Implement debug strings

[vext01]

[This requires a compiler change and a llvm sync]

@vext01
Implement debug strings.
6fb181e

This allows the interpreter author to insert special comments into the
trace IR for helping with debugging.

In the interpreter source code you can add a call like
yk_debug_str(debug_string) and when it is traced, a special DebugStr
bytecode is inserted into the JIT IR, which when printed, displays the
dynamic value of debug_string at the time the instruction was traced.

Under the hood, this works a lot like promotion. Calls to yk_debug_str
(when tracing is enabled) causes debug_string to be copied into a
vector which is passed down the pipeline.

Here's an example for yklua:

--- a/src/lvm.c
+++ b/src/lvm.c
@@ -27,6 +27,7 @@
 #include "lgc.h"
 #include "lobject.h"
 #include "lopcodes.h"
+#include "lopnames.h"
 #include "lstate.h"
 #include "lstring.h"
 #include "ltable.h"
@@ -1234,6 +1235,10 @@ void luaV_execute (lua_State *L, CallInfo *ci) {
     lua_assert(base <= L->top.p && L->top.p <= L->stack_last.p);
     /* invalidate top for instructions not expecting it */
     lua_assert(isIT(i) || (cast_void(L->top.p = base), 1));
+    char dbgmsg[128];
+    snprintf(dbgmsg, 128, "%s", opnames[GET_OPCODE(i)]);
+    yk_debug_str(dbgmsg);
+
     vmdispatch (GET_OPCODE(i)) {
       vmcase(OP_MOVE) {
         StkId ra = RA(i);

Will give a trace with debug strings inside like:

    ...
    %408: ptr = ptr_add %361, 152
    %409: ptr = load %408
    %410: i32 = call @snprintf(%320, 128i64, %364, %409)
    ; debug_str: NEWTABLE
    %412: ptr = ptr_add %351, 592
    %413: i32 = shl 1i32, 4294967295i32
    ...

The strings are formatted as a comment.

(Obviously you'd probably want to guard the debug string logic in the
interpreter somehow so the overhead doesn't arise in production)

For this facility to be useful, the DebugStr bytecode must never be
optimised out. We tell the trace optimiser that this bytecode is an
"internal" bytecode, which stops DCE from killing it.

When adding the code to outline the contents of yk_debug_str() I
needed to copy logic from handle_promote() and found several other
copies of that same code in other places in the trace builder. Thic
change also de-duplicates the copies into a function
outline_until_after_call().

[vext01]

Force pushed the renaming and fixed conflicts.

[vext01]

That last push syncs with master and adds the missing outlining logic (and a test for it).

[vext01]

If you are OK with the deubgging logic, OK to squash this?

[ltratt]

Please squash.

[vext01]

That last force push syncs ykllvm and this is now ready to merge.

[vext01]

Force pushed to fix clippy warnings.