Make the compile time flag actually atomic, not just data race free.

src/aotcompile.cpp

@@ -286,7 +286,8 @@ void *jl_create_native(jl_array_t *methods, const jl_cgparams_t cgparams, int _p
  JL_GC_PUSH1(&src);
  JL_LOCK(&codegen_lock);
  uint64_t compiler_start_time = 0;
- if (jl_atomic_load(&jl_measure_compile_time))
+ uint8_t measure_compile_time = jl_atomic_load(&jl_measure_compile_time);
+ if (measure_compile_time)
  compiler_start_time = jl_hrtime();
 
  CompilationPolicy policy = (CompilationPolicy) _policy;
@@ -414,7 +415,7 @@ void *jl_create_native(jl_array_t *methods, const jl_cgparams_t cgparams, int _p
  }
 
  data->M = std::move(clone);
- if (jl_atomic_load(&jl_measure_compile_time))
+ if (measure_compile_time)
  jl_atomic_fetch_add(&jl_measure_compile_time, (jl_hrtime() - compiler_start_time));
  if (policy == CompilationPolicy::ImagingMode)
  imaging_mode = 0;
@@ -915,7 +916,8 @@ void *jl_get_llvmf_defn(jl_method_instance_t *mi, size_t world, char getwrapper,
  jl_llvm_functions_t decls;
  JL_LOCK(&codegen_lock);
  uint64_t compiler_start_time = 0;
- if (jl_atomic_load(&jl_measure_compile_time))
+ uint8_t measure_compile_time = jl_atomic_load(&jl_measure_compile_time);
+ if (measure_compile_time)
  compiler_start_time = jl_hrtime();
  std::tie(m, decls) = jl_emit_code(mi, src, jlrettype, output);
 
@@ -940,7 +942,7 @@ void *jl_get_llvmf_defn(jl_method_instance_t *mi, size_t world, char getwrapper,
  m.release(); // the return object `llvmf` will be the owning pointer
  }
  JL_GC_POP();
- if (jl_atomic_load(&jl_measure_compile_time))
+ if (measure_compile_time)
  jl_atomic_fetch_add(&jl_measure_compile_time, (jl_hrtime() - compiler_start_time));
  JL_UNLOCK(&codegen_lock); // Might GC
  if (F)

src/gf.c

@@ -3160,18 +3160,23 @@ int jl_has_concrete_subtype(jl_value_t *typ)
 #define typeinf_lock codegen_lock
 
 static uint64_t inference_start_time = 0;
+static uint8_t inference_is_measuring_compile_time = 0;
 
 JL_DLLEXPORT void jl_typeinf_begin(void)
 {
  JL_LOCK(&typeinf_lock);
- if (jl_atomic_load(&jl_measure_compile_time))
+ if (jl_atomic_load(&jl_measure_compile_time)) {
  inference_start_time = jl_hrtime();
+ inference_is_measuring_compile_time = 1;
+ }
 }
 
 JL_DLLEXPORT void jl_typeinf_end(void)
 {
- if (typeinf_lock.count == 1 && jl_atomic_load(&jl_measure_compile_time))
- jl_atomic_fetch_add(&jl_measure_compile_time, (jl_hrtime() - inference_start_time));
+ if (typeinf_lock.count == 1 && inference_is_measuring_compile_time) {
+ jl_atomic_fetch_add(&jl_cumulative_compile_time, (jl_hrtime() - inference_start_time));
+ inference_is_measuring_compile_time = 0;
+ }
  JL_UNLOCK(&typeinf_lock);
 }
 

src/jitlayers.cpp

@@ -238,7 +238,8 @@ int jl_compile_extern_c(void *llvmmod, void *p, void *sysimg, jl_value_t *declrt
 {
  JL_LOCK(&codegen_lock);
  uint64_t compiler_start_time = 0;
- if (jl_atomic_load(&jl_measure_compile_time))
+ uint8_t measure_compile_time = jl_atomic_load(&jl_measure_compile_time);
+ if (measure_compile_time)
  compiler_start_time = jl_hrtime();
  jl_codegen_params_t params;
  jl_codegen_params_t *pparams = (jl_codegen_params_t*)p;
@@ -262,7 +263,7 @@ int jl_compile_extern_c(void *llvmmod, void *p, void *sysimg, jl_value_t *declrt
  if (success && llvmmod == NULL)
  jl_add_to_ee(std::unique_ptr<Module>(into));
  }
- if (codegen_lock.count == 1 && jl_atomic_load(&jl_measure_compile_time))
+ if (codegen_lock.count == 1 && measure_compile_time)
  jl_atomic_fetch_add(&jl_cumulative_compile_time, (jl_hrtime() - compiler_start_time));
  JL_UNLOCK(&codegen_lock);
  return success;
@@ -319,7 +320,8 @@ jl_code_instance_t *jl_generate_fptr(jl_method_instance_t *mi JL_PROPAGATES_ROOT
 {
  JL_LOCK(&codegen_lock); // also disables finalizers, to prevent any unexpected recursion
  uint64_t compiler_start_time = 0;
- if (jl_atomic_load(&jl_measure_compile_time))
+ uint8_t measure_compile_time = jl_atomic_load(&jl_measure_compile_time);
+ if (measure_compile_time)
  compiler_start_time = jl_hrtime();
  // if we don't have any decls already, try to generate it now
  jl_code_info_t *src = NULL;
@@ -357,7 +359,7 @@ jl_code_instance_t *jl_generate_fptr(jl_method_instance_t *mi JL_PROPAGATES_ROOT
  else {
  codeinst = NULL;
  }
- if (codegen_lock.count == 1 && jl_atomic_load(&jl_measure_compile_time))
+ if (codegen_lock.count == 1 && measure_compile_time)
  jl_atomic_fetch_add(&jl_cumulative_compile_time, (jl_hrtime() - compiler_start_time));
  JL_UNLOCK(&codegen_lock);
  JL_GC_POP();
@@ -372,7 +374,8 @@ void jl_generate_fptr_for_unspecialized(jl_code_instance_t *unspec)
  }
  JL_LOCK(&codegen_lock);
  uint64_t compiler_start_time = 0;
- if (jl_atomic_load(&jl_measure_compile_time))
+ uint8_t measure_compile_time = jl_atomic_load(&jl_measure_compile_time);
+ if (measure_compile_time)
  compiler_start_time = jl_hrtime();
  if (unspec->invoke == NULL) {
  jl_code_info_t *src = NULL;
@@ -400,7 +403,7 @@ void jl_generate_fptr_for_unspecialized(jl_code_instance_t *unspec)
  }
  JL_GC_POP();
  }
- if (codegen_lock.count == 1 && jl_atomic_load(&jl_measure_compile_time))
+ if (codegen_lock.count == 1 && measure_compile_time)
  jl_atomic_fetch_add(&jl_cumulative_compile_time, (jl_hrtime() - compiler_start_time));
  JL_UNLOCK(&codegen_lock); // Might GC
 }
@@ -424,7 +427,8 @@ jl_value_t *jl_dump_method_asm(jl_method_instance_t *mi, size_t world,
  // so create an exception here so we can print pretty our lies
  JL_LOCK(&codegen_lock); // also disables finalizers, to prevent any unexpected recursion
  uint64_t compiler_start_time = 0;
- if (jl_atomic_load(&jl_measure_compile_time))
+ uint8_t measure_compile_time = jl_atomic_load(&jl_measure_compile_time);
+ if (measure_compile_time)
  compiler_start_time = jl_hrtime();
  specfptr = (uintptr_t)codeinst->specptr.fptr;
  if (specfptr == 0) {
@@ -449,7 +453,7 @@ jl_value_t *jl_dump_method_asm(jl_method_instance_t *mi, size_t world,
  }
  JL_GC_POP();
  }
- if (jl_atomic_load(&jl_measure_compile_time))
+ if (measure_compile_time)
  jl_atomic_fetch_add(&jl_cumulative_compile_time, (jl_hrtime() - compiler_start_time));
  JL_UNLOCK(&codegen_lock);
  }