Fix deferred_codegen registration

[vchuravy]

According to @gbaraldi this was running into issues on 1.11 when code-coverage was enable (since Base dropped the registration of the ccallable)
and on 1.12 this is going to be broken by JuliaLang/julia#56987

[github-actions]

Your PR requires formatting changes to meet the project's style guidelines.
Please consider running Runic (git runic master) to apply these changes.

Click here to view the suggested changes.

diff --git a/src/GPUCompiler.jl b/src/GPUCompiler.jl
index ec33e9a..8f2cf75 100644
--- a/src/GPUCompiler.jl
+++ b/src/GPUCompiler.jl
@@ -66,7 +66,7 @@ function __init__()
     global compile_cache = dir
 
     Tracy.@register_tracepoints()
-    register_deferred_codegen()
+    return register_deferred_codegen()
 end
 
 end # module
diff --git a/src/driver.jl b/src/driver.jl
index 8dbff91..7b0c318 100644
--- a/src/driver.jl
+++ b/src/driver.jl
@@ -154,13 +154,15 @@ end
 # Called from __init__
 # On 1.11+ this is needed due to a Julia bug that drops the pointer when code-coverage is enabled.
 function register_deferred_codegen()
-    @dispose jljit=JuliaOJIT() begin
+    @dispose jljit = JuliaOJIT() begin
         jd = JITDylib(jljit)
 
         address = LLVM.API.LLVMOrcJITTargetAddress(
-            reinterpret(UInt, @cfunction(deferred_codegen, Ptr{Cvoid}, (Ptr{Cvoid},))))
+            reinterpret(UInt, @cfunction(deferred_codegen, Ptr{Cvoid}, (Ptr{Cvoid},)))
+        )
         flags = LLVM.API.LLVMJITSymbolFlags(
-            LLVM.API.LLVMJITSymbolGenericFlagsExported, 0)
+            LLVM.API.LLVMJITSymbolGenericFlagsExported, 0
+        )
         name = mangle(jljit, "deferred_codegen")
         symbol = LLVM.API.LLVMJITEvaluatedSymbol(address, flags)
         map = if LLVM.version() >= v"15"

[vchuravy]

@gbaraldi

With this change: @time_imports

               ┌ 24.8 ms GPUCompiler.__init__() 98.57% compilation time (100% recompilation)
     91.5 ms  GPUCompiler 26.69% compilation time (100% recompilation)

Before this change:

               ┌ 14.9 ms GPUCompiler.__init__() 97.75% compilation time (100% recompilation)
     83.2 ms  GPUCompiler 17.52% compilation time (100% recompilation)

[codecov]

Codecov Report

❌ Patch coverage is 86.66667% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 75.43%. Comparing base (32b4fc8) to head (477be80).

Files with missing lines	Patch %	Lines
src/GPUCompiler.jl	85.71%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #711      +/-   ##
==========================================
+ Coverage   75.27%   75.43%   +0.15%     
==========================================
  Files          24       24              
  Lines        3523     3537      +14     
==========================================
+ Hits         2652     2668      +16     
+ Misses        871      869       -2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[gbaraldi]

Should we ccall it here just to check if it's working? I guess the address is enough but not sure. I guess the test is enough

[vchuravy]

Yeah I think the ccall would not add more information, and if failed lead to a more confusing error.

[gbaraldi]

This isn't pretty, but it's also not very different from what was there before.