MIPS: The new ABI should be allowed to use r8 and r9 for syscalls.

linux_syscall_support.h

Index: linux_syscall_support.h
===================================================================
--- linux_syscall_support.h	(revision 28)
+++ linux_syscall_support.h	(working copy)
@@ -2566,6 +2566,7 @@
     #define LSS_REG(r,a) register unsigned long __r##r __asm__("$"#r) =       \
                                  (unsigned long)(a)
     #undef  LSS_BODY
+    #if _MIPS_SIM == _MIPS_SIM_ABI32
     #define LSS_BODY(type,name,r7,...)                                        \
           register unsigned long __v0 __asm__("$2") = __NR_##name;            \
           __asm__ __volatile__ ("syscall\n"                                   \
@@ -2575,6 +2576,16 @@
                                   "$13", "$14", "$15", "$24", "$25",          \
                                   "memory");                                  \
           LSS_RETURN(type, __v0, __r7)
+    #else
+    #define LSS_BODY(type,name,r7,...)                                        \
+          register unsigned long __v0 __asm__("$2") = __NR_##name;            \
+          __asm__ __volatile__ ("syscall\n"                                   \
+                                : "+r"(__v0), r7 (__r7)                       \
+                                : "0"(__v0), ##__VA_ARGS__                    \
+                                : "$10", "$11", "$12", "$13", "$14",          \

[Mark Seaborn, 2014/07/12 23:58:08]

So, you're only changing the list of clobbered registers, to remove registers
that are used as input constraints if _MIPS_SIM != _MIPS_SIM_ABI32 below.

Would it make sense to conditionally #define a LSS_SYSCALL_CLOBBERS to reduce
the duplication here?

Does that mean that compilation was failing with the error "‘asm’ operand has
impossible constraints" from GCC?  If so, can you say that in the commit
message?

It looks like the N32/N64 support in this file wouldn't have worked before,
unless the compiler was more lax before.

Does your TEST= line mean you only compile-tested this, and didn't run the code?

For comparison, I noticed that glibc uses:

#define __SYSCALL_CLOBBERS "$1", "$3", "$10", "$11", "$12", "$13", \
        "$14", "$15", "$24", "$25", "hi", "lo", "memory"

(from ports/sysdeps/unix/sysv/linux/mips/mips64/n32/sysdep.h)

Does that mean this code is missing clobbers for the "hi" and "lo" registers?

See:
https://sourceware.org/ml/libc-ports/2009-09/msg00006.html ("Avoid reuse of MIPS
"hi" and "lo"")
http://www.linux-mips.org/wiki/Syscall

[Lei Zhang, 2014/07/14 19:39:28]

Disclaimer: I've never worked on this before. So it's your call.
Done. It was "error: asm-specifier for variable '__r8' conflicts with asm
clobber list"
No idea if anyone's ever tested this before.
There's other issues in the Breakpad build, and I don't have a MIPS VM set up. I
can if you'd like.
I defer to petarj.

[petarj, 2014/07/16 02:23:24]

breakpad does not support mips64 yet.
True. Mimicking __SYSCALL_CLOBBERS from
./sysdeps/unix/sysv/linux/mips/mips64/n64/sysdep.h
is the way to go, I agree with Mark.

[Lei Zhang, 2014/07/16 23:42:07]

Done.

+                                  "$15", "$24", "$25", "memory");             \
+          LSS_RETURN(type, __v0, __r7)
+    #endif
     #undef _syscall0
     #define _syscall0(type, name)                                             \
       type LSS_NAME(name)(void) {                                             \