Merge pull request #597 from Keno/kf/noplrdtsc

Use patchable rdtsc sequence to avoid slowdowns under rr

CMakeLists.txt

@@ -78,6 +78,7 @@ set_option(TRACY_NO_VERIFY "Disable zone validation for C API" OFF)
 set_option(TRACY_NO_VSYNC_CAPTURE "Disable capture of hardware Vsync events" OFF)
 set_option(TRACY_NO_FRAME_IMAGE  "Disable the frame image support and its thread" OFF)
 set_option(TRACY_NO_SYSTEM_TRACING  "Disable systrace sampling" OFF)
+set_option(TRACY_PATCHABLE_NOPSLEDS  "Enable nopsleds for efficient patching by system-level tools (e.g. rr)" OFF)
 set_option(TRACY_DELAYED_INIT "Enable delayed initialization of the library (init on first call)" OFF)
 set_option(TRACY_MANUAL_LIFETIME "Enable the manual lifetime management of the profile" OFF)
 set_option(TRACY_FIBERS "Enable fibers support" OFF)

meson.build

@@ -64,6 +64,10 @@ if get_option('tracy_no_system_tracing')
   add_project_arguments('-DTRACY_NO_SYSTEM_TRACING', language : 'cpp')
 endif
 
+if get_option('tracy_no_extra_nopsleds')
+  add_project_arguments('-DTRACY_PATCHABLE_NOPSLEDS', language : 'cpp')
+endif
+
 if get_option('tracy_delayed_init')
   add_project_arguments('-DTRACY_DELAYED_INIT', language : 'cpp')
 endif

public/client/TracyProfiler.hpp

@@ -209,7 +209,22 @@ class Profiler
         if( HardwareSupportsInvariantTSC() )
         {
             uint64_t rax, rdx;
+#ifdef TRACY_PATCHABLE_NOPSLEDS
+            // Some external tooling (such as rr) wants to patch our rdtsc and replace it by a
+            // branch to control the external input seen by a program. This kind of patching is
+            // not generally possible depending on the surrounding code and can lead to significant
+            // slowdowns if the compiler generated unlucky code and rr and tracy are used together.
+            // To avoid this, use the rr-safe `nopl 0(%rax, %rax, 1); rdtsc` instruction sequence,
+            // which rr promises will be patchable independent of the surrounding code.
+            asm volatile (
+                    // This is nopl 0(%rax, %rax, 1), but assemblers are inconsistent about whether
+                    // they emit that as a 4 or 5 byte sequence and we need to be guaranteed to use
+                    // the 5 byte one.
+                    ".byte 0x0f, 0x1f, 0x44, 0x00, 0x00\n\t"
+                    "rdtsc" : "=a" (rax), "=d" (rdx) );
+#else
             asm volatile ( "rdtsc" : "=a" (rax), "=d" (rdx) );
+#endif
             return (int64_t)(( rdx << 32 ) + rax);
         }
 #  else