From 4e4c266632d64437e17c2c7f815734b952a2a387 Mon Sep 17 00:00:00 2001
From: Richard Thier <magosit@outlook.hu>
Date: Mon, 21 Oct 2024 14:21:34 +0200
Subject: [PATCH] RAW (can-fail) flags for optimization and non-failing
 implementation added

---
 main.cpp |  2 +-
 simap.h  | 62 ++++++++++++++++++++++++++++++++++++++++++--------------
 2 files changed, 48 insertions(+), 16 deletions(-)

diff --git a/main.cpp b/main.cpp
index ae68fe1..8cce812 100644
--- a/main.cpp
+++ b/main.cpp
@@ -135,7 +135,7 @@ int main() {
 	test_basics(unomap, &umi);
 
 	/* Performance tests */
-	int i = 100;
+	int i = 10000;
 	keystore(i, true);
 	datastore(i, true);
 
diff --git a/simap.h b/simap.h
index c6b6a99..6d82168 100644
--- a/simap.h
+++ b/simap.h
@@ -7,8 +7,18 @@
 #include "amap.h"
 #include "arena.h/arena.h"
 
+/* Possible optimizations, but they mean there can be lookup / insert errors (very rarely)
+*/
+#define SIMAP_AVX2_RAW
+#define SIMAP_RAW
+
 /* Perf trickery */
 
+/* This unifies the ifdefs but separates code paths when needed */
+#ifdef SIMAP_AVX2_RAW
+#define SIMAP_RAW
+#endif
+
 /* I have no idea what MSVC has instead... */
 #ifdef _MSC_VER
 #define SM_THREAD_LOCAL __declspec(thread)
@@ -41,6 +51,17 @@ union simap_ptr64 {
 };
 typedef union simap_ptr64 simap_ptr64;
 
+struct elem_nonkey_prefix {
+	/** The value (ptr) */
+	simap_ptr64 value;
+
+	/** Previous element index from base (full offset) */
+	uint32_t previndex;
+	/** Next element index from base (full offset) */
+	uint32_t nextindex;
+};
+typedef struct elem_nonkey_prefix elem_nonkey_prefix;
+
 /**
  * The per-element storage layout
  *
@@ -61,15 +82,10 @@ typedef union simap_ptr64 simap_ptr64;
  * and if you want check extra validity by jumping back&forth in it.
  */
 struct elem_prefix {
-	/** The value (ptr) */
-	simap_ptr64 value;
+	/** Value and meta-data - divisible by 8bytes */
+	elem_nonkey_prefix nonkey_prefix;
 
-	/** Previous element index from base (full offset) */
-	uint32_t previndex;
-	/** Next element index from base (full offset) */
-	uint32_t nextindex;
-
-	/** The prefix of the key */
+	/** The prefix of the key - divisible by 8bytes padded string after this (inlined) */
 	simap_c64 key_prefix;
 };
 typedef struct elem_prefix elem_prefix;
@@ -107,10 +123,24 @@ static inline simap_instance simap_create() {
 static inline void* simap(void *amap_instance, AMAP_OP op, const char *key, void *ptr);
 
 // TODO: We can possibly hand-optimise this with intrinsics maybe - but I hope autovectorization (does not seem to happen???)
-static inline SM_ALWAYS_INLINE auint64 *make_tipp(auint64 *tip, auint64 prefix, auint64 *end) {
-	#pragma GCC unroll 4
+static inline SM_ALWAYS_INLINE auint64 *make_tipp(auint64 *base, auint64 *tip, auint64 prefix, auint64 *end) {
+#ifdef SIMAP_AVX2_RAW
+	/* TODO: Implement */
+#endif
+#ifdef SIMAP_RAW
+	#pragma GCC unroll 16
 	while((++tip < end) && (*tip != prefix));
 	return tip;
+#endif
+	/* XXX: This only works because of (***) because reading -1 tips makes tip >= end for sure here and back */
+	elem_nonkey_prefix *pre = (elem_nonkey_prefix *)((uint8_t *)tip - sizeof(elem_nonkey_prefix));
+	tip = (auint64 *) ((uint8_t *)base + pre->nextindex + sizeof(elem_nonkey_prefix));
+	#pragma GCC unroll 16
+	while((tip < end) && (*tip != prefix)) {
+		pre = (elem_nonkey_prefix *)((uint8_t *)tip - sizeof(elem_nonkey_prefix));
+		tip = (auint64 *) ((uint8_t *)base + pre->nextindex + sizeof(elem_nonkey_prefix));
+	}
+	return tip;
 }
 
 static inline simap_ptr64 *simap_search_internal(simap_instance *map, const char *key) {
@@ -128,8 +158,8 @@ static inline simap_ptr64 *simap_search_internal(simap_instance *map, const char
 	/* Lookup prefix (fast-key) - hopefully this gets vectorized (should be)!!! */
 	auint64 *base = (auint64 *) (map->base);
 	auint64 *end = (auint64 *)((uint8_t *)base + (map->usage_end));
-	auint64 *tipp = make_tipp(base, prefix.u64, end);
-	while(tipp < end) {
+	auint64 *tipp = make_tipp(base, base, prefix.u64, end);
+	while(tipp < end) { // XXX: (***)
 
 		/* Need detailed lookup, because found the prefix */
 		assert((*tipp == prefix.u64));
@@ -138,13 +168,14 @@ static inline simap_ptr64 *simap_search_internal(simap_instance *map, const char
 		if(!is_smallkey) {
 			char *tippremains = (char *)((uint8_t *)tipp + sizeof(uint64_t));
 			if(strcmp(keyremains, tippremains) != 0) {
-				tipp = make_tipp(tipp, prefix.u64, end);
+				tipp = make_tipp(base, tipp, prefix.u64, end);
 				continue;
 			}
 		}
 
 		simap_ptr64 *ptr = (simap_ptr64 *)((uint8_t *) (tipp - 2));
 
+#ifdef SIMAP_RAW
 		/* Check back & forth (jump validation) */
 		uint32_t previ = *((uint32_t *)(tipp - 1));
 		if(previ == (uint32_t) -1) {
@@ -160,9 +191,10 @@ static inline simap_ptr64 *simap_search_internal(simap_instance *map, const char
 				+ sizeof(uint32_t));
 
 		if(retipp != tipp) {
-			tipp = make_tipp(tipp, prefix.u64, end);
+			tipp = make_tipp(base, tipp, prefix.u64, end);
 			continue;
 		}
+#endif /* SIMAP_RAW */
 
 		/* Can have the (statistically checked) pointer */
 		return ptr;
@@ -229,7 +261,7 @@ static inline void *simap_force_add(simap_instance *map, const char *key, void *
 	uint32_t *usprev = (uint32_t *)((uint8_t *)(map->base) + usi +
 		sizeof(simap_ptr64));
 	*usprev = previ;
-	*(usprev + 1) = (uint32_t) -1;
+	*(usprev + 1) = (uint32_t) -1; /* XXX: (***): ensures the "not < end" here! */
 
 	/* 8byte: First 8 char */
 	simap_c64 *start_str = (simap_c64 *)(usprev + 2);