Register aliasing in load_traceinput raises an exception

[ltratt]

For boring reasons I changed yklua's vmfetch thus:

diff --git src/lvm.c src/lvm.c
index f796a756..aa351a0b 100644
--- src/lvm.c
+++ src/lvm.c
@@ -1139,7 +1139,7 @@ void luaV_finishOp (lua_State *L) {
     trap = luaG_traceexec(L, pc);  /* handle hooks */ \
     updatebase(ci);  /* correct stack */ \
   } \
-  i = (Instruction) yk_promote(*(pc++)); \
+  i = (Instruction) yk_promote((uintptr_t) *(pc++)); \
 }
 #else
 #  define vmfetch()	{ \

This then causes the resulting trace to alias registers: one register (I think) has a the uintptr_t value in it (8 bytes) and the other the Instruction value (4 bytes). The start of the resulting trace thus looks as follows:

    %45: i64 = load_ti Register(0, 8, 0, [])
    %46: i32 = load_ti Register(0, 4, 0, [])
    %47: i32 = load_ti Register(3, 4, 0, [])
    %48: ptr = load_ti Register(4, 8, 0, [])
    %49: ptr = load_ti Register(5, 8, 0, [])

Notice that %45 and %46 both reference register 0. This then causes our register allocator to explode as it expects a register to reference a single value.

I'm not sure that fixing this is very high priority, but it does show that LLVM can generate a style of IR that we haven't considered before. It might be worth keeping this in the back of our mind.

[ltratt]

We have now seen this happen again, and happen more often.