6912064: type profiles need to be exploited more for dynamic language support

Reviewed-by: kvn
This commit is contained in:
John R Rose 2010-08-20 23:40:30 -07:00
parent a54b1ff70e
commit c7e50e8305
8 changed files with 204 additions and 64 deletions

View file

@ -892,6 +892,62 @@ bool Parse::seems_never_taken(float prob) {
return prob < PROB_MIN;
}
// True if the comparison seems to be the kind that will not change its
// statistics from true to false. See comments in adjust_map_after_if.
// This question is only asked along paths which are already
// classifed as untaken (by seems_never_taken), so really,
// if a path is never taken, its controlling comparison is
// already acting in a stable fashion. If the comparison
// seems stable, we will put an expensive uncommon trap
// on the untaken path. To be conservative, and to allow
// partially executed counted loops to be compiled fully,
// we will plant uncommon traps only after pointer comparisons.
bool Parse::seems_stable_comparison(BoolTest::mask btest, Node* cmp) {
for (int depth = 4; depth > 0; depth--) {
// The following switch can find CmpP here over half the time for
// dynamic language code rich with type tests.
// Code using counted loops or array manipulations (typical
// of benchmarks) will have many (>80%) CmpI instructions.
switch (cmp->Opcode()) {
case Op_CmpP:
// A never-taken null check looks like CmpP/BoolTest::eq.
// These certainly should be closed off as uncommon traps.
if (btest == BoolTest::eq)
return true;
// A never-failed type check looks like CmpP/BoolTest::ne.
// Let's put traps on those, too, so that we don't have to compile
// unused paths with indeterminate dynamic type information.
if (ProfileDynamicTypes)
return true;
return false;
case Op_CmpI:
// A small minority (< 10%) of CmpP are masked as CmpI,
// as if by boolean conversion ((p == q? 1: 0) != 0).
// Detect that here, even if it hasn't optimized away yet.
// Specifically, this covers the 'instanceof' operator.
if (btest == BoolTest::ne || btest == BoolTest::eq) {
if (_gvn.type(cmp->in(2))->singleton() &&
cmp->in(1)->is_Phi()) {
PhiNode* phi = cmp->in(1)->as_Phi();
int true_path = phi->is_diamond_phi();
if (true_path > 0 &&
_gvn.type(phi->in(1))->singleton() &&
_gvn.type(phi->in(2))->singleton()) {
// phi->region->if_proj->ifnode->bool->cmp
BoolNode* bol = phi->in(0)->in(1)->in(0)->in(1)->as_Bool();
btest = bol->_test._test;
cmp = bol->in(1);
continue;
}
}
}
return false;
}
}
return false;
}
//-------------------------------repush_if_args--------------------------------
// Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
inline int Parse::repush_if_args() {
@ -1137,19 +1193,22 @@ void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob,
bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
int cop = c->Opcode();
if (seems_never_taken(prob) && cop == Op_CmpP && btest == BoolTest::eq) {
// (An earlier version of do_if omitted '&& btest == BoolTest::eq'.)
//
if (seems_never_taken(prob) && seems_stable_comparison(btest, c)) {
// If this might possibly turn into an implicit null check,
// and the null has never yet been seen, we need to generate
// an uncommon trap, so as to recompile instead of suffering
// with very slow branches. (We'll get the slow branches if
// the program ever changes phase and starts seeing nulls here.)
//
// The tests we worry about are of the form (p == null).
// We do not simply inspect for a null constant, since a node may
// We do not inspect for a null constant, since a node may
// optimize to 'null' later on.
//
// Null checks, and other tests which expect inequality,
// show btest == BoolTest::eq along the non-taken branch.
// On the other hand, type tests, must-be-null tests,
// and other tests which expect pointer equality,
// show btest == BoolTest::ne along the non-taken branch.
// We prune both types of branches if they look unused.
repush_if_args();
// We need to mark this branch as taken so that if we recompile we will
// see that it is possible. In the tiered system the interpreter doesn't