guix-play/gnu/packages/patches/jemalloc-arm-address-bits.patch
Eric Bavier aa5c206348
gnu: jemalloc: Fix tests for aarch64.
* gnu/packages/patches/jemalloc-arm-address-bits.patch: New patch.
* gnu/packages/jemalloc.scm (jemalloc)[source]: Use it.
* gnu/local.mk (dist_patch_DATA): Add it.
2017-12-10 16:04:16 -06:00

40 lines
1.5 KiB
Diff

From 8cfc9dec37b312a2686f602bbcdd102ca07cca99 Mon Sep 17 00:00:00 2001
From: David Goldblatt <davidgoldblatt@fb.com>
Date: Fri, 29 Sep 2017 13:54:08 -0700
Subject: [PATCH] ARM: Don't extend bit LG_VADDR to compute high address bits.
In userspace ARM on Linux, zero-ing the high bits is the correct way to do this.
This doesn't fix the fact that we currently set LG_VADDR to 48 on ARM, when in
fact larger virtual address sizes are coming soon. We'll cross that bridge when
we come to it.
---
include/jemalloc/internal/rtree.h | 12 ++++++++++++
1 file changed, 12 insertions(+)
diff --git a/include/jemalloc/internal/rtree.h b/include/jemalloc/internal/rtree.h
index b5d4db39..4563db23 100644
--- a/include/jemalloc/internal/rtree.h
+++ b/include/jemalloc/internal/rtree.h
@@ -178,9 +178,21 @@ rtree_leaf_elm_bits_read(tsdn_t *tsdn, rtree_t *rtree, rtree_leaf_elm_t *elm,
JEMALLOC_ALWAYS_INLINE extent_t *
rtree_leaf_elm_bits_extent_get(uintptr_t bits) {
+# ifdef __aarch64__
+ /*
+ * aarch64 doesn't sign extend the highest virtual address bit to set
+ * the higher ones. Instead, the high bits gets zeroed.
+ */
+ uintptr_t high_bit_mask = ((uintptr_t)1 << LG_VADDR) - 1;
+ /* Mask off the slab bit. */
+ uintptr_t low_bit_mask = ~(uintptr_t)1;
+ uintptr_t mask = high_bit_mask & low_bit_mask;
+ return (extent_t *)(bits & mask);
+# else
/* Restore sign-extended high bits, mask slab bit. */
return (extent_t *)((uintptr_t)((intptr_t)(bits << RTREE_NHIB) >>
RTREE_NHIB) & ~((uintptr_t)0x1));
+# endif
}
JEMALLOC_ALWAYS_INLINE szind_t