runtime: delete MHeapMapCache, which is useless

because free needs to mark the block as freed to
coordinate with the garbage collector.

(in C++ free can blindly put the block on the free list,
no questions asked, so the cache saves some work.)

R=iant
CC=golang-dev
https://golang.org/cl/206069

src/pkg/runtime/malloc.cgo

@@ -48,6 +48,12 @@ mallocgc(uintptr size, uint32 refflag, int32 dogc, int32 zeroed)
 		mstats.alloc += size;
 		mstats.total_alloc += size;
 		mstats.by_size[sizeclass].nmalloc++;
+
+		if(!mlookup(v, nil, nil, nil, &ref)) {
+			printf("malloc %D; mlookup failed\n", (uint64)size);
+			throw("malloc mlookup");
+		}
+		*ref = RefNone | refflag;
 	} else {
 		// TODO(rsc): Report tracebacks for very large allocations.
 
@@ -61,14 +67,10 @@ mallocgc(uintptr size, uint32 refflag, int32 dogc, int32 zeroed)
 		mstats.alloc += npages<<PageShift;
 		mstats.total_alloc += npages<<PageShift;
 		v = (void*)(s->start << PageShift);
-	}
 
-	// setup for mark sweep
-	if(!mlookup(v, nil, nil, &ref)) {
-		printf("malloc %D; mlookup failed\n", (uint64)size);
-		throw("malloc mlookup");
+		// setup for mark sweep
+		s->gcref0 = RefNone | refflag;
 	}
-	*ref = RefNone | refflag;
 
 	m->mallocing = 0;
 
@@ -88,7 +90,6 @@ void
 free(void *v)
 {
 	int32 sizeclass, size;
-	uintptr page, tmp;
 	MSpan *s;
 	MCache *c;
 	uint32 *ref;
@@ -100,46 +101,34 @@ free(void *v)
 		throw("malloc/free - deadlock");
 	m->mallocing = 1;
 
-	if(!mlookup(v, nil, nil, &ref)) {
+	if(!mlookup(v, nil, nil, &s, &ref)) {
 		printf("free %p: not an allocated block\n", v);
 		throw("free mlookup");
 	}
 	*ref = RefFree;
 
 	// Find size class for v.
-	page = (uintptr)v >> PageShift;
-	sizeclass = MHeapMapCache_GET(&mheap.mapcache, page, tmp);
+	sizeclass = s->sizeclass;
 	if(sizeclass == 0) {
-		// Missed in cache.
-		s = MHeap_Lookup(&mheap, page);
-		if(s == nil)
-			throw("free - invalid pointer");
-		sizeclass = s->sizeclass;
-		if(sizeclass == 0) {
-			// Large object.
-			mstats.alloc -= s->npages<<PageShift;
-			runtime_memclr(v, s->npages<<PageShift);
-			MHeap_Free(&mheap, s);
-			goto out;
-		}
-		MHeapMapCache_SET(&mheap.mapcache, page, sizeclass);
+		// Large object.
+		mstats.alloc -= s->npages<<PageShift;
+		runtime_memclr(v, s->npages<<PageShift);
+		MHeap_Free(&mheap, s);
+	} else {
+		// Small object.
+		c = m->mcache;
+		size = class_to_size[sizeclass];
+		if(size > sizeof(uintptr))
+			((uintptr*)v)[1] = 1;	// mark as "needs to be zeroed"
+		mstats.alloc -= size;
+		mstats.by_size[sizeclass].nfree++;
+		MCache_Free(c, v, sizeclass, size);
 	}
-
-	// Small object.
-	c = m->mcache;
-	size = class_to_size[sizeclass];
-	if(size > sizeof(uintptr))
-		((uintptr*)v)[1] = 1;	// mark as "needs to be zeroed"
-	mstats.alloc -= size;
-	mstats.by_size[sizeclass].nfree++;
-	MCache_Free(c, v, sizeclass, size);
-
-out:
 	m->mallocing = 0;
 }
 
 int32
-mlookup(void *v, byte **base, uintptr *size, uint32 **ref)
+mlookup(void *v, byte **base, uintptr *size, MSpan **sp, uint32 **ref)
 {
 	uintptr n, nobj, i;
 	byte *p;
@@ -147,6 +136,8 @@ mlookup(void *v, byte **base, uintptr *size, uint32 **ref)
 
 	mstats.nlookup++;
 	s = MHeap_LookupMaybe(&mheap, (uintptr)v>>PageShift);
+	if(sp)
+		*sp = s;
 	if(s == nil) {
 		if(base)
 			*base = nil;
@@ -256,7 +247,7 @@ stackalloc(uint32 n)
 		return v;
 	}
 	v = mallocgc(n, 0, 0, 0);
-	if(!mlookup(v, nil, nil, &ref))
+	if(!mlookup(v, nil, nil, nil, &ref))
 		throw("stackalloc mlookup");
 	*ref = RefStack;
 	return v;
@@ -283,7 +274,7 @@ func Free(p *byte) {
 }
 
 func Lookup(p *byte) (base *byte, size uintptr) {
-	mlookup(p, &base, &size, nil);
+	mlookup(p, &base, &size, nil, nil);
 }
 
 func GC() {
@@ -306,7 +297,7 @@ func SetFinalizer(obj Eface, finalizer Eface) {
 		printf("runtime.SetFinalizer: first argument is %S, not pointer\n", *obj.type->string);
 		goto throw;
 	}
-	if(!mlookup(obj.data, &base, &size, nil) || obj.data != base) {
+	if(!mlookup(obj.data, &base, &size, nil, nil) || obj.data != base) {
 		printf("runtime.SetFinalizer: pointer not at beginning of allocated block\n");
 		goto throw;
 	}

src/pkg/runtime/malloc.h

@@ -20,8 +20,6 @@
 //	MHeap: the malloc heap, managed at page (4096-byte) granularity.
 //	MSpan: a run of pages managed by the MHeap.
 //	MHeapMap: a mapping from page IDs to MSpans.
-//	MHeapMapCache: a small cache of MHeapMap mapping page IDs
-//		to size classes for pages used for small objects.
 //	MCentral: a shared free list for a given size class.
 //	MCache: a per-thread (in Go, per-M) cache for small objects.
 //	MStats: allocation statistics.
@@ -87,7 +85,6 @@ typedef struct FixAlloc	FixAlloc;
 typedef struct MCentral	MCentral;
 typedef struct MHeap	MHeap;
 typedef struct MHeapMap	MHeapMap;
-typedef struct MHeapMapCache	MHeapMapCache;
 typedef struct MSpan	MSpan;
 typedef struct MStats	MStats;
 typedef struct MLink	MLink;
@@ -296,7 +293,6 @@ struct MHeap
 
 	// span lookup
 	MHeapMap map;
-	MHeapMapCache mapcache;
 
 	// range of addresses we might see in the heap
 	byte *min;
@@ -324,7 +320,7 @@ MSpan*	MHeap_LookupMaybe(MHeap *h, PageID p);
 void	MGetSizeClassInfo(int32 sizeclass, int32 *size, int32 *npages, int32 *nobj);
 
 void*	mallocgc(uintptr size, uint32 flag, int32 dogc, int32 zeroed);
-int32	mlookup(void *v, byte **base, uintptr *size, uint32 **ref);
+int32	mlookup(void *v, byte **base, uintptr *size, MSpan **s, uint32 **ref);
 void	gc(int32 force);
 
 void*	SysAlloc(uintptr);

src/pkg/runtime/mfinal.c

@@ -97,7 +97,7 @@ addfinalizer(void *p, void (*f)(void*), int32 nret)
 	uint32 *ref;
 	byte *base;
 
-	if(!mlookup(p, &base, nil, &ref) || p != base)
+	if(!mlookup(p, &base, nil, nil, &ref) || p != base)
 		throw("addfinalizer on invalid pointer");
 	if(f == nil) {
 		if(*ref & RefHasFinalizer) {

src/pkg/runtime/mgc0.c

@@ -65,7 +65,7 @@ scanblock(int32 depth, byte *b, int64 n)
 		obj = vp[i];
 		if(obj == nil || (byte*)obj < mheap.min || (byte*)obj >= mheap.max)
 			continue;
-		if(mlookup(obj, &obj, &size, &refp)) {
+		if(mlookup(obj, &obj, &size, nil, &refp)) {
 			ref = *refp;
 			switch(ref & ~(RefNoPointers|RefHasFinalizer)) {
 			case RefFinalize:

src/pkg/runtime/mheap.c

@@ -108,27 +108,11 @@ HaveSpan:
 		MHeap_FreeLocked(h, t);
 	}
 
-	// If span is being used for small objects, cache size class.
-	// No matter what, cache span info, because gc needs to be
+	// Record span info, because gc needs to be
 	// able to map interior pointer to containing span.
 	s->sizeclass = sizeclass;
 	for(n=0; n<npage; n++)
 		MHeapMap_Set(&h->map, s->start+n, s);
-	if(sizeclass == 0) {
-		uintptr tmp;
-
-		// If there are entries for this span, invalidate them,
-		// but don't blow out cache entries about other spans.
-		for(n=0; n<npage; n++)
-			if(MHeapMapCache_GET(&h->mapcache, s->start+n, tmp) != 0)
-				MHeapMapCache_SET(&h->mapcache, s->start+n, 0);
-	} else {
-		// Save cache entries for this span.
-		// If there's a size class, there aren't that many pages.
-		for(n=0; n<npage; n++)
-			MHeapMapCache_SET(&h->mapcache, s->start+n, sizeclass);
-	}
-
 	return s;
 }
 

src/pkg/runtime/mheapmap32.h

@@ -39,38 +39,3 @@ MSpan*	MHeapMap_GetMaybe(MHeapMap *m, PageID k);
 void	MHeapMap_Set(MHeapMap *m, PageID k, MSpan *v);
 
 
-// Much of the time, free(v) needs to know only the size class for v,
-// not which span it came from.  The MHeapMap finds the size class
-// by looking up the span.
-//
-// An MHeapMapCache is a simple direct-mapped cache translating
-// page numbers to size classes.  It avoids the expensive MHeapMap
-// lookup for hot pages.
-//
-// The cache entries are 32 bits, with the page number in the low part
-// and the value at the top.
-//
-// NOTE(rsc): On a machine with 32-bit addresses (= 20-bit page numbers),
-// we can use a 16-bit cache entry by not storing the redundant 12 bits
-// of the key that are used as the entry index.  For now, keep it simple.
-enum
-{
-	MHeapMapCache_HashBits = 12
-};
-
-struct MHeapMapCache
-{
-	uint32 array[1<<MHeapMapCache_HashBits];
-};
-
-// All macros for speed (sorry).
-#define HMASK	((1<<MHeapMapCache_HashBits)-1)
-#define KBITS	MHeapMap_TotalBits
-#define KMASK	((1LL<<KBITS)-1)
-
-#define MHeapMapCache_SET(cache, key, value) \
-	((cache)->array[(key) & HMASK] = (key) | ((uintptr)(value) << KBITS))
-
-#define MHeapMapCache_GET(cache, key, tmp) \
-	(tmp = (cache)->array[(key) & HMASK], \
-	 (tmp & KMASK) == (key) ? (tmp >> KBITS) : 0)

src/pkg/runtime/mheapmap64.h

@@ -58,39 +58,3 @@ MSpan*	MHeapMap_GetMaybe(MHeapMap *m, PageID k);
 void	MHeapMap_Set(MHeapMap *m, PageID k, MSpan *v);
 
 
-// Much of the time, free(v) needs to know only the size class for v,
-// not which span it came from.  The MHeapMap finds the size class
-// by looking up the span.
-//
-// An MHeapMapCache is a simple direct-mapped cache translating
-// page numbers to size classes.  It avoids the expensive MHeapMap
-// lookup for hot pages.
-//
-// The cache entries are 64 bits, with the page number in the low part
-// and the value at the top.
-//
-// NOTE(rsc): On a machine with 32-bit addresses (= 20-bit page numbers),
-// we can use a 16-bit cache entry by not storing the redundant 12 bits
-// of the key that are used as the entry index.  Here in 64-bit land,
-// that trick won't work unless the hash table has 2^28 entries.
-enum
-{
-	MHeapMapCache_HashBits = 12
-};
-
-struct MHeapMapCache
-{
-	uintptr array[1<<MHeapMapCache_HashBits];
-};
-
-// All macros for speed (sorry).
-#define HMASK	((1<<MHeapMapCache_HashBits)-1)
-#define KBITS	MHeapMap_TotalBits
-#define KMASK	((1LL<<KBITS)-1)
-
-#define MHeapMapCache_SET(cache, key, value) \
-	((cache)->array[(key) & HMASK] = (key) | ((uintptr)(value) << KBITS))
-
-#define MHeapMapCache_GET(cache, key, tmp) \
-	(tmp = (cache)->array[(key) & HMASK], \
-	 (tmp & KMASK) == (key) ? (tmp >> KBITS) : 0)