This fixes the Alpha target for PowerPC hosts. It is merely a part of the original gcc4 patch posted by Michael Matz.
--- qemu-0.9.0.cvs/target-alpha/op_template.h.mm 2007-08-22 03:17:57.000000000 +0000 +++ qemu-0.9.0.cvs/target-alpha/op_template.h 2007-08-22 03:15:49.000000000 +0000 @@ -28,7 +28,26 @@ void OPPROTO glue(op_reset_T, REG) (void void OPPROTO glue(op_reset_FT, REG) (void) { +#ifdef HOST_PPC + /* We have a problem with HOST_PPC here: + We want this code: + glue(FT, REG) = 0; + unfortunately GCC4 notices that this stores (double)0.0 into + env->ft0 and emits that constant into the .rodata, and instructions + to load that zero from there. But that construct can't be parsed by dyngen. + We could add -ffast-math for compiling op.c, that would just make it generate + two stores of zeros into both words of ft0. But -ffast-math may have other + side-effects regarding the emulation. We could use __builtin_memset, + which perhaps would be the sanest. That relies on -O2 and our other options + to inline that memset, which currently it does, but who knows for how long. + So, we simply do that by hand, and a barely typesafe way :-/ */ + union baeh { double d; unsigned int i[2];}; + union baeh *p = (union baeh*)&(glue(FT, REG)); + p->i[0] = 0; + p->i[1] = 0; +#else glue(FT, REG) = 0; +#endif RETURN(); }