From 04f49d1ca41f3706bf74f867c775797cf3bd2037 Mon Sep 17 00:00:00 2001
From: Bill Currie <bill@taniwha.org>
Date: Sun, 1 Oct 2023 21:30:41 +0900
Subject: [PATCH] [qfcc] Commit common scale terms

While it works, and does improve the code slightly, it could do better
by favoring constants over variables for the common factor.
---
 tools/qfcc/include/expr.h         |   5 +
 tools/qfcc/source/expr_algebra.c  |  30 +++---
 tools/qfcc/source/expr_optimize.c | 159 ++++++++++++++++++++++++++++--
 tools/qfcc/test/motor-point.r     |   2 +-
 4 files changed, 173 insertions(+), 23 deletions(-)

diff --git a/tools/qfcc/include/expr.h b/tools/qfcc/include/expr.h
index 4a8a8ccd6..8a5b042ba 100644
--- a/tools/qfcc/include/expr.h
+++ b/tools/qfcc/include/expr.h
@@ -924,6 +924,7 @@ const expr_t *edag_add_expr (const expr_t *e);
 bool is_scale (const expr_t *expr) __attribute__((pure));
 bool is_cross (const expr_t *expr) __attribute__((pure));
 bool is_sum (const expr_t *expr) __attribute__((pure));
+bool is_mult (const expr_t *expr) __attribute__((pure));
 bool is_neg (const expr_t *expr) __attribute__((pure));
 
 const expr_t *neg_expr (const expr_t *e);
@@ -941,6 +942,10 @@ void scatter_terms (const expr_t *sum,
 					const expr_t **adds, const expr_t **subs);
 const expr_t *gather_terms (type_t *type,
 							const expr_t **adds, const expr_t **subs);
+int count_factors (const expr_t *expr) __attribute__((pure));
+void scatter_factors (const expr_t *prod, const expr_t **factors);
+const expr_t *gather_factors (type_t *type, int op, const expr_t **factors,
+							  int count);
 
 ///@}
 
diff --git a/tools/qfcc/source/expr_algebra.c b/tools/qfcc/source/expr_algebra.c
index 1fb6be8e1..c0c3419db 100644
--- a/tools/qfcc/source/expr_algebra.c
+++ b/tools/qfcc/source/expr_algebra.c
@@ -425,7 +425,7 @@ is_sum (const expr_t *expr)
 			&& (expr->expr.op == '+' || expr->expr.op == '-'));
 }
 
-static bool __attribute__((pure))
+bool
 is_mult (const expr_t *expr)
 {
 	return (expr && expr->type == ex_expr
@@ -450,7 +450,7 @@ count_terms (const expr_t *expr)
 	return terms;
 }
 
-static int __attribute__((pure))
+int __attribute__((pure))
 count_factors (const expr_t *expr)
 {
 	if (!is_mult (expr)) {
@@ -475,34 +475,34 @@ is_cross (const expr_t *expr)
 }
 
 static void
-distribute_factors_core (const expr_t *prod, const expr_t **factors, int *ind)
+scatter_factors_core (const expr_t *prod, const expr_t **factors, int *ind)
 {
 	auto e1 = prod->expr.e1;
 	auto e2 = prod->expr.e2;
 	if (is_mult (e1)) {
-		distribute_factors_core (e1, factors, ind);
+		scatter_factors_core (e1, factors, ind);
 	} else {
 		factors[(*ind)++] = e1;
 	}
 	if (is_mult (e2)) {
-		distribute_factors_core (e2, factors, ind);
+		scatter_factors_core (e2, factors, ind);
 	} else {
 		factors[(*ind)++] = e2;
 	}
 }
 
-static void
-distribute_factors (const expr_t *prod, const expr_t **factors)
+void
+scatter_factors (const expr_t *prod, const expr_t **factors)
 {
 	if (!is_mult (prod)) {
-		internal_error (prod, "distribute_factors with no product");
+		internal_error (prod, "scatter_factors with no product");
 	}
 	int         ind = 0;
-	distribute_factors_core (prod, factors, &ind);
+	scatter_factors_core (prod, factors, &ind);
 }
 
-static const expr_t *
-collect_factors (type_t *type, int op, const expr_t **factors, int count)
+const expr_t *
+gather_factors (type_t *type, int op, const expr_t **factors, int count)
 {
 	if (!count) {
 		internal_error (0, "no factors to collect");
@@ -516,8 +516,8 @@ collect_factors (type_t *type, int op, const expr_t **factors, int count)
 		b = factors[1];
 	} else {
 		int mid = (count + 1) / 2;
-		a = collect_factors (type, op, factors, mid);
-		b = collect_factors (type, op, factors + mid, count - mid);
+		a = gather_factors (type, op, factors, mid);
+		b = gather_factors (type, op, factors + mid, count - mid);
 	}
 	auto prod = typed_binary_expr (type, op, a, b);
 	return edag_add_expr (prod);
@@ -551,9 +551,9 @@ sort_factors (type_t *type, const expr_t *e)
 
 	int count = count_factors (e);
 	const expr_t *factors[count + 1] = {};
-	distribute_factors (e, factors);
+	scatter_factors (e, factors);
 	heapsort (factors, count, sizeof (factors[0]), expr_ptr_cmp);
-	auto mult = collect_factors (type, '*', factors, count);
+	auto mult = gather_factors (type, '*', factors, count);
 	return mult;
 }
 
diff --git a/tools/qfcc/source/expr_optimize.c b/tools/qfcc/source/expr_optimize.c
index 315ca17a4..ed941ecf3 100644
--- a/tools/qfcc/source/expr_optimize.c
+++ b/tools/qfcc/source/expr_optimize.c
@@ -45,6 +45,18 @@
 static const expr_t *optimize_core (const expr_t *expr);
 static const expr_t skip;
 
+static void
+clean_skips (const expr_t **expr_list)
+{
+	auto dst = expr_list;
+	for (auto src = dst; *src; src++) {
+		if (*src != &skip) {
+			*dst++ = *src;
+		}
+	}
+	*dst = 0;
+}
+
 static const expr_t *
 rescale (const expr_t *expr, const expr_t *target, const expr_t *remove)
 {
@@ -62,6 +74,37 @@ rescale (const expr_t *expr, const expr_t *target, const expr_t *remove)
 	return scale_expr (type, rescale (expr->expr.e1, target, remove), scale);
 }
 
+static const expr_t *
+remult (const expr_t *expr, const expr_t *remove)
+{
+	if (!is_mult (expr)) {
+		internal_error (expr, "not a mult expression");
+	}
+	auto type = get_type (expr);
+	int count = count_factors (expr);
+	const expr_t *factors[count + 1] = {};
+	scatter_factors (expr, factors);
+	for (auto f = factors; *f; f++) {
+		if (*f == remove) {
+			*f = &skip;
+			break;
+		}
+	}
+	clean_skips (factors);
+	auto new = gather_factors (type, expr->expr.op, factors, count - 1);
+	return new;
+}
+
+static const expr_t *
+remult_scale (const expr_t *expr, const expr_t *remove)
+{
+	auto mult = remult (expr->expr.e2, remove);
+	auto scalee = expr->expr.e1;
+	auto type = get_type (expr);
+	auto new = typed_binary_expr (type, SCALE, scalee, mult);
+	return edag_add_expr (new);
+}
+
 static const expr_t *
 optimize_cross (const expr_t *expr, const expr_t **adds, const expr_t **subs)
 {
@@ -170,16 +213,95 @@ optimize_cross (const expr_t *expr, const expr_t **adds, const expr_t **subs)
 	return cross;
 }
 
-static void
-clean_skips (const expr_t **expr_list)
+static bool __attribute__((pure))
+mult_has_factor (const expr_t *mult, const expr_t *factor)
 {
-	auto dst = expr_list;
-	for (auto src = dst; *src; src++) {
-		if (*src != &skip) {
-			*dst++ = *src;
+	if (!is_mult (mult)) {
+		return false;
+	}
+	if (mult->expr.e1 == factor || mult->expr.e2 == factor) {
+		return true;
+	}
+	bool has_factor = mult_has_factor (mult->expr.e1, factor);
+	if (!has_factor) {
+		has_factor = mult_has_factor (mult->expr.e2, factor);
+	}
+	return has_factor;
+}
+
+static const expr_t *
+optimize_scale (const expr_t *expr, const expr_t **adds, const expr_t **subs)
+{
+	auto mult = expr->expr.e2;
+	int num_factors = count_factors (mult);
+	int total = 0;
+	int fac_counts[num_factors + 1] = {};
+	const expr_t *factors[num_factors + 2] = {};
+	if (is_mult (mult)) {
+		scatter_factors (mult, factors);
+	} else {
+		factors[0] = mult;
+	}
+
+	for (auto search = adds; *search; search++) {
+		if (is_scale (*search)) {
+			for (auto f = factors; *f; f++) {
+				if (mult_has_factor ((*search)->expr.e2, *f)) {
+					fac_counts[f - factors]++;
+					total++;
+				}
+			}
 		}
 	}
-	*dst = 0;
+	for (auto search = subs; *search; search++) {
+		if (is_scale (*search)) {
+			for (auto f = factors; *f; f++) {
+				if (mult_has_factor ((*search)->expr.e2, *f)) {
+					fac_counts[f - factors]++;
+					total++;
+				}
+			}
+		}
+	}
+	if (!total) {
+		return expr;
+	}
+
+	const expr_t *common = 0;
+	int count = 0;
+	for (auto f = factors; *f; f++) {
+		if (fac_counts[f - factors] > count) {
+			common = *f;
+			count = fac_counts[f - factors];
+		}
+	}
+
+	const expr_t *com_adds[count + 2] = {};
+	const expr_t *com_subs[count + 2] = {};
+	auto dst = com_adds;
+	*dst++ = remult_scale (expr, common);
+	for (auto src = adds; *src; src++) {
+		if (is_scale (*src) && mult_has_factor ((*src)->expr.e2, common)) {
+			*dst++ = remult_scale (*src, common);
+			*src = &skip;
+		}
+	}
+	dst = com_subs;
+	for (auto src = subs; *src; src++) {
+		if (is_scale (*src) && mult_has_factor ((*src)->expr.e2, common)) {
+			*dst++ = remult_scale (*src, common);
+			*src = &skip;
+		}
+	}
+
+	auto type = get_type (expr);
+	auto scale = expr->expr.e1;
+	auto col = gather_terms (type, com_adds, com_subs);
+	col = optimize_core (col);
+
+	scale = typed_binary_expr (type, SCALE, col, common);
+	scale = edag_add_expr (scale);
+	return scale;
 }
 
 static void
@@ -216,6 +338,28 @@ optimize_cross_products (const expr_t **adds, const expr_t **subs)
 	clean_skips (subs);
 }
 
+static void
+optimize_scale_products (const expr_t **adds, const expr_t **subs)
+{
+	for (auto scan = adds; *scan; scan++) {
+		if (is_scale (*scan)) {
+			auto e = *scan;
+			*scan = &skip;
+			*scan = optimize_scale (e, adds, subs);
+		}
+	}
+	for (auto scan = subs; *scan; scan++) {
+		if (is_scale (*scan)) {
+			auto e = *scan;
+			*scan = &skip;
+			*scan = optimize_scale (e, subs, adds);
+		}
+	}
+
+	clean_skips (adds);
+	clean_skips (subs);
+}
+
 static int
 expr_ptr_cmp (const void *_a, const void *_b)
 {
@@ -269,6 +413,7 @@ optimize_core (const expr_t *expr)
 		optimize_adds (subs);
 
 		optimize_cross_products (adds, subs);
+		optimize_scale_products (adds, subs);
 
 		expr = gather_terms (type, adds, subs);
 	}
diff --git a/tools/qfcc/test/motor-point.r b/tools/qfcc/test/motor-point.r
index a2a3602b7..56bfbef58 100644
--- a/tools/qfcc/test/motor-point.r
+++ b/tools/qfcc/test/motor-point.r
@@ -23,7 +23,7 @@ main (void)
 	point_t p = (point_t)'10 4 -1.5 1'f;
 	point_t n = apply_motor (m, p);
 	printf ("n: %.9q\n", n);
-	if ((vec4)n != '9.99999905 -3.99999952 -1.49999988 0.99999994'f) {
+	if ((vec4)n != '10 -3.99999952 -1.49999988 0.99999994'f) {
 		return 1;
 	}
 	return 0;