transition @zone in Core.Compiler to directly use jl_timing (#58213)

This builds on the prior work from Kristoffer and Prem to implement a
Julia-side `@zone` macro that interacts with the JL_TIMING system to
provide profiling events to Tracy, VTune, and/or the TIMING_COUNTS
back-end.

This implementation includes some extra tricks to make sure that the
timing block is stack-allocated, which can be important for performance
if we start to use `@zone` for high-frequency events.

Co-authored-by: Cody Tapscott <[email protected]>
Co-authored-by: Kristoffer Carlsson <[email protected]>
Co-authored-by: Prem Chintalapudi <[email protected]>

Author: 4 people

Compiler/src/Compiler.jl

@@ -120,7 +120,7 @@ function is_return_type(Core.@nospecialize(f))
     return false
 end
 
-include("profiling.jl")
+include("timing.jl")
 include("sort.jl")
 
 # We don't include some.jl, but this definition is still useful.

Compiler/src/profiling.jl

@@ -0,0 +1,38 @@
+if ccall(:jl_timing_enabled, Cint, ()) != 0
+    function getzonedexpr(name::Union{Symbol, String}, ex::Expr, func::Symbol, file::Symbol, line::Integer, color::Integer)
+        event = RefValue{Ptr{Cvoid}}(C_NULL)
+        name = QuoteNode(Symbol(name))
+        func = QuoteNode(func)
+        file = QuoteNode(file)
+
+        # XXX: This buffer must be large enough to store any jl_timing_block_t (runtime-checked)
+        buffer = (0, 0, 0, 0, 0, 0, 0)
+        buffer_size = Core.sizeof(buffer)
+        return quote
+            if $event[] === C_NULL
+                $event[] = ccall(:_jl_timing_event_create, Ptr{Cvoid},
+                                 (Ptr{UInt8}, Ptr{UInt8}, Ptr{UInt8}, Ptr{UInt8}, Cint, Cint),
+                                 :CORE_COMPILER, $name, $func, $file, $line, $color)
+            end
+            timing_block = RefValue($buffer)
+            block_ptr = pointer_from_objref(timing_block)
+            $(Expr(:gc_preserve, quote
+                ccall(:_jl_timing_block_init, Cvoid, (Ptr{Cvoid}, Csize_t, Ptr{Cvoid}), block_ptr, $buffer_size, $event[])
+                ccall(:_jl_timing_block_start, Cvoid, (Ptr{Cvoid},), block_ptr)
+                $(Expr(:tryfinally,
+                    :($(Expr(:escape, ex))),
+                    quote
+                        ccall(:_jl_timing_block_end, Cvoid, (Ptr{Cvoid},), block_ptr)
+                    end
+                ))
+            end, :timing_block))
+        end
+    end
+    macro zone(name, ex::Expr)
+        return getzonedexpr(name, ex, :unknown_julia_function, __source__.file, __source__.line, 0)
+    end
+else
+    macro zone(name, ex::Expr)
+        return esc(ex)
+    end
+end

Compiler/src/profiling/ittapi.jl

@@ -16,29 +16,14 @@ jl_module_t *jl_module_root(jl_module_t *m);
 extern "C" {
 #endif
 
-JL_DLLEXPORT int jl_tracy_enabled(void) {
-#ifdef USE_TRACY
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-JL_DLLEXPORT int jl_ittapi_enabled(void) {
-#ifdef USE_ITTAPI
+JL_DLLEXPORT int jl_timing_enabled(void) {
+#ifdef ENABLE_TIMINGS
     return 1;
 #else
     return 0;
 #endif
 }
 
-JL_DLLEXPORT int jl_nvtx_enabled(void) {
-#ifdef USE_NVTX
-    return 1;
-#else
-    return 0;
-#endif
-}
 
 #ifdef ENABLE_TIMINGS
 

Compiler/src/profiling/tracy.jl

@@ -55,9 +55,7 @@ JL_DLLEXPORT jl_timing_event_t *_jl_timing_event_create(const char *subsystem, c
 JL_DLLEXPORT void _jl_timing_block_init(char *buf, size_t size, jl_timing_event_t *event);
 JL_DLLEXPORT void _jl_timing_block_start(jl_timing_block_t *cur_block);
 JL_DLLEXPORT void _jl_timing_block_end(jl_timing_block_t *cur_block);
-JL_DLLEXPORT int jl_tracy_enabled(void);
-JL_DLLEXPORT int jl_ittapi_enabled(void);
-JL_DLLEXPORT int jl_nvtx_enabled(void);
+JL_DLLEXPORT int jl_timing_enabled(void);
 
 
 #ifdef __cplusplus
@@ -191,6 +189,7 @@ JL_DLLEXPORT void jl_timing_puts(jl_timing_block_t *cur_block, const char *str);
         X(STACKWALK)             \
         X(DL_OPEN)               \
         X(JULIA_INIT)            \
+        X(CORE_COMPILER)        \
 
 
 #define JL_TIMING_COUNTERS \