compat flags explanation

2 files changed, 36 insertions, 6 deletions

diff --git a/pu2/pu2.c b/pu2/pu2.c
index 076261c..85c28f4 100644
--- a/pu2/pu2.c
+++ b/pu2/pu2.c
@@ -647,12 +647,20 @@ unsigned long long rdtsc(void) {
 #endif
 }
 
+#ifndef PU2_BAD_CLOCK
 clock_t ntime() {
     struct timespec t;
     clock_gettime(CLOCK_MONOTONIC_RAW, &t);
 
     return t.tv_nsec + t.tv_sec * 1e9;
 }
+#else
+clock_t ntime() {
+    struct timeval tv;
+    gettimeofday(&tv, NULL);
+    return tv.tv_sec * 1000000000 + tv.tv_usec * 1000;
+}
+#endif
 
 // bench: measure using clock()
 double bench_cycle(int instr) {
@@ -729,8 +737,14 @@ sleep_nano(int nanoseconds) {
     }
 }
 
-[[clang::always_inline]] sleep_micro(int microseconds) { sleep_nano(microseconds * 1e3); }
-[[clang::always_inline]] sleep_milli(int milliseconds) { sleep_nano(milliseconds * 1e6); }
+#ifdef PU2_OLD_CLANG
+    #define inline() __attribute__((always_inline))
+#else
+    #define inline() [[clang::always_inline]]
+#endif
+
+inline() sleep_micro(int microseconds) { sleep_nano(microseconds * 1e3); }
+inline() sleep_milli(int milliseconds) { sleep_nano(milliseconds * 1e6); }
 
 // benchmark: run target for all opcodes and print summary
 benchmark(double (*target)(int), const char *mode_label) {
diff --git a/pu2/pu2.out.c b/pu2/pu2.out.c
index 98f0554..cd403f5 100644
--- a/pu2/pu2.out.c
+++ b/pu2/pu2.out.c
@@ -1,10 +1,26 @@
+// clang-format off
+    // COMPILE TIME FLAGS EXPLANATION
+    // - NO_MMAP                        Prevents usage of mmap+hugepages, which is only supported on linux
+    // - NO_ASM                         Prevents usage of inline assembly, useful for non-x86 platforms
+    // - BAD_CLOCK                      Use if your system clock does not support nanoseconds. Uses microseconds instead.
+    // - OLD_CLANG                      Use if your compiler version is not very up-to-date.
+    // - SPEED_IMPROVEMENT              Almost doubles clock rate for the VM with a few optimizations.
+    // - EXTENDED_INSTRUCTION_SET       Adds a few instructions which decrease token throughput for ir2. The high-level IR emitter will not work otherwise.
+// clang-format on
+
 /* -- BEGIN COMPILE TIME FLAGS -- */
 // #define PU2_NO_MMAP
 // #define PU2_NO_ASM
-#define __PU2_SPEED_IMPROVEMENT
-#define __PU2_EXTENDED_INSTRUCTION_SET
+// #define PU2_BAD_CLOCK
+// #define PU2_OLD_CLANG
+#define PU2_SPEED_IMPROVEMENT
+#define PU2_EXTENDED_INSTRUCTION_SET
 /* -- END COMPILE TIME FLAGS --*/
 
+#ifdef PU2_NO_ASM
+    #warning Enabling this flag will invalidate all clock speed and cycle count measurements (thus, your CPU clock rate will be reported as 0MHz)
+#endif
+
 #define _GNU_SOURCE
 #include <ctype.h>
 #include <sched.h>
@@ -32,9 +48,9 @@
 #define J while
 #define H !--c
 
-#ifndef __PU2_SPEED_IMPROVEMENT
+#ifndef PU2_SPEED_IMPROVEMENT
 
-    #ifndef __PU2_EXTENDED_INSTRUCTION_SET
+    #ifndef PU2_EXTENDED_INSTRUCTION_SET
 
 // clang-format off
     m[S],c,v,a,b,u;pu2(){J(c=m[m[7]++],v=m[m[7]++],c){H?a=v:H?a=m[b]:H?u?m[a]=b:0:H?v=a,a=b,b=v:H?a+=b:H?a-=b:H?a&=b:H?a=!a:H?u=v:H?u=a:H?a=u:H?u=b:H?b=u:A;}}