Class Profile

java.lang.Object
com.oracle.truffle.api.nodes.NodeCloneable
com.oracle.truffle.api.profiles.Profile
All Implemented Interfaces:
Cloneable
Direct Known Subclasses:
BranchProfile, ByteValueProfile, ConditionProfile, CountingConditionProfile, DoubleValueProfile, FloatValueProfile, IntValueProfile, LongValueProfile, ValueProfile

public abstract class Profile extends NodeCloneable

A profile is a Truffle utility class that uses the Truffle compiler directives to guard for and/or forward runtime information to the compiler. Whenever Truffle DSL can be used inlined profiles subclasses should be used instead of regular profile subclasses.

Usage: Profiles should be stored in final or compilation final fields of node classes to ensure that they can get optimized properly. Profiles must not be shared between ASTs. Using the same profile multiple times in a single node or in multiple nodes which link to the same root is allowed. Never store profiles inside runtime values that leave the scope of the originating AST. This limitation exists because the used mechanism to invalidate compiled code performs local invalidations only. For global speculations use assumptions instead.

Compilation: Some profiles like branch profiles do not induce additional overhead in compiled code. Others like value profiles might require a runtime check to verify their assumptions which are forwarded to the compiler. Even if profiles do not induce direct overhead in compiled code it still might get invalidated as a result of using profiles. Invalidating profiles will result in the invalidation of compiled code. It is therefore essential to place these profiles in way that is neither too aggressive nor too conservative.

Footprint: Whether profiling information can be forwarded to the compiler depends on the capabilities of the runtime system. If the runtime returns true in TruffleRuntime.isProfilingEnabled() then runtime information will get collected. This comes at at the cost of additional overhead and footprint in interpreted mode. Thats why the factory methods of profiles can return implementations where profiling is disabled. Using disabled profiles makes sense for runtimes that are unable to use the collected profiling information. Even runtime implementations that are able to use this information might decide to turn off profiling for benchmarking purposes.

Profile subclasses:

Since:
0.10
See Also:
  • Method Details

    • disable

      public void disable()
      Disables this profile by setting it to its generic state. After disabling it is guaranteed to never deoptimize on any invocation of a profile method.

      This method must not be called on compiled code paths. Note that disabling the profile will not invalidate existing compiled code that uses this profile.

      Since:
      21.1
    • reset

      public void reset()
      Resets this profile to its uninitialized state. Has no effect if this profile is already in its uninitialized state or a disabled version of this profile is used.

      This method must not be called on compiled code paths. Note that disabling the profile will not invalidate existing compiled code that uses this profile.

      Since:
      21.1