use Tracy instead of TimerOutputs

Project.toml

@@ -14,7 +14,7 @@ Preferences = "21216c6a-2e73-6563-6e65-726566657250"
 Scratch = "6c6a2e73-6563-6170-7368-637461726353"
 Serialization = "9e88b42a-f829-5b0c-bbe9-9e923198166b"
 TOML = "fa267f1f-6049-4f14-aa54-33bafae1ed76"
-TimerOutputs = "a759f4b9-e2f1-59dc-863e-4aeb61b1ea8f"
+Tracy = "e689c965-62c8-4b79-b2c5-8359227902fd"
 UUIDs = "cf7118a7-6976-5b1a-9a39-7adc72f591a4"
 
 [compat]
@@ -28,6 +28,6 @@ Preferences = "1"
 Scratch = "1"
 Serialization = "1"
 TOML = "1"
-TimerOutputs = "0.5"
+Tracy = "0.1.4"
 UUIDs = "1"
 julia = "1.10"

src/GPUCompiler.jl

@@ -3,7 +3,7 @@ module GPUCompiler
 using LLVM
 using LLVM.Interop
 
-using TimerOutputs
+using Tracy
 
 using ExprTools: splitdef, combinedef
 
@@ -64,6 +64,8 @@ function __init__()
     end
     mkpath(dir)
     global compile_cache = dir
+
+    Tracy.@register_tracepoints()
 end
 
 end # module

src/driver.jl

@@ -82,7 +82,7 @@ function compile_unhooked(output::Symbol, @nospecialize(job::CompilerJob); kwarg
         error("No active LLVM context. Use `JuliaContext()` do-block syntax to create one.")
     end
 
-    @timeit_debug to "Validation" begin
+    @tracepoint "Validation" begin
         check_method(job)   # not optional
         job.config.validate && check_invocation(job)
     end
@@ -96,7 +96,7 @@ function compile_unhooked(output::Symbol, @nospecialize(job::CompilerJob); kwarg
 
     if output == :llvm
         if job.config.strip
-            @timeit_debug to "strip debug info" strip_debuginfo!(ir)
+            @tracepoint "strip debug info" strip_debuginfo!(ir)
         end
 
         return ir, ir_meta
@@ -168,7 +168,7 @@ const __llvm_initialized = Ref(false)
         __llvm_initialized[] = true
     end
 
-    @timeit_debug to "IR generation" begin
+    @tracepoint "IR generation" begin
         ir, compiled = irgen(job)
         if job.config.entry_abi === :specfunc
             entry_fn = compiled[job.source].specfunc
@@ -267,21 +267,21 @@ const __llvm_initialized = Ref(false)
             runtime_intrinsics = ["julia.gc_alloc_obj"]
         end
 
-        @timeit_debug to "Library linking" begin
+        @tracepoint "Library linking" begin
             # target-specific libraries
             undefined_fns = LLVM.name.(decls(ir))
-            @timeit_debug to "target libraries" link_libraries!(job, ir, undefined_fns)
+            @tracepoint "target libraries" link_libraries!(job, ir, undefined_fns)
 
             # GPU run-time library
             if !uses_julia_runtime(job) && any(fn -> fn in runtime_fns ||
                                                         fn in runtime_intrinsics,
                                                 undefined_fns)
-                @timeit_debug to "runtime library" link_library!(ir, runtime)
+                @tracepoint "runtime library" link_library!(ir, runtime)
             end
         end
     end
 
-    @timeit_debug to "IR post-processing" begin
+    @tracepoint "IR post-processing" begin
         # mark everything internal except for entrypoints and any exported
         # global variables. this makes sure that the optimizer can, e.g.,
         # rewrite function signatures.
@@ -312,7 +312,7 @@ const __llvm_initialized = Ref(false)
         end
 
         if job.config.toplevel && job.config.optimize
-            @timeit_debug to "optimization" begin
+            @tracepoint "optimization" begin
                 optimize!(job, ir; job.config.opt_level)
 
                 # deferred codegen has some special optimization requirements,
@@ -339,7 +339,7 @@ const __llvm_initialized = Ref(false)
         end
 
         if job.config.toplevel && job.config.cleanup
-            @timeit_debug to "clean-up" begin
+            @tracepoint "clean-up" begin
                 @dispose pb=NewPMPassBuilder() begin
                     add!(pb, RecomputeGlobalsAAPass())
                     add!(pb, GlobalOptPass())
@@ -379,13 +379,13 @@ const __llvm_initialized = Ref(false)
     end
 
     if job.config.toplevel && job.config.validate
-        @timeit_debug to "Validation" begin
+        @tracepoint "Validation" begin
             check_ir(job, ir)
         end
     end
 
     if should_verify()
-        @timeit_debug to "verification" verify(ir)
+        @tracepoint "verification" verify(ir)
     end
 
     return ir, (; entry, compiled)
@@ -395,13 +395,13 @@ end
                           format::LLVM.API.LLVMCodeGenFileType)
     # NOTE: strip after validation to get better errors
     if job.config.strip
-        @timeit_debug to "Debug info removal" strip_debuginfo!(ir)
+        @tracepoint "Debug info removal" strip_debuginfo!(ir)
     end
 
-    @timeit_debug to "LLVM back-end" begin
-        @timeit_debug to "preparation" prepare_execution!(job, ir)
+    @tracepoint "LLVM back-end" begin
+        @tracepoint "preparation" prepare_execution!(job, ir)
 
-        code = @timeit_debug to "machine-code generation" mcgen(job, ir, format)
+        code = @tracepoint "machine-code generation" mcgen(job, ir, format)
     end
 
     return code, ()

src/gcn.jl

@@ -61,7 +61,7 @@ end
 function lower_throw_extra!(mod::LLVM.Module)
     job = current_job::CompilerJob
     changed = false
-    @timeit_debug to "lower throw (extra)" begin
+    @tracepoint "lower throw (extra)" begin
 
     throw_functions = [
         r"julia_bounds_error.*",

src/irgen.jl

@@ -1,7 +1,7 @@
 # LLVM IR generation
 
 function irgen(@nospecialize(job::CompilerJob))
-    mod, compiled = @timeit_debug to "emission" compile_method_instance(job)
+    mod, compiled = @tracepoint "emission" compile_method_instance(job)
     if job.config.entry_abi === :specfunc
         entry_fn = compiled[job.source].specfunc
     else
@@ -11,7 +11,7 @@ function irgen(@nospecialize(job::CompilerJob))
     entry = functions(mod)[entry_fn]
 
     # clean up incompatibilities
-    @timeit_debug to "clean-up" begin
+    @tracepoint "clean-up" begin
         for llvmf in functions(mod)
             if Base.isdebugbuild()
                 # only occurs in debug builds
@@ -81,7 +81,7 @@ function irgen(@nospecialize(job::CompilerJob))
         (; compiled[job.source].ci, func, specfunc)
 
     # minimal required optimization
-    @timeit_debug to "rewrite" begin
+    @tracepoint "rewrite" begin
         if job.config.kernel && needs_byval(job)
             # pass all bitstypes by value; by default Julia passes aggregates by reference
             # (this improves performance, and is mandated by certain back-ends like SPIR-V).
@@ -136,7 +136,7 @@ end
 function lower_throw!(mod::LLVM.Module)
     job = current_job::CompilerJob
     changed = false
-    @timeit_debug to "lower throw" begin
+    @tracepoint "lower throw" begin
 
     throw_functions = [
         # unsupported runtime functions that are used to throw specific exceptions
@@ -370,7 +370,7 @@ end
 # https://reviews.llvm.org/D79744
 function lower_byval(@nospecialize(job::CompilerJob), mod::LLVM.Module, f::LLVM.Function)
     ft = function_type(f)
-    @timeit_debug to "lower byval" begin
+    @tracepoint "lower byval" begin
 
     # classify the arguments
     args = classify_arguments(job, ft)

src/ptx.jl

@@ -387,7 +387,7 @@ function nvvm_reflect!(fun::LLVM.Function)
     job = current_job::CompilerJob
     mod = LLVM.parent(fun)
     changed = false
-    @timeit_debug to "nvvmreflect" begin
+    @tracepoint "nvvmreflect" begin
 
     # find and sanity check the nnvm-reflect function
     # TODO: also handle the llvm.nvvm.reflect intrinsic

src/spirv.jl

@@ -212,7 +212,7 @@ end
 function rm_trap!(mod::LLVM.Module)
     job = current_job::CompilerJob
     changed = false
-    @timeit_debug to "remove trap" begin
+    @tracepoint "remove trap" begin
 
     if haskey(functions(mod), "llvm.trap")
         trap = functions(mod)["llvm.trap"]
@@ -238,7 +238,7 @@ end
 function rm_freeze!(mod::LLVM.Module)
     job = current_job::CompilerJob
     changed = false
-    @timeit_debug to "remove freeze" begin
+    @tracepoint "remove freeze" begin
 
     for f in functions(mod), bb in blocks(f), inst in instructions(bb)
         if inst isa LLVM.FreezeInst

src/utils.jl

@@ -2,15 +2,6 @@ defs(mod::LLVM.Module)  = filter(f -> !isdeclaration(f), collect(functions(mod))
 decls(mod::LLVM.Module) = filter(f ->  isdeclaration(f) && !LLVM.isintrinsic(f),
                                  collect(functions(mod)))
 
-## timings
-
-const to = TimerOutput()
-
-timings() = (TimerOutputs.print_timer(to); println())
-
-enable_timings() = (TimerOutputs.enable_debug_timings(GPUCompiler); return)
-
-
 ## debug verification
 
 should_verify() = ccall(:jl_is_debugbuild, Cint, ()) == 1 ||