cheaper mul by small constants on amd64

1 files changed, 56 insertions, 0 deletions

diff --git a/amd64/emit.c b/amd64/emit.c
index 00dd80f..bbc7def 100644
--- a/amd64/emit.c
+++ b/amd64/emit.c
@@ -360,6 +360,54 @@ Next:
 	goto Next;
 }
 
+static int
+mulops(int depth, int *ops, int64_t k)
+{
+	static int tab[] = {9, 8, 5, 4, 3, 2, 1, 0};
+	int *pm, m;
+
+	while (depth && k>1) {
+		for (pm=tab; (m=*pm); pm++)
+			if (k % m == 0) {
+				*ops++ = m;
+				depth -= 1;
+				k /= m;
+				goto next;
+			}
+		return 0;
+	next:;
+	}
+	*ops = 1;
+	return k == 1;
+}
+
+static int
+emitmulk(Ref to, Ref r, int64_t k, Ins *i, Fn *fn, FILE *f)
+{
+	int ops[3], *pm, m;
+	char fmt[32];
+
+	if (k == 1 || !mulops(2, ops, k))
+		return 0;
+
+	for (pm=ops; (m=*pm) != 1; pm++) {
+		if (m == 4 || m == 2) {
+			if (!req(r, to))
+				emitf("mov%k %1, %=", i, fn, f);
+			emitf("add%k %=, %=", i, fn, f);
+			if (m == 4)
+				emitf("add%k %=, %=", i, fn, f);
+		} else {
+			sprintf(fmt, "lea%%k (%s, %%1, %d), %%=",
+				(m & 1) ? "%1" : "", m & ~1);
+			emitf(fmt, i, fn, f);
+		}
+		r = to;
+		i->arg[1] = to;
+	}
+	return 1;
+}
+
 static void *negmask[4] = {
 	[Ks] = (uint32_t[4]){ 0x80000000 },
 	[Kd] = (uint64_t[2]){ 0x8000000000000000 },
@@ -407,6 +455,14 @@ emitins(Ins i, Fn *fn, FILE *f)
 			i.arg[0] = i.arg[1];
 			i.arg[1] = r;
 		}
+		if (KBASE(i.cls) == 0)
+		if (rtype(i.arg[0]) == RCon) {
+			con = &fn->con[i.arg[0].val];
+			val = con->bits.i;
+			if (con->type == CBits)
+			if (emitmulk(i.to, i.arg[1], val, &i, fn, f))
+				break;
+		}
 		if (KBASE(i.cls) == 0 /* only available for ints */
 		&& rtype(i.arg[0]) == RCon
 		&& rtype(i.arg[1]) == RTmp) {