Ticket #2087 (new bug)

Opened 11 years ago

Last modified 11 years ago

Register allocator and lexicals bug

Reported by: jonathan Owned by:
Priority: blocker Milestone:
Component: imcc Version: 3.2.0
Severity: high Keywords:
Cc: Language:
Patch status: new Platform:

Description (last modified by jonathan) (diff)


.sub ''
$P42 = new ['Integer']
.lex 'i', $P42
$P69 = new ['Integer']
.lex 'j', $P69
.lex 'foo', $P10
.lex 'bar', $P11
.lex 'baz', $P12
$P3 = new ['String']
.lex 'k', $P3

$P100 = new ['String']
$P100 = "oh no"
store_lex 'foo', $P100
$P101 = find_lex 'bar'
say $P101

Should be a null PMC access, but prints "oh no". It looks like

.lex 'foo', $P0

isn't considered an instruction, but just saying "this lexical goes in this register". If that's the only mention of $P0, then there's no "first instruction". It builds a du-chain, then based on that throws away $P0 from the "needs allocation" list because it confuses it with an unused register (e.g. it thinks it got optimized out, which is not the case at all here).

This patch makes the issue go away:

diff --git a/compilers/imcc/reg_alloc.c b/compilers/imcc/reg_alloc.c
index de0f3e0..300216d 100644
--- a/compilers/imcc/reg_alloc.c
+++ b/compilers/imcc/reg_alloc.c
@@ -391,6 +391,9 @@ reg_sort_f(ARGIN(const void *a), ARGIN(const void *b))
     const SymReg * const ra = *(const SymReg * const *)a;
     const SymReg * const rb = *(const SymReg * const *)b;
+    if (!ra->first_ins || !rb->first_ins)
+        return 0;
     if (ra->first_ins->index < rb->first_ins->index)
         return -1;
@@ -473,7 +476,7 @@ build_reglist(PARROT_INTERP, ARGMOD(IMC_Unit *unit))
     /* we might have unused symbols here, from optimizations */
     for (i = count = unused = 0; i < n_symbols; i++) {
-        if (!unit->reglist[i]->first_ins)
+        if (!unit->reglist[i]->first_ins && !(unit->reglist[i]->usage & U_LEXICAL))
         else if (i == count)
@@ -716,7 +719,7 @@ vanilla_reg_alloc(PARROT_INTERP, ARGMOD(IMC_Unit *unit))
         SymReg *r;
         for (r = hsh->data[i]; r; r = r->next) {
             /* TODO Ignore non-volatiles */
-            if (REG_NEEDS_ALLOC(r) && r->use_count)
+            if (REG_NEEDS_ALLOC(r))
                 r->color = -1;
@@ -732,7 +735,7 @@ vanilla_reg_alloc(PARROT_INTERP, ARGMOD(IMC_Unit *unit))
             for (r = hsh->data[i]; r; r = r->next) {
                 if (r->set != reg_set)
-                if (REG_NEEDS_ALLOC(r) && (r->color == -1) && r->use_count) {
+                if (REG_NEEDS_ALLOC(r) && (r->color == -1)) {
                     if (set_contains(avail, first_reg))
                         first_reg = first_avail(interp, unit, reg_set, NULL);

But it's kinda hacky. OTOH, so is the register allocator as a whole, and this is blocking me on NQP, so provided we can be sure we don't make the quicksort of symbols to allocate unstable I'd be OK with something like this going in. Will leave it until tomorrow to see if anyone has any better ideas.

Change History

Changed 11 years ago by jonathan

  • description modified (diff)

Changed 11 years ago by dukeleto

Is this still blocking you, Jonathan? If parrot fulltest and rakudo spectest pass with this patch, then we should apply it, barring horrible performance problems.

Note: See TracTickets for help on using tickets.