Mesa (master): freedreno/a3xx/compiler: scheduler vs pred reg

[Rob Clark]

Module: Mesa
Branch: master
Commit: faaeddb55e090ecc43d56529e99d56962a18ad9d
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=faaeddb55e090ecc43d56529e99d56962a18ad9d

Author: Rob Clark <robclark at freedesktop.org>
Date:   Fri Jul 25 09:50:34 2014 -0400

freedreno/a3xx/compiler: scheduler vs pred reg

The scheduler also needs to be aware of predicate register (p0) in
addition to address register (a0).

Signed-off-by: Rob Clark <robclark at freedesktop.org>

---

 src/gallium/drivers/freedreno/a3xx/ir3.h       |    9 +++++
 src/gallium/drivers/freedreno/a3xx/ir3_sched.c |   45 ++++++++++++++++++++++--
 2 files changed, 51 insertions(+), 3 deletions(-)

diff --git a/src/gallium/drivers/freedreno/a3xx/ir3.h b/src/gallium/drivers/freedreno/a3xx/ir3.h
index 0a8e538..9f95abd 100644
--- a/src/gallium/drivers/freedreno/a3xx/ir3.h
+++ b/src/gallium/drivers/freedreno/a3xx/ir3.h
@@ -368,6 +368,15 @@ static inline bool writes_addr(struct ir3_instruction *instr)
 	return false;
 }
 
+static inline bool writes_pred(struct ir3_instruction *instr)
+{
+	if (instr->regs_count > 0) {
+		struct ir3_register *dst = instr->regs[0];
+		return reg_num(dst) == REG_P0;
+	}
+	return false;
+}
+
 static inline bool reg_gpr(struct ir3_register *r)
 {
 	if (r->flags & (IR3_REG_CONST | IR3_REG_IMMED | IR3_REG_RELATIV | IR3_REG_SSA | IR3_REG_ADDR))
diff --git a/src/gallium/drivers/freedreno/a3xx/ir3_sched.c b/src/gallium/drivers/freedreno/a3xx/ir3_sched.c
index ef84c54..b396284 100644
--- a/src/gallium/drivers/freedreno/a3xx/ir3_sched.c
+++ b/src/gallium/drivers/freedreno/a3xx/ir3_sched.c
@@ -48,11 +48,21 @@ enum {
  * reach the end of depth sorted list without being able to insert any
  * instruction, insert nop's.  Repeat until no more unscheduled
  * instructions.
+ *
+ * There are a few special cases that need to be handled, since sched
+ * is currently independent of register allocation.  Usages of address
+ * register (a0.x) or predicate register (p0.x) must be serialized.  Ie.
+ * if you have two pairs of instructions that write the same special
+ * register and then read it, then those pairs cannot be interleaved.
+ * To solve this, when we are in such a scheduling "critical section",
+ * and we encounter a conflicting write to a special register, we try
+ * to schedule any remaining instructions that use that value first.
  */
 
 struct ir3_sched_ctx {
 	struct ir3_instruction *scheduled; /* last scheduled instr */
 	struct ir3_instruction *addr;      /* current a0.x user, if any */
+	struct ir3_instruction *pred;      /* current p0.x user, if any */
 	unsigned cnt;
 };
 
@@ -134,6 +144,11 @@ static void schedule(struct ir3_sched_ctx *ctx,
 		ctx->addr = instr;
 	}
 
+	if (writes_pred(instr)) {
+		assert(ctx->pred == NULL);
+		ctx->pred = instr;
+	}
+
 	instr->flags |= IR3_INSTR_MARK;
 
 	instr->next = ctx->scheduled;
@@ -224,13 +239,18 @@ static int trysched(struct ir3_sched_ctx *ctx,
 	if (delay)
 		return delay;
 
-	/* if this is a write to address register, and addr register
-	 * is currently in use, we need to defer until it is free:
+	/* if this is a write to address/predicate register, and that
+	 * register is currently in use, we need to defer until it is
+	 * free:
 	 */
 	if (writes_addr(instr) && ctx->addr) {
 		assert(ctx->addr != instr);
 		return DELAYED;
 	}
+	if (writes_pred(instr) && ctx->pred) {
+		assert(ctx->pred != instr);
+		return DELAYED;
+	}
 
 	schedule(ctx, instr, true);
 	return SCHEDULED;
@@ -266,6 +286,18 @@ static bool uses_current_addr(struct ir3_sched_ctx *ctx,
 	return false;
 }
 
+static bool uses_current_pred(struct ir3_sched_ctx *ctx,
+		struct ir3_instruction *instr)
+{
+	unsigned i;
+	for (i = 1; i < instr->regs_count; i++) {
+		struct ir3_register *reg = instr->regs[i];
+		if ((reg->flags & IR3_REG_SSA) && (ctx->pred == reg->instr))
+				return true;
+	}
+	return false;
+}
+
 /* when we encounter an instruction that writes to the address register
  * when it is in use, we delay that instruction and try to schedule all
  * other instructions using the current address register:
@@ -275,13 +307,15 @@ static int block_sched_undelayed(struct ir3_sched_ctx *ctx,
 {
 	struct ir3_instruction *instr = block->head;
 	bool addr_in_use = false;
+	bool pred_in_use = false;
 	unsigned cnt = ~0;
 
 	while (instr) {
 		struct ir3_instruction *next = instr->next;
 		bool addr = uses_current_addr(ctx, instr);
+		bool pred = uses_current_pred(ctx, instr);
 
-		if (addr) {
+		if (addr || pred) {
 			int ret = trysched(ctx, instr);
 			if (ret == SCHEDULED)
 				cnt = 0;
@@ -289,6 +323,8 @@ static int block_sched_undelayed(struct ir3_sched_ctx *ctx,
 				cnt = MIN2(cnt, ret);
 			if (addr)
 				addr_in_use = true;
+			if (pred)
+				pred_in_use = true;
 		}
 
 		instr = next;
@@ -297,6 +333,9 @@ static int block_sched_undelayed(struct ir3_sched_ctx *ctx,
 	if (!addr_in_use)
 		ctx->addr = NULL;
 
+	if (!pred_in_use)
+		ctx->pred = NULL;
+
 	return cnt;
 }