Use rr-safe nopl; rdtsc sequence

When running under `rr`, it needs to patch out `rdtsc` to record the values returned.
If this is not possible, `rr` falls back to an expensive signal-based emulation.
As of rr master, a specific `nopl; rdtsc` sequence may be used to guarantee that
`rdtsc` patching is always possible. Use this sequence for uses of rdtsc in our
runtime.

Author: Keno

src/julia_internal.h

@@ -209,8 +209,17 @@ JL_DLLEXPORT void jl_unlock_profile_wr(void) JL_NOTSAFEPOINT JL_NOTSAFEPOINT_LEA
 static inline uint64_t cycleclock(void) JL_NOTSAFEPOINT
 {
 #if defined(_CPU_X86_64_)
+    // This is nopl 0(%rax, %rax, 1), but assembler are incosistent about whether
+    // they emit that as a 4 or 5 byte sequence and we need to be guaranteed to use
+    // the 5 byte one.
+#define NOP5_OVERRIDE_NOP ".byte 0x0f, 0x1f, 0x44, 0x00, 0x00\n\t"
     uint64_t low, high;
-    __asm__ volatile("rdtsc" : "=a"(low), "=d"(high));
+    // This instruction sequence is promised by rr to be patchable. rr can usually
+    // also patch `rdtsc` in regular code, but without the preceeding nop, there could
+    // be an interfering branch into the middle of rr's patch region. Using this
+    // sequence prevents a massive rr-induced slowdown if the compiler happens to emit
+    // an unlucky pattern. See https://github.com/rr-debugger/rr/pull/3580.
+    __asm__ volatile(NOP5_OVERRIDE_NOP "rdtsc" : "=a"(low), "=d"(high));
     return (high << 32) | low;
 #elif defined(_CPU_X86_)
     int64_t ret;