diff options
author | Aart Bik <ajcbik@google.com> | 2017-05-01 16:50:45 -0700 |
---|---|---|
committer | Aart Bik <ajcbik@google.com> | 2017-05-01 16:50:45 -0700 |
commit | 65ffd8ef044465c47d4f97ab2556310f9ee30a01 (patch) | |
tree | b520b0071f80d6c6f9c4cc649a079969ef3e0d20 /compiler/optimizing/loop_optimization.cc | |
parent | c009d14ce52bca0de34b441f8a575b0d36c1d69e (diff) |
Bug fix on shift that exceeds "lane width".
Rationale:
ARM is a bit less forgiving on shifting more than
the lane width of the SIMD instruction (rejecting
such cases is no loss, since it yields 0 anyway
and should be optimized differently).
Bug: 37776122
Test: test-art-target, test-art-host
Change-Id: I22d04afbfce82b4593f17c2f48c1fd5a0805d305
Diffstat (limited to 'compiler/optimizing/loop_optimization.cc')
-rw-r--r-- | compiler/optimizing/loop_optimization.cc | 25 |
1 files changed, 15 insertions, 10 deletions
diff --git a/compiler/optimizing/loop_optimization.cc b/compiler/optimizing/loop_optimization.cc index c783ddecf5..bbc55dd16f 100644 --- a/compiler/optimizing/loop_optimization.cc +++ b/compiler/optimizing/loop_optimization.cc @@ -833,17 +833,22 @@ bool HLoopOptimization::VectorizeUse(LoopNode* node, // TODO: accept symbolic, albeit loop invariant shift factors. HInstruction* opa = instruction->InputAt(0); HInstruction* opb = instruction->InputAt(1); - if (VectorizeUse(node, opa, generate_code, type, restrictions) && opb->IsIntConstant()) { - if (generate_code) { - // Make sure shift factor only looks at lower bits, as defined for sequential shifts. - // Note that even the narrower SIMD shifts do the right thing after that. - int32_t mask = (instruction->GetType() == Primitive::kPrimLong) - ? kMaxLongShiftDistance - : kMaxIntShiftDistance; - HInstruction* s = graph_->GetIntConstant(opb->AsIntConstant()->GetValue() & mask); - GenerateVecOp(instruction, vector_map_->Get(opa), s, type); + int64_t value = 0; + if (VectorizeUse(node, opa, generate_code, type, restrictions) && IsInt64AndGet(opb, &value)) { + // Make sure shift distance only looks at lower bits, as defined for sequential shifts. + int64_t mask = (instruction->GetType() == Primitive::kPrimLong) + ? kMaxLongShiftDistance + : kMaxIntShiftDistance; + int64_t distance = value & mask; + // Restrict shift distance to packed data type width. + int64_t max_distance = Primitive::ComponentSize(type) * 8; + if (0 <= distance && distance < max_distance) { + if (generate_code) { + HInstruction* s = graph_->GetIntConstant(distance); + GenerateVecOp(instruction, vector_map_->Get(opa), s, type); + } + return true; } - return true; } } else if (instruction->IsInvokeStaticOrDirect()) { // Accept particular intrinsics. |