix the assembler code in interpret.s to handle
|
|
unsigned short offsets.
|
|
|
|
NeXTStep 3.2 has a broken memcmp(), that will give different return values
|
|
depending on alignment.
|
|
The library that contained this broken memcmp() looks like this:
|
|
-r-xr-xr-x 1 root 1012100 Sep 1 1993 libsys_s.B.shlib*
|
|
quickfix: #define memcmp(a,b,l) (*(int*)(a) - *(int*)(b)) in closure.c and
|
|
prolang.y, or a file included by these.
|
|
|
|
ULTRIX dec1 4.3 0 RISC needs <sys/types.h> to be included before <sys/stat.h>
|