Fast-path improvement: reduce # of branches and unnecessary operations.

- Combine multiple runtime branches into a single malloc_slow check.
- Avoid calling arena_choose / size2index / index2size on fast path.
- A few micro optimizations.

4 files changed, 150 insertions, 97 deletions

diff --git a/include/jemalloc/internal/arena.h b/include/jemalloc/internal/arena.h
index 9e2375c..9715ad9 100644
--- a/include/jemalloc/internal/arena.h
+++ b/include/jemalloc/internal/arena.h
@@ -461,8 +461,10 @@ extern arena_dalloc_junk_small_t *arena_dalloc_junk_small;
 void	arena_dalloc_junk_small(void *ptr, arena_bin_info_t *bin_info);
 #endif
 void	arena_quarantine_junk_small(void *ptr, size_t usize);
-void	*arena_malloc_small(arena_t *arena, size_t size, bool zero);
-void	*arena_malloc_large(arena_t *arena, size_t size, bool zero);
+void	*arena_malloc_small(arena_t *arena, size_t size, szind_t ind,
+    bool zero);
+void	*arena_malloc_large(arena_t *arena, size_t size, szind_t ind,
+    bool zero);
 void	*arena_palloc(tsd_t *tsd, arena_t *arena, size_t usize,
     size_t alignment, bool zero, tcache_t *tcache);
 void	arena_prof_promoted(const void *ptr, size_t size);
@@ -558,11 +560,11 @@ prof_tctx_t	*arena_prof_tctx_get(const void *ptr);
 void	arena_prof_tctx_set(const void *ptr, size_t usize, prof_tctx_t *tctx);
 void	arena_prof_tctx_reset(const void *ptr, size_t usize,
     const void *old_ptr, prof_tctx_t *old_tctx);
-void	*arena_malloc(tsd_t *tsd, arena_t *arena, size_t size, bool zero,
-    tcache_t *tcache);
+void	*arena_malloc(tsd_t *tsd, arena_t *arena, size_t size, szind_t ind,
+    bool zero, tcache_t *tcache, bool slow_path);
 arena_t	*arena_aalloc(const void *ptr);
 size_t	arena_salloc(const void *ptr, bool demote);
-void	arena_dalloc(tsd_t *tsd, void *ptr, tcache_t *tcache);
+void	arena_dalloc(tsd_t *tsd, void *ptr, tcache_t *tcache, bool slow_path);
 void	arena_sdalloc(tsd_t *tsd, void *ptr, size_t size, tcache_t *tcache);
 #endif
 
@@ -1158,34 +1160,34 @@ arena_prof_tctx_reset(const void *ptr, size_t usize, const void *old_ptr,
 }
 
 JEMALLOC_ALWAYS_INLINE void *
-arena_malloc(tsd_t *tsd, arena_t *arena, size_t size, bool zero,
-    tcache_t *tcache)
+arena_malloc(tsd_t *tsd, arena_t *arena, size_t size, szind_t ind,
+    bool zero, tcache_t *tcache, bool slow_path)
 {
 
 	assert(size != 0);
 
+	if (likely(tcache != NULL)) {
+		if (likely(size <= SMALL_MAXCLASS)) {
+			return (tcache_alloc_small(tsd, arena, tcache, size,
+			    ind, zero, slow_path));
+		}
+		if (likely(size <= tcache_maxclass)) {
+			return (tcache_alloc_large(tsd, arena, tcache, size,
+			    ind, zero, slow_path));
+		}
+		/* (size > tcache_maxclass) case falls through. */
+		assert(size > tcache_maxclass);
+	}
+
 	arena = arena_choose(tsd, arena);
 	if (unlikely(arena == NULL))
 		return (NULL);
 
-	if (likely(size <= SMALL_MAXCLASS)) {
-		if (likely(tcache != NULL)) {
-			return (tcache_alloc_small(tsd, arena, tcache, size,
-			    zero));
-		} else
-			return (arena_malloc_small(arena, size, zero));
-	} else if (likely(size <= large_maxclass)) {
-		/*
-		 * Initialize tcache after checking size in order to avoid
-		 * infinite recursion during tcache initialization.
-		 */
-		if (likely(tcache != NULL) && size <= tcache_maxclass) {
-			return (tcache_alloc_large(tsd, arena, tcache, size,
-			    zero));
-		} else
-			return (arena_malloc_large(arena, size, zero));
-	} else
-		return (huge_malloc(tsd, arena, size, zero, tcache));
+	if (likely(size <= SMALL_MAXCLASS))
+		return (arena_malloc_small(arena, size, ind, zero));
+	if (likely(size <= large_maxclass))
+		return (arena_malloc_large(arena, size, ind, zero));
+	return (huge_malloc(tsd, arena, size, zero, tcache));
 }
 
 JEMALLOC_ALWAYS_INLINE arena_t *
@@ -1251,7 +1253,7 @@ arena_salloc(const void *ptr, bool demote)
 }
 
 JEMALLOC_ALWAYS_INLINE void
-arena_dalloc(tsd_t *tsd, void *ptr, tcache_t *tcache)
+arena_dalloc(tsd_t *tsd, void *ptr, tcache_t *tcache, bool slow_path)
 {
 	arena_chunk_t *chunk;
 	size_t pageind, mapbits;
@@ -1268,7 +1270,8 @@ arena_dalloc(tsd_t *tsd, void *ptr, tcache_t *tcache)
 			if (likely(tcache != NULL)) {
 				szind_t binind = arena_ptr_small_binind_get(ptr,
 				    mapbits);
-				tcache_dalloc_small(tsd, tcache, ptr, binind);
+				tcache_dalloc_small(tsd, tcache, ptr, binind,
+				    slow_path);
 			} else {
 				arena_dalloc_small(extent_node_arena_get(
 				    &chunk->node), chunk, ptr, pageind);
@@ -1283,7 +1286,7 @@ arena_dalloc(tsd_t *tsd, void *ptr, tcache_t *tcache)
 			if (likely(tcache != NULL) && size - large_pad <=
 			    tcache_maxclass) {
 				tcache_dalloc_large(tsd, tcache, ptr, size -
-				    large_pad);
+				    large_pad, slow_path);
 			} else {
 				arena_dalloc_large(extent_node_arena_get(
 				    &chunk->node), chunk, ptr);
@@ -1319,7 +1322,7 @@ arena_sdalloc(tsd_t *tsd, void *ptr, size_t size, tcache_t *tcache)
 			/* Small allocation. */
 			if (likely(tcache != NULL)) {
 				szind_t binind = size2index(size);
-				tcache_dalloc_small(tsd, tcache, ptr, binind);
+				tcache_dalloc_small(tsd, tcache, ptr, binind, true);
 			} else {
 				size_t pageind = ((uintptr_t)ptr -
 				    (uintptr_t)chunk) >> LG_PAGE;
@@ -1331,7 +1334,7 @@ arena_sdalloc(tsd_t *tsd, void *ptr, size_t size, tcache_t *tcache)
 			    PAGE_MASK) == 0);
 
 			if (likely(tcache != NULL) && size <= tcache_maxclass)
-				tcache_dalloc_large(tsd, tcache, ptr, size);
+				tcache_dalloc_large(tsd, tcache, ptr, size, true);
 			else {
 				arena_dalloc_large(extent_node_arena_get(
 				    &chunk->node), chunk, ptr);
diff --git a/include/jemalloc/internal/jemalloc_internal.h.in b/include/jemalloc/internal/jemalloc_internal.h.in
index 654cd08..d31da4c 100644
--- a/include/jemalloc/internal/jemalloc_internal.h.in
+++ b/include/jemalloc/internal/jemalloc_internal.h.in
@@ -437,7 +437,7 @@ extern unsigned		ncpus;
  * index2size_tab encodes the same information as could be computed (at
  * unacceptable cost in some code paths) by index2size_compute().
  */
-extern size_t const	index2size_tab[NSIZES];
+extern size_t const	index2size_tab[NSIZES+1];
 /*
  * size2index_tab is a compact lookup table that rounds request sizes up to
  * size classes.  In order to reduce cache footprint, the table is compressed,
@@ -624,7 +624,7 @@ JEMALLOC_ALWAYS_INLINE size_t
 index2size(szind_t index)
 {
 
-	assert(index < NSIZES);
+	assert(index <= NSIZES);
 	return (index2size_lookup(index));
 }
 
@@ -823,12 +823,14 @@ arena_get(tsd_t *tsd, unsigned ind, bool init_if_missing,
 #ifndef JEMALLOC_ENABLE_INLINE
 arena_t	*iaalloc(const void *ptr);
 size_t	isalloc(const void *ptr, bool demote);
-void	*iallocztm(tsd_t *tsd, size_t size, bool zero, tcache_t *tcache,
-    bool is_metadata, arena_t *arena);
-void	*imalloct(tsd_t *tsd, size_t size, tcache_t *tcache, arena_t *arena);
-void	*imalloc(tsd_t *tsd, size_t size);
-void	*icalloct(tsd_t *tsd, size_t size, tcache_t *tcache, arena_t *arena);
-void	*icalloc(tsd_t *tsd, size_t size);
+void	*iallocztm(tsd_t *tsd, size_t size, szind_t ind, bool zero,
+    tcache_t *tcache, bool is_metadata, arena_t *arena, bool slow_path);
+void	*imalloct(tsd_t *tsd, size_t size, szind_t ind, tcache_t *tcache,
+    arena_t *arena);
+void	*imalloc(tsd_t *tsd, size_t size, szind_t ind, bool slow_path);
+void	*icalloct(tsd_t *tsd, size_t size, szind_t ind, tcache_t *tcache,
+    arena_t *arena);
+void	*icalloc(tsd_t *tsd, size_t size, szind_t ind);
 void	*ipallocztm(tsd_t *tsd, size_t usize, size_t alignment, bool zero,
     tcache_t *tcache, bool is_metadata, arena_t *arena);
 void	*ipalloct(tsd_t *tsd, size_t usize, size_t alignment, bool zero,
@@ -837,10 +839,11 @@ void	*ipalloc(tsd_t *tsd, size_t usize, size_t alignment, bool zero);
 size_t	ivsalloc(const void *ptr, bool demote);
 size_t	u2rz(size_t usize);
 size_t	p2rz(const void *ptr);
-void	idalloctm(tsd_t *tsd, void *ptr, tcache_t *tcache, bool is_metadata);
+void	idalloctm(tsd_t *tsd, void *ptr, tcache_t *tcache, bool is_metadata,
+    bool slow_path);
 void	idalloct(tsd_t *tsd, void *ptr, tcache_t *tcache);
 void	idalloc(tsd_t *tsd, void *ptr);
-void	iqalloc(tsd_t *tsd, void *ptr, tcache_t *tcache);
+void	iqalloc(tsd_t *tsd, void *ptr, tcache_t *tcache, bool slow_path);
 void	isdalloct(tsd_t *tsd, void *ptr, size_t size, tcache_t *tcache);
 void	isqalloc(tsd_t *tsd, void *ptr, size_t size, tcache_t *tcache);
 void	*iralloct_realign(tsd_t *tsd, void *ptr, size_t oldsize, size_t size,
@@ -881,14 +884,14 @@ isalloc(const void *ptr, bool demote)
 }
 
 JEMALLOC_ALWAYS_INLINE void *
-iallocztm(tsd_t *tsd, size_t size, bool zero, tcache_t *tcache, bool is_metadata,
-    arena_t *arena)
+iallocztm(tsd_t *tsd, size_t size, szind_t ind, bool zero, tcache_t *tcache,
+    bool is_metadata, arena_t *arena, bool slow_path)
 {
 	void *ret;
 
 	assert(size != 0);
 
-	ret = arena_malloc(tsd, arena, size, zero, tcache);
+	ret = arena_malloc(tsd, arena, size, ind, zero, tcache, slow_path);
 	if (config_stats && is_metadata && likely(ret != NULL)) {
 		arena_metadata_allocated_add(iaalloc(ret), isalloc(ret,
 		    config_prof));
@@ -897,31 +900,33 @@ iallocztm(tsd_t *tsd, size_t size, bool zero, tcache_t *tcache, bool is_metadata
 }
 
 JEMALLOC_ALWAYS_INLINE void *
-imalloct(tsd_t *tsd, size_t size, tcache_t *tcache, arena_t *arena)
+imalloct(tsd_t *tsd, size_t size, szind_t ind, tcache_t *tcache, arena_t *arena)
 {
 
-	return (iallocztm(tsd, size, false, tcache, false, arena));
+	return (iallocztm(tsd, size, ind, false, tcache, false, arena, true));
 }
 
 JEMALLOC_ALWAYS_INLINE void *
-imalloc(tsd_t *tsd, size_t size)
+imalloc(tsd_t *tsd, size_t size, szind_t ind, bool slow_path)
 {
 
-	return (iallocztm(tsd, size, false, tcache_get(tsd, true), false, NULL));
+	return (iallocztm(tsd, size, ind, false, tcache_get(tsd, true), false,
+	    NULL, slow_path));
 }
 
 JEMALLOC_ALWAYS_INLINE void *
-icalloct(tsd_t *tsd, size_t size, tcache_t *tcache, arena_t *arena)
+icalloct(tsd_t *tsd, size_t size, szind_t ind, tcache_t *tcache, arena_t *arena)
 {
 
-	return (iallocztm(tsd, size, true, tcache, false, arena));
+	return (iallocztm(tsd, size, ind, true, tcache, false, arena, true));
 }
 
 JEMALLOC_ALWAYS_INLINE void *
-icalloc(tsd_t *tsd, size_t size)
+icalloc(tsd_t *tsd, size_t size, szind_t ind)
 {
 
-	return (iallocztm(tsd, size, true, tcache_get(tsd, true), false, NULL));
+	return (iallocztm(tsd, size, ind, true, tcache_get(tsd, true), false,
+	    NULL, true));
 }
 
 JEMALLOC_ALWAYS_INLINE void *
@@ -997,7 +1002,8 @@ p2rz(const void *ptr)
 }
 
 JEMALLOC_ALWAYS_INLINE void
-idalloctm(tsd_t *tsd, void *ptr, tcache_t *tcache, bool is_metadata)
+idalloctm(tsd_t *tsd, void *ptr, tcache_t *tcache, bool is_metadata,
+    bool slow_path)
 {
 
 	assert(ptr != NULL);
@@ -1006,31 +1012,31 @@ idalloctm(tsd_t *tsd, void *ptr, tcache_t *tcache, bool is_metadata)
 		    config_prof));
 	}
 
-	arena_dalloc(tsd, ptr, tcache);
+	arena_dalloc(tsd, ptr, tcache, slow_path);
 }
 
 JEMALLOC_ALWAYS_INLINE void
 idalloct(tsd_t *tsd, void *ptr, tcache_t *tcache)
 {
 
-	idalloctm(tsd, ptr, tcache, false);
+	idalloctm(tsd, ptr, tcache, false, true);
 }
 
 JEMALLOC_ALWAYS_INLINE void
 idalloc(tsd_t *tsd, void *ptr)
 {
 
-	idalloctm(tsd, ptr, tcache_get(tsd, false), false);
+	idalloctm(tsd, ptr, tcache_get(tsd, false), false, true);
 }
 
 JEMALLOC_ALWAYS_INLINE void
-iqalloc(tsd_t *tsd, void *ptr, tcache_t *tcache)
+iqalloc(tsd_t *tsd, void *ptr, tcache_t *tcache, bool slow_path)
 {
 
-	if (config_fill && unlikely(opt_quarantine))
+	if (slow_path && config_fill && unlikely(opt_quarantine))
 		quarantine(tsd, ptr);
 	else
-		idalloctm(tsd, ptr, tcache, false);
+		idalloctm(tsd, ptr, tcache, false, slow_path);
 }
 
 JEMALLOC_ALWAYS_INLINE void
diff --git a/include/jemalloc/internal/prof.h b/include/jemalloc/internal/prof.h
index e5198c3..a25502a 100644
--- a/include/jemalloc/internal/prof.h
+++ b/include/jemalloc/internal/prof.h
@@ -436,16 +436,16 @@ prof_sample_accum_update(tsd_t *tsd, size_t usize, bool update,
 	cassert(config_prof);
 
 	tdata = prof_tdata_get(tsd, true);
-	if ((uintptr_t)tdata <= (uintptr_t)PROF_TDATA_STATE_MAX)
+	if (unlikely((uintptr_t)tdata <= (uintptr_t)PROF_TDATA_STATE_MAX))
 		tdata = NULL;
 
 	if (tdata_out != NULL)
 		*tdata_out = tdata;
 
-	if (tdata == NULL)
+	if (unlikely(tdata == NULL))
 		return (true);
 
-	if (tdata->bytes_until_sample >= usize) {
+	if (likely(tdata->bytes_until_sample >= usize)) {
 		if (update)
 			tdata->bytes_until_sample -= usize;
 		return (true);
diff --git a/include/jemalloc/internal/tcache.h b/include/jemalloc/internal/tcache.h
index 5079cd2..c292140 100644
--- a/include/jemalloc/internal/tcache.h
+++ b/include/jemalloc/internal/tcache.h
@@ -70,6 +70,13 @@ struct tcache_bin_s {
 	int		low_water;	/* Min # cached since last GC. */
 	unsigned	lg_fill_div;	/* Fill (ncached_max >> lg_fill_div). */
 	unsigned	ncached;	/* # of cached objects. */
+	/*
+	 * To make use of adjacent cacheline prefetch, the items in the avail
+	 * stack goes to higher address for newer allocations.  avail points
+	 * just above the available space, which means that
+	 * avail[-ncached, ... 1] are available items and the lowest item will
+	 * be allocated first.
+	 */
 	void		**avail;	/* Stack of available objects. */
 };
 
@@ -126,7 +133,7 @@ extern tcaches_t	*tcaches;
 size_t	tcache_salloc(const void *ptr);
 void	tcache_event_hard(tsd_t *tsd, tcache_t *tcache);
 void	*tcache_alloc_small_hard(tsd_t *tsd, arena_t *arena, tcache_t *tcache,
-    tcache_bin_t *tbin, szind_t binind);
+    tcache_bin_t *tbin, szind_t binind, bool *tcache_success);
 void	tcache_bin_flush_small(tsd_t *tsd, tcache_t *tcache, tcache_bin_t *tbin,
     szind_t binind, unsigned rem);
 void	tcache_bin_flush_large(tsd_t *tsd, tcache_bin_t *tbin, szind_t binind,
@@ -155,15 +162,15 @@ void	tcache_flush(void);
 bool	tcache_enabled_get(void);
 tcache_t *tcache_get(tsd_t *tsd, bool create);
 void	tcache_enabled_set(bool enabled);
-void	*tcache_alloc_easy(tcache_bin_t *tbin);
+void	*tcache_alloc_easy(tcache_bin_t *tbin, bool *tcache_success);
 void	*tcache_alloc_small(tsd_t *tsd, arena_t *arena, tcache_t *tcache,
-    size_t size, bool zero);
+    size_t size, szind_t ind, bool zero, bool slow_path);
 void	*tcache_alloc_large(tsd_t *tsd, arena_t *arena, tcache_t *tcache,
-    size_t size, bool zero);
+    size_t size, szind_t ind, bool zero, bool slow_path);
 void	tcache_dalloc_small(tsd_t *tsd, tcache_t *tcache, void *ptr,
-    szind_t binind);
+    szind_t binind, bool slow_path);
 void	tcache_dalloc_large(tsd_t *tsd, tcache_t *tcache, void *ptr,
-    size_t size);
+    size_t size, bool slow_path);
 tcache_t	*tcaches_get(tsd_t *tsd, unsigned ind);
 #endif
 
@@ -247,44 +254,69 @@ tcache_event(tsd_t *tsd, tcache_t *tcache)
 }
 
 JEMALLOC_ALWAYS_INLINE void *
-tcache_alloc_easy(tcache_bin_t *tbin)
+tcache_alloc_easy(tcache_bin_t *tbin, bool *tcache_success)
 {
 	void *ret;
 
 	if (unlikely(tbin->ncached == 0)) {
 		tbin->low_water = -1;
+		*tcache_success = false;
 		return (NULL);
 	}
+	/*
+	 * tcache_success (instead of ret) should be checked upon the return of
+	 * this function.  We avoid checking (ret == NULL) because there is
+	 * never a null stored on the avail stack (which is unknown to the
+	 * compiler), and eagerly checking ret would cause pipeline stall
+	 * (waiting for the cacheline).
+	 */
+	*tcache_success = true;
+	ret = *(tbin->avail - tbin->ncached);
 	tbin->ncached--;
+
 	if (unlikely((int)tbin->ncached < tbin->low_water))
 		tbin->low_water = tbin->ncached;
-	ret = tbin->avail[tbin->ncached];
+
 	return (ret);
 }
 
 JEMALLOC_ALWAYS_INLINE void *
 tcache_alloc_small(tsd_t *tsd, arena_t *arena, tcache_t *tcache, size_t size,
-    bool zero)
+    szind_t binind, bool zero, bool slow_path)
 {
 	void *ret;
-	szind_t binind;
-	size_t usize;
 	tcache_bin_t *tbin;
+	bool tcache_success;
+	size_t usize JEMALLOC_CC_SILENCE_INIT(0);
 
-	binind = size2index(size);
 	assert(binind < NBINS);
 	tbin = &tcache->tbins[binind];
-	usize = index2size(binind);
-	ret = tcache_alloc_easy(tbin);
-	if (unlikely(ret == NULL)) {
-		ret = tcache_alloc_small_hard(tsd, arena, tcache, tbin, binind);
-		if (ret == NULL)
+	ret = tcache_alloc_easy(tbin, &tcache_success);
+	assert(tcache_success == (ret != NULL));
+	if (unlikely(!tcache_success)) {
+		bool tcache_hard_success;
+		arena = arena_choose(tsd, arena);
+		if (unlikely(arena == NULL))
+			return (NULL);
+
+		ret = tcache_alloc_small_hard(tsd, arena, tcache, tbin, binind,
+			&tcache_hard_success);
+		if (tcache_hard_success == false)
 			return (NULL);
 	}
-	assert(tcache_salloc(ret) == usize);
+
+	assert(ret);
+	/*
+	 * Only compute usize if required.  The checks in the following if
+	 * statement are all static.
+	 */
+	if (config_prof || (slow_path && config_fill) || unlikely(zero)) {
+		usize = index2size(binind);
+		assert(tcache_salloc(ret) == usize);
+	}
 
 	if (likely(!zero)) {
-		if (config_fill) {
+		if (slow_path && config_fill) {
 			if (unlikely(opt_junk_alloc)) {
 				arena_alloc_junk_small(ret,
 				    &arena_bin_info[binind], false);
@@ -292,7 +324,7 @@ tcache_alloc_small(tsd_t *tsd, arena_t *arena, tcache_t *tcache, size_t size,
 				memset(ret, 0, usize);
 		}
 	} else {
-		if (config_fill && unlikely(opt_junk_alloc)) {
+		if (slow_path && config_fill && unlikely(opt_junk_alloc)) {
 			arena_alloc_junk_small(ret, &arena_bin_info[binind],
 			    true);
 		}
@@ -309,28 +341,38 @@ tcache_alloc_small(tsd_t *tsd, arena_t *arena, tcache_t *tcache, size_t size,
 
 JEMALLOC_ALWAYS_INLINE void *
 tcache_alloc_large(tsd_t *tsd, arena_t *arena, tcache_t *tcache, size_t size,
-    bool zero)
+    szind_t binind, bool zero, bool slow_path)
 {
 	void *ret;
-	szind_t binind;
-	size_t usize;
 	tcache_bin_t *tbin;
+	bool tcache_success;
+	size_t usize JEMALLOC_CC_SILENCE_INIT(0);
 
-	binind = size2index(size);
-	usize = index2size(binind);
-	assert(usize <= tcache_maxclass);
 	assert(binind < nhbins);
 	tbin = &tcache->tbins[binind];
-	ret = tcache_alloc_easy(tbin);
-	if (unlikely(ret == NULL)) {
+	ret = tcache_alloc_easy(tbin, &tcache_success);
+	assert(tcache_success == (ret != NULL));
+	if (unlikely(!tcache_success)) {
 		/*
 		 * Only allocate one large object at a time, because it's quite
 		 * expensive to create one and not use it.
 		 */
-		ret = arena_malloc_large(arena, usize, zero);
+		arena = arena_choose(tsd, arena);
+		if (unlikely(arena == NULL))
+			return (NULL);
+
+		usize = index2size(binind);
+		assert(usize <= tcache_maxclass);
+		ret = arena_malloc_large(arena, usize, binind, zero);
 		if (ret == NULL)
 			return (NULL);
 	} else {
+		/* Only compute usize on demand */
+		if (config_prof || (slow_path && config_fill) || unlikely(zero)) {
+			usize = index2size(binind);
+			assert(usize <= tcache_maxclass);
+		}
+
 		if (config_prof && usize == LARGE_MINCLASS) {
 			arena_chunk_t *chunk =
 			    (arena_chunk_t *)CHUNK_ADDR2BASE(ret);
@@ -340,7 +382,7 @@ tcache_alloc_large(tsd_t *tsd, arena_t *arena, tcache_t *tcache, size_t size,
 			    BININD_INVALID);
 		}
 		if (likely(!zero)) {
-			if (config_fill) {
+			if (slow_path && config_fill) {
 				if (unlikely(opt_junk_alloc))
 					memset(ret, 0xa5, usize);
 				else if (unlikely(opt_zero))
@@ -360,14 +402,15 @@ tcache_alloc_large(tsd_t *tsd, arena_t *arena, tcache_t *tcache, size_t size,
 }
 
 JEMALLOC_ALWAYS_INLINE void
-tcache_dalloc_small(tsd_t *tsd, tcache_t *tcache, void *ptr, szind_t binind)
+tcache_dalloc_small(tsd_t *tsd, tcache_t *tcache, void *ptr, szind_t binind,
+    bool slow_path)
 {
 	tcache_bin_t *tbin;
 	tcache_bin_info_t *tbin_info;
 
 	assert(tcache_salloc(ptr) <= SMALL_MAXCLASS);
 
-	if (config_fill && unlikely(opt_junk_free))
+	if (slow_path && config_fill && unlikely(opt_junk_free))
 		arena_dalloc_junk_small(ptr, &arena_bin_info[binind]);
 
 	tbin = &tcache->tbins[binind];
@@ -377,14 +420,15 @@ tcache_dalloc_small(tsd_t *tsd, tcache_t *tcache, void *ptr, szind_t binind)
 		    (tbin_info->ncached_max >> 1));
 	}
 	assert(tbin->ncached < tbin_info->ncached_max);
-	tbin->avail[tbin->ncached] = ptr;
 	tbin->ncached++;
+	*(tbin->avail - tbin->ncached) = ptr;
 
 	tcache_event(tsd, tcache);
 }
 
 JEMALLOC_ALWAYS_INLINE void
-tcache_dalloc_large(tsd_t *tsd, tcache_t *tcache, void *ptr, size_t size)
+tcache_dalloc_large(tsd_t *tsd, tcache_t *tcache, void *ptr, size_t size,
+    bool slow_path)
 {
 	szind_t binind;
 	tcache_bin_t *tbin;
@@ -396,7 +440,7 @@ tcache_dalloc_large(tsd_t *tsd, tcache_t *tcache, void *ptr, size_t size)
 
 	binind = size2index(size);
 
-	if (config_fill && unlikely(opt_junk_free))
+	if (slow_path && config_fill && unlikely(opt_junk_free))
 		arena_dalloc_junk_large(ptr, size);
 
 	tbin = &tcache->tbins[binind];
@@ -406,8 +450,8 @@ tcache_dalloc_large(tsd_t *tsd, tcache_t *tcache, void *ptr, size_t size)
 		    (tbin_info->ncached_max >> 1), tcache);
 	}
 	assert(tbin->ncached < tbin_info->ncached_max);
-	tbin->avail[tbin->ncached] = ptr;
 	tbin->ncached++;
+	*(tbin->avail - tbin->ncached) = ptr;
 
 	tcache_event(tsd, tcache);
 }