add missing arena.h

TurboList.hpp

@@ -8,6 +8,9 @@
 #ifndef TL_NOINLINE
 #define TL_NOINLINE __attribute__((noinline))
 #endif /* TL_NOINLINE */
+#ifndef TL_INLINE
+#define TL_INLINE __attribute__((always_inline))
+#endif /* TL_INLINE */
 
 #ifndef TL_LIKELY
 #define TL_LIKELY(x) __builtin_expect(!!(x), 1)
@@ -16,6 +19,10 @@
 #define TL_UNLIKELY(x) __builtin_expect(!!(x), 0)
 #endif /* TL_UNLIKELY */
 
+#ifndef TL_GROWTH_RATE
+#define TL_GROWTH_RATE 2
+#endif /* TL_GROWTH_RATE */
+
 typedef void*(MallocLike)(size_t);
 typedef void(FreeLike)(void*);
 
@@ -29,39 +36,63 @@ class TurboList {
 
 	uint32_t capacity;
 
-	TL_NOINLINE T& grow_and_insert(T elem) noexcept {
+	TL_NOINLINE void grow_and_insert(T elem) noexcept {
 			// assert(mid == 0);
 			if(old) FREE(old);
 			old = nex;
 			mid = end;
-			capacity *= 2;
-			nex = (int *) MALLOC(this->capacity * sizeof(T));
+			capacity *= TL_GROWTH_RATE;
+			nex = (T *) MALLOC(this->capacity * sizeof(T));
 
 			// Will go into the INSERT code path here
-			return insert(elem);
+			insert(elem);
 	}
 
+	template<typename... Args>
+	TL_NOINLINE T& grow_and_emplace(Args&&... args) noexcept {
+			// assert(mid == 0);
+			if(old) FREE(old);
+			old = nex;
+			mid = end;
+			capacity *= TL_GROWTH_RATE;
+			nex = (T *) MALLOC(this->capacity * sizeof(T));
+
+			// Will go into the INSERT code path here
+			return emplace_back(std::forward<Args>(args)...);
+	}
 public:
-	inline TurboList(uint32_t initial_size = 0, uint32_t initial_cap = 16) noexcept :
+	TL_INLINE TurboList(uint32_t initial_size = 0, uint32_t initial_cap = 16) noexcept :
 			old(nullptr),
 			mid(0),
 			end(initial_size),
 			capacity(initial_cap) {
 
-		nex = (int *) MALLOC(this->capacity * sizeof(T));
+		nex = (T *) MALLOC(this->capacity * sizeof(T));
 	}
 
-	inline ~TurboList() noexcept {
+	TL_INLINE ~TurboList() noexcept {
 		if(nex) FREE(nex);
 		if(old) FREE(old);
 	}
 
-	inline T& operator[](uint32_t i) noexcept {
-		return (i < mid) ? old[i] : nex[i];
+	TL_INLINE T& operator[](uint32_t i) const noexcept {
+		// This seem to be more often compiled to cmov
+		// branchless conditional codes this way..
+		//
+		T *base = (i < mid) ? old : nex;
+		return base[i];
+		//
+		// if(i < mid)	return old[i];
+		// else return nex[i];
+		//
+		// T* loc = (T*) ((i < mid) * (size_t)old +
+		// 	(i>=mid) * (size_t)nex +
+		// 	i* sizeof(T));
+		// return *loc;
 	}
 
 	/** This is much faster than operator[] if you do small amounts of work per access */
-	inline void iterate(void(callback)(T&)) noexcept {
+	TL_INLINE void iterate(void(callback)(T&)) noexcept {
 		// old
 		for(uint32_t i = 0; i < mid; ++i) {
 			callback(old[i]);
@@ -72,35 +103,54 @@ public:
 		}
 	}
 
-	inline T& insert(T elem) noexcept {
+	/** Vector compatibility: Use insert() if you want the inserted thing as reference too */
+	TL_INLINE void push_back(T elem) noexcept {
+		this->insert(elem);
+	}
+
+	/** Vector compatibility: Use pop() if you want the popped thing out as copy too */
+	TL_INLINE void pop_back() noexcept {
+		if(end > 0) {
+			--end;
+			if(end < mid) { // end > 0 here!
+				end = mid;
+				mid = 0;
+				FREE(nex);
+				nex = old;
+				old = nullptr;
+			}
+		}
+	}
+
+	TL_INLINE void insert(T elem) noexcept {
 		if(TL_LIKELY(end < capacity)) {
 
 			// INSERT
 
-			/* Same as this:
+			/* Same as this - but in this case it measures as faster:
 			if(mid > 0) {
-				nex[mid - 1] = old[mid - 1];
 				--mid;
+				nex[mid] = old[mid];
 			}
 			*/
 			bool hasmid = (mid > 0);
 			mid -= hasmid;
 			nex[mid] = hasmid ? old[mid] : nex[mid];
 
-			return (nex[end++] = elem);
+			nex[end++] = elem;
 		} else {
 			// GROW
-			return grow_and_insert(elem);
+			grow_and_insert(elem);
 		}
 	}
 
 	template<typename... Args>
-	inline T& emplace(Args&&... args) {
+	TL_INLINE T& emplace_back(Args&&... args) {
 		if(TL_LIKELY(end < capacity)) {
 
 			// INSERT
 
-			/* Same as this:
+			/* Same as this - but in this case it measures as faster:
 			if(mid > 0) {
 				nex[mid - 1] = old[mid - 1];
 				--mid;
@@ -114,21 +164,13 @@ public:
 			return *new (nex + end++) T(std::forward<Args>(args)...);
 		} else {
 			// GROW
-			//
-			// Rem.: I just chose this to be less optimized than
-			//       it is possible by making a copy and reusing
-			//       the existing grow and insert code instead of
-			//       writing a new "grow_and_emplace" again.
-			//
-			//       This happens rarely so its probably fine and
-			//       makes less template instantiations, smaller binary.
-			return grow_and_insert(T(std::forward<Args>(args)...));
+			return grow_and_emplace(std::forward<Args>(args)...);
 		}
 	}
 
 	// TODO: finalize() call which memcpy remaining elements of old into nex and then frees old and sets nullptr + mid = 0;
 
-	inline uint32_t size() noexcept {
+	TL_INLINE uint32_t size() noexcept {
 		return end;
 	}
 };

VirtualMemList.hpp

@@ -0,0 +1,66 @@
+#ifndef VIRTUAL_MEM_LIST_H
+#define VIRTUAL_MEM_LIST_H
+#include "arena.h"
+
+#ifndef VML_NOINLINE
+#define VML_NOINLINE __attribute__((noinline))
+#endif /* VML_NOINLINE */
+#ifndef VML_INLINE
+#define VML_INLINE __attribute__((always_inline))
+#endif /* VML_INLINE */
+
+#ifndef VML_LIKELY
+#define VML_LIKELY(x) __builtin_expect(!!(x), 1)
+#endif /* VML_LIKELY */
+#ifndef VML_UNLIKELY
+#define VML_UNLIKELY(x) __builtin_expect(!!(x), 0)
+#endif /* VML_UNLIKELY */
+
+#ifndef VML_GROWTH_RATE
+#define VML_GROWTH_RATE 2
+#endif /* VML_GROWTH_RATE */
+
+template<typename T>
+class VirtualMemList {
+	arena a;
+	T *base;
+	uint32_t end;
+
+public:
+	VML_INLINE VirtualMemList(uint32_t initial_size = 0) noexcept {
+		a = newarena((ptrdiff_t)1 << 33);
+		base = ((T*) alloc(&a, sizeof(T), sizeof(T), 1)) + 1;
+		end = initial_size;
+	}
+
+	VML_INLINE ~VirtualMemList() noexcept {
+		// TODO: arena free currently not implemented
+	}
+
+	VML_INLINE T& operator[](uint32_t i) const noexcept {
+		return base[i];
+	}
+
+	VML_INLINE void push_back(T elem) noexcept {
+		// Smallest solution:
+		// base[end++] = elem;
+
+		// Non-working, but logically better solution:
+		// T *value = (T*) alloc(&a, sizeof(T), sizeof(T), 1);
+		// *value = elem;
+
+		// Longer solution (more correct - but I let it waste a bit of memory)
+		T *value = (T*) alloc(&a, sizeof(T), sizeof(T), 1);
+		base[end++] = elem;
+	}
+
+	VML_INLINE void pop_back() noexcept {
+		--end;
+	}
+
+	VML_INLINE uint32_t size() noexcept {
+		return end;
+	}
+};
+
+#endif // VIRTUAL_MEM_LIST_H

arena.h

@@ -0,0 +1,103 @@
+// Gradual-commit arena demonstration
+// This is free and unencumbered software released into the public domain.
+/* Usage:
+ * 
+ *    arena a = newarena((ptrdiff_t)1 << 33);
+ *    if (!alloc(a, size, align, count)) {
+ *      break;
+ *    }
+ *    total += size * count;
+ */
+#ifndef ARENA_H
+#define ARENA_H
+
+#include <stddef.h>
+#include <string.h>
+
+static void *os_reserve(ptrdiff_t);
+static char os_commit(void *, ptrdiff_t);
+
+#define ARENA_PAGESIZE ((ptrdiff_t)1<<26)
+
+typedef struct {
+    char *begin;
+    char *commit;
+    char *end;
+} arena;
+
+static arena newarena(ptrdiff_t cap)
+{
+    arena a = {0};
+    cap += -cap & (ARENA_PAGESIZE - 1);
+    a.begin = a.commit = a.end = (char*) os_reserve(cap);
+    if (a.begin) {
+        a.end += cap;
+    }
+    return a;
+}
+
+static void *alloc(arena *a, ptrdiff_t size, ptrdiff_t align, ptrdiff_t count)
+{
+    ptrdiff_t padding   = -(size_t)a->begin & (align - 1);
+    ptrdiff_t committed = a->commit - a->begin;
+    if (count > (committed-padding)/size) {
+        ptrdiff_t reserved = a->end - a->begin;
+        if (count > (reserved-padding)/size) {
+            return 0;
+        }
+
+        ptrdiff_t needed = size*count + padding - committed;
+        needed += -needed & (ARENA_PAGESIZE - 1);
+        if (!os_commit(a->commit, needed)) {
+            return 0;
+        }
+        a->commit += needed;
+    }
+
+    void *ptr = a->begin + padding;
+    a->begin += padding + size*count;
+
+	// Change to this instead if you want zero-inited (but this gets slow with many arenas)
+    //return memset(ptr, 0, size*count);
+	return ptr;
+}
+
+#ifdef _WIN32
+// $ cc -g3 -nostartfiles -o arena.exe arena.c
+// $ cl /Z7 arena.c /link /subsystem:console kernel32.lib libvcruntime.lib
+#define W32(r) __declspec(dllimport) r __stdcall
+W32(void)   ExitProcess(int);
+W32(void *) VirtualAlloc(void *, ptrdiff_t, int, int);
+
+#define MEM_COMMIT     0x1000
+#define MEM_RESERVE    0x2000
+#define PAGE_NOACCESS  0x0001
+#define PAGE_READWRITE 0x0004
+
+static void *os_reserve(ptrdiff_t cap)
+{
+    return VirtualAlloc(0, cap, MEM_RESERVE, PAGE_NOACCESS);
+}
+
+static char os_commit(void *ptr, ptrdiff_t len)
+{
+    return VirtualAlloc(ptr, len, MEM_COMMIT, PAGE_READWRITE);
+}
+
+#else  // POSIX
+// $ cc -g3 -o arena arena.c
+#include <sys/mman.h>
+
+static void *os_reserve(ptrdiff_t cap)
+{
+    void *r = mmap(0, cap, PROT_NONE, MAP_ANON|MAP_PRIVATE, -1, 0);
+    return r==MAP_FAILED ? 0 : r;
+}
+
+static char os_commit(void *ptr, ptrdiff_t len)
+{
+    return !mprotect(ptr, len, PROT_READ|PROT_WRITE);
+}
+#endif // POSIX
+
+#endif /* ARENA_H */

main.cpp

@@ -6,7 +6,10 @@
 #include<vector>
 #include<stdlib.h>
 #include"TurboList.hpp"
+
+#define TLT int
 #include "turbolist.h"
+
 #include"arena.h"
 
 // #define PRINT_DBG

turbolist.h

@@ -1,6 +1,9 @@
 #ifndef TURBO_LIST_H
 #define TURBO_LIST_H
 
+#ifndef TL_NO_CSTDLIB
+#include<cstdlib>
+#endif /* TL_NO_CSTDLIB */
 #include<stdint.h>
 #include<assert.h>
 
@@ -56,11 +59,11 @@ static inline turbolist turbolist_create_adv(void* (*malloc_like)(size_t size),
 	return tl;
 }
 
-#ifndef NO_CSTDLIB
+#ifndef TL_NO_CSTDLIB
 static inline turbolist turbolist_create() {
 	return turbolist_create_adv(malloc, free, 0, 16);
 }
-#endif /* NO_CSTDLIB */
+#endif /* TL_NO_CSTDLIB */
 
 static inline void turbolist_delete(turbolist *tl) {
 	if(tl->nex) tl->free(tl->nex);