From 5de6be0c9120550aaa15534d0a1466018eff137a Mon Sep 17 00:00:00 2001
From: Fabian Thorand <fabian@channable.com>
Date: Fri, 14 Jan 2022 16:54:02 +0100
Subject: [PATCH] Add note about inefficiency in returnMemoryToOS
---
rts/sm/BlockAlloc.c | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/rts/sm/BlockAlloc.c b/rts/sm/BlockAlloc.c
index d5651db6d8e8..d99d01d743bb 100644
--- a/rts/sm/BlockAlloc.c
+++ b/rts/sm/BlockAlloc.c
@@ -1190,6 +1190,14 @@ uint32_t returnMemoryToOS(uint32_t n /* megablocks */)
uint32_t init_n;
init_n = n;
+ // TODO: This is inefficient because this loop will essentially result in
+ // quadratic runtime behavior: for each call to `freeMBlocks`, the
+ // USE_LARGE_ADDRESS_SPACE implementation of it will walk the internal free
+ // list to insert it at the right position, and thus traverse all previously
+ // inserted values to get to it. We can do better though: both the internal
+ // free list and the `free_mblock_list` here are sorted, so one walk should
+ // be enough.
+
// ToDo: not fair, we free all the memory starting with node 0.
for (node = 0; n > 0 && node < n_numa_nodes; node++) {
bd = free_mblock_list[node];
--
GitLab