add pdl.ntrans_children to speed up destroying ndarrays - #421 #451

Basic/Core/pdl.h.PL

@@ -549,6 +549,7 @@ struct pdl {
 
    void *hdrsv; /* "header", settable from Perl */
    PDL_Value value; /* to store at least one value */
+   PDL_Indx ntrans_children; /* CORE21 put next to other trans-tracking stuff */
 };
 
 typedef struct pdl_slice_args {
@@ -563,20 +564,22 @@ typedef struct pdl_slice_args {
  * Some macros for looping over the trans_children of a given PDL
  */
 #define PDL_DECL_CHILDLOOP(p) \
-                int p##__i; pdl_trans_children *p##__c;
+  int p##__i; pdl_trans_children *p##__c; PDL_Indx p##__unfound;
 #define PDL_START_CHILDLOOP(p) \
-                p##__c = &p->trans_children; \
-                do { \
-                        for(p##__i=0; p##__i<PDL_NCHILDREN; p##__i++) { \
-                                if(p##__c->trans[p##__i]) {
+  p##__c = &p->trans_children; p##__unfound = p->ntrans_children; \
+  do { \
+    if (p##__unfound <= 0) break; \
+    for (p##__i=0; p##__i<PDL_NCHILDREN; p##__i++) { \
+      if (p##__c->trans[p##__i]) { \
+        p##__unfound--;
 #define PDL_CHILDLOOP_THISCHILD(p) p##__c->trans[p##__i]
 #define PDL_END_CHILDLOOP(p) \
-                                } \
-                        } \
-                        if(!p##__c) break; \
-                        if(!p##__c->next) break; \
-                        p##__c=p##__c->next; \
-                } while(1);
+      } \
+    } \
+    if (!p##__c) break; \
+    if (!p##__c->next) break; \
+    p##__c=p##__c->next; \
+  } while (1);
 
 #define PDL_USESTRUCTVALUE(it) \
         (it->nbytes <= sizeof(it->value))

Basic/Core/pdlapi.c

@@ -193,6 +193,7 @@ pdl* pdl_pdlnew() {
      PDL_Indx i;
      for(i=0; i<PDL_NCHILDREN; i++) {it->trans_children.trans[i]=NULL;}
      it->trans_children.next = NULL;
+     it->ntrans_children = 0;
      it->magic = 0;
      it->hdrsv = 0;
      PDLDEBUG_f(printf("pdl_pdlnew %p (size=%zu)\n",(void*)it,sizeof(pdl)));
@@ -303,6 +304,7 @@ void pdl__removetrans_children(pdl *it,pdl_trans *trans)
 		if (PDL_CHILDLOOP_THISCHILD(it) != trans) continue;
 		PDL_CHILDLOOP_THISCHILD(it) = NULL;
 		flag = 1;
+		it->ntrans_children--;
 		/* Can't return; might be many times (e.g. $x+$x) */
 	PDL_END_CHILDLOOP(it)
 	/* this might be due to a croak when performing the trans; so
@@ -602,6 +604,7 @@ pdl_error pdl__addchildtrans(pdl *it,pdl_trans *trans)
 	    if (c->next) { c=c->next; continue; } else {
 		for(i=0; i<PDL_NCHILDREN; i++)
 		    if(! c->trans[i]) {
+			it->ntrans_children++;
 			c->trans[i] = trans; return PDL_err;
 		    }
 		break;
@@ -610,6 +613,7 @@ pdl_error pdl__addchildtrans(pdl *it,pdl_trans *trans)
 	c = c->next = malloc(sizeof(pdl_trans_children));
 	if (!c) return pdl_make_error_simple(PDL_EFATAL, "Out of Memory\n");
 	c->trans[0] = trans;
+	it->ntrans_children++;
 	for(i=1; i<PDL_NCHILDREN; i++)
 		c->trans[i] = 0;
 	c->next = 0;

Changes

@@ -1,4 +1,5 @@
 - fix 64-bit problem in Minuit revealed by LTO (#468) - thanks @eli-schwartz for report
+- add pdl.ntrans_children to speed up destroying ndarrays (#421,#451)
 
 2.085_02 2024-03-25
 - PP add loop(n=value) idiom to start not at 0