Emit safepoints at function entry (JuliaLang#41616)

* Emit safepoints at function entry

* Make safepoint emission on function entry a codegen feature

* Hoist signal page lookup outside fence

* Update src/cgutils.cpp

* Fix rebase

base/reflection.jl

@@ -1092,6 +1092,7 @@ struct CodegenParams
     prefer_specsig::Cint
     gnu_pubnames::Cint
     debug_info_kind::Cint
+    safepoint_on_entry::Cint
 
     lookup::Ptr{Cvoid}
 
@@ -1100,12 +1101,14 @@ struct CodegenParams
     function CodegenParams(; track_allocations::Bool=true, code_coverage::Bool=true,
                    prefer_specsig::Bool=false,
                    gnu_pubnames=true, debug_info_kind::Cint = default_debug_info_kind(),
+                   safepoint_on_entry::Bool=true,
                    lookup::Ptr{Cvoid}=cglobal(:jl_rettype_inferred),
                    generic_context = nothing)
         return new(
             Cint(track_allocations), Cint(code_coverage),
             Cint(prefer_specsig),
             Cint(gnu_pubnames), debug_info_kind,
+            Cint(safepoint_on_entry),
             lookup, generic_context)
     end
 end

src/cgutils.cpp

@@ -3933,7 +3933,6 @@ static Value *emit_defer_signal(jl_codectx_t &ctx)
     return ctx.builder.CreateInBoundsGEP(ctx.types().T_sigatomic, ptls, ArrayRef<Value*>(offset), "jl_defer_signal");
 }
 
-
 #ifndef JL_NDEBUG
 static int compare_cgparams(const jl_cgparams_t *a, const jl_cgparams_t *b)
 {

src/codegen.cpp

@@ -1279,6 +1279,7 @@ extern "C" {
         1,
 #endif
         (int) DICompileUnit::DebugEmissionKind::FullDebug,
+        1,
         jl_rettype_inferred, NULL };
 }
 
@@ -7805,7 +7806,11 @@ static jl_llvm_functions_t
         ctx.builder.CreateAlignedStore(load_world, world_age_field, Align(sizeof(size_t)));
     }
 
-    // step 11b. Do codegen in control flow order
+    // step 11b. Emit the entry safepoint
+    if (JL_FEAT_TEST(ctx, safepoint_on_entry))
+        emit_gc_safepoint(ctx.builder, get_current_ptls(ctx), ctx.tbaa().tbaa_const);
+
+    // step 11c. Do codegen in control flow order
     std::vector<int> workstack;
     std::map<int, BasicBlock*> BB;
     std::map<size_t, BasicBlock*> come_from_bb;

src/julia.h

@@ -2247,9 +2247,11 @@ typedef struct {
 
     // controls the emission of debug-info. mirrors the clang options
     int gnu_pubnames;       // can we emit the gnu pubnames debuginfo
-    int debug_info_kind; // Enum for line-table-only, line-directives-only,
+    int debug_info_kind;    // Enum for line-table-only, line-directives-only,
                             // limited, standalone
 
+    int safepoint_on_entry; // Emit a safepoint on entry to each function
+
     // Cache access. Default: jl_rettype_inferred.
     jl_codeinstance_lookup_t lookup;
 

test/compiler/codegen.jl

@@ -15,9 +15,12 @@ function libjulia_codegen_name()
     is_debug_build ? "libjulia-codegen-debug" : "libjulia-codegen"
 end
 
-# `_dump_function` might be more efficient but it doesn't really matter here...
-get_llvm(@nospecialize(f), @nospecialize(t), raw=true, dump_module=false, optimize=true) =
-    sprint(code_llvm, f, t, raw, dump_module, optimize)
+# The tests below assume a certain format and safepoint_on_entry=true breaks that.
+function get_llvm(@nospecialize(f), @nospecialize(t), raw=true, dump_module=false, optimize=true)
+    params = Base.CodegenParams(safepoint_on_entry=false)
+    d = InteractiveUtils._dump_function(f, t, false, false, !raw, dump_module, :att, optimize, :none, false, params)
+    sprint(print, d)
+end
 
 if !is_debug_build && opt_level > 0
     # Make sure getptls call is removed at IR level with optimization on