py, gc: Further reduce heap fragmentation with new, faster gc alloc.

The heap allocation is now exactly as it was before the "faster gc
alloc" patch, but it's still nearly as fast.  It is fixed by being
careful to always update the "last free block" pointer whenever the heap
changes (eg free or realloc).

Tested on all tests by enabling EXTENSIVE_HEAP_PROFILING in py/gc.c:
old and new allocator have exactly the same behaviour, just the new one
is much faster.

py/gc.c

@@ -46,6 +46,9 @@
 #define DEBUG_printf(...) (void)0
 #endif
 
+// make this 1 to dump the heap each time it changes
+#define EXTENSIVE_HEAP_PROFILING (0)
+
 #define WORDS_PER_BLOCK (4)
 #define BYTES_PER_BLOCK (WORDS_PER_BLOCK * BYTES_PER_WORD)
 #define STACK_SIZE (64) // tunable; minimum is 1
@@ -405,7 +408,8 @@ found:
     // Set last free ATB index to block after last block we found, for start of
     // next scan.  To reduce fragmentation, we only do this if we were looking
     // for a single free block, which guarantees that there are no free blocks
-    // before this one.
+    // before this one.  Also, whenever we free or shink a block we must check
+    // if this index needs adjusting (see gc_realloc and gc_free).
     if (n_free == 1) {
         gc_last_free_atb_index = (i + 1) / BLOCKS_PER_ATB;
     }
@@ -439,6 +443,10 @@ found:
     }
 #endif
 
+    #if EXTENSIVE_HEAP_PROFILING
+    gc_dump_alloc_table();
+    #endif
+
     return ret_ptr;
 }
 
@@ -465,11 +473,20 @@ void gc_free(void *ptr_in) {
     if (VERIFY_PTR(ptr)) {
         mp_uint_t block = BLOCK_FROM_PTR(ptr);
         if (ATB_GET_KIND(block) == AT_HEAD) {
+            // set the last_free pointer to this block if it's earlier in the heap
+            if (block / BLOCKS_PER_ATB < gc_last_free_atb_index) {
+                gc_last_free_atb_index = block / BLOCKS_PER_ATB;
+            }
+
             // free head and all of its tail blocks
             do {
                 ATB_ANY_TO_FREE(block);
                 block += 1;
             } while (ATB_GET_KIND(block) == AT_TAIL);
+
+            #if EXTENSIVE_HEAP_PROFILING
+            gc_dump_alloc_table();
+            #endif
         }
     }
 }
@@ -581,6 +598,16 @@ void *gc_realloc(void *ptr_in, mp_uint_t n_bytes) {
         for (mp_uint_t bl = block + new_blocks; ATB_GET_KIND(bl) == AT_TAIL; bl++) {
             ATB_ANY_TO_FREE(bl);
         }
+
+        // set the last_free pointer to end of this block if it's earlier in the heap
+        if ((block + new_blocks) / BLOCKS_PER_ATB < gc_last_free_atb_index) {
+            gc_last_free_atb_index = (block + new_blocks) / BLOCKS_PER_ATB;
+        }
+
+        #if EXTENSIVE_HEAP_PROFILING
+        gc_dump_alloc_table();
+        #endif
+
         return ptr_in;
     }
 
@@ -595,6 +622,10 @@ void *gc_realloc(void *ptr_in, mp_uint_t n_bytes) {
         // zero out the additional bytes of the newly allocated blocks (see comment above in gc_alloc)
         memset((byte*)ptr_in + n_bytes, 0, new_blocks * BYTES_PER_BLOCK - n_bytes);
 
+        #if EXTENSIVE_HEAP_PROFILING
+        gc_dump_alloc_table();
+        #endif
+
         return ptr_in;
     }
 
@@ -628,9 +659,34 @@ void gc_dump_info() {
 }
 
 void gc_dump_alloc_table(void) {
+    static const mp_uint_t DUMP_BYTES_PER_LINE = 64;
+    #if !EXTENSIVE_HEAP_PROFILING
+    // When comparing heap output we don't want to print the starting
+    // pointer of the heap because it changes from run to run.
     printf("GC memory layout; from %p:", gc_pool_start);
+    #endif
     for (mp_uint_t bl = 0; bl < gc_alloc_table_byte_len * BLOCKS_PER_ATB; bl++) {
-        if (bl % 64 == 0) {
+        if (bl % DUMP_BYTES_PER_LINE == 0) {
+            // a new line of blocks
+            #if EXTENSIVE_HEAP_PROFILING
+            {
+                // check if this line contains only free blocks
+                bool only_free_blocks = true;
+                for (mp_uint_t bl2 = bl; bl2 < gc_alloc_table_byte_len * BLOCKS_PER_ATB && bl2 < bl + DUMP_BYTES_PER_LINE; bl2++) {
+                    if (ATB_GET_KIND(bl2) != AT_FREE) {
+
+                        only_free_blocks = false;
+                        break;
+                    }
+                }
+                if (only_free_blocks) {
+                    // line contains only free blocks, so skip printing it
+                    bl += DUMP_BYTES_PER_LINE - 1;
+                    continue;
+                }
+            }
+            #endif
+            // print header for new line of blocks
             printf("\n%04x: ", (uint)bl);
         }
         int c = ' ';