schwab: full branchless parts on partitioning now - speed gain but minimal... need like 5-6 of these

This reverts commit bf9c22f4ecd37039fe69142e9b72db6eb50b82a9.

schwab_sort.h

@@ -158,11 +158,22 @@ static inline int schwab_partition(
 		/* TODO: should be copy of whole element when not just uint32s! */
 		uint32_t curr = arr[b3];
 
-		/* TODO: We can do "ILP-memcpy"s here:
-		 *
-		 * Key from b2->b3, value from b2->b3, key from b1->b2, value from b1... etc
-		 * This is likely faster than calling a memcpy if we code this for not just uint32s!
-		 */
+		/* Full branchless - see below for branch alternative */
+		int where = (curr < klo) ? 0 :
+			((curr < kmid) ? 1 : 2);
+		int target = (curr < klo) ? b0 :
+			((curr < kmid) ? b1 : b2);
+
+		arr[b3] = arr[b2];
+		arr[b2] = (where == 2) ? arr[b2] : arr[b1];
+		arr[b1] = (where == 1) ? arr[b1] : arr[b0];
+
+		++b2;
+		b1 += (where < 2);
+		b0 += (where < 1);
+		arr[target] = curr;
+
+		/* Same as this would have been:
 		if(curr < klo) {
 			arr[b3] = arr[b2];
 			arr[b2] = arr[b1];
@@ -176,12 +187,13 @@ static inline int schwab_partition(
 			arr[b3] = arr[b2];
 			arr[b2] = arr[b1];
 			arr[b1] = curr;
-			++b1; ++b2;
+			++b1;
 		} else {
 			arr[b3] = arr[b2];
 			arr[b2] = curr;
-			++b2;
 		}
+		++b2;
+		*/
 	}
 
 	/* [*] Swap the chosen pivot to begin of last block */