Tcache flush: keep common path state in registers.

By carefully force-inlining the division constants and the operation sum count,
we can eliminate redundant operations in the arena-level dalloc function.  Do
so.

include/jemalloc/internal/arena_externs.h

@@ -2,6 +2,7 @@
 #define JEMALLOC_INTERNAL_ARENA_EXTERNS_H
 
 #include "jemalloc/internal/bin.h"
+#include "jemalloc/internal/div.h"
 #include "jemalloc/internal/extent_dss.h"
 #include "jemalloc/internal/hook.h"
 #include "jemalloc/internal/pages.h"
@@ -13,6 +14,8 @@ extern ssize_t opt_muzzy_decay_ms;
 extern percpu_arena_mode_t opt_percpu_arena;
 extern const char *percpu_arena_mode_names[];
 
+extern div_info_t arena_binind_div_info[SC_NBINS];
+
 extern const uint64_t h_steps[SMOOTHSTEP_NSTEPS];
 extern malloc_mutex_t arenas_lock;
 extern emap_t arena_emap_global;
@@ -29,9 +32,6 @@ void arena_stats_merge(tsdn_t *tsdn, arena_t *arena, unsigned *nthreads,
     bin_stats_data_t *bstats, arena_stats_large_t *lstats,
     pac_estats_t *estats, hpa_shard_stats_t *hpastats, sec_stats_t *secstats);
 void arena_handle_new_dirty_pages(tsdn_t *tsdn, arena_t *arena);
-#ifdef JEMALLOC_JET
-size_t arena_slab_regind(edata_t *slab, szind_t binind, const void *ptr);
-#endif
 edata_t *arena_extent_alloc_large(tsdn_t *tsdn, arena_t *arena,
     size_t usize, size_t alignment, bool zero);
 void arena_extent_dalloc_large_prep(tsdn_t *tsdn, arena_t *arena,
@@ -59,8 +59,11 @@ void arena_prof_promote(tsdn_t *tsdn, void *ptr, size_t usize);
 void arena_dalloc_promoted(tsdn_t *tsdn, void *ptr, tcache_t *tcache,
     bool slow_path);
 void arena_slab_dalloc(tsdn_t *tsdn, arena_t *arena, edata_t *slab);
-bool arena_dalloc_bin_locked(tsdn_t *tsdn, arena_t *arena, bin_t *bin,
+
-    szind_t binind, edata_t *edata, void *ptr);
+void arena_dalloc_bin_locked_handle_newly_empty(tsdn_t *tsdn, arena_t *arena,
+    edata_t *slab, bin_t *bin);
+void arena_dalloc_bin_locked_handle_newly_nonempty(tsdn_t *tsdn, arena_t *arena,
+    edata_t *slab, bin_t *bin);
 void arena_dalloc_small(tsdn_t *tsdn, void *ptr);
 bool arena_ralloc_no_move(tsdn_t *tsdn, void *ptr, size_t oldsize, size_t size,
     size_t extra, bool zero, size_t *newsize);

include/jemalloc/internal/arena_inlines_b.h

@@ -1,6 +1,7 @@
 #ifndef JEMALLOC_INTERNAL_ARENA_INLINES_B_H
 #define JEMALLOC_INTERNAL_ARENA_INLINES_B_H
 
+#include "jemalloc/internal/div.h"
 #include "jemalloc/internal/emap.h"
 #include "jemalloc/internal/jemalloc_internal_types.h"
 #include "jemalloc/internal/mutex.h"
@@ -441,4 +442,96 @@ arena_cache_oblivious_randomize(tsdn_t *tsdn, arena_t *arena, edata_t *edata,
 	}
 }
 
+/*
+ * The dalloc bin info contains just the information that the common paths need
+ * during tcache flushes.  By force-inlining these paths, and using local copies
+ * of data (so that the compiler knows it's constant), we avoid a whole bunch of
+ * redundant loads and stores by leaving this information in registers.
+ */
+typedef struct arena_dalloc_bin_locked_info_s arena_dalloc_bin_locked_info_t;
+struct arena_dalloc_bin_locked_info_s {
+	div_info_t div_info;
+	uint32_t nregs;
+	uint64_t ndalloc;
+};
+
+JEMALLOC_ALWAYS_INLINE size_t
+arena_slab_regind(arena_dalloc_bin_locked_info_t *info, szind_t binind,
+    edata_t *slab, const void *ptr) {
+	size_t diff, regind;
+
+	/* Freeing a pointer outside the slab can cause assertion failure. */
+	assert((uintptr_t)ptr >= (uintptr_t)edata_addr_get(slab));
+	assert((uintptr_t)ptr < (uintptr_t)edata_past_get(slab));
+	/* Freeing an interior pointer can cause assertion failure. */
+	assert(((uintptr_t)ptr - (uintptr_t)edata_addr_get(slab)) %
+	    (uintptr_t)bin_infos[binind].reg_size == 0);
+
+	diff = (size_t)((uintptr_t)ptr - (uintptr_t)edata_addr_get(slab));
+
+	/* Avoid doing division with a variable divisor. */
+	regind = div_compute(&info->div_info, diff);
+
+	assert(regind < bin_infos[binind].nregs);
+
+	return regind;
+}
+
+JEMALLOC_ALWAYS_INLINE void
+arena_dalloc_bin_locked_begin(arena_dalloc_bin_locked_info_t *info,
+    szind_t binind) {
+	info->div_info = arena_binind_div_info[binind];
+	info->nregs = bin_infos[binind].nregs;
+	info->ndalloc = 0;
+}
+
+/*
+ * Does the deallocation work associated with freeing a single pointer (a
+ * "step") in between a arena_dalloc_bin_locked begin and end call.
+ *
+ * Returns true if arena_slab_dalloc must be called on slab.  Doesn't do
+ * stats updates, which happen during finish (this lets running counts get left
+ * in a register).
+ */
+JEMALLOC_ALWAYS_INLINE bool
+arena_dalloc_bin_locked_step(tsdn_t *tsdn, arena_t *arena, bin_t *bin,
+    arena_dalloc_bin_locked_info_t *info, szind_t binind, edata_t *slab,
+    void *ptr) {
+	const bin_info_t *bin_info = &bin_infos[binind];
+	size_t regind = arena_slab_regind(info, binind, slab, ptr);
+	slab_data_t *slab_data = edata_slab_data_get(slab);
+
+	assert(edata_nfree_get(slab) < bin_info->nregs);
+	/* Freeing an unallocated pointer can cause assertion failure. */
+	assert(bitmap_get(slab_data->bitmap, &bin_info->bitmap_info, regind));
+
+	bitmap_unset(slab_data->bitmap, &bin_info->bitmap_info, regind);
+	edata_nfree_inc(slab);
+
+	if (config_stats) {
+		info->ndalloc++;
+	}
+
+	unsigned nfree = edata_nfree_get(slab);
+	if (nfree == bin_info->nregs) {
+		arena_dalloc_bin_locked_handle_newly_empty(tsdn, arena, slab,
+		    bin);
+		return true;
+	} else if (nfree == 1 && slab != bin->slabcur) {
+		arena_dalloc_bin_locked_handle_newly_nonempty(tsdn, arena, slab,
+		    bin);
+	}
+	return false;
+}
+
+JEMALLOC_ALWAYS_INLINE void
+arena_dalloc_bin_locked_finish(tsdn_t *tsdn, arena_t *arena, bin_t *bin,
+    arena_dalloc_bin_locked_info_t *info) {
+	if (config_stats) {
+		bin->stats.ndalloc += info->ndalloc;
+		assert(bin->stats.curregs >= (size_t)info->ndalloc);
+		bin->stats.curregs -= (size_t)info->ndalloc;
+	}
+}
+
 #endif /* JEMALLOC_INTERNAL_ARENA_INLINES_B_H */

src/arena.c

@@ -3,7 +3,6 @@
 
 #include "jemalloc/internal/assert.h"
 #include "jemalloc/internal/decay.h"
-#include "jemalloc/internal/div.h"
 #include "jemalloc/internal/ehooks.h"
 #include "jemalloc/internal/extent_dss.h"
 #include "jemalloc/internal/extent_mmap.h"
@@ -45,7 +44,7 @@ const uint64_t h_steps[SMOOTHSTEP_NSTEPS] = {
 #undef STEP
 };
 
-static div_info_t arena_binind_div_info[SC_NBINS];
+div_info_t arena_binind_div_info[SC_NBINS];
 
 size_t opt_oversize_threshold = OVERSIZE_THRESHOLD_DEFAULT;
 size_t oversize_threshold = OVERSIZE_THRESHOLD_DEFAULT;
@@ -260,44 +259,6 @@ arena_slab_reg_alloc_batch(edata_t *slab, const bin_info_t *bin_info,
 	edata_nfree_sub(slab, cnt);
 }
 
-#ifndef JEMALLOC_JET
-static
-#endif
-size_t
-arena_slab_regind(edata_t *slab, szind_t binind, const void *ptr) {
-	size_t diff, regind;
-
-	/* Freeing a pointer outside the slab can cause assertion failure. */
-	assert((uintptr_t)ptr >= (uintptr_t)edata_addr_get(slab));
-	assert((uintptr_t)ptr < (uintptr_t)edata_past_get(slab));
-	/* Freeing an interior pointer can cause assertion failure. */
-	assert(((uintptr_t)ptr - (uintptr_t)edata_addr_get(slab)) %
-	    (uintptr_t)bin_infos[binind].reg_size == 0);
-
-	diff = (size_t)((uintptr_t)ptr - (uintptr_t)edata_addr_get(slab));
-
-	/* Avoid doing division with a variable divisor. */
-	regind = div_compute(&arena_binind_div_info[binind], diff);
-
-	assert(regind < bin_infos[binind].nregs);
-
-	return regind;
-}
-
-static void
-arena_slab_reg_dalloc(edata_t *slab, slab_data_t *slab_data, void *ptr) {
-	szind_t binind = edata_szind_get(slab);
-	const bin_info_t *bin_info = &bin_infos[binind];
-	size_t regind = arena_slab_regind(slab, binind, ptr);
-
-	assert(edata_nfree_get(slab) < bin_info->nregs);
-	/* Freeing an unallocated pointer can cause assertion failure. */
-	assert(bitmap_get(slab_data->bitmap, &bin_info->bitmap_info, regind));
-
-	bitmap_unset(slab_data->bitmap, &bin_info->bitmap_info, regind);
-	edata_nfree_inc(slab);
-}
-
 static void
 arena_large_malloc_stats_update(tsdn_t *tsdn, arena_t *arena, size_t usize) {
 	szind_t index, hindex;
@@ -1189,39 +1150,20 @@ arena_dalloc_bin_slab_prepare(tsdn_t *tsdn, edata_t *slab, bin_t *bin) {
 	}
 }
 
-/* Returns true if arena_slab_dalloc must be called on slab */
+void
-static bool
+arena_dalloc_bin_locked_handle_newly_empty(tsdn_t *tsdn, arena_t *arena,
-arena_dalloc_bin_locked_impl(tsdn_t *tsdn, arena_t *arena, bin_t *bin,
+    edata_t *slab, bin_t *bin) {
-    szind_t binind, edata_t *slab, void *ptr) {
-	const bin_info_t *bin_info = &bin_infos[binind];
-	arena_slab_reg_dalloc(slab, edata_slab_data_get(slab), ptr);
-
-	bool ret = false;
-	unsigned nfree = edata_nfree_get(slab);
-	if (nfree == bin_info->nregs) {
 	arena_dissociate_bin_slab(arena, slab, bin);
 	arena_dalloc_bin_slab_prepare(tsdn, slab, bin);
-		ret = true;
+}
-	} else if (nfree == 1 && slab != bin->slabcur) {
+
+void
+arena_dalloc_bin_locked_handle_newly_nonempty(tsdn_t *tsdn, arena_t *arena,
+    edata_t *slab, bin_t *bin) {
 	arena_bin_slabs_full_remove(arena, bin, slab);
 	arena_bin_lower_slab(tsdn, arena, slab, bin);
 }
 
-	if (config_stats) {
-		bin->stats.ndalloc++;
-		bin->stats.curregs--;
-	}
-
-	return ret;
-}
-
-bool
-arena_dalloc_bin_locked(tsdn_t *tsdn, arena_t *arena, bin_t *bin,
-szind_t binind, edata_t *edata, void *ptr) {
-	return arena_dalloc_bin_locked_impl(tsdn, arena, bin, binind, edata,
-	    ptr);
-}
-
 static void
 arena_dalloc_bin(tsdn_t *tsdn, arena_t *arena, edata_t *edata, void *ptr) {
 	szind_t binind = edata_szind_get(edata);
@@ -1229,8 +1171,11 @@ arena_dalloc_bin(tsdn_t *tsdn, arena_t *arena, edata_t *edata, void *ptr) {
 	bin_t *bin = &arena->bins[binind].bin_shards[binshard];
 
 	malloc_mutex_lock(tsdn, &bin->lock);
-	bool ret = arena_dalloc_bin_locked_impl(tsdn, arena, bin, binind, edata,
+	arena_dalloc_bin_locked_info_t info;
-	    ptr);
+	arena_dalloc_bin_locked_begin(&info, binind);
+	bool ret = arena_dalloc_bin_locked_step(tsdn, arena, bin,
+	    &info, binind, edata, ptr);
+	arena_dalloc_bin_locked_finish(tsdn, arena, bin, &info);
 	malloc_mutex_unlock(tsdn, &bin->lock);
 
 	if (ret) {

src/tcache.c

@@ -399,6 +399,10 @@ tcache_bin_flush_impl(tsd_t *tsd, tcache_t *tcache, cache_bin_t *cache_bin,
 
 		/* Deallocate whatever we can. */
 		unsigned ndeferred = 0;
+		arena_dalloc_bin_locked_info_t dalloc_bin_info;
+		if (small) {
+			arena_dalloc_bin_locked_begin(&dalloc_bin_info, binind);
+		}
 		for (unsigned i = 0; i < nflush; i++) {
 			void *ptr = cache_bin_ptr_array_get(&ptrs, i);
 			edata = item_edata[i].edata;
@@ -417,8 +421,9 @@ tcache_bin_flush_impl(tsd_t *tsd, tcache_t *tcache, cache_bin_t *cache_bin,
 				continue;
 			}
 			if (small) {
-				if (arena_dalloc_bin_locked(tsdn, cur_arena,
+				if (arena_dalloc_bin_locked_step(tsdn,
-				    cur_bin, binind, edata, ptr)) {
+				    cur_arena, cur_bin, &dalloc_bin_info,
+				    binind, edata, ptr)) {
 					dalloc_slabs[dalloc_count] = edata;
 					dalloc_count++;
 				}
@@ -432,6 +437,8 @@ tcache_bin_flush_impl(tsd_t *tsd, tcache_t *tcache, cache_bin_t *cache_bin,
 		}
 
 		if (small) {
+			arena_dalloc_bin_locked_finish(tsdn, cur_arena, cur_bin,
+			    &dalloc_bin_info);
 			malloc_mutex_unlock(tsdn, &cur_bin->lock);
 		}
 		arena_decay_ticks(tsdn, cur_arena, nflush - ndeferred);

test/unit/slab.c

@@ -16,10 +16,13 @@ TEST_BEGIN(test_arena_slab_regind) {
 		    EXTENT_NOT_HEAD);
 		expect_ptr_not_null(edata_addr_get(&slab),
 		    "Unexpected malloc() failure");
+		arena_dalloc_bin_locked_info_t dalloc_info;
+		arena_dalloc_bin_locked_begin(&dalloc_info, binind);
 		for (regind = 0; regind < bin_info->nregs; regind++) {
 			void *reg = (void *)((uintptr_t)edata_addr_get(&slab) +
 			    (bin_info->reg_size * regind));
-			expect_zu_eq(arena_slab_regind(&slab, binind, reg),
+			expect_zu_eq(arena_slab_regind(&dalloc_info, binind,
+			    &slab, reg),
 			    regind,
 			    "Incorrect region index computed for size %zu",
 			    bin_info->reg_size);