I generated a gcc crosscompiler for PPC 7400 using a sparc Solaris 2.8 system as the host. For the record, in order to compile the all tool chain I used Dan kegel "crosstool" script (crosstool-0.28-rc25). This script applies some patches to GCC 3.4.0. in particular it applies a patch to gcc/config/rs6000/rs6000.h where it adds (amon other things) an unconditional -many option to be passed to the assembler. By browsing the GCC CVS it appears that this patch is now part of mainline. I selected the following components to build my crosstool chain: GCC 3.4.O BINUTILS 2.14 GLIBC 2.3.2 My crosstool chain was configured with the following "configure" command: /export/home/jdubois/crosstool-0.28-rc25/build/powerpc-7400-linux-gnu/gcc-3.4.0-glibc-2.3.2/gcc-3.4.0/configure --target=powerpc-7400-linux-gnu --host=sparc-host_sun-solaris2.8 --prefix=/vobs/target_tools/sparc-unknown-solaris-gnu/powerpc-7400-linux-gnu/gcc-3.4.0-glibc-2.3.2 --with-cpu=7400 --enable-altivec --with-headers=/vobs/target_tools/sparc-unknown-solaris-gnu/powerpc-7400-linux-gnu/gcc-3.4.0-glibc-2.3.2/powerpc-7400-linux-gnu/include --with-local-prefix=/vobs/target_tools/sparc-unknown-solaris-gnu/powerpc-7400-linux-gnu/gcc-3.4.0-glibc-2.3.2/powerpc-7400-linux-gnu --disable-nls --enable-threads=posix --enable-symvers=gnu --enable-__cxa_atexit --enable-languages=c,c++ --enable-shared --enable-c99 --enable-long-long Thread model: posix gcc version 3.4.0 When compiling the Linux kernel (2.6.2) with this crosstool chain from Solaris for PPC 7400 target, the compilation will fail on arch/ppc/kernel/time.c. This is because the get_tbl() and get_tbu() functions (defined in include/asm-ppc/time.h) are PPC specific. They are defiend as follow: extern __inline__ unsigned long get_tbl(void) { unsigned long tbl; #if defined(CONFIG_403GCX) asm volatile("mfspr %0, 0x3dd" : "=r" (tbl)); #else asm volatile("mftb %0" : "=r" (tbl)); #endif return tbl; } extern __inline__ unsigned long get_tbu(void) { unsigned long tbl; #if defined(CONFIG_403GCX) asm volatile("mfspr %0, 0x3dc" : "=r" (tbl)); #else asm volatile("mftbu %0" : "=r" (tbl)); #endif return tbl; } These function are "translated" in the .i file as follow: extern __inline__ __attribute__((always_inline)) unsigned long get_tbl(void) { unsigned long tbl; asm volatile("mftb %0" : "=r" (tbl)); return tbl; } extern __inline__ __attribute__((always_inline)) unsigned long get_tbu(void) { unsigned long tbl; asm volatile("mftbu %0" : "=r" (tbl)); return tbl; } And then the .s file generated from this will contain some fragments looking as follow: #APP mftb 0 #NO_APP .... #APP mftbu 0 mftb 6 mftbu 11 #NO_APP So it looks like the "mftbX" opcode are not valid on PPC403. When compiling the Linux kernel for PPC7400, the "mftbX" version of the functions are selected and the resulting assembly code (generated by cc1) is passed to the assembler with the following options: "-mppc -maltivec -many". The 2 first options are OK but the last one seems to prevent the assembler to interpret the "mftbX" opcode and the assembler will fail with some messages similar to this: time.s: Assembler messages: time.s:158: Error: Unrecognized opcode: `mftb' time.s:376: Error: Unrecognized opcode: `mftb' time.s:500: Error: Unrecognized opcode: `mftb' time.s:685: Error: Unrecognized opcode: `mftb' time.s:802: Error: Unrecognized opcode: `mftb' time.s:1016: Error: Unrecognized opcode: `mftbu' time.s:1017: Error: Unrecognized opcode: `mftb' time.s:1018: Error: Unrecognized opcode: `mftbu' I regenerated the GCC 3.4.0 tool chain without the -many and the compiler was then able to compile the time.c file. So while the official GCC 3.4 tool chain should not contain this "fix" it seems to be quite an official one anyway as it is now part of mainline (3.5?). However it seems to me it causes some trouble in the Linux kernel compilation for PPC when the C code select itself the assembly code to use for a function. Therefore, I'd like to ask if this unconditional "-many" is really required. If it is, is there some kind of incompatibilty/misbehavior in the 2.14 binutils I selscted for the task? I don't see a way to attach the .c, .i or .s to this web form. So please let me know where/how to send it and I'll do it.
Created attachment 6646 [details] files that fail to compile on GCC 3.4 + -many patch These are the source code (with genrated .i and .s files) that show the bug. They are From Linux 2.6.2 source tree.
Not a gcc bug, there was a bug in binutils before 2.15 which causes gas to change what was accepted for -many.
(In reply to comment #2) > Not a gcc bug, there was a bug in binutils before 2.15 which causes gas to change what was accepted > for -many. Then, I think gcc 3.4 should refuse to build against a binutils version lesser than 2.15. So there is still a bug and configure should check for this. Also, I would gladly use binutils 2.15 but the binaries generated on my Solaris 2.8 system for the cross tools are crashing on me. So I had to downgrade to 2.14 to get something working. OK, this is my problem and not a gcc bug but still... Also, please, note also that removing the -many from the argument passed to gas makes the all thing work correctly even with 2.14. So what is the rational for adding this flag when it seems to cause trouble with all but one version of binutils.
See http://gcc.gnu.org/ml/gcc-patches/2004-05/msg01244.html
(In reply to comment #4) > See http://gcc.gnu.org/ml/gcc-patches/2004-05/msg01244.html This is the patch that is applied (automatically) to GCC 3.4.0 in Dan kegel's crosstool and I am fine with all of it except the last unconditional "-many" that cause trouble to (at least) gas 2.14. BTW the "comment" in the applied patch file is as follow: <quote> (I'm not so sure about the unconditional -many it sends to binutils; that seems redundant?) </quote> It seems that gas 2.14 will reduce the number of valid instruction opcode to a "common" subset when -many is passed to it even if -mppc is passed prior to it. Therefore gas fails to interpret the mftbX opcodes. If the answer is: use 2.15 because we know 2.14 has a bug, and 2.15 handles the "-mppc -maltivec -many" correctly, I am fine with this (even if I will now have to find out why 2.15 crashes on me :( ). But then we should try to check this condition in the configure script and prevent gcc 3.4.0 (and later) to build against 2.14 or prior.
(In reply to comment #5) This is just to give a status on my progress in case it could help somebody else. I upgraded my tool selection (for crosstool.sh) to: gcc 3.4.1 binutils 2.15 glibc 2.3.2 In order to get binutils to run on Solaris 2.8 without crashing, I had to apply the patch described on the following link (now part of CVS): http://sources.redhat.com/ml/binutils/2004-06/msg00114.html Once done I was able to cross compile a Linux 2.6.7 kernel for PPC on my Solaris 2.8 host (with few additional patches to the Linux source tree). It still seems to me that binutils 2.14 should not be a valid choice (at configure time) for gcc 3.4.0 and above.