bench: add heavy AST load test forcing full PCH deserialization

Adds a "heavy" source file that references types from all 70 headers (vector, map, optional, regex, thread, chrono, etc.) to stress-test PCH chain deserialization. Even under full deserialization the chained PCH overhead is only +6% vs monolithic. Also adds an end-to-end benchmark simulating the real workflow: 1. Monolithic PCH build (what user waits for): 1233ms 2. Background chain split (async): 2672ms (user doesn't wait) 3. Incremental append of <chrono>: 28ms vs 1251ms rebuild (44x speedup) 4. Correctness verification: PASS Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-18 15:03:08 +08:00 · 2026-04-18 15:02:21 +08:00 · 2026-04-18 15:02:21 +08:00 · 2026-04-18 15:02:21 +08:00
2 changed files with 1064 additions and 0 deletions
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -200,6 +200,14 @@ if(CLICE_ENABLE_BENCHMARK)
        "${PROJECT_SOURCE_DIR}/src"
    )
    target_link_libraries(scan_benchmark PRIVATE clice::core kota::deco)
+
+    add_executable(pch_chain_benchmark
+        "${PROJECT_SOURCE_DIR}/benchmarks/pch_chain/pch_chain_benchmark.cpp"
+    )
+    target_include_directories(pch_chain_benchmark PRIVATE
+        "${PROJECT_SOURCE_DIR}/src"
+    )
+    target_link_libraries(pch_chain_benchmark PRIVATE clice::core)
 endif()

 if(CLICE_RELEASE)
--- a/benchmarks/pch_chain/pch_chain_benchmark.cpp
+++ b/benchmarks/pch_chain/pch_chain_benchmark.cpp
Author	SHA1	Message	Date
ykiko	f19e75e2fb	bench: add heavy AST load test forcing full PCH deserialization Adds a "heavy" source file that references types from all 70 headers (vector, map, optional, regex, thread, chrono, etc.) to stress-test PCH chain deserialization. Even under full deserialization the chained PCH overhead is only +6% vs monolithic. Also adds an end-to-end benchmark simulating the real workflow: 1. Monolithic PCH build (what user waits for): 1233ms 2. Background chain split (async): 2672ms (user doesn't wait) 3. Incremental append of <chrono>: 28ms vs 1251ms rebuild (44x speedup) 4. Correctness verification: PASS Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	2026-04-18 15:03:08 +08:00
ykiko	ba191943b6	bench: add heavy AST load test forcing full PCH deserialization Adds a second scenario that references symbols from all 70 headers, forcing clang to fully deserialize the entire PCH chain. Results: - Light source (3 types): chained +2% overhead - Heavy source (all 70 headers): chained +6% overhead Even in the worst case, chained PCH adds only 28ms to compilation. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	2026-04-18 15:02:21 +08:00
ykiko	90f3d34768	bench: add AST load latency comparison for chained vs monolithic PCH Measures the cost of compiling a source file against a chained PCH (70 links) vs a monolithic PCH. Result: chained PCH adds only ~4% overhead to AST load, making it practical for production use. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	2026-04-18 15:02:21 +08:00
ykiko	5b24dac6c3	bench: add chained PCH benchmark Adds a benchmark comparing monolithic PCH vs chained PCH (one link per Key findings on 70 C++ stdlib headers: - Monolithic full build: ~1300ms - Chained full build: ~2600ms (2x, expected serialization overhead) - Incremental append-one-link: ~37ms vs ~1300ms monolithic rebuild (36x speedup) - All 70 chain links compile and verify successfully Also documents that PrecompiledPreambleBytes bound must be 0 for chained PCH (each link is a separate file, previous PCH doesn't cover current file). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	2026-04-18 15:02:21 +08:00