Mercurial > hg > graal-jvmci-8
diff src/share/vm/opto/loopTransform.cpp @ 2412:f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
Summary: In Ideal() method of String intrinsics nodes look for TypeAryPtr::CHARS memory slice if memory is MergeMem. Do not unroll a loop with String intrinsics code.
Reviewed-by: never
author | kvn |
---|---|
date | Wed, 30 Mar 2011 12:08:49 -0700 |
parents | 1927db75dd85 |
children | cb162b348743 |
line wrap: on
line diff
--- a/src/share/vm/opto/loopTransform.cpp Wed Mar 30 07:47:19 2011 -0700 +++ b/src/share/vm/opto/loopTransform.cpp Wed Mar 30 12:08:49 2011 -0700 @@ -396,16 +396,16 @@ // Return exact loop trip count, or 0 if not maximally unrolling bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { CountedLoopNode *cl = _head->as_CountedLoop(); - assert( cl->is_normal_loop(), "" ); + assert(cl->is_normal_loop(), ""); Node *init_n = cl->init_trip(); Node *limit_n = cl->limit(); // Non-constant bounds - if( init_n == NULL || !init_n->is_Con() || + if (init_n == NULL || !init_n->is_Con() || limit_n == NULL || !limit_n->is_Con() || // protect against stride not being a constant - !cl->stride_is_con() ) { + !cl->stride_is_con()) { return false; } int init = init_n->get_int(); @@ -428,7 +428,25 @@ uint unroll_limit = (uint)LoopUnrollLimit * 4; assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); cl->set_trip_count(trip_count); - if( trip_count <= unroll_limit && body_size <= unroll_limit ) { + if (trip_count > unroll_limit || body_size > unroll_limit) { + return false; + } + + // Do not unroll a loop with String intrinsics code. + // String intrinsics are large and have loops. + for (uint k = 0; k < _body.size(); k++) { + Node* n = _body.at(k); + switch (n->Opcode()) { + case Op_StrComp: + case Op_StrEquals: + case Op_StrIndexOf: + case Op_AryEq: { + return false; + } + } // switch + } + + if (body_size <= unroll_limit) { uint new_body_size = body_size * trip_count; if (new_body_size <= unroll_limit && body_size == new_body_size / trip_count && @@ -448,13 +466,13 @@ bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const { CountedLoopNode *cl = _head->as_CountedLoop(); - assert( cl->is_normal_loop() || cl->is_main_loop(), "" ); + assert(cl->is_normal_loop() || cl->is_main_loop(), ""); // protect against stride not being a constant - if( !cl->stride_is_con() ) return false; + if (!cl->stride_is_con()) return false; // protect against over-unrolling - if( cl->trip_count() <= 1 ) return false; + if (cl->trip_count() <= 1) return false; int future_unroll_ct = cl->unrolled_count() * 2; @@ -485,21 +503,21 @@ // Non-constant bounds. // Protect against over-unrolling when init or/and limit are not constant // (so that trip_count's init value is maxint) but iv range is known. - if( init_n == NULL || !init_n->is_Con() || - limit_n == NULL || !limit_n->is_Con() ) { + if (init_n == NULL || !init_n->is_Con() || + limit_n == NULL || !limit_n->is_Con()) { Node* phi = cl->phi(); - if( phi != NULL ) { + if (phi != NULL) { assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); int next_stride = cl->stride_con() * 2; // stride after this unroll - if( next_stride > 0 ) { - if( iv_type->_lo + next_stride <= iv_type->_lo || // overflow - iv_type->_lo + next_stride > iv_type->_hi ) { + if (next_stride > 0) { + if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow + iv_type->_lo + next_stride > iv_type->_hi) { return false; // over-unrolling } - } else if( next_stride < 0 ) { - if( iv_type->_hi + next_stride >= iv_type->_hi || // overflow - iv_type->_hi + next_stride < iv_type->_lo ) { + } else if (next_stride < 0) { + if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow + iv_type->_hi + next_stride < iv_type->_lo) { return false; // over-unrolling } } @@ -511,24 +529,33 @@ // Key test to unroll CaffeineMark's Logic test int xors_in_loop = 0; // Also count ModL, DivL and MulL which expand mightly - for( uint k = 0; k < _body.size(); k++ ) { - switch( _body.at(k)->Opcode() ) { - case Op_XorI: xors_in_loop++; break; // CaffeineMark's Logic test - case Op_ModL: body_size += 30; break; - case Op_DivL: body_size += 30; break; - case Op_MulL: body_size += 10; break; - } + for (uint k = 0; k < _body.size(); k++) { + Node* n = _body.at(k); + switch (n->Opcode()) { + case Op_XorI: xors_in_loop++; break; // CaffeineMark's Logic test + case Op_ModL: body_size += 30; break; + case Op_DivL: body_size += 30; break; + case Op_MulL: body_size += 10; break; + case Op_StrComp: + case Op_StrEquals: + case Op_StrIndexOf: + case Op_AryEq: { + // Do not unroll a loop with String intrinsics code. + // String intrinsics are large and have loops. + return false; + } + } // switch } // Check for being too big - if( body_size > (uint)LoopUnrollLimit ) { - if( xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; + if (body_size > (uint)LoopUnrollLimit) { + if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; // Normal case: loop too big return false; } // Check for stride being a small enough constant - if( abs(cl->stride_con()) > (1<<3) ) return false; + if (abs(cl->stride_con()) > (1<<3)) return false; // Unroll once! (Each trip will soon do double iterations) return true;