Bug 5264 – x86_64 changes for druntime 2

Created attachment 830 most of the changes to get druntime 2 to build with dmd -m64 I spent a couple hours today working on getting druntime 2 to build with the x86_64 dmd compiler. It's not 100% so far, but almost there. What's broken: 1) lifetime.d has 4 va_arg calls commented out. The compiler croaks with: src/core/stdc/stdarg.d(229): Error: static assert "not a valid argument type for va_arg" src/rt/lifetime.d(814): instantiated from here: va_arg!(ulong) 2) lifetime.d _d_arrayappendcT needs to be altered to work with varargs properly. The 64 bit support is fragile/hacky/whatever and I haven't pieced together how this code evolved between d1 and d2's runtimes. 3) lifetime.d _d_arrayshrinkfit() crashes the backend of the compiler: function _d_arrayshrinkfit cod2.c:1769: assert(stackpush == stackpushsave) // 0xfffffffc == 0 Normally I'd create 3 bugs since there's 3 different issues, but I don't think that'd be useful here.

Ok, for #2, I think the right answer is to change _d_arrayappendcT to this: extern (C) void[] _d_arrayappendcT(TypeInfo ti, Array *x, ...) { version(X86) { byte *argp = cast(byte*)(&ti + 2); return _d_arrayappendT(ti, x, argp[0..1]); } else version(X86_64) { va_list ap; va_start(ap, __va_argsave); byte[] argp; va_arg(ap, ti.next, argp); return _d_arrayappendT(ti, x, argp); } else static assert(false, "unknown version"); } Builds at least, but can't test it yet.

ok, debugging #3. There's actually a couple of different assert locations, though I suspect they're all related. If the fix for this reduction doesn't solve all of them, I'll re-reduce the next one. ========= module hrm; struct BlkInfo { size_t size; } extern (C) BlkInfo gc_qalloc(size_t sz, uint ba); void __setArrayAllocLength(ref BlkInfo info) { if (info.size) {} } extern (C) void foo() { BlkInfo info = gc_qalloc(0, 0); __setArrayAllocLength(info); } ====== $ ../dmd-trunk/src/dmd -c -m64 hrm.d Internal error: backend/cod1.c 2554 Building in 32 bit mode (without -m64) works just fine.

for #3, call cleanup code in cod1.c funccall(): 2936 if (tym1 == TYhfunc) 2937 { // Hidden parameter is popped off by the callee 2938 c = genadjesp(c, -4); 2939 stackpush -= 4; This code isn't aware of the 32 vs 64 bit calling convention differences yet?

I'm splitting #3 into a standalone bug report. dmd-1.x exhibits the same problem and it's unrelated to druntime at all. New bug 5275.

Looks like the cause of #1 is that only dmd 1 got one of the big vararg related change sets: http://www.dsource.org/projects/dmd/changeset/717 Walter's merging those into dmd2 soon.

Created attachment 835 remaining changes to pass semantic checks of druntime (actually not all, see also bug 5263 for intrinsic cleanup) This time I ran the druntime and phobos unittests on 32 bit. They pass. The 64 bit tests can't be run yet as there's still dmd codgen bugs while building druntime.

All of these changes have been checked in. This doesn't mean that it work, just that it builds now. :)