Profile-Guided Optimization¶
Use measured program behavior to choose later or adaptive transformations at the code sites that generated the profile.
Core Idea¶
Profile-guided optimization measures how a program actually runs, maps the observations back to program sites, and lets a compiler or runtime use them to choose transformations. The profile informs decisions that static code alone cannot settle confidently, but any gain depends on the measured workload being relevant to later use.[ref-3cb4a5f53a31][ref-deb6f6cd3e7a]
Scope of Application¶
An ahead-of-time compiler may build an instrumented program, run representative inputs and rebuild using the profile. An adaptive JIT can gather hotness and branch behavior while running. AutoFetch applies a related profile-to-transformation idea to ORM prefetching, but is a separate named system, not an alias for compiler PGO.[ref-3cb4a5f53a31][ref-889824462f58][^ref-85b73e9369f8]
Clarity¶
A profiler report is measurement, not optimization. The optimizer must consume the profile and alter code or data-access choices. Better performance is then an empirical outcome to test; a stale or mismatched profile can regress results.[^ref-deb6f6cd3e7a]
For a constructed Clang training profile, a parser branch taken 9,900 times versus 100 can favor basic-block ordering around the hot edge; Clang documents that choice, but no speedup is asserted. In a different constructed HotSpot profile, a shape.area() call with 9,500 Circle versus 500 Square receivers can favor guarded Circle-method inlining with a fallback; OpenJDK's implementation consults receiver profiles for such decisions. Both profiles must be matched to their program sites, and neither count is a published measurement.[ref-3cb4a5f53a31][ref-9fc4047bdd37]
Manages Complexity¶
Instead of treating every path equally, a profile summarizes which paths or sites mattered during execution. The optimizer can focus effort there. The summary is incomplete: instrumentation costs time, sampling misses detail, and training behavior may differ from deployment.[^ref-3cb4a5f53a31]
Abstract Reasoning¶
Choose a runtime-sensitive optimization, collect a representative profile, verify its site mapping, apply it through a compiler or adaptive runtime, then test on intended workloads. If no measured data are consumed, the process is static optimization; if no transformation follows, it is profiling alone.[ref-3cb4a5f53a31][ref-deb6f6cd3e7a]
Knowledge Transfer¶
The observation-to-transformation relation carries across offline compiler PGO, JIT feedback and some data-access strategies. The method presupposes Feedback from measured execution, but is not itself generic feedback. Profile units and optimizations differ, so branch weights cannot be substituted for ORM traversal data; AutoFetch is related, not an exact alias, and a profile does not guarantee speedup.
[^ref-3cb4a5f53a31]: Clang Compiler User's Manual, PGO sections, official documentation.
[^ref-deb6f6cd3e7a]: GCC Optimize Options, official profile-use documentation.
[^ref-889824462f58]: Oracle, Java Virtual Machine Technology Overview, adaptive JIT overview.
[^ref-85b73e9369f8]: “Automatic Prefetching by Traversal”, original AutoFetch paper.
[^ref-9fc4047bdd37]: OpenJDK, HotSpot doCall.cpp reviewed source snapshot, receiver-profile inlining logic.
Relationships to Other Abstractions¶
Current abstraction Profile-Guided Optimization Domain-specific
Parents (1) — more general patterns this builds on
-
Profile-Guided Optimization presupposes Feedback Prime
PGO presupposes measured feedback.
Hierarchy path (1) — routes to 1 parentless root
- Profile-Guided Optimization → Feedback
Neighborhood in Abstraction Space¶
Profile-Guided Optimization sits in a moderately populated region (58th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.
Family — Program Execution & Runtime Concepts (27 abstractions)
Nearest neighbors
- Program Profiling — 0.88
- Polyvariance — 0.85
- Release Early, Release Often — 0.85
- Duck Typing — 0.84
- Reconfigurable Computing — 0.84
Computed from structural-signature embeddings · 2026-10-08