From 3f8d86814ad40daf5c8cb583f0f66585b057c7b8 Mon Sep 17 00:00:00 2001
From: CodSpeed Bot <no-reply@codspeed.io>
Date: Fri, 24 Jul 2026 15:21:05 +0000
Subject: [PATCH] perf(callgrind): inline CLG_(get_fn_entry) to remove a
 per-basic-block call

CLG_(setup_bbcc) runs on every basic-block execution and called the
out-of-line CLG_(get_fn_entry) in fn.c once per BB (the separate_recursions
> 1 path, active by default). Because the accessor lived in a different
translation unit, the compiler could not inline what is a single array index.

Expose current_fn_active as CLG_(current_fn_active) and move get_fn_entry into
global.h as a static __inline__ accessor after CLG_ASSERT. Disassembly of the
built tool confirms zero remaining call sites; Callgrind output is byte-for-byte
identical (default and --separate-recs=3).
---
 callgrind/fn.c     | 11 ++++-------
 callgrind/global.h | 18 +++++++++++++++++-
 2 files changed, 21 insertions(+), 8 deletions(-)

diff --git a/callgrind/fn.c b/callgrind/fn.c
index f1d96aca1..a117bae08 100644
--- a/callgrind/fn.c
+++ b/callgrind/fn.c
@@ -28,7 +28,10 @@
 
 #define N_INITIAL_FN_ARRAY_SIZE 10071
 
-static fn_array current_fn_active;
+/* The active function array. Defined here but declared in global.h so the
+ * hot-path accessor CLG_(get_fn_entry) can be inlined. */
+fn_array CLG_(current_fn_active);
+#define current_fn_active CLG_(current_fn_active)
 
 /* x86_64 defines 4 variants.  */
 #define MAX_RESOLVE_ADDRS 4
@@ -783,12 +786,6 @@ fn_node* CLG_(get_fn_node)(BB* bb)
  * levels should be separated.
  */
 
-UInt* CLG_(get_fn_entry)(Int n)
-{
-  CLG_ASSERT(n < current_fn_active.size);
-  return current_fn_active.array + n;
-}
-
 void CLG_(init_fn_array)(fn_array* a)
 {
   Int i;
diff --git a/callgrind/global.h b/callgrind/global.h
index 10d4c53b1..bf3dc664f 100644
--- a/callgrind/global.h
+++ b/callgrind/global.h
@@ -730,7 +730,12 @@ void CLG_(init_fn_array)(fn_array*);
 void CLG_(copy_current_fn_array)(fn_array* dst);
 fn_array* CLG_(get_current_fn_array)(void);
 void CLG_(set_current_fn_array)(fn_array*);
-UInt* CLG_(get_fn_entry)(Int n);
+
+/* The active function array (function number -> active count). This is on
+ * the per-BB hot path (setup_bbcc, get_cxt, ...), so the accessor is inlined
+ * (see CLG_(get_fn_entry) below) to avoid a cross-TU function call for what
+ * is a single array index. */
+extern fn_array CLG_(current_fn_active);
 
 void      CLG_(init_obj_table)(void);
 obj_node* CLG_(get_obj_node)(DebugInfo* si);
@@ -853,6 +858,17 @@ extern ULong* CLG_(cost_base);
 #define CLG_ASSERT(cond) tl_assert(cond);
 #endif
 
+/* Inlined hot-path accessor for the active function array. Defined after
+ * CLG_ASSERT so the assertion is available. Called once per basic block from
+ * setup_bbcc() (and from get_cxt/push_cxt), previously as an out-of-line
+ * cross-TU call in fn.c. */
+static __inline__
+UInt* CLG_(get_fn_entry)(Int n)
+{
+  CLG_ASSERT(n < CLG_(current_fn_active).size);
+  return CLG_(current_fn_active).array + n;
+}
+
 /* from debug.c */
 void CLG_(print_bbno)(void);
 void CLG_(print_context)(void);