make regalloc work with ssa form

+ Still not ideal register allocator, but can be improved with some
  heuristics and potentially completely replaced in the future if so
  desired with graph coloring or something.

6 files changed, 439 insertions, 178 deletions

diff --git a/src/abi.c b/src/abi.c
index cb92190..80f34e4 100644
--- a/src/abi.c
+++ b/src/abi.c
@@ -19,7 +19,7 @@ static struct insn rewrite_param(struct insn n)
 	assert(n.in[1].class == IMM);
 	int64_t nth_param = n.in[1].v;
 	assert(nth_param >= 0 && nth_param < 25
-			&& "stack argument passing not yet supported");
+	       && "stack argument passing not yet supported");
 
 	return insn_create(MOVE, I27, n.out, nth_ar(nth_param), noclass());
 }
@@ -41,9 +41,11 @@ static struct insn rewrite_arg(struct insn n)
 	assert(nth_arg >= 0 && nth_arg < 25);
 
 	if (n.in[0].class == TMP || n.in[0].class == REG)
-		return insn_create(MOVE, I27, nth_ar(nth_arg), n.in[0], noclass());
+		return insn_create(MOVE, I27, nth_ar(nth_arg), n.in[0],
+		                   noclass());
 	else if (n.in[0].class == IMM || n.in[0].class == REF)
-		return insn_create(COPY, I27, nth_ar(nth_arg), n.in[0], noclass());
+		return insn_create(COPY, I27, nth_ar(nth_arg), n.in[0],
+		                   noclass());
 
 	assert("illegal arg type");
 	abort();
@@ -57,9 +59,11 @@ static struct insn rewrite_retarg(struct insn n)
 	assert(nth_arg >= 0 && nth_arg < 25);
 
 	if (n.in[0].class == TMP || n.in[0].class == REG)
-		return insn_create(MOVE, I27, nth_ar(nth_arg), n.in[0], noclass());
+		return insn_create(MOVE, I27, nth_ar(nth_arg), n.in[0],
+		                   noclass());
 	else if (n.in[0].class == IMM || n.in[0].class == REF)
-		return insn_create(COPY, I27, nth_ar(nth_arg), n.in[0], noclass());
+		return insn_create(COPY, I27, nth_ar(nth_arg), n.in[0],
+		                   noclass());
 
 	assert("illegal retval type");
 	abort();
diff --git a/src/asm.c b/src/asm.c
index e177c4e..fd4d68a 100644
--- a/src/asm.c
+++ b/src/asm.c
@@ -8,87 +8,87 @@ static const char *rname(struct val v)
 	assert(v.class == REG);
 	/** @todo use ABI names? */
 	switch (v.r) {
-		case RX0: return "x0";
-		case RX1: return "x1";
-		case RX2: return "x2";
-		case RX3: return "x3";
-		case RX4: return "x4";
-		case RX5: return "x5";
-		case RX6: return "x6";
-		case RX7: return "x7";
-		case RX8: return "x8";
-		case RX9: return "x9";
-		case RX10: return "x10";
-		case RX11: return "x11";
-		case RX12: return "x12";
-		case RX13: return "x13";
-		case RX14: return "x14";
-		case RX15: return "x15";
-		case RX16: return "x16";
-		case RX17: return "x17";
-		case RX18: return "x18";
-		case RX19: return "x19";
-		case RX20: return "x20";
-		case RX21: return "x21";
-		case RX22: return "x22";
-		case RX23: return "x23";
-		case RX24: return "x24";
-		case RX25: return "x25";
-		case RX26: return "x26";
-		case RX27: return "x27";
-		case RX28: return "x28";
-		case RX29: return "x29";
-		case RX30: return "x30";
-		case RX31: return "x31";
-		case RX32: return "x32";
-		case RX33: return "x33";
-		case RX34: return "x34";
-		case RX35: return "x35";
-		case RX36: return "x36";
-		case RX37: return "x37";
-		case RX38: return "x38";
-		case RX39: return "x39";
-		case RX40: return "x40";
-		case RX41: return "x41";
-		case RX42: return "x42";
-		case RX43: return "x43";
-		case RX44: return "x44";
-		case RX45: return "x45";
-		case RX46: return "x46";
-		case RX47: return "x47";
-		case RX48: return "x48";
-		case RX49: return "x49";
-		case RX50: return "x50";
-		case RX51: return "x51";
-		case RX52: return "x52";
-		case RX53: return "x53";
-		case RX54: return "x54";
-		case RX55: return "x55";
-		case RX56: return "x56";
-		case RX57: return "x57";
-		case RX58: return "x58";
-		case RX59: return "x59";
-		case RX60: return "x60";
-		case RX61: return "x61";
-		case RX62: return "x62";
-		case RX63: return "x63";
-		case RX64: return "x64";
-		case RX65: return "x65";
-		case RX66: return "x66";
-		case RX67: return "x67";
-		case RX68: return "x68";
-		case RX69: return "x69";
-		case RX70: return "x70";
-		case RX71: return "x71";
-		case RX72: return "x72";
-		case RX73: return "x73";
-		case RX74: return "x74";
-		case RX75: return "x75";
-		case RX76: return "x76";
-		case RX77: return "x77";
-		case RX78: return "x78";
-		case RX79: return "x79";
-		case RX80: return "x80";
+	case RX0: return "x0";
+	case RX1: return "x1";
+	case RX2: return "x2";
+	case RX3: return "x3";
+	case RX4: return "x4";
+	case RX5: return "x5";
+	case RX6: return "x6";
+	case RX7: return "x7";
+	case RX8: return "x8";
+	case RX9: return "x9";
+	case RX10: return "x10";
+	case RX11: return "x11";
+	case RX12: return "x12";
+	case RX13: return "x13";
+	case RX14: return "x14";
+	case RX15: return "x15";
+	case RX16: return "x16";
+	case RX17: return "x17";
+	case RX18: return "x18";
+	case RX19: return "x19";
+	case RX20: return "x20";
+	case RX21: return "x21";
+	case RX22: return "x22";
+	case RX23: return "x23";
+	case RX24: return "x24";
+	case RX25: return "x25";
+	case RX26: return "x26";
+	case RX27: return "x27";
+	case RX28: return "x28";
+	case RX29: return "x29";
+	case RX30: return "x30";
+	case RX31: return "x31";
+	case RX32: return "x32";
+	case RX33: return "x33";
+	case RX34: return "x34";
+	case RX35: return "x35";
+	case RX36: return "x36";
+	case RX37: return "x37";
+	case RX38: return "x38";
+	case RX39: return "x39";
+	case RX40: return "x40";
+	case RX41: return "x41";
+	case RX42: return "x42";
+	case RX43: return "x43";
+	case RX44: return "x44";
+	case RX45: return "x45";
+	case RX46: return "x46";
+	case RX47: return "x47";
+	case RX48: return "x48";
+	case RX49: return "x49";
+	case RX50: return "x50";
+	case RX51: return "x51";
+	case RX52: return "x52";
+	case RX53: return "x53";
+	case RX54: return "x54";
+	case RX55: return "x55";
+	case RX56: return "x56";
+	case RX57: return "x57";
+	case RX58: return "x58";
+	case RX59: return "x59";
+	case RX60: return "x60";
+	case RX61: return "x61";
+	case RX62: return "x62";
+	case RX63: return "x63";
+	case RX64: return "x64";
+	case RX65: return "x65";
+	case RX66: return "x66";
+	case RX67: return "x67";
+	case RX68: return "x68";
+	case RX69: return "x69";
+	case RX70: return "x70";
+	case RX71: return "x71";
+	case RX72: return "x72";
+	case RX73: return "x73";
+	case RX74: return "x74";
+	case RX75: return "x75";
+	case RX76: return "x76";
+	case RX77: return "x77";
+	case RX78: return "x78";
+	case RX79: return "x79";
+	case RX80: return "x80";
 	}
 
 	assert(0 && "illegal register");
@@ -114,7 +114,7 @@ static void save_state(struct fn *f, FILE *o)
 
 	for (size_t i = 0; i < f->max_callee_save; ++i) {
 		fprintf(o, "st w s%zi, -%zi(fp)\n",
-				i, 3 * i + 9);
+		        i, 3 * i + 9);
 	}
 }
 
@@ -122,7 +122,7 @@ static void restore_state(struct fn *f, FILE *o)
 {
 	for (size_t i = 0; i < f->max_callee_save; ++i) {
 		fprintf(o, "ld w s%zi, -%zi(fp)\n",
-				i, 3 * i + 9);
+		        i, 3 * i + 9);
 	}
 
 	if (f->has_calls)
@@ -134,22 +134,23 @@ static void restore_state(struct fn *f, FILE *o)
 
 static void output_move(struct insn n, FILE *o)
 {
-	fprintf(o, "mv %s, %s\n",
-			rname(n.out), rname(n.in[0]));
+	if (!same_val(n.out, n.in[0]))
+		fprintf(o, "mv %s, %s\n",
+		        rname(n.out), rname(n.in[0]));
 }
 
 static void output_add(struct insn n, FILE *o)
 {
 	if (n.in[1].class == REG) {
 		fprintf(o, "add %s, %s, %s\n",
-				rname(n.out), rname(n.in[0]), rname(n.in[1]));
+		        rname(n.out), rname(n.in[0]), rname(n.in[1]));
 		return;
 	}
 	else if (n.in[1].class == IMM) {
 		/** @todo fix for values larger than what addi allows, mark one
 		 * temporary register reserved for the compiler? */
 		fprintf(o, "addi %s, %s, %lli\n",
-				rname(n.out), rname(n.in[0]), (long long int)n.in[1].v);
+		        rname(n.out), rname(n.in[0]), (long long int)n.in[1].v);
 		return;
 	}
 
@@ -161,16 +162,18 @@ static void output_sub(struct insn n, FILE *o)
 {
 	if (n.in[1].class == REG) {
 		fprintf(o, "sub %s, %s, %s\n",
-				rname(n.out), rname(n.in[0]), rname(n.in[1]));
+		        rname(n.out), rname(n.in[0]), rname(n.in[1]));
 		return;
 	}
 	else if (n.in[1].class == IMM) {
 		if (n.in[1].v >= 0)
 			fprintf(o, "addi %s, %s, -%lli\n",
-				rname(n.out), rname(n.in[0]), (long long int)n.in[1].v);
+			        rname(n.out), rname(n.in[0]),
+			        (long long int)n.in[1].v);
 		else /* double negative */
 			fprintf(o, "addi %s, %s, %lli\n",
-				rname(n.out), rname(n.in[0]), (long long int)n.in[1].v);
+			        rname(n.out), rname(n.in[0]),
+			        (long long int)n.in[1].v);
 		return;
 	}
 
@@ -181,7 +184,7 @@ static void output_sub(struct insn n, FILE *o)
 static void output_copy(struct insn n, FILE *o)
 {
 	fprintf(o, "li %s, %lli\n",
-			rname(n.out), (long long int)n.in[0].v);
+	        rname(n.out), (long long int)n.in[0].v);
 }
 
 static void output_call(struct insn n, FILE *o)
@@ -205,13 +208,13 @@ static void output_insn(struct insn n, FILE *o)
 	 * we may be missing out on certain optimizations by not using some kind
 	 * of matching here but good enough for now */
 	switch (n.type) {
-		case MOVE: output_move(n, o); break;
-		case ADD: output_add(n, o); break;
-		case SUB: output_sub(n, o); break;
-		case COPY: output_copy(n, o); break;
-		case CALL: output_call(n, o); break;
-		default: fprintf(stderr, "unimplemented insn: %s\n", op_str(n.type));
-			 abort();
+	case MOVE: output_move(n, o); break;
+	case ADD: output_add(n, o); break;
+	case SUB: output_sub(n, o); break;
+	case COPY: output_copy(n, o); break;
+	case CALL: output_call(n, o); break;
+	default: fprintf(stderr, "unimplemented insn: %s\n", op_str(n.type));
+		abort();
 	}
 }
 
@@ -219,16 +222,16 @@ static void output_blt(struct blk *b, struct fn *f, FILE *o)
 {
 	assert(b->s2);
 	fprintf(o, "blt %s, %s, .%s.%lli\n",
-			rname(b->cmp[0]), rname(b->cmp[1]),
-				f->name, (long long int)b->s2->id);
+	        rname(b->cmp[0]), rname(b->cmp[1]),
+	        f->name, (long long int)b->s2->id);
 }
 
 static void output_ble(struct blk *b, struct fn *f, FILE *o)
 {
 	assert(b->s2);
 	fprintf(o, "ble %s, %s, .%s.%lli\n",
-			rname(b->cmp[0]), rname(b->cmp[1]),
-				f->name, (long long int)b->s2->id);
+	        rname(b->cmp[0]), rname(b->cmp[1]),
+	        f->name, (long long int)b->s2->id);
 }
 
 static void output_ret(struct fn *f, FILE *o)
@@ -249,12 +252,25 @@ static void output_j(struct blk *b, struct fn *f, FILE *o)
 static void output_branch(struct blk *b, struct fn *f, FILE *o)
 {
 	switch (b->btype) {
-		case RET: output_ret(f, o); break;
-		case J:   output_j(b, f, o); break;
-		case BLT: output_blt(b, f, o); break;
-		case BLE: output_ble(b, f, o); break;
-		default: fprintf(stderr, "unimplemented branch: %s\n", op_str(b->btype));
-			 abort();
+	case RET: output_ret(f, o); break;
+	case J:   output_j(b, f, o); break;
+	case BLT: output_blt(b, f, o); break;
+	case BLE: output_ble(b, f, o); break;
+	default: fprintf(stderr, "unimplemented branch: %s\n",
+		         op_str(b->btype));
+		abort();
+	}
+}
+
+static void output_moves(struct vec *params, struct vec *args, FILE *o)
+{
+	/* move arguments to parameters */
+	assert(vec_len(params) == vec_len(args));
+	foreach_blk_param(pi, *params) {
+		struct val p = blk_param_at(*params, pi);
+		struct val a = blk_param_at(*args, pi);
+		struct insn i = insn_create(MOVE, NOTYPE, p, a, noclass());
+		output_move(i, o);
 	}
 }
 
@@ -265,13 +281,20 @@ void output(struct fn *f, FILE *o)
 	foreach_blk(bi, f->blks) {
 		struct blk *b = blk_at(f->blks, bi);
 		fprintf(o, ".%s.%lli:\n",
-				f->name, (long long int)b->id);
+		        f->name, (long long int)b->id);
 
 		foreach_insn(i, b->insns) {
 			struct insn n = insn_at(b->insns, i);
 			output_insn(n, o);
 		}
 
+		if (b->s2)
+			output_moves(&b->s2->params, &b->args2, o);
+
 		output_branch(b, f, o);
+
+		/* J is again the special case */
+		if (b->s1 && b->btype != J)
+			output_moves(&b->s1->params, &b->args1, o);
 	}
 }
diff --git a/src/nodes.c b/src/nodes.c
index a6ff0b8..05f96fe 100644
--- a/src/nodes.c
+++ b/src/nodes.c
@@ -6,7 +6,8 @@
 #include <qbt/debug.h>
 #include <qbt/vec.h>
 
-void insadd(struct blk *b, enum insn_type o, enum val_type t, struct val r, struct val a0, struct val a1)
+void insadd(struct blk *b, enum insn_type o, enum val_type t, struct val r,
+            struct val a0, struct val a1)
 {
 	struct insn i = insn_create(o, t, r, a0, a1);
 	vec_append(&b->insns, &i);
@@ -35,7 +36,8 @@ int64_t idmatch(struct fn *f, const char *id)
 	return -1;
 }
 
-void finish_block(struct blk *b, enum insn_type cmp, struct val a0, struct val a1, const char *label)
+void finish_block(struct blk *b, enum insn_type cmp, struct val a0,
+                  struct val a1, const char *label)
 {
 	assert(cmp >= BEQ && cmp <= RET && "illegal comparison type for block");
 	b->btype = cmp;
@@ -198,7 +200,7 @@ void dump_block(struct blk *b)
 {
 	printf("//\t/*** block %lld ", (long long)b->id);
 	if (b->name) printf("\"%s\" ", b->name);
-	
+
 	printf("(");
 	foreach_blk_param(pi, b->params) {
 		struct val v = blk_param_at(b->params, pi);
diff --git a/src/regalloc.c b/src/regalloc.c
index 474e5a4..68e68a2 100644
--- a/src/regalloc.c
+++ b/src/regalloc.c
@@ -1,90 +1,317 @@
 #include <assert.h>
+#include <stdlib.h>
 
 #include <qbt/regalloc.h>
 #include <qbt/vec.h>
 #include <qbt/abi.h>
 
+/* first try to use temporaries, then callee-save, then finally args.
+ * Idea is that callee-save cost is paid once at the start of the function call,
+ * after that they're free. Argument registers are generally important to keep
+ * free to make sure function calls in loops etc. don't have to shuffle
+ * registers around as much */
 static const int64_t tr_map[] = {
+	RT0, RT1, RT2, RT3, RT4, RT5, RT6, RT7, RT8, RT9,
+	RT10, RT11, RT12, RT13, RT14, RT15, RT16, RT17, RT18, RT19,
+	RT20, RT21, RT22, RT23,
+
 	RS0, RS1, RS2, RS3, RS4, RS5, RS6, RS7, RS8, RS9,
 	RS10, RS11, RS12, RS13, RS14, RS15, RS16, RS17, RS18, RS19,
-	RS20, RS21, RS22, RS23, RS24
+	RS20, RS21, RS22, RS23,
+
+	RA0, RA1, RA2, RA3, RA4, RA5, RA6, RA7, RA8, RA9,
+	RA10, RA11, RA12, RA13, RA14, RA15, RA16, RA17, RA18, RA19,
+	RA20, RA21, RA22, RA23,
 };
 
-#define reg_at(v, i)\
-	vect_at(int64_t, v, i)
+#define reg_at(v, i) \
+	vect_at(struct val, v, i)
 
-static struct val rewrite_tmp(struct vec rmap, struct val t)
+static bool has_rewrite_rule(struct vec *rmap, struct val t)
 {
-	assert(t.class == TMP);
-	assert(t.r < (int64_t)vec_len(&rmap));
-	int64_t r = reg_at(rmap, t.r);
-	assert(r);
-	return reg_val(r);
+	if (t.r >= (int64_t)vec_len(rmap))
+		return false;
+
+	struct val r = reg_at(*rmap, t.r);
+	return r.class != NOCLASS;
 }
 
-static size_t cur_reg = 0;
-static void add_rewrite_rule(struct vec *rmap, struct val t)
+static struct val rewrite_tmp(struct vec *rmap, struct val t)
 {
 	assert(t.class == TMP);
-	assert(cur_reg < 25
-			&& "ran out of temp registers, time to implement proper regalloc!");
+	assert(t.r < (int64_t)vec_len(rmap));
+	struct val r = reg_at(*rmap, t.r);
+	assert(r.class != NOCLASS);
+	return r;
+}
+
+static void add_rewrite_rule(struct vec *rmap, struct val from, struct val to)
+{
+	assert(from.class == TMP);
+	assert(to.class == REG);
 	/* note <=, we go one 'beyond' just to make sure that 0 fits */
-	while ((int64_t)vec_len(rmap) <= t.r) {
-		int64_t zero = 0;
-		vec_append(rmap, &zero);
+	while ((int64_t)vec_len(rmap) <= from.r) {
+		struct val no = noclass();
+		vec_append(rmap, &no);
 	}
 
-	if (reg_at(*rmap, t.r) == 0)
-		reg_at(*rmap, t.r) = tr_map[cur_reg++];
+	reg_at(*rmap, from.r) = to;
 }
 
-void regalloc(struct fn *f)
+static void add_hint(struct vec *hints, struct val from, struct val to)
 {
-	/* very ugly, fix once we get the proper regalloc implemented */
-	cur_reg = 0;
-	struct vec rmap = vec_create(sizeof(int64_t));
+	/* kind of a hack */
+	add_rewrite_rule(hints, from, to);
+}
 
-	/* here we would ideally do some kind of lifetime checking, to start
-	 * with we only assign to different temporary registers, we have enough
-	 * of them to work for some smaller test functions */
-	/* lifetime info would probably also be useful for saving callee-save
-	 * registers during calls, so that should probably also be done during
-	 * register allocation? maybe? */
-	foreach_blk(bi, f->blks) {
-		struct blk *b = blk_at(f->blks, bi);
-		foreach_insn(i, b->insns) {
-			struct insn n = insn_at(b->insns, i);
-			if (n.in[0].class == TMP) {
-				n.in[0] = rewrite_tmp(rmap, n.in[0]);
-			}
+static struct val get_hint(struct vec *hints, struct val from)
+{
+	if ((int64_t)vec_len(hints) <= from.r)
+		return noclass();
 
-			if (n.in[1].class == TMP) {
-				n.in[1] = rewrite_tmp(rmap, n.in[1]);
-			}
+	return rewrite_tmp(hints, from);
+}
 
-			if (n.out.class == TMP) {
-				add_rewrite_rule(&rmap, n.out);
-				n.out = rewrite_tmp(rmap, n.out);
-			}
+struct lifetime {
+	struct val v;
+	size_t start;
+	size_t end;
+	size_t used;
+};
+
+#define lifetime_at(lifetimes, i) \
+	vect_at(struct lifetime, lifetimes, i)
+
+#define foreach_lifetime(iter, lifetimes) \
+	foreach_vec(iter, lifetimes)
+
+static void add_def(struct vec *lifetimes, struct val v, size_t i)
+{
+	while ((int64_t)vec_len(lifetimes) <= v.r) {
+		vec_append(lifetimes, &(struct lifetime){noclass(), 0, 0, 0});
+	}
+
+	/* I guess a def isn't techincally a 'use' but good enough, we assume
+	 * that unused variables have already been removed by some earlier stage
+	 * (not currently true but that's the intention) */
+	lifetime_at(*lifetimes, v.r) = (struct lifetime){v, i, 0, 1};
+}
 
-			/* write back changes */
-			insn_at(b->insns, i) = n;
+static void add_use(struct vec *lifetimes, struct val v, size_t i)
+{
+	assert((int64_t)vec_len(lifetimes) > v.r);
+	struct lifetime l = lifetime_at(*lifetimes, v.r);
+	assert(l.used);
+	l.used++;
+	l.end = i;
+	lifetime_at(*lifetimes, v.r) = l;
+}
+
+static void collect_lifetimes(struct blk *b, struct vec *hints,
+                              struct vec *lifetimes)
+{
+	foreach_blk_param(pi, b->params) {
+		struct val v = blk_param_at(b->params, pi);
+		add_def(lifetimes, v, 0);
+	}
+
+	size_t pos = 1;
+	foreach_insn(ii, b->insns) {
+		struct insn i = insn_at(b->insns, ii);
+		if (i.in[0].class == TMP)
+			add_use(lifetimes, i.in[0], pos);
 
-			/* do this here since the register allocation is the
-			 * last stage before actually lowering to assembly, so
-			 * no possibility of dead code elimination or stuff like
-			 * that */
-			if (n.type == CALL)
-				f->has_calls = true;
+		if (i.in[1].class == TMP)
+			add_use(lifetimes, i.in[1], pos);
+
+		if (i.out.class == TMP)
+			add_def(lifetimes, i.out, pos);
+
+		/* collect some early hints */
+		if (i.type == MOVE) {
+			/* input arguments, retvals */
+			if (i.out.class == TMP && i.in[0].class == REG)
+				add_hint(hints, i.out, i.in[0]);
+
+			if (i.out.class == REG && i.in[0].class == TMP)
+				add_hint(hints, i.in[0], i.out);
 		}
 
-		if (b->cmp[0].class == TMP)
-			b->cmp[0] = rewrite_tmp(rmap, b->cmp[0]);
+		pos++;
+	}
+
+	if (b->cmp[0].class == TMP)
+		add_use(lifetimes, b->cmp[0], pos);
+
+	if (b->cmp[1].class == TMP)
+		add_use(lifetimes, b->cmp[1], pos);
+
+	foreach_blk_param(pi, b->args1) {
+		struct val v = blk_param_at(b->args1, pi);
+		add_use(lifetimes, v, pos);
+	}
+
+	foreach_blk_param(pi, b->args2) {
+		struct val v = blk_param_at(b->args2, pi);
+		add_use(lifetimes, v, pos);
+	}
+}
+
+static void build_active(struct vec *active, struct vec *lifetimes, size_t i)
+{
+	struct lifetime ref = lifetime_at(*lifetimes, i);
+	foreach_lifetime(li, *lifetimes) {
+		struct lifetime l = lifetime_at(*lifetimes, li);
+		if (l.used == 0)
+			continue;
+
+		if (l.end < ref.start)
+			continue;
+
+		if (l.start > ref.end)
+			continue;
+
+		vec_append(active, &l);
+	}
+}
 
-		if (b->cmp[1].class == TMP)
-			b->cmp[1] = rewrite_tmp(rmap, b->cmp[1]);
+static void build_reserved(struct vec *reserved, struct vec *active,
+                           struct vec *rmap)
+{
+	foreach_lifetime(li, *active) {
+		struct lifetime l = lifetime_at(*active, li);
+		if (l.used == 0)
+			continue;
+
+		if (has_rewrite_rule(rmap, l.v)) {
+			struct val act = rewrite_tmp(rmap, l.v);
+			vec_append(reserved, &act);
+		}
+	}
+}
+
+static bool reg_free(struct vec *reserved, struct val f)
+{
+	assert(f.class == REG);
+	/* not the fastest way in the world, but good enough for now */
+	foreach_val(ri, *reserved) {
+		struct val r = val_at(*reserved, ri);
+		if (same_val(r, f))
+			return false;
+	}
+
+	return true;
+}
+
+static struct val find_free_reg(struct vec *reserved)
+{
+	for (size_t i = 0; i < sizeof(tr_map) / sizeof(tr_map[0]); ++i) {
+		struct val r = reg_val(tr_map[i]);
+		if (reg_free(reserved, r))
+			return r;
+	}
+
+	/* handle spill case later */
+	assert(0 &&
+	       "ran out of registers, time to implement proper spill handling");
+	abort();
+}
+
+static void build_rmap(struct vec *hints, struct vec *lifetimes,
+                       struct vec *rmap)
+{
+	struct vec active = vec_create(sizeof(struct lifetime));
+	struct vec reserved = vec_create(sizeof(struct val));
+
+	foreach_lifetime(li, *lifetimes) {
+		struct lifetime l = lifetime_at(*lifetimes, li);
+		if (l.used == 0)
+			continue;
+
+		vec_reset(&active);
+		build_active(&active, lifetimes, li);
+		build_reserved(&reserved, &active, rmap);
+
+		struct val h = get_hint(hints, l.v);
+		if (h.class != NOCLASS) {
+			if (reg_free(&reserved, h)) {
+				add_rewrite_rule(rmap, l.v, h);
+				continue;
+			}
+		}
+
+		/* eventually we should select the spill register only for the
+		 * least used register, but that's a bit more complicated than
+		 * just this linear scan */
+		struct val f = find_free_reg(&reserved);
+		add_rewrite_rule(rmap, l.v, f);
+	}
+}
+
+/* has_calls is kind of in an iffy place but I guess it's fine */
+static void do_rewrites(struct blk *b, struct vec *rmap, struct fn *f)
+{
+	foreach_blk_param(pi, b->params) {
+		struct val p = blk_param_at(b->params, pi);
+		blk_param_at(b->params, pi) = rewrite_tmp(rmap, p);
+	}
+
+	foreach_insn(ii, b->insns) {
+		struct insn i = insn_at(b->insns, ii);
+		if (i.in[0].class == TMP)
+			i.in[0] = rewrite_tmp(rmap, i.in[0]);
+
+		if (i.in[1].class == TMP)
+			i.in[1] = rewrite_tmp(rmap, i.in[1]);
+
+		if (i.out.class == TMP)
+			i.out = rewrite_tmp(rmap, i.out);
+
+		/** @todo save caller-save at callsites here */
+		if (i.type == CALL)
+			f->has_calls = true;
+
+		/* write back changes, christ I forget this a lot */
+		insn_at(b->insns, ii) = i;
+	}
+
+	if (b->cmp[0].class == TMP)
+		b->cmp[0] = rewrite_tmp(rmap, b->cmp[0]);
+
+	if (b->cmp[1].class == TMP)
+		b->cmp[1] = rewrite_tmp(rmap, b->cmp[1]);
+
+	foreach_blk_param(pi, b->args1) {
+		struct val p = blk_param_at(b->args1, pi);
+		blk_param_at(b->args1, pi) = rewrite_tmp(rmap, p);
+	}
+
+	foreach_blk_param(pi, b->args2) {
+		struct val p = blk_param_at(b->args2, pi);
+		blk_param_at(b->args2, pi) = rewrite_tmp(rmap, p);
+	}
+}
+
+/* assumes ssa form */
+void regalloc(struct fn *f)
+{
+	struct vec hints = vec_create(sizeof(struct val));
+	struct vec lifetimes = vec_create(sizeof(struct lifetime));
+	struct vec rmap = vec_create(sizeof(struct val));
+
+	/* here it would probably make sense to iterate through lifetimes in
+	 * order of priority, like loop nesting/count etc */
+	foreach_blk(bi, f->blks) {
+		struct blk *b = blk_at(f->blks, bi);
+		vec_reset(&lifetimes);
+		/** @todo collect hints from args/params */
+		collect_lifetimes(b, &hints, &lifetimes);
+		build_rmap(&hints, &lifetimes, &rmap);
+		do_rewrites(b, &rmap, f);
+		/** @todo forward_hints(b, &hints, &rmap) */
 	}
 
-	f->max_callee_save = cur_reg;
+	vec_destroy(&hints);
+	vec_destroy(&lifetimes);
 	vec_destroy(&rmap);
 }
diff --git a/src/ssa.c b/src/ssa.c
index 59e7254..96d1eb1 100644
--- a/src/ssa.c
+++ b/src/ssa.c
@@ -78,7 +78,7 @@ static void build_params(struct blk *b, int visited)
 		struct val in2 = i.in[1];
 		if (in2.class == TMP && !has_val(&generated, in2))
 			add_val(&required, in2);
-			
+
 		struct val out = i.out;
 		if (out.class == TMP) {
 			add_val(&generated, out);
@@ -147,7 +147,7 @@ static void collect_params(struct blk *b, int visited)
 	}
 }
 
-#define tmpval_at(rmap, i)\
+#define tmpval_at(rmap, i) \
 	vect_at(struct val, rmap, i)
 
 static void add_rewrite_rule(struct vec *rmap, struct val from, struct val to)
diff --git a/src/vec.c b/src/vec.c
index 4ba92fb..e487162 100644
--- a/src/vec.c
+++ b/src/vec.c
@@ -7,13 +7,18 @@
 struct vec vec_create(size_t ns)
 {
 	return (struct vec) {
-		.n = 0,
-		.s = 1,
-		.ns = ns,
-		.buf = malloc(ns),
+		       .n = 0,
+		       .s = 1,
+		       .ns = ns,
+		       .buf = malloc(ns),
 	};
 }
 
+void vec_reset(struct vec *v)
+{
+	v->n = 0;
+}
+
 size_t vec_len(struct vec *v)
 {
 	return v->n;