nv50: use BREAKADDR and BREAK for loops

This is easier, we don't have to record all the BRK branches
to fill in the address at ENDLOOP - which wasn't done before
and thus made all but the last BRK branch jump to the start.

Also, we don't risk reactivting threads that already left
the loop if we join threads after a conditional with a BRK
(although it might be wiser to remove the join altogether
in such cases).

1 files changed, 75 insertions, 47 deletions

diff --git a/src/gallium/drivers/nv50/nv50_program.c b/src/gallium/drivers/nv50/nv50_program.c
index 0b78a7112b..9ad0677978 100644
--- a/src/gallium/drivers/nv50/nv50_program.c
+++ b/src/gallium/drivers/nv50/nv50_program.c
@@ -98,9 +98,17 @@ struct nv50_reg {
 #define NV50_MOD_ABS 2
 #define NV50_MOD_SAT 4
 
-/* arbitrary limits */
-#define MAX_IF_DEPTH 4
-#define MAX_LOOP_DEPTH 4
+/* STACK: Conditionals and loops have to use the (per warp) stack.
+ * Stack entries consist of an entry type (divergent path, join at),
+ * a mask indicating the active threads of the warp, and an address.
+ * MPs can store 12 stack entries internally, if we need more (and
+ * we probably do), we have to create a stack buffer in VRAM.
+ */
+/* impose low limits for now */
+#define NV50_MAX_COND_NESTING 4
+#define NV50_MAX_LOOP_NESTING 3
+
+#define JOIN_ON(e) e; pc->p->exec_tail->inst[1] |= 2
 
 struct nv50_pc {
 	struct nv50_program *p;
@@ -139,12 +147,11 @@ struct nv50_pc {
 	struct nv50_reg *iv_p;
 	struct nv50_reg *iv_c;
 
-	struct nv50_program_exec *if_cond;
-	struct nv50_program_exec *if_insn[MAX_IF_DEPTH];
-	struct nv50_program_exec *br_join[MAX_IF_DEPTH];
-	struct nv50_program_exec *br_loop[MAX_LOOP_DEPTH]; /* for BRK branch */
+	struct nv50_program_exec *if_insn[NV50_MAX_COND_NESTING];
+	struct nv50_program_exec *if_join[NV50_MAX_COND_NESTING];
+	struct nv50_program_exec *loop_brka[NV50_MAX_LOOP_NESTING];
 	int if_lvl, loop_lvl;
-	unsigned loop_pos[MAX_LOOP_DEPTH];
+	unsigned loop_pos[NV50_MAX_LOOP_NESTING];
 
 	/* current instruction and total number of insns */
 	unsigned insn_cur;
@@ -1184,7 +1191,6 @@ emit_set(struct nv50_pc *pc, unsigned ccode, struct nv50_reg *dst, int wp,
 	set_src_1(pc, src1, e);
 
 	emit(pc, e);
-	pc->if_cond = pc->p->exec_tail; /* record for OPCODE_IF */
 
 	/* cvt.f32.u32/s32 (?) if we didn't only write the predicate */
 	if (rdst)
@@ -1326,21 +1332,49 @@ emit_kil(struct nv50_pc *pc, struct nv50_reg *src)
 }
 
 static struct nv50_program_exec *
-emit_branch(struct nv50_pc *pc, int pred, unsigned cc,
-	    struct nv50_program_exec **join)
+emit_breakaddr(struct nv50_pc *pc)
 {
 	struct nv50_program_exec *e = exec(pc);
 
-	if (join) {
-		set_long(pc, e);
-		e->inst[0] |= 0xa0000002;
-		emit(pc, e);
-		*join = e;
-		e = exec(pc);
-	}
+	e->inst[0] = 0x40000002;
+	set_long(pc, e);
+
+	emit(pc, e);
+	return e;
+}
+
+static void
+emit_break(struct nv50_pc *pc, int pred, unsigned cc)
+{
+	struct nv50_program_exec *e = exec(pc);
+
+	e->inst[0] = 0x50000002;
+	set_long(pc, e);
+	if (pred >= 0)
+		set_pred(pc, cc, pred, e);
+
+	emit(pc, e);
+}
+
+static struct nv50_program_exec *
+emit_joinat(struct nv50_pc *pc)
+{
+	struct nv50_program_exec *e = exec(pc);
 
+	e->inst[0] = 0xa0000002;
+	set_long(pc, e);
+
+	emit(pc, e);
+	return e;
+}
+
+static struct nv50_program_exec *
+emit_branch(struct nv50_pc *pc, int pred, unsigned cc)
+{
+	struct nv50_program_exec *e = exec(pc);
+
+	e->inst[0] = 0x10000002;
 	set_long(pc, e);
-	e->inst[0] |= 0x10000002;
 	if (pred >= 0)
 		set_pred(pc, cc, pred, e);
 	emit(pc, e);
@@ -1504,20 +1538,20 @@ emit_texlod_sequence(struct nv50_pc *pc, struct nv50_reg *tlod,
 	/* Subtract lod of each pixel from lod of top left pixel, jump
 	 * texlod insn if result is 0, then repeat for 2 other pixels.
 	 */
+	join_at = emit_joinat(pc);
 	emit_quadop(pc, NULL, 0, 0, tlod, tlod, 0x55);
-	emit_branch(pc, 0, 2, &join_at)->param.index = target;
+	emit_branch(pc, 0, 2)->param.index = target;
 
 	for (i = 1; i < 4; ++i) {
 		emit_quadop(pc, NULL, 0, i, tlod, tlod, 0x55);
-		emit_branch(pc, 0, 2, NULL)->param.index = target;
+		emit_branch(pc, 0, 2)->param.index = target;
 	}
 
 	emit_mov(pc, tlod, src); /* target */
 	emit(pc, tex); /* texlod */
 
 	join_at->param.index = target + 2 * 2;
-	emit_nop(pc);
-	pc->p->exec_tail->inst[1] |= 2; /* join _after_ tex */
+	JOIN_ON(emit_nop(pc)); /* join _after_ tex */
 }
 
 static void
@@ -2058,22 +2092,19 @@ nv50_kill_branch(struct nv50_pc *pc)
 	 */
 	if (has_pred(pc->if_insn[lvl], 0xf))
 		return FALSE;
-	assert(pc->if_insn[lvl] && pc->br_join[lvl]);
+	assert(pc->if_insn[lvl] && pc->if_join[lvl]);
 
-	/* We'll use the exec allocated for JOIN_AT (as we can't easily
-	 * update prev's next); if exec_tail is BRK, update the pointer.
+	/* We'll use the exec allocated for JOIN_AT (we can't easily
+	 * access nv50_program_exec's prev).
 	 */
-	if (pc->loop_lvl && pc->br_loop[pc->loop_lvl - 1] == pc->p->exec_tail)
-		pc->br_loop[pc->loop_lvl - 1] = pc->br_join[lvl];
-
 	pc->p->exec_size -= 4; /* remove JOIN_AT and BRA */
 
-	*pc->br_join[lvl] = *pc->p->exec_tail;
+	*pc->if_join[lvl] = *pc->p->exec_tail;
 
 	FREE(pc->if_insn[lvl]);
 	FREE(pc->p->exec_tail);
 
-	pc->p->exec_tail = pc->br_join[lvl];
+	pc->p->exec_tail = pc->if_join[lvl];
 	pc->p->exec_tail->next = NULL;
 	set_pred(pc, 0xd, 0, pc->p->exec_tail);
 
@@ -2184,13 +2215,13 @@ nv50_program_tx_insn(struct nv50_pc *pc,
 		emit_arl(pc, dst[0], temp, 4);
 		break;
 	case TGSI_OPCODE_BGNLOOP:
+		pc->loop_brka[pc->loop_lvl] = emit_breakaddr(pc);
 		pc->loop_pos[pc->loop_lvl++] = pc->p->exec_size;
 		terminate_mbb(pc);
 		break;
 	case TGSI_OPCODE_BRK:
-		emit_branch(pc, -1, 0, NULL);
 		assert(pc->loop_lvl > 0);
-		pc->br_loop[pc->loop_lvl - 1] = pc->p->exec_tail;
+		emit_break(pc, -1, 0);
 		break;
 	case TGSI_OPCODE_CEIL:
 		for (c = 0; c < 4; c++) {
@@ -2266,7 +2297,7 @@ nv50_program_tx_insn(struct nv50_pc *pc,
 			emit_mov_immdval(pc, dst[0], 1.0f);
 		break;
 	case TGSI_OPCODE_ELSE:
-		emit_branch(pc, -1, 0, NULL);
+		emit_branch(pc, -1, 0);
 		pc->if_insn[--pc->if_lvl]->param.index = pc->p->exec_size;
 		pc->if_insn[pc->if_lvl++] = pc->p->exec_tail;
 		terminate_mbb(pc);
@@ -2278,21 +2309,20 @@ nv50_program_tx_insn(struct nv50_pc *pc,
 		if (nv50_kill_branch(pc) == TRUE)
 			break;
 
-		if (pc->br_join[pc->if_lvl]) {
-			pc->br_join[pc->if_lvl]->param.index = pc->p->exec_size;
-			pc->br_join[pc->if_lvl] = NULL;
+		if (pc->if_join[pc->if_lvl]) {
+			pc->if_join[pc->if_lvl]->param.index = pc->p->exec_size;
+			pc->if_join[pc->if_lvl] = NULL;
 		}
 		terminate_mbb(pc);
 		/* emit a NOP as join point, we could set it on the next
 		 * one, but would have to make sure it is long and !immd
 		 */
-		emit_nop(pc);
-		pc->p->exec_tail->inst[1] |= 2;
+		JOIN_ON(emit_nop(pc));
 		break;
 	case TGSI_OPCODE_ENDLOOP:
-		emit_branch(pc, -1, 0, NULL);
-		pc->p->exec_tail->param.index = pc->loop_pos[--pc->loop_lvl];
-		pc->br_loop[pc->loop_lvl]->param.index = pc->p->exec_size;
+		emit_branch(pc, -1, 0)->param.index =
+			pc->loop_pos[--pc->loop_lvl];
+		pc->loop_brka[pc->loop_lvl]->param.index = pc->p->exec_size;
 		terminate_mbb(pc);
 		break;
 	case TGSI_OPCODE_EX2:
@@ -2316,13 +2346,11 @@ nv50_program_tx_insn(struct nv50_pc *pc,
 		}
 		break;
 	case TGSI_OPCODE_IF:
-		/* emitting a join_at may not be necessary */
-		assert(pc->if_lvl < MAX_IF_DEPTH);
-		/* set_pred_wr(pc, 1, 0, pc->if_cond); */
+		assert(pc->if_lvl < NV50_MAX_COND_NESTING);
 		emit_cvt(pc, NULL, src[0][0], 0, CVTOP_ABS | CVTOP_RN,
 			 CVT_F32_F32);
-		emit_branch(pc, 0, 2, &pc->br_join[pc->if_lvl]);
-		pc->if_insn[pc->if_lvl++] = pc->p->exec_tail;
+		pc->if_join[pc->if_lvl] = emit_joinat(pc);
+		pc->if_insn[pc->if_lvl++] = emit_branch(pc, 0, 2);;
 		terminate_mbb(pc);
 		break;
 	case TGSI_OPCODE_KIL: