From fa8294ebe79adf52f9bf940512caf0023958c065 Mon Sep 17 00:00:00 2001
From: Patric Stout <truebrain@openttd.org>
Date: Sun, 21 Jan 2024 16:38:15 +0100
Subject: [PATCH] Remove: rdtsc and TIC/TOC based on CPU ticks (#11856)

Use TIC/TOC based on std::chrono instead. This information is also
easier to compare with others, as although it depends on CPU, it
means a bit more if "yours takes 4ms and mine takes 10ms".
---
 src/cpu.cpp | 80 -----------------------------------------------------
 src/cpu.h   |  6 ----
 src/debug.h | 23 +++------------
 3 files changed, 4 insertions(+), 105 deletions(-)

diff --git a/src/cpu.cpp b/src/cpu.cpp
index 239c0e2347..f1b498b6c1 100644
--- a/src/cpu.cpp
+++ b/src/cpu.cpp
@@ -12,86 +12,6 @@
 
 #include "safeguards.h"
 
-#undef RDTSC_AVAILABLE
-
-/* rdtsc for MSC_VER, uses simple inline assembly, or _rdtsc
- * from external win64.asm because VS2005 does not support inline assembly */
-#if defined(_MSC_VER) && !defined(RDTSC_AVAILABLE)
-#include <intrin.h>
-#include <windows.h>
-uint64_t ottd_rdtsc()
-{
-#if defined(_M_ARM)
-	return __rdpmccntr64();
-#elif defined(_M_ARM64)
-	return _ReadStatusReg(ARM64_PMCCNTR_EL0);
-#else
-	return __rdtsc();
-#endif
-}
-#define RDTSC_AVAILABLE
-#endif
-
-/* rdtsc for all other *nix-en (hopefully). Use GCC syntax */
-#if (defined(__i386__) || defined(__x86_64__)) && !defined(RDTSC_AVAILABLE)
-uint64_t ottd_rdtsc()
-{
-	uint32_t high, low;
-	__asm__ __volatile__ ("rdtsc" : "=a" (low), "=d" (high));
-	return ((uint64_t)high << 32) | low;
-}
-# define RDTSC_AVAILABLE
-#endif
-
-/* rdtsc for PPC which has this not */
-#if (defined(__POWERPC__) || defined(__powerpc__)) && !defined(RDTSC_AVAILABLE)
-uint64_t ottd_rdtsc()
-{
-	uint32_t high = 0, high2 = 0, low;
-	/* PPC does not have rdtsc, so we cheat by reading the two 32-bit time-counters
-	 * it has, 'Move From Time Base (Upper)'. Since these are two reads, in the
-	 * very unlikely event that the lower part overflows to the upper part while we
-	 * read it; we double-check and reread the registers */
-	asm volatile (
-				  "mftbu %0\n"
-				  "mftb %1\n"
-				  "mftbu %2\n"
-				  "cmpw %3,%4\n"
-				  "bne- $-16\n"
-				  : "=r" (high), "=r" (low), "=r" (high2)
-				  : "0" (high), "2" (high2)
-				  );
-	return ((uint64_t)high << 32) | low;
-}
-# define RDTSC_AVAILABLE
-#endif
-
-/* rdtsc for MCST Elbrus 2000 */
-#if defined(__e2k__) && !defined(RDTSC_AVAILABLE)
-uint64_t ottd_rdtsc()
-{
-	uint64_t dst;
-# pragma asm_inline
-	asm("rrd %%clkr, %0" : "=r" (dst));
-	return dst;
-}
-# define RDTSC_AVAILABLE
-#endif
-
-#if defined(__EMSCRIPTEN__) && !defined(RDTSC_AVAILABLE)
-/* On emscripten doing TIC/TOC would be ill-advised */
-uint64_t ottd_rdtsc() {return 0;}
-# define RDTSC_AVAILABLE
-#endif
-
-/* In all other cases we have no support for rdtsc. No major issue,
- * you just won't be able to profile your code with TIC()/TOC() */
-#if !defined(RDTSC_AVAILABLE)
-#warning "(non-fatal) No support for rdtsc(), you won't be able to profile with TIC/TOC"
-uint64_t ottd_rdtsc() {return 0;}
-#endif
-
-
 /**
  * Definitions for CPU detection:
  *
diff --git a/src/cpu.h b/src/cpu.h
index edaa8d7c17..116ec8b0ea 100644
--- a/src/cpu.h
+++ b/src/cpu.h
@@ -10,12 +10,6 @@
 #ifndef CPU_H
 #define CPU_H
 
-/**
- * Get the tick counter from the CPU (high precision timing).
- * @return The count.
- */
-uint64_t ottd_rdtsc();
-
 /**
  * Get the CPUID information from the CPU.
  * @param info The retrieved info. All zeros on architectures without CPUID.
diff --git a/src/debug.h b/src/debug.h
index a7e6300683..c6433df546 100644
--- a/src/debug.h
+++ b/src/debug.h
@@ -63,7 +63,8 @@ std::string GetDebugString();
 /* Shorter form for passing filename and linenumber */
 #define FILE_LINE __FILE__, __LINE__
 
-/* Used for profiling
+/**
+ * Used for profiling.
  *
  * Usage:
  * TIC();
@@ -84,30 +85,14 @@ std::string GetDebugString();
  * TIC() / TOC() creates its own block, so make sure not the mangle
  *  it with another block.
  *
- * The output is counted in CPU cycles, and not comparable across
- *  machines. Mainly useful for local optimisations.
+ * The output is counted in microseconds. Mainly useful for local optimisations.
  **/
 #define TIC() {\
-	uint64_t _xxx_ = ottd_rdtsc();\
-	static uint64_t _sum_ = 0;\
-	static uint32_t _i_ = 0;
-
-#define TOC(str, count)\
-	_sum_ += ottd_rdtsc() - _xxx_;\
-	if (++_i_ == count) {\
-		Debug(misc, 0, "[{}] {} [avg: {:.1f}]", str, _sum_, _sum_/(double)_i_);\
-		_i_ = 0;\
-		_sum_ = 0;\
-	}\
-}
-
-/* Chrono based version. The output is in microseconds. */
-#define TICC() {\
 	auto _start_ = std::chrono::high_resolution_clock::now();\
 	static uint64_t _sum_ = 0;\
 	static uint32_t _i_ = 0;
 
-#define TOCC(str, _count_)\
+#define TOC(str, _count_)\
 	_sum_ += (std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::high_resolution_clock::now() - _start_)).count();\
 	if (++_i_ == _count_) {\
 		Debug(misc, 0, "[{}] {} us [avg: {:.1f} us]", str, _sum_, _sum_/(double)_i_);\