C2

Inlining

• Predicate: allow_inline parameter to Compile::call_generator()

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/b3d08045220f/src/share/vm/opto/doCall.cpp#l65

Inlining is controlled by the product flag Inline. There are many other inlining related flags which control inlining depth, size and other heuristics.

---

Incremental Inlining

• Predicate: global flag IncrementalInline and Compile::_inlining_incrementally which is set during compilation

If the inlining budgets are exhausted methods are queue to be inlined later if possible, like here:

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/b3d08045220f/src/share/vm/opto/bytecodeInfo.cpp#l328

---

Escape Analysis

• Predicate: do_escape_analysis parameter to Compile::Compile()

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/b3d08045220f/src/share/vm/opto/compile.cpp#l610

Compile::_do_escape_analysis is required in many places in the compiler.

It depends on DoEscapeAnalysis flag, JVMTI enabled events and recompilation. It is set in C2Compiler::compile_method().

Additional conditions are checked in ConnectionGraph::has_candidates() - presence of allocations, locks, boxing methods (valueOf).

---

Autobox Elimination

• Predicate: eliminate_boxing parameter to Compile::Compile()

which is set by the global flag EliminateAutoBox in C2Compiler::compile_method:

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/b3d08045220f/src/share/vm/opto/c2compiler.cpp#l103

Compile::_eliminate_boxing is required from many places in the compiler. Compile::aggressive_unboxing() is also controlled by the global flag AggressiveUnboxing.

---

Loop unrolling

• Predicate: IdealLoopTree::policy_unroll() which is required here:

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/b3d08045220f/src/share/vm/opto/loopTransform.cpp#l2333

Checks number of unrolls and predicted size of loop after unroll and other restrictions.

• Transformation: PhaseIdealLoop::do_unroll()

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/b3d08045220f/src/share/vm/opto/loopTransform.cpp#l1195

Does one unroll per PhaseIdealLoop invocation. Unroll only main loop after a counted loop is split into pre-main-post.

---

C1

Inlining

• predicate: Inline (product)
• trace: PrintInlining (diagnostic)

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/a184ee1d7172/src/share/vm/c1/c1_GraphBuilder.cpp#l3300

• Happens during parsing
• Depth limited (MaxInlineLevel, MaxRecursiveInlineLevel), bytecode size limited (MaxInlineSize, MaxTrivialSize, NestedInliningSizeRatio) inlining of static calls, and monomorphic virtual/interface calls (either final method/class or discovered via class hierarchy analysis).
• Inlining of method handle intrinsics
• Inlining of intrinsics

Canonicalizer

• predicate: CanonicalizeNodes (develop)
• trace: PrintCanonicalization (develop)

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/a184ee1d7172/src/share/vm/c1/c1_Canonicalizer.hpp#l53

• Applied during parsing
• IR instruction specific simplification

Local Value Numbering

• predicate: UseLocalValueNumbering (develop)
• trace: PrintValueNumbering (develop)

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/a184ee1d7172/src/share/vm/c1/c1_GraphBuilder.cpp#l2215

• Applied during parsing
• Remove redundant IR instructions using value numbering in current block

Conditional Expressions Eliminations

• predicate: DoCEE (develop), disabled when C1 does profiling of branches
• trace: PrintCEE (develop)

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/a184ee1d7172/src/share/vm/c1/c1_Optimizer.cpp#l305

• Independent pass
• Convert if/then/else control flow to single IR instruction of the form x {cond} y ? tval : fval to then take advantage of conditional moves

Basic Block Eliminations

• predicate: EliminateBlocks (develop), disabled when C1 does profiling of branches
• trace: PrintBlockElimination (develop)

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/a184ee1d7172/src/share/vm/c1/c1_Optimizer.cpp#l470

• Independent pass
• Merge block with single successor

Null Check Elimination

• predicate: EliminateNullChecks (develop)
• trace: PrintNullCheckElimination (develop)

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/a184ee1d7172/src/share/vm/c1/c1_Optimizer.cpp#l1152

• Independent pass
• Eliminate null checks that can be proved redundant by data flow analysis (iterate until fix point)

Global Value Numbering

• predicate: UseGlobalValueNumbering (develop)
• trace: PrintValueNumbering (develop)

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/a184ee1d7172/src/share/vm/c1/c1_ValueMap.cpp#l487

• Independent pass
• Remove redundant IR instructions using dominator based value numbering

Loop Invariant Code Motion

• predicate: UseLoopInvariantCodeMotion (product), disabled if tiered is on or a previous compilation of this method saw a deoptimization due to this optimization or bound check elimination
• trace: PrintValueNumbering (develop)

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/a184ee1d7172/src/share/vm/c1/c1_ValueMap.cpp#l330

• Independent pass
• Move some IR instructions out of loops if their operands are loop invariant. If the IR instruction traps, deoptimization is needed.

Bound Check Elimination

• predicate: RangeCheckElimination (product), more aggressive if tiered is off and a previous compilation of this method didn't see a deoptimization due to this optimization or loop invariant code motion
• trace: TraceRangeCheckElimination (develop)

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/a184ee1d7172/src/share/vm/c1/c1_RangeCheckElimination.cpp#l46

• Independent pass
• Follow dominator tree and propagate knowledge about ranges of integer values, use it to remove useless array bound checks
• Attempts to reorder array bound checks in a block to minimize the number of checks
• Aggressive version of the optimization can cause deoptimization

Register Allocation

• predicate: none
• trace: TraceLinearScanLevel (develop)

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/a184ee1d7172/src/share/vm/c1/c1_LinearScan.cpp#l3016

• operates on the LIR as independent pass
• linear scan register allocation

Peephole

• predicate: sparc only, LIRFillDelaySlots
• trace: LIRTracePeephole (develop)

http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/a184ee1d7172/src/cpu/sparc/vm/c1_LIRAssembler_sparc.cpp#l3515

• description

• operates on the LIR during code emission

• attempts to fill delay slot of sparc branch/call instructions with a useful instruction