Measure MILP overhead without trip count annotations

results/trip_count/milp.csv

@@ -0,0 +1,14 @@
+benchmark,capacitor,status,basic_blocks,edges,abstract_cfg_blocks,abstract_cfg_edges,region_boundaries,compilation_time_ms,peak_rss_kb,profiling_time_ms,execution_time_us,runtime_region_boundary_calls,runtime_debug_store_mem_calls,runtime_debug_restore_mem_calls,candidate_globals,milp_allocation_mode,milp_variables,milp_constraints,milp_presolved_variables,milp_presolved_constraints,optimal_solution,distributed_checkpoints_inserted,vm_placed_globals,vm_placed_global_names,milp_solve_time_ms,result,run_attempts
+activity_recognition-10uF,10uF,ok,16,22,8,9,4,45,40640,439,34048.7,13,0,0,9,regional,533,1986,147,323,yes,6,8,lfsr_state:4;__local_main.features.i.sroa.0:4;__local_main.features.i.sroa.4:4;__local_main.features.i4.sroa.0:4;__local_main.features.i4.sroa.4:4;__local_main.stats.sroa.0.i:8;__local_main.features.i13.sroa.0:4;__local_main.features.i13.sroa.5:4,26,64,1
+aes-10uF,10uF,ok,26,35,9,10,4,208,48624,559,104854.28,20,0,0,3,regional,2475,10033,61,133,yes,49,0,,26,107,1
+bitcount-10uF,10uF,ok,64,95,52,75,22,400,77056,488,6178386.14,80096,0,0,2,regional,8126,35533,1885,5944,yes,49,0,,230,52382,1
+chacha20-10uF,10uF,ok,11,15,4,4,2,26,37632,536,50559.28,8,0,0,3,regional,495,1907,23,46,yes,30,0,,4,79,1
+crc-10uF,10uF,ok,9,12,6,6,3,10,35936,423,29828.77,5,0,0,2,regional,117,414,18,37,yes,0,1,__local_main.result:6,3,20431,1
+cuckoo_filter-10uF,10uF,ok,23,33,15,20,4,203,47840,536,72434.9,2052,0,0,3,regional,1009,4162,344,951,yes,12,0,,166,4048,1
+dijkstra-10uF,10uF,ok,17,24,11,15,7,138,44768,429,1126514.82,3787,0,0,5,regional,1229,4972,223,550,yes,48,1,__local_main.nodes.i:5,99,33447,1
+poly1305-10uF,10uF,ok,7,9,4,4,2,19,36880,472,478591.92,86,0,0,3,regional,559,2135,31,68,yes,0,0,,4,68,1
+qsort-10uF,10uF,ok,17,24,12,15,5,18,37776,421,1091951.53,643,0,0,2,regional,324,1264,36,75,yes,7,0,,6,27023,1
+rsa-10uF,10uF,ok,144,209,86,119,43,4652,146896,658,74903.7,324,0,0,8,regional,21292,91133,3214,9905,yes,59,3,g_product:54;__local_main.qn_arr.i:11;__local_main.qn_arr.i34:8,4182,32,1
+sensor_fusion-10uF,10uF,ok,177,268,4,4,2,18,38944,419,25604.81,17,0,0,2,regional,431,1696,17,34,yes,0,1,g_output:4,5,613,1
+sha256-10uF,10uF,ok,38,54,14,18,6,164,58848,399,94440.73,18,0,0,6,regional,4083,16818,307,830,yes,79,0,,89,131,1
+stringsearch-10uF,10uF,ok,41,58,31,43,12,545,61632,556,82545.06,2270,0,0,4,regional,2200,9143,1070,3127,yes,21,3,__local_main.total_found:7;len:7;findme:7,473,20,1

results/trip_count/milp_debug.csv

@@ -0,0 +1,14 @@
+benchmark,capacitor,status,basic_blocks,edges,abstract_cfg_blocks,abstract_cfg_edges,region_boundaries,compilation_time_ms,peak_rss_kb,profiling_time_ms,execution_time_us,runtime_region_boundary_calls,runtime_debug_store_mem_calls,runtime_debug_restore_mem_calls,candidate_globals,milp_allocation_mode,milp_variables,milp_constraints,milp_presolved_variables,milp_presolved_constraints,optimal_solution,distributed_checkpoints_inserted,vm_placed_globals,vm_placed_global_names,milp_solve_time_ms,result,run_attempts
+activity_recognition-10uF,10uF,ok,16,22,8,9,4,38,40624,381,34054.13,13,0,0,9,regional,533,1986,147,323,yes,6,8,lfsr_state:4;__local_main.features.i.sroa.0:4;__local_main.features.i.sroa.4:4;__local_main.features.i4.sroa.0:4;__local_main.features.i4.sroa.4:4;__local_main.stats.sroa.0.i:8;__local_main.features.i13.sroa.0:4;__local_main.features.i13.sroa.5:4,23,64,1
+aes-10uF,10uF,ok,26,35,9,10,4,210,46384,393,104872.82,20,0,0,3,regional,2475,10033,61,133,yes,49,0,,25,107,1
+bitcount-10uF,10uF,ok,64,95,52,75,22,370,75344,412,6179187.38,80096,0,0,2,regional,8126,35533,1885,5944,yes,49,0,,202,52382,1
+chacha20-10uF,10uF,ok,11,15,4,4,2,20,37760,395,50568.56,8,0,0,3,regional,495,1907,23,46,yes,30,0,,3,79,1
+crc-10uF,10uF,ok,9,12,6,6,3,10,35888,404,29834.06,5,0,0,2,regional,117,414,18,37,yes,0,1,__local_main.result:6,2,20431,1
+cuckoo_filter-10uF,10uF,ok,23,33,15,20,4,202,47184,414,72446.45,2052,0,0,3,regional,1009,4162,344,951,yes,12,0,,166,4048,1
+dijkstra-10uF,10uF,ok,17,24,11,15,7,139,45344,405,1126739.77,3787,0,0,5,regional,1229,4972,223,550,yes,48,1,__local_main.nodes.i:5,100,33447,1
+poly1305-10uF,10uF,ok,7,9,4,4,2,14,36752,374,478678.17,86,0,0,3,regional,559,2135,31,68,yes,0,0,,3,68,1
+qsort-10uF,10uF,ok,17,24,12,15,5,18,38208,398,1092206.94,643,0,0,2,regional,324,1264,36,75,yes,7,0,,6,27023,1
+rsa-10uF,10uF,ok,144,209,86,119,43,4645,146048,407,74921.11,324,0,0,8,regional,21292,91133,3214,9905,yes,59,3,g_product:54;__local_main.qn_arr.i:11;__local_main.qn_arr.i34:8,4174,32,1
+sensor_fusion-10uF,10uF,ok,177,268,4,4,2,18,38928,418,25607.51,17,0,0,2,regional,431,1696,17,34,yes,0,1,g_output:4,5,613,1
+sha256-10uF,10uF,ok,38,54,14,18,6,166,55520,409,94456.81,18,0,0,6,regional,4083,16818,307,830,yes,79,0,,85,131,1
+stringsearch-10uF,10uF,ok,41,58,31,43,12,543,62832,409,82554.74,2270,0,0,4,regional,2200,9143,1070,3127,yes,21,3,__local_main.total_found:7;len:7;findme:7,473,20,1

scripts/ckpt/bench/all.py

@@ -133,6 +133,7 @@ def _run_step(
             energy_config=opts.energy_config,
             cpu_freq=opts.cpu_freq,
             coarse_allocation=opts.coarse_allocation,
+            tripcount_annotations=True,
             milp_gap=opts.milp_gap,
             pass_log_level=opts.pass_log_level,
             accumulate_keys_file=None,

scripts/ckpt/bench/milp.py

@@ -130,6 +130,7 @@ def run_milp_benchmarks(
     energy_config: Path | None,
     cpu_freq: int,
     coarse_allocation: bool,
+    tripcount_annotations: bool,
     milp_gap: float,
     pass_log_level: str,
     accumulate_keys_file: Path | None,
@@ -202,6 +203,7 @@ def compile_fn(bench_path: Path, cap: CapacitorConfig) -> CompileResult:
                 milp_gap=milp_gap,
                 milp_log_file="",
                 coarse_allocation=coarse_allocation,
+                tripcount_annotations=tripcount_annotations,
                 save_temps=False,
             )
 

scripts/ckpt/cli.py

@@ -154,6 +154,11 @@ def _cap_callback(
     is_flag=True,
     help="Use one MILP placement variable per eligible value instead of per-region placement.",
 )
+_no_tripcount_flag = click.option(
+    "--no-tripcount",
+    is_flag=True,
+    help="Ignore __loop_tripcount annotations; loops with unknown bounds are chunked.",
+)
 _milp_gap_option = click.option(
     "--milp-gap",
     type=float,
@@ -420,6 +425,7 @@ def _handle_accumulate_keys(result, accumulate_keys: str | None) -> None:
     "--milp-log-file", type=click.Path(), default="", help="Gurobi log file path."
 )
 @_coarse_allocation_flag
+@_no_tripcount_flag
 @_accumulate_keys_option
 @click.pass_context
 def compile_milp_cmd(
@@ -442,6 +448,7 @@ def compile_milp_cmd(
     milp_gap: float,
     milp_log_file: str,
     coarse_allocation: bool,
+    no_tripcount: bool,
     accumulate_keys: str | None,
     csv_path: str | None,
 ) -> None:
@@ -498,6 +505,7 @@ def compile_milp_cmd(
             milp_gap=milp_gap,
             milp_log_file=milp_log_file,
             coarse_allocation=coarse_allocation,
+            tripcount_annotations=not no_tripcount,
             save_temps=save_temps,
         ),
     )
@@ -979,6 +987,7 @@ def bench() -> None:
 @_energy_override_option
 @_cpu_freq_option("16")
 @_coarse_allocation_flag
+@_no_tripcount_flag
 @_milp_gap_option
 @_accumulate_keys_option
 @_saleae_timeout_option
@@ -994,6 +1003,7 @@ def bench_milp_cmd(
     energy_config: str | None,
     cpu_freq: str,
     coarse_allocation: bool,
+    no_tripcount: bool,
     milp_gap: float,
     accumulate_keys: str | None,
     timeout: float,
@@ -1014,6 +1024,7 @@ def bench_milp_cmd(
         energy_config=_path_or_none(energy_config),
         cpu_freq=_mhz_to_hz(cpu_freq),
         coarse_allocation=coarse_allocation,
+        tripcount_annotations=not no_tripcount,
         milp_gap=milp_gap,
         pass_log_level=ctx.obj["pass_log_level"],
         accumulate_keys_file=_path_or_none(accumulate_keys),
@@ -1597,6 +1608,7 @@ def verify_rockclimb_cmd(
 @_estimator_mode_option
 @_cpu_freq_option("16")
 @_coarse_allocation_flag
+@_no_tripcount_flag
 @_saleae_timeout_option
 @click.pass_context
 def verify_milp_cmd(
@@ -1608,6 +1620,7 @@ def verify_milp_cmd(
     estimator_mode: str,
     cpu_freq: str,
     coarse_allocation: bool,
+    no_tripcount: bool,
     timeout: float,
 ) -> None:
     """Verify semantic correctness of MILP checkpoint insertion."""
@@ -1625,6 +1638,7 @@ def verify_milp_cmd(
         estimator_mode=estimator_mode,
         cpu_freq=_mhz_to_hz(cpu_freq),
         coarse_allocation=coarse_allocation,
+        tripcount_annotations=not no_tripcount,
         pass_log_level=ctx.obj["pass_log_level"],
     )
     if not all_ok(results):

scripts/ckpt/compile/chunked.py

@@ -89,6 +89,7 @@ def compile_chunked(
             cpu_freq=opts.cpu_freq,
             extra_includes=opts.extra_includes,
             extra_defines=opts.extra_defines,
+            tripcount_annotations=True,
         )
 
         # Middle-end optimization

scripts/ckpt/compile/common.py

@@ -122,12 +122,17 @@ def compile_annotated_ir(
     cpu_freq: int,
     extra_includes: list[str],
     extra_defines: list[str],
+    tripcount_annotations: bool,
 ) -> Path:
     """Shared phase 1 of the instrumented pipelines: C -> raw IR -> tripcounts.
 
     No LLVM pass has run on the returned IR, so the ``__loop_tripcount``
     markers still sit inside their source loops for exact trip-count
     annotation. Returns the annotated IR path (tmp/tripcount.ll).
+
+    With ``tripcount_annotations=False`` the annotation pass is skipped and
+    the raw IR is returned; the no-op marker calls are inlined away by the
+    middle-end optimization, so the passes see no trip-count metadata.
     """
     input_ll = tmp / "input.ll"
     includes = list(extra_includes)
@@ -144,6 +149,9 @@ def compile_annotated_ir(
         extra_defines=[f"F_CPU={cpu_freq}", *extra_defines],
     )
 
+    if not tripcount_annotations:
+        return input_ll
+
     tripcount_ll = tmp / "tripcount.ll"
     annotate_tripcounts(tc, env, input_ll, tripcount_ll)
     return tripcount_ll

scripts/ckpt/compile/milp.py

@@ -49,6 +49,7 @@ class MilpCompileOptions:
     milp_log_file: str
     coarse_allocation: bool
     save_temps: bool
+    tripcount_annotations: bool
     extra_includes: list[str] = field(default_factory=list)
     extra_defines: list[str] = field(default_factory=list)
 
@@ -105,6 +106,7 @@ def compile_milp(
             cpu_freq=opts.cpu_freq,
             extra_includes=opts.extra_includes,
             extra_defines=opts.extra_defines,
+            tripcount_annotations=opts.tripcount_annotations,
         )
 
         # Middle-end optimization

scripts/ckpt/compile/rockclimb.py

@@ -100,6 +100,7 @@ def compile_rockclimb(
             cpu_freq=opts.cpu_freq,
             extra_includes=[],
             extra_defines=opts.extra_defines,
+            tripcount_annotations=True,
         )
 
         # Step 1c: Middle-end optimization (the second half of what a plain

scripts/ckpt/compile/schematic.py

@@ -110,6 +110,7 @@ def compile_schematic(
             cpu_freq=opts.cpu_freq,
             extra_includes=opts.extra_includes,
             extra_defines=opts.extra_defines,
+            tripcount_annotations=True,
         )
 
         # Middle-end optimization

scripts/ckpt/intermittent/runner.py

@@ -184,6 +184,7 @@ def compile_fn(bench_path: Path, cap: CapacitorConfig) -> CompileResult:
                 milp_gap=0.0,
                 milp_log_file="",
                 coarse_allocation=False,
+                tripcount_annotations=True,
                 save_temps=False,
                 extra_defines=list(_SHRINK_DEFINES),
             )

scripts/ckpt/verify/all.py

@@ -37,6 +37,7 @@ def verify_all(
             energy_config=energy_config,
             estimator_mode=estimator_mode,
             coarse_allocation=False,
+            tripcount_annotations=True,
             pass_log_level=pass_log_level,
         ),
         rockclimb_spec(

scripts/ckpt/verify/milp.py

@@ -22,6 +22,7 @@ def milp_spec(
     energy_config: Path | None,
     estimator_mode: str,
     coarse_allocation: bool,
+    tripcount_annotations: bool,
     pass_log_level: str,
 ) -> AlgorithmSpec:
     """Build the MILP verification spec."""
@@ -60,6 +61,7 @@ def compile_instrumented(
                 milp_gap=0.0,
                 milp_log_file="",
                 coarse_allocation=coarse_allocation,
+                tripcount_annotations=tripcount_annotations,
                 save_temps=False,
                 extra_defines=extra_defines,
             ),
@@ -88,6 +90,7 @@ def verify_milp(
     cpu_freq: int,
     capture_timeout_seconds: float,
     coarse_allocation: bool,
+    tripcount_annotations: bool,
     pass_log_level: str,
 ) -> list[BenchResult]:
     """Verify semantic correctness of MILP checkpoint insertion."""
@@ -96,6 +99,7 @@ def verify_milp(
         energy_config=energy_config,
         estimator_mode=estimator_mode,
         coarse_allocation=coarse_allocation,
+        tripcount_annotations=tripcount_annotations,
         pass_log_level=pass_log_level,
     )
     return verify_algorithms(