From 89fdb8cda43807a267dc00b76d92cc4f924a68df Mon Sep 17 00:00:00 2001
From: a <a@b.c>
Date: Thu, 6 Oct 2016 16:35:33 +0200
Subject: [PATCH] Add the cpuid serializing instruction to rdtscp. Thanks to
 intel docs.

---
 m4/cflagz.m4 | 10 +++++++---
 src/cpu.c    | 13 +++++++++++--
 2 files changed, 18 insertions(+), 5 deletions(-)

diff --git a/m4/cflagz.m4 b/m4/cflagz.m4
index 49e38bf..de78c52 100644
--- a/m4/cflagz.m4
+++ b/m4/cflagz.m4
@@ -54,9 +54,13 @@ AC_DEFUN([TEST_ASSEMBLY],[
 
   AC_COMPILE_IFELSE([
     AC_LANG_PROGRAM([[
-      #include <stdint.h>]],[[
-      uintmax_t x;
-      __asm__ __volatile__ (".byte 0x0f, 0x31" : "=A" (x));
+      #include <stdio.h>]],[[
+      unsigned int x = 1, z = 2;
+      __asm__ __volatile__ (
+        "addl %%ebx, %%eax"
+        : "=a"(x)
+        : "a"(x), "b"(z)
+      );
     ]])
   ],[supportz_assembly=yes],[supportz_assembly=no])
 
diff --git a/src/cpu.c b/src/cpu.c
index 09ddc38..96e05bf 100644
--- a/src/cpu.c
+++ b/src/cpu.c
@@ -244,6 +244,12 @@ get_cpu_clock_speed(char *str1) {
 
 
 #elif defined(__x86_64__)
+/* 
+ * Thanks to Intel docs for pointing out
+ * "Out of Order Execution", added
+ * cpuid after reading it and edited the rdtscp
+ * code according to the docs
+ * https://www-ssl.intel.com/content/dam/www/public/us/en/documents/white-papers/ia-32-ia-64-benchmark-code-execution-paper.pdf */
 static __inline__ uintmax_t 
 rdtsc(void) {
   unsigned int tickhi = 0, ticklo = 0;
@@ -274,8 +280,11 @@ rdtsc(void) {
   if (0 != (edx & (1 << 27))) {
     __asm__ __volatile__ (
       "rdtscp\n\t"
-      : "=a"(ticklo), "=d"(tickhi)
-      :: "%rbx", "%rcx"
+      "mov %%edx, %0\n\t"
+      "mov %%eax, %1\n\t"
+      "cpuid\n\t"
+      : "=r"(tickhi), "=r"(ticklo)
+      :: "%rax", "%rbx", "%rcx", "%rdx"
     );
   }
 
-- 
2.11.0