perf_counters: always-export symbols + benchmark BENCH_STATE knob

Two small follow-ups to the perf_counters baseline:

* Move the perf_counters_register/dump/reset declarations out of
  the BENCH_PROFILE ifdef in the header, and provide always-defined
  no-op bodies in perf_counters.c.  This lets the test ABI export
  the symbols unconditionally (added back to exports-test.list /
  link-test.T) so test_benchmark can dlsym them in any build flavor;
  in non-BENCH_PROFILE builds the bodies are empty.

* Makefile benchmark target gains a BENCH_STATE knob:
    make benchmark BENCH_ROM=foo.j64 BENCH_STATE=foo.state6
  Plumbed through to test_benchmark --load-state, which already
  supports both raw retro_serialize payloads and RetroArch RASTATE
  containers (lands via PR #128).  Lets us profile actual gameplay
  scenes instead of boot/menu idle.

Co-Authored-By: Claude Opus 4.7 <[email protected]>

Author: JoeMatt

Makefile

@@ -893,7 +893,8 @@ benchmark:
 		-o test/tools/test_benchmark test/tools/test_benchmark.c \
 		$(if $(filter Linux,$(shell uname -s)),-ldl)
 	./test/tools/test_benchmark ./$(TARGET) "$(BENCH_ROM)" $(BENCH_FRAMES) \
-		--warmup $(BENCH_WARMUP) --blitter $(BENCH_BLITTER)
+		--warmup $(BENCH_WARMUP) --blitter $(BENCH_BLITTER) \
+		$(if $(BENCH_STATE),--load-state "$(BENCH_STATE)")
 
 print-%:
 	@echo '$*=$($*)'

src/core/perf_counters.c

@@ -1,31 +1,41 @@
 /*
  * perf_counters.c - registry + dump for opt-in instrumentation counters.
- * Only compiled into the program when BENCH_PROFILE is defined; the header
- * provides no-op stubs otherwise.
+ *
+ * The register/dump/reset functions are *always* defined so they can be
+ * exported through the test ABI without conditional linker scripts.
+ * In !BENCH_PROFILE builds the bodies are no-ops and no PERF_COUNTER
+ * calls perf_counters_register, so the registry stays empty.
  */
 #include "perf_counters.h"
 
 #ifdef BENCH_PROFILE
-
 static perf_counter_entry_t *perf_head = (perf_counter_entry_t *)0;
+#endif
 
 void perf_counters_register(perf_counter_entry_t *entry)
 {
+#ifdef BENCH_PROFILE
    if (!entry || entry->next)
       return; /* already linked */
    entry->next = perf_head;
    perf_head = entry;
+#else
+   (void)entry;
+#endif
 }
 
 void perf_counters_reset(void)
 {
+#ifdef BENCH_PROFILE
    perf_counter_entry_t *e;
    for (e = perf_head; e; e = e->next)
       *e->value = 0;
+#endif
 }
 
 void perf_counters_dump(FILE *out)
 {
+#ifdef BENCH_PROFILE
    perf_counter_entry_t *e;
    if (!out)
       out = stderr;
@@ -36,6 +46,7 @@ void perf_counters_dump(FILE *out)
    fprintf(out, "[perf] counter dump:\n");
    for (e = perf_head; e; e = e->next)
       fprintf(out, "[perf]   %-40s %llu\n", e->name, *e->value);
+#else
+   (void)out;
+#endif
 }
-
-#endif /* BENCH_PROFILE */

src/core/perf_counters.h

@@ -33,7 +33,10 @@
 extern "C" {
 #endif
 
-#ifdef BENCH_PROFILE
+/* Registry types and entry points are *always* declared so the test
+ * ABI can export them unconditionally.  When BENCH_PROFILE is undefined
+ * the bodies (in perf_counters.c) become no-ops and no PERF_COUNTER
+ * macro registers anything, so the registry stays empty. */
 
 typedef struct perf_counter_entry
 {
@@ -46,6 +49,8 @@ void perf_counters_register(perf_counter_entry_t *entry);
 void perf_counters_dump(FILE *out);
 void perf_counters_reset(void);
 
+#ifdef BENCH_PROFILE
+
 #define PERF_COUNTER(name) \
    static unsigned long long perf_##name = 0; \
    static perf_counter_entry_t perf_entry_##name = \
@@ -72,10 +77,6 @@ void perf_counters_reset(void);
 #define PERF_INC(name)        (0)
 #define PERF_ADD(name, n)     ((void)(n), 0)
 
-/* Stubs so callers don't need their own #ifdef around dump/reset. */
-static __inline void perf_counters_dump(FILE *out)  { (void)out; }
-static __inline void perf_counters_reset(void)      { }
-
 #endif /* BENCH_PROFILE */
 
 #ifdef __cplusplus