From ac3a91f4f173e7a38c70d5cd0caf5a025642312b Mon Sep 17 00:00:00 2001
From: Sebastiano Tronto <sebastiano@tronto.net>
Date: Wed, 17 Dec 2025 10:51:34 +0100
Subject: Improve pruning value fetching and re-do benchmarks

---
 src/solvers/h48/gendata_h48.h | 21 +++++++++++++++++++++
 src/solvers/h48/solve.h       | 36 ++++++++++++++++--------------------
 2 files changed, 37 insertions(+), 20 deletions(-)

(limited to 'src')

diff --git a/src/solvers/h48/gendata_h48.h b/src/solvers/h48/gendata_h48.h
index 15b7589..ef8a854 100644
--- a/src/solvers/h48/gendata_h48.h
+++ b/src/solvers/h48/gendata_h48.h
@@ -18,6 +18,8 @@ STATIC_INLINE uint8_t get_h48_pval(const unsigned char *, uint64_t);
 STATIC_INLINE void set_h48_pval(unsigned char *, uint64_t, uint8_t);
 STATIC_INLINE uint8_t get_h48_pvalmin(const unsigned char *, uint64_t);
 STATIC_INLINE void set_h48_pvalmin(unsigned char *, uint64_t, uint8_t);
+STATIC_INLINE uint8_t get_h48_pval_and_min(
+    const unsigned char *, uint64_t, uint8_t [static 1]);
 
 STATIC long long
 gendata_h48_dispatch(
@@ -526,3 +528,22 @@ set_h48_pvalmin(unsigned char *table, uint64_t i, uint8_t val)
 	t = table[H48_INDEX(i)];
 	table[H48_INDEX(i)] = (t & UINT8_C(0xF)) | (v << UINT8_C(4));
 }
+
+STATIC_INLINE uint8_t
+get_h48_pval_and_min(
+	const unsigned char *table,
+	uint64_t coord_noext,
+	uint8_t pval_min[static 1]
+)
+{
+	uint64_t iext, imin;
+	uint8_t t, tmin;
+
+	iext = H48_LINE_EXT(coord_noext);
+	imin = H48_LINE_MIN(coord_noext);
+	t = table[H48_INDEX(iext)];
+	tmin = table[H48_INDEX(imin)];
+
+	*pval_min = tmin >> UINT8_C(4);
+	return (t & H48_MASK(iext)) >> H48_SHIFT(iext);
+}
diff --git a/src/solvers/h48/solve.h b/src/solvers/h48/solve.h
index 5fd3dd9..2920fa7 100644
--- a/src/solvers/h48/solve.h
+++ b/src/solvers/h48/solve.h
@@ -108,9 +108,9 @@ STATIC_INLINE bool
 solve_h48_stop(dfsarg_solve_h48_t arg[static 1])
 {
 	uint32_t data, data_inv;
-	int64_t coord, coordext, coordmin;
+	int64_t coord;
 	int8_t target, nh, n;
-	uint8_t pval_min, pval_eoesep;
+	uint8_t pval, pval_min, pval_eoesep;
 
 	arg->movemask_normal = arg->movemask_inverse = MM18_ALLMOVES;
 	arg->nodes_visited++;
@@ -132,25 +132,23 @@ solve_h48_stop(dfsarg_solve_h48_t arg[static 1])
 	/* Inverse probing */
 
 	if (!arg->use_lb_inverse) {
+		arg->table_lookups++;
+		arg->use_lb_inverse = true;
 		coord = coord_h48_edges(
 		    arg->inverse, COCLASS(data_inv), TTREP(data_inv), arg->h);
-		coordext = H48_LINE_EXT(coord);
-		arg->lb_inverse = get_h48_pval(arg->h48data, coordext);
-		arg->table_lookups++;
+		pval = get_h48_pval_and_min(arg->h48data, coord, &pval_min);
 
-		if (arg->lb_inverse == 0) {
+		if (pval == 0) {
 			arg->table_fallbacks++;
 
-			coordmin = H48_LINE_MIN(coord);
-			pval_min = get_h48_pvalmin(arg->h48data, coordmin);
 			pval_eoesep = get_eoesep_pval_cube(
 			    arg->h48data_fallback_eoesep, arg->inverse);
-			arg->lb_inverse = MAX(pval_min, pval_eoesep);
+			pval = MAX(pval_min, pval_eoesep);
 		} else {
-			arg->lb_inverse += arg->base;
+			pval += arg->base;
 		}
 
-		arg->use_lb_inverse = true;
+		arg->lb_inverse = pval;
 	}
 
 	if (arg->lb_inverse > target)
@@ -161,25 +159,23 @@ solve_h48_stop(dfsarg_solve_h48_t arg[static 1])
 	/* Normal probing */
 
 	if (!arg->use_lb_normal) {
+		arg->table_lookups++;
+		arg->use_lb_normal = true;
 		coord = coord_h48_edges(
 		    arg->cube, COCLASS(data), TTREP(data), arg->h);
-		coordext = H48_LINE_EXT(coord);
-		arg->lb_normal = get_h48_pval(arg->h48data, coordext);
-		arg->table_lookups++;
+		pval = get_h48_pval_and_min(arg->h48data, coord, &pval_min);
 
-		if (arg->lb_normal == 0) {
+		if (pval == 0) {
 			arg->table_fallbacks++;
 
-			coordmin = H48_LINE_MIN(coord);
-			pval_min = get_h48_pval(arg->h48data, coordmin);
 			pval_eoesep = get_eoesep_pval_cube(
 			    arg->h48data_fallback_eoesep, arg->cube);
-			arg->lb_normal = MAX(pval_min, pval_eoesep);
+			pval = MAX(pval_min, pval_eoesep);
 		} else {
-			arg->lb_normal += arg->base;
+			pval += arg->base;
 		}
 
-		arg->use_lb_normal = true;
+		arg->lb_normal = pval;
 	}
 
 	if (arg->lb_normal > target)
-- 
cgit v1.3