[Bug c++/29139] New: GCC produces wrong code with SSE / inline functions
Output of gcc -v:
Using built-in specs.
Target: i686-pc-cygwin
Configured with: ../../source/gcc-4.1.1/configure --enable-threads=posix
--enable-languages=c,c++ --disable-win32-registry
Thread model: posix
gcc version 4.1.1
I wrote a Vector class that I wanted to optimize by using some inline assembly
with SSE instructions. I noticed that with certain optimizations on, GCC
version 4.1.1 produces wrong code. The source code consists of my Vector class
and some test code.
When I compile this with the following command line I get the error:
g++ -Wall -O2 -msse -finline-functions gccbug.cpp -o gccbug
Output of the program:
Start
FAILED: expected (1.5;3;4.5) but value is (1;3;4.5): bug in operator*
FAILED: expected (1.5;3;4.5) but value is (1;3;4.5): bug in operator*
FAILED: expected (0.5;1;1.5) but value is (1;1;1.5): bug in operator/
FAILED: expected (1.5;3;4.5) but value is (1;3;4.5): bug in operator*
FAILED: expected (1.5;3;4.5) but value is (1;3;4.5): bug in operator*
FAILED: expected (0.5;1;1.5) but value is (1;1;1.5): bug in operator/
Finished in 14338932 ticks
Ther are some computations with vectors, and when I compile it with these
settings, there are errors in the compilation. Note that the first element of
the result vector is 1 instead of 1.5 (for the multiplication) or 0.5 (for the
division).
When I compile this code with different options, for example:
1. Without "-msse": g++ -Wall -O2 -finline-functions gccbug.cpp -o gccbug
2. Without "-finline-functions": g++ -Wall -O2 -msse gccbug.cpp -o gccbug
Then the program works correctly!
Conclusion: GCC produces wrong code with this combination of optimization
options.
--
Summary: GCC produces wrong code with SSE / inline functions
Product: gcc
Version: 4.1.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: jespdj at hotmail dot com
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29139
[Bug c++/29139] GCC produces wrong code with SSE / inline functions
--- Comment #1 from jespdj at hotmail dot com 2006-09-19 11:33 --- Created an attachment (id=12294) --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=12294&action=view) Source code that demonstrates the bug This is the source code that demonstrates bug 29139. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29139
[Bug c++/29139] GCC produces wrong code with SSE / inline functions
--- Comment #2 from jespdj at hotmail dot com 2006-09-19 11:34 --- Created an attachment (id=12295) --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=12295&action=view) Precompiled source (*.ii) Precompiled code for gccbug.cpp of bug 29139. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29139
[Bug target/29139] GCC produces wrong code with SSE / inline functions
--- Comment #3 from jespdj at hotmail dot com 2006-09-22 06:51 --- I've been looking at my code some more time, and I noticed that there are bugs in it. For example, I'm not using the '&' early clobber modifier in the embedded asm, which is necessary if the output operand is modified before the input operands are used. I don't believe it's a bug in GCC after all. Sorry. Please don't use the code I added in the attachments, it has more issues besides what I mentioned above. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29139
[Bug target/29139] GCC produces wrong code with SSE / inline functions
--- Comment #4 from jespdj at hotmail dot com 2006-09-22 06:51 --- Marking this bug as invalid. -- jespdj at hotmail dot com changed: What|Removed |Added Status|UNCONFIRMED |RESOLVED Resolution||INVALID http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29139
[Bug c++/29249] New: Alignment error with static const variable in inline function
Look at the following source code:
#include
typedef int __attribute__ ((aligned (16))) aint;
inline void function() {
static const aint x = 123;
unsigned long a = (unsigned long) &x;
printf("Address of x: 0x%lX - %s\n", a, a & 15L ? "ALIGNMENT ERROR" :
"ok");
}
int main() {
function();
return 0;
}
The output on my system is:
Address of x: 0x40E3E8 - ALIGNMENT ERROR
The variable x is supposed to be aligned on a 16 byte boundary, but it is not.
Note that this only happens when the function is inline and the variable is
static. If I remove either "inline" or "static", the alignment seems to work
properly.
I am aware of bug #15795, but this is a different issue, since I'm not
dynamically allocating memory here.
Compiled like this:
g++ tst.cpp -o tst
Output of gcc -v:
Using built-in specs.
Target: i686-pc-cygwin
Configured with: ../../source/gcc-4.1.1/configure --enable-threads=posix
--enable-languages=c,c++ --disable-win32-registry
Thread model: posix
gcc version 4.1.1
I am using Cygwin on Windows 2000 Professional and Windows XP, and compiled GCC
4.1.1 from source using the configuration as specified above.
--
Summary: Alignment error with static const variable in inline
function
Product: gcc
Version: 4.1.1
Status: UNCONFIRMED
Severity: major
Priority: P3
Component: c++
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: jespdj at hotmail dot com
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29249
[Bug c++/29249] Alignment error with static const variable in inline function
--- Comment #1 from jespdj at hotmail dot com 2006-09-27 08:55 --- Created an attachment (id=12336) --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=12336&action=view) Source code that demonstrates the bug -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29249
[Bug c++/29249] Alignment error with static const variable in inline function
--- Comment #2 from jespdj at hotmail dot com 2006-09-27 08:55 --- Created an attachment (id=12337) --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=12337&action=view) Precompiled source (*.ii) -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29249
[Bug target/29249] Alignment error with static const variable in inline function
--- Comment #5 from jespdj at hotmail dot com 2006-09-27 21:11 --- Created an attachment (id=12342) --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=12342&action=view) output of g++ tst.cpp -S Here is the assembler source produced with g++ tst.cpp -S At the bottom of the file is this: .globl __ZZ8functionvE1x .section.rdata$_ZZ8functionvE1x,"dr" .linkonce same_size .align 16 __ZZ8functionvE1x: .long 123 .def_printf;.scl2; .type 32; .endef Indeed there is an ".align 16", but despite this, I get the alignment error. When I comment out "inline" on the function, the assembly source looks slightly different. I will attach the assembly file in a separate attachment. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29249
[Bug target/29249] Alignment error with static const variable in inline function
--- Comment #6 from jespdj at hotmail dot com 2006-09-27 21:12 ---
Created an attachment (id=12343)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=12343&action=view)
assembly source without "inline"
This is the assembly file I get when I compile without "inline", so the source
looks like:
#include
typedef int __attribute__ ((aligned (16))) aint;
// no inline
void function() {
static const aint x = 123;
unsigned long a = (unsigned long) &x;
printf("Address of x: 0x%lX - %s\n", a, a & 15L ? "ALIGNMENT ERROR" :
"ok");
}
int main() {
function();
return 0;
}
The bottom of the assembly file now looks like this:
popl%ebp
leal-4(%ecx), %esp
ret
.section .rdata,"dr"
.align 16
__ZZ8functionvE1x:
.long 123
.def_printf;.scl2; .type 32; .endef
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29249
[Bug target/29249] Alignment error with static const variable in inline function
--- Comment #8 from jespdj at hotmail dot com 2006-09-28 07:45 --- Ok, I filed a bug report for binutils about this. http://sourceware.org/bugzilla/show_bug.cgi?id=3276 -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29249
