builtin, cgen: implement chunked allocation of 16MB blocks with -prealloc (#10130)

vlib/builtin/builtin.c.v

@@ -217,6 +217,7 @@ pub fn println(s string) {
 	}
 }
 
+__global total_m = i64(0)
 // malloc dynamically allocates a `n` bytes block of memory on the heap.
 // malloc returns a `byteptr` pointing to the memory address of the allocated space.
 // unlike the `calloc` family of functions - malloc will not zero the memory block.
@@ -237,13 +238,8 @@ pub fn malloc(n int) &byte {
 	}
 	mut res := &byte(0)
 	$if prealloc {
-		res = g_m2_ptr
+		res = unsafe { prealloc_malloc(n) }
-		unsafe {
+	} $else $if gcboehm ? {
-			g_m2_ptr += n
-		}
-		nr_mallocs++
-	} $else {
-		$if gcboehm ? {
 		unsafe {
 			res = C.GC_MALLOC(n)
 		}
@@ -261,7 +257,6 @@ pub fn malloc(n int) &byte {
 	if res == 0 {
 		panic('malloc($n) failed')
 	}
-	}
 	$if debug_malloc ? {
 		// Fill in the memory with something != 0, so it is easier to spot
 		// when the calling code wrongly relies on it being zeroed.
@@ -310,12 +305,7 @@ pub fn v_realloc(b &byte, n int) &byte {
 [unsafe]
 pub fn realloc_data(old_data &byte, old_size int, new_size int) &byte {
 	$if prealloc {
-		unsafe {
+		return unsafe { prealloc_realloc(old_data, old_size, new_size) }
-			new_ptr := malloc(new_size)
-			min_size := if old_size < new_size { old_size } else { new_size }
-			C.memcpy(new_ptr, old_data, min_size)
-			return new_ptr
-		}
 	}
 	$if debug_realloc ? {
 		// NB: this is slower, but helps debugging memory problems.
@@ -356,16 +346,22 @@ pub fn vcalloc(n int) &byte {
 	} else if n == 0 {
 		return &byte(0)
 	}
+	$if prealloc {
+		return unsafe { prealloc_calloc(n) }
+	}
 	$if gcboehm ? {
-		return &byte(C.GC_MALLOC(n))
+		return unsafe { &byte(C.GC_MALLOC(n)) }
 	} $else {
-		return C.calloc(1, n)
+		return unsafe { C.calloc(1, n) }
 	}
 }
 
 // special versions of the above that allocate memory which is not scanned
 // for pointers (but is collected) when the Boehm garbage collection is used
 pub fn vcalloc_noscan(n int) &byte {
+	$if prealloc {
+		return unsafe { prealloc_calloc(n) }
+	}
 	$if gcboehm ? {
 		$if vplayground ? {
 			if n > 10000 {
@@ -375,7 +371,7 @@ pub fn vcalloc_noscan(n int) &byte {
 		if n < 0 {
 			panic('calloc(<0)')
 		}
-		return &byte(unsafe { C.memset(C.GC_MALLOC_ATOMIC(n), 0, n) })
+		return unsafe { &byte(C.memset(C.GC_MALLOC_ATOMIC(n), 0, n)) }
 	} $else {
 		return unsafe { vcalloc(n) }
 	}
@@ -394,7 +390,7 @@ pub fn free(ptr voidptr) {
 		//
 		// The exception is doing leak detection for manual memory management:
 		$if gcboehm_leak ? {
-			C.GC_FREE(ptr)
+			unsafe { C.GC_FREE(ptr) }
 		}
 		return
 	}

vlib/builtin/builtin.v

@@ -3,11 +3,6 @@
 // that can be found in the LICENSE file.
 module builtin
 
-__global (
-	g_m2_buf &byte
-	g_m2_ptr &byte
-)
-
 // isnil returns true if an object is nil (only for C objects).
 [inline]
 pub fn isnil(v voidptr) bool {
@@ -45,12 +40,8 @@ struct VCastTypeIndexName {
 	tname  string
 }
 
-__global (
-	total_m              = i64(0)
-	nr_mallocs           = int(0)
 // will be filled in cgen
-	as_cast_type_indexes []VCastTypeIndexName
+__global as_cast_type_indexes []VCastTypeIndexName
-)
 
 fn __as_cast(obj voidptr, obj_type int, expected_type int) voidptr {
 	if obj_type != expected_type {

vlib/builtin/prealloc.c.v

@@ -0,0 +1,114 @@
+module builtin
+
+// With -prealloc, V calls libc's malloc to get chunks, each at least 16MB
+// in size, as needed. Once a chunk is available, all malloc() calls within
+// V code, that can fit inside the chunk, will use it instead, each bumping a
+// pointer, till the chunk is filled. Once a chunk is filled, a new chunk will
+// be allocated by calling libc's malloc, and the process continues.
+// Each new chunk has a pointer to the old one, and at the end of the program,
+// the entire linked list of chunks is freed.
+// The goal of all this is to amortize the cost of calling libc's malloc,
+// trading higher memory usage for a compiler (or any single threaded batch
+// mode program), for a ~8-10% speed increase.
+// NB: `-prealloc` is NOT safe to be used for multithreaded programs!
+
+// size of the preallocated chunk
+const prealloc_block_size = 16 * 1024 * 1024
+
+__global g_memory_block &VMemoryBlock
+[heap]
+struct VMemoryBlock {
+mut:
+	id        int
+	cap       int
+	start     &byte = 0
+	previous  &VMemoryBlock = 0
+	remaining int
+	current   &byte = 0
+	mallocs   int
+}
+
+[unsafe]
+fn vmemory_block_new(prev &VMemoryBlock, at_least int) &VMemoryBlock {
+	mut v := unsafe { &VMemoryBlock(C.calloc(1, sizeof(VMemoryBlock))) }
+	if prev != 0 {
+		v.id = prev.id + 1
+	}
+	v.previous = prev
+	block_size := if at_least < prealloc_block_size { prealloc_block_size } else { at_least }
+	v.start = unsafe { C.malloc(block_size) }
+	v.cap = block_size
+	v.remaining = block_size
+	v.current = v.start
+	return v
+}
+
+[unsafe]
+fn vmemory_block_malloc(n int) &byte {
+	unsafe {
+		if g_memory_block.remaining < n {
+			g_memory_block = vmemory_block_new(g_memory_block, n)
+		}
+		mut res := &byte(0)
+		res = g_memory_block.current
+		g_memory_block.remaining -= n
+		g_memory_block.mallocs++
+		g_memory_block.current += n
+		return res
+	}
+}
+
+/////////////////////////////////////////////////
+
+[unsafe]
+fn prealloc_vinit() {
+	unsafe {
+		g_memory_block = vmemory_block_new(voidptr(0), prealloc_block_size)
+		$if !freestanding {
+			C.atexit(prealloc_vcleanup)
+		}
+	}
+}
+
+[unsafe]
+fn prealloc_vcleanup() {
+	$if prealloc_stats ? {
+		// NB: we do 2 loops here, because string interpolation
+		// in the first loop may still use g_memory_block
+		// The second loop however should *not* allocate at all.
+		mut nr_mallocs := i64(0)
+		mut mb := g_memory_block
+		for mb != 0 {
+			nr_mallocs += mb.mallocs
+			eprintln('> freeing mb.id: ${mb.id:3} | cap: ${mb.cap:7} | rem: ${mb.remaining:7} | start: ${voidptr(mb.start)} | current: ${voidptr(mb.current)} | diff: ${u64(mb.current) - u64(mb.start):7} bytes | mallocs: $mb.mallocs')
+			mb = mb.previous
+		}
+		eprintln('> nr_mallocs: $nr_mallocs')
+	}
+	unsafe {
+		for g_memory_block != 0 {
+			C.free(g_memory_block.start)
+			g_memory_block = g_memory_block.previous
+		}
+	}
+}
+
+[unsafe]
+fn prealloc_malloc(n int) &byte {
+	return unsafe { vmemory_block_malloc(n) }
+}
+
+[unsafe]
+fn prealloc_realloc(old_data &byte, old_size int, new_size int) &byte {
+	new_ptr := unsafe { vmemory_block_malloc(new_size) }
+	min_size := if old_size < new_size { old_size } else { new_size }
+	unsafe { C.memcpy(new_ptr, old_data, min_size) }
+	return new_ptr
+}
+
+[unsafe]
+fn prealloc_calloc(n int) &byte {
+	new_ptr := unsafe { vmemory_block_malloc(n) }
+	unsafe { C.memset(new_ptr, 0, n) }
+	return new_ptr
+}

vlib/v/gen/c/cgen.v

@@ -5535,8 +5535,7 @@ fn (mut g Gen) write_init_function() {
 		// g.writeln('g_str_buf = malloc( ${mb_size} * 1024 * 1000 );')
 	}
 	if g.pref.prealloc {
-		g.writeln('g_m2_buf = malloc(150 * 1000 * 1000);')
+		g.writeln('prealloc_vinit();')
-		g.writeln('g_m2_ptr = g_m2_buf;')
 	}
 	// NB: the as_cast table should be *before* the other constant initialize calls,
 	// because it may be needed during const initialization of builtin and during

vlib/v/gen/c/cheaders.v

@@ -66,11 +66,7 @@ void _STR_PRINT_ARG(const char *fmt, char** refbufp, int *nbytes, int *memsize,
 		}
 		// increase buffer (somewhat exponentially)
 		*memsize += guess + 3 * (*memsize) / 2;
-#ifdef _VGCBOEHM
+		*refbufp = (char*)v_realloc((void*)*refbufp, *memsize);
-		*refbufp = (char*)GC_REALLOC((void*)*refbufp, *memsize);
-#else
-		*refbufp = (char*)realloc((void*)*refbufp, *memsize);
-#endif
 	}
 	va_end(args);
 }
@@ -79,11 +75,7 @@ string _STR(const char *fmt, int nfmts, ...) {
 	va_list argptr;
 	int memsize = 128;
 	int nbytes = 0;
-#ifdef _VGCBOEHM
+	char* buf = (char*)v_malloc(memsize);
-	char* buf = (char*)GC_MALLOC(memsize);
-#else
-	char* buf = (char*)malloc(memsize);
-#endif
 	va_start(argptr, nfmts);
 	for (int i=0; i<nfmts; ++i) {
 		int k = strlen(fmt);

vlib/v/markused/markused.v

@@ -18,6 +18,8 @@ pub fn mark_used(mut table ast.Table, pref &pref.Preferences, ast_files []ast.Fi
 		'__new_array',
 		'__new_array_with_default',
 		'__new_array_with_array_default',
+		'v_realloc' /* needed for _STR */,
+		'v_malloc' /* needed for _STR */,
 		'new_array_from_c_array',
 		'v_fixed_index',
 		'memdup',
@@ -162,6 +164,10 @@ pub fn mark_used(mut table ast.Table, pref &pref.Preferences, ast_files []ast.Fi
 			all_fn_root_names << k
 			continue
 		}
+		if pref.prealloc && k.starts_with('prealloc_') {
+			all_fn_root_names << k
+			continue
+		}
 	}
 
 	// handle assertions and testing framework callbacks: