From 1d3df9f7dd82fe49001e714a4c31962387b526f6 Mon Sep 17 00:00:00 2001
From: Dmitry Stogov <dmitry@zend.com>
Date: Wed, 13 Mar 2024 19:57:53 +0300
Subject: [PATCH] Switch folding engine to semi-perfect hash function

---
 gen_ir_fold_hash.c | 53 ++++++++++++++++++++++++++++++++++++++++------
 ir.c               |  6 +++++-
 2 files changed, 52 insertions(+), 7 deletions(-)

diff --git a/gen_ir_fold_hash.c b/gen_ir_fold_hash.c
index 52205297..efd656f5 100644
--- a/gen_ir_fold_hash.c
+++ b/gen_ir_fold_hash.c
@@ -15,6 +15,10 @@
 #define MAX_RULES 2048
 #define MAX_SLOTS (MAX_RULES * 4)
 
+#define USE_SEMI_PERFECT_HASH 1
+#define USE_SHL_HASH 1
+#define USE_ROL_HASH 0
+
 static ir_strtab strtab;
 
 void print_hash(uint32_t *mask, uint32_t count)
@@ -28,12 +32,14 @@ void print_hash(uint32_t *mask, uint32_t count)
 	printf("};\n\n");
 }
 
-#if 0
+#if USE_SHL_HASH
 static uint32_t hash_shl2(uint32_t mask, uint32_t r1, uint32_t r2)
 {
 	return ((mask << r1) - mask) << r2;
 }
-#else
+#endif
+
+#if USE_ROL_HASH
 #define ir_rol(x, n)	(((x)<<(n)) | ((x)>>(-(int)(n)&(8*sizeof(x)-1))))
 #define ir_ror(x, n)	(((x)<<(-(int)(n)&(8*sizeof(x)-1))) | ((x)>>(n)))
 
@@ -50,29 +56,64 @@ int find_hash(uint32_t *mask, uint32_t count)
 	uint32_t n, r1, r2, i, h;
 
 	for (n = (count | 1); n < MAX_SLOTS; n += 2) {
+#if USE_SEMI_PERFECT_HASH
+		int semi_perfect = 0;
+#endif
+
 		for (r1 = 0; r1 < 31; r1++) {
 			for (r2 = 0; r2 < 32; r2++) {
-#if 0
+#if USE_SHL_HASH
 				memset(hash, 0, n * sizeof(uint32_t));
 				for (i = 0; i < count; i++) {
 					h = hash_shl2(mask[i] & 0x1fffff, r1, r2) % n;
-					if (hash[h]) break; /* collision */
+					if (hash[h]) {
+#if USE_SEMI_PERFECT_HASH
+						h++;
+						if (!hash[h]) {
+							hash[h] = mask[i];
+							semi_perfect = 1;
+							continue;
+						}
+#endif
+						break; /* collision */
+					}
 					hash[h] = mask[i];
 				}
 				if (i == count) {
 					print_hash(hash, n);
+#if USE_SEMI_PERFECT_HASH
+					if (semi_perfect) {
+						printf("#define IR_FOLD_SEMI_PERFECT_HASH\n\n");
+					}
+#endif
 					printf("static uint32_t _ir_fold_hashkey(uint32_t h)\n{\n\treturn (((h << %d) - h) << %d) %% %d;\n}\n", r1, r2, n);
 					return 1;
 				}
-#else
+#endif
+#if USE_ROL_HASH
 				memset(hash, 0, n * sizeof(uint32_t));
 				for (i = 0; i < count; i++) {
 					h = hash_rol2(mask[i] & 0x1fffff, r1, r2) % n;
-					if (hash[h]) break; /* collision */
+					if (hash[h]) {
+#if USE_SEMI_PERFECT_HASH
+						h++;
+						if (!hash[h]) {
+							hash[h] = mask[i];
+							semi_perfect = 1;
+							continue;
+						}
+#endif
+						break; /* collision */
+					}
 					hash[h] = mask[i];
 				}
 				if (i == count) {
 					print_hash(hash, n);
+#if USE_SEMI_PERFECT_HASH
+					if (semi_perfect) {
+						printf("#define IR_FOLD_SEMI_PERFECT_HASH\n\n");
+					}
+#endif
 					printf("static uint32_t _ir_fold_hashkey(uint32_t h)\n{\nreturn ir_rol32((ir_rol32(h, %d) - h), %d) %% %d;\n}\n", r1, r2, n);
 					return 1;
 				}
diff --git a/ir.c b/ir.c
index 32527e45..d4e3314f 100644
--- a/ir.c
+++ b/ir.c
@@ -926,7 +926,11 @@ ir_ref ir_folding(ir_ctx *ctx, uint32_t opt, ir_ref op1, ir_ref op2, ir_ref op3,
 		uint32_t k = key & any;
 		uint32_t h = _ir_fold_hashkey(k);
 		uint32_t fh = _ir_fold_hash[h];
-		if (IR_FOLD_KEY(fh) == k /*|| (fh = _ir_fold_hash[h+1], (fh & 0x1fffff) == k)*/) {
+		if (IR_FOLD_KEY(fh) == k
+#ifdef IR_FOLD_SEMI_PERFECT_HASH
+		 || (fh = _ir_fold_hash[h+1], (fh & 0x1fffff) == k)
+#endif
+		) {
 			switch (IR_FOLD_RULE(fh)) {
 #include "ir_fold.h"
 				default: