KMP_HW_SUBSET vs KMP_PLACE_THREADS rival envirables fix

[jlpeyton]

Summary

if both KMP_HW_SUBSET and KMP_PLACE_THREADS are set and KMP_PLACE_THREADS gets
parsed first, then the current environment variable parser rejects both and
neither get used. This patch uses the rivals mechanism that is used for other
environment variable groups (e.g., KMP_STACKSIZE, GOMP_STACKSIZE, OMP_STACKSIZE).
If both are set, then it tells the user that it is ignoring KMP_PLACE_THREADS in
favor of KMP_HW_SUBSET. The message about deprecating KMP_PLACE_THREADS when it
is set is still printed regardless.

[AndreyChurbanov]

runtime/src/kmp_settings.cpp

4749 ↗

(On Diff #116703)

This cast (void*) removes the volatile qualifier, that will make the compiler a bit unhappy.

I think we need const_cast here, e.g. CCAST(kmp_setting_t **, rivals), then it will be implicitly converted to (void*).

[jlpeyton]

Updated with Andrey's suggestion

[AndreyChurbanov]

LGTM

openmp/trunk/runtime/src/kmp_settings.cpp

 // So, let's limit to 5 levels for now
 // The input string is usually short enough, let's use 512 limit for now
 #define MAX_T_LEVEL 5
 #define MAX_STR_LEN 512
 static void __kmp_stg_parse_hw_subset(char const *name, char const *value,
                                       void *data) {
   // Value example: 1s,5c@3,2T
   // Which means "use 1 socket, 5 cores with offset 3, 2 threads per core"
-  static int parsed = 0;
+  kmp_setting_t **rivals = (kmp_setting_t **)data;
   if (strcmp(name, "KMP_PLACE_THREADS") == 0) {
     KMP_INFORM(EnvVarDeprecated, name, "KMP_HW_SUBSET");
-    if (parsed == 1) {
-      return; // already parsed KMP_HW_SUBSET
-    }
+  if (__kmp_stg_check_rivals(name, value, rivals)) {
+    return;
   }
-  parsed = 1;
 
   char *components[MAX_T_LEVEL];
   char const *digits = "0123456789";
   char input[MAX_STR_LEN];
   size_t len = 0, mlen = MAX_STR_LEN;
   int level = 0;
   // Canonize the string (remove spaces, unify delimiters, etc.)
   char *pos = CCAST(char *, value);
#endif
       rivals[i++] = kmp_device_thread_limit;
       rivals[i++] = kmp_all_threads;
       rivals[i++] = NULL;
 
       kmp_device_thread_limit->data = CCAST(kmp_setting_t **, rivals);
       kmp_all_threads->data = CCAST(kmp_setting_t **, rivals);
     }
 
+    { // Initialize KMP_HW_SUBSET and KMP_PLACE_THREADS
+      // 1st priority
+      kmp_setting_t *kmp_hw_subset = __kmp_stg_find("KMP_HW_SUBSET");
+      // 2nd priority
+      kmp_setting_t *kmp_place_threads = __kmp_stg_find("KMP_PLACE_THREADS");
+
+      // !!! volatile keyword is Intel (R) C Compiler bug CQ49908 workaround.
+      static kmp_setting_t *volatile rivals[3];
+      int i = 0;
+
+      rivals[i++] = kmp_hw_subset;
+      rivals[i++] = kmp_place_threads;
+      rivals[i++] = NULL;
+
+      kmp_hw_subset->data = CCAST(kmp_setting_t **, rivals);
+      kmp_place_threads->data = CCAST(kmp_setting_t **, rivals);
+    }
+
 #if KMP_AFFINITY_SUPPORTED
     { // Initialize KMP_AFFINITY, GOMP_CPU_AFFINITY, and OMP_PROC_BIND data.
       kmp_setting_t *kmp_affinity =
           __kmp_stg_find("KMP_AFFINITY"); // 1st priority.
       KMP_DEBUG_ASSERT(kmp_affinity != NULL);
 
 #ifdef KMP_GOMP_COMPAT
       kmp_setting_t *gomp_cpu_affinity =