Opened 15 years ago

Closed 15 years ago

Last modified 17 months ago

#2069 closed defect (fixed)

MT32Emu does not compile with gcc4/fc4

Reported by: SF/awjb Owned by: eriktorbjorn
Priority: high Component: Audio: MT32
Keywords: Cc:


g++ -Wp,-MMD,"backends/midi/mt32/.deps/i386.d",-MQ,"backends/midi/mt32/i386.o",-MP -O2 -g -pipe -Wp,-D_FORTIFY_SOURCE=2 -fexceptions -m32 -march=i386 -mtune=pentium4 -fasynchronous-unwind-tables -DHAVE_CONFIG_H -DUNIX -DDISABLE_SAGA -I. -I. -I./common -I./backends/midi/mt32 -I/usr/include/SDL -D_REENTRANT -Itest/cxxtest -c backends/midi/mt32/i386.cpp -o backends/midi/mt32/i386.o backends/midi/mt32/i386.cpp: In function 'float MT32Emu::iir_filter_sse(float, float*, float*, int)': backends/midi/mt32/i386.cpp:142: error: unknown register name 'xmm3' in 'asm' backends/midi/mt32/i386.cpp:142: error: unknown register name 'xmm2' in 'asm' backends/midi/mt32/i386.cpp:142: error: unknown register name 'xmm1' in 'asm'

Ticket imported from: #1224138. Ticket imported from: bugs/2069.

Change History (8)

comment:1 by fingolfin, 15 years ago

Owner: set to kingguppy
Priority: normalhigh

comment:2 by fingolfin, 15 years ago

This seems related to bug #1221549 and apparently is due to a problem with the inline assembler?

comment:3 by SF/awjb, 15 years ago

from my pov it is... I don't know asm well enough to actually deal with the problem but thats what it looks like... seems to be changes in gcc4 as other software has asm issues with gcc4 as well (but get fixed so it is not a gcc4 error).

comment:4 by eriktorbjorn, 15 years ago

I don't know much about assembly language either, but from what I understand it's complaining about the list of "clobbered" registers, which I assume are the ones that have their values changed as a result of running that piece of code.

This file takes advantage of SSE and MMX instructions, if the CPU it's running on has support for it. But apparently GCC 4 does not recognize the SSE and MMX registers unless told to.

If I add -msse to the command line when compiling this particular file, I no longer get the compiler errors described in the bug report. If I add -mno-mmx I get a set of new errors, so I guess we should also add -mmmx. From what I understand, these flags only tells GCC to recognize SSE and MMX, not to generate such instructions on its own. But I could be wrong about that.

Another way might be to change the code to make sure that the registers aren't clobbered. That should just be a matter of pushing them on the stack at the beginning and pulling them off the stack at the end, wouldn't it? Or is MMX/SSE too magic for such tricks?

comment:5 by eriktorbjorn, 15 years ago

Here's a post about how Red Hat handled a similar (I think) problem in The GIMP:

I.e. surround the clobber lists with either #ifdef __SSE__ ... #endif or #ifdef __MMX__ ... #endif. I guess the reasoning is that if the compiler isn't going to generate SSE or MMX instructions on its own, it doesn't need to know whether or not these registers are clobbered. Debian appears to use the same solution.

I don't have GCC 4 on this computer, but I'll look at it later.

comment:6 by eriktorbjorn, 15 years ago

I believe it's fixed in CVS now, but since I can't actually run the MT-32 emulator (my computer is too slow, and I don't have the necessary data files), I haven't tested it other than by trying to compile it.

comment:7 by eriktorbjorn, 15 years ago

Owner: changed from kingguppy to eriktorbjorn
Resolution: fixed
Status: newclosed

comment:8 by digitall, 17 months ago

Component: Audio: MT32
Note: See TracTickets for help on using tickets.