freedreno/ir3: ra debug

Some compile time RA debug

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

src/gallium/drivers/freedreno/ir3/ir3.h

@@ -27,6 +27,8 @@
 #include <stdint.h>
 #include <stdbool.h>
 
+#include "util/u_debug.h"
+
 #include "instr-a3xx.h"
 #include "disasm.h"  /* TODO move 'enum shader_t' somewhere else.. */
 
@@ -283,7 +285,7 @@ static inline bool ir3_instr_check_mark(struct ir3_instruction *instr)
 {
 	if (instr->flags & IR3_INSTR_MARK)
 		return true;  /* already visited */
-	instr->flags ^= IR3_INSTR_MARK;
+	instr->flags |= IR3_INSTR_MARK;
 	return false;
 }
 
@@ -405,7 +407,7 @@ static inline bool writes_pred(struct ir3_instruction *instr)
 
 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))
+	if (r->flags & (IR3_REG_CONST | IR3_REG_IMMED | IR3_REG_RELATIV | IR3_REG_ADDR))
 		return false;
 	if ((reg_num(r) == REG_A0) || (reg_num(r) == REG_P0))
 		return false;
@@ -455,7 +457,7 @@ typedef uint8_t regmask_t[2 * MAX_REG / 8];
 static inline unsigned regmask_idx(struct ir3_register *reg)
 {
 	unsigned num = reg->num;
-	assert(num < MAX_REG);
+	debug_assert(num < MAX_REG);
 	if (reg->flags & IR3_REG_HALF)
 		num += MAX_REG;
 	return num;

src/gallium/drivers/freedreno/ir3/ir3_dump.c

@@ -108,7 +108,7 @@ static void dump_instr_name(struct ir3_dump_ctx *ctx,
 }
 
 static void dump_reg_name(struct ir3_dump_ctx *ctx,
-		struct ir3_register *reg)
+		struct ir3_register *reg, bool followssa)
 {
 	if ((reg->flags & IR3_REG_ABS) && (reg->flags & IR3_REG_NEGATE))
 		fprintf(ctx->f, "(absneg)");
@@ -121,9 +121,12 @@ static void dump_reg_name(struct ir3_dump_ctx *ctx,
 		fprintf(ctx->f, "imm[%f,%d,0x%x]", reg->fim_val, reg->iim_val, reg->iim_val);
 	} else if (reg->flags & IR3_REG_SSA) {
 		if (ctx->verbose) {
-			fprintf(ctx->f, "_[");
-			dump_instr_name(ctx, reg->instr);
-			fprintf(ctx->f, "]");
+			fprintf(ctx->f, "_");
+			if (followssa) {
+				fprintf(ctx->f, "[");
+				dump_instr_name(ctx, reg->instr);
+				fprintf(ctx->f, "]");
+			}
 		}
 	} else {
 		if (reg->flags & IR3_REG_HALF)
@@ -282,7 +285,7 @@ static void ir3_instr_dump(struct ir3_dump_ctx *ctx,
 		if (reg->flags & IR3_REG_SSA)
 			fprintf(ctx->f, "<src%u> ", (i - 1));
 
-		dump_reg_name(ctx, reg);
+		dump_reg_name(ctx, reg, true);
 	}
 
 	fprintf(ctx->f, "}\"];\n");
@@ -406,14 +409,19 @@ ir3_dump_instr_single(struct ir3_instruction *instr)
 	for (i = 0; i < instr->regs_count; i++) {
 		struct ir3_register *reg = instr->regs[i];
 		printf(i ? ", " : " ");
-		dump_reg_name(&ctx, reg);
+		dump_reg_name(&ctx, reg, !!i);
 	}
+
+	if (is_meta(instr) && (instr->opc == OPC_META_FO))
+		printf(", off=%d", instr->fo.off);
+
 	printf("\n");
 }
 
 void
 ir3_dump_instr_list(struct ir3_instruction *instr)
 {
+	struct ir3_block *block = instr->block;
 	unsigned n = 0;
 
 	while (instr) {
@@ -423,4 +431,11 @@ ir3_dump_instr_list(struct ir3_instruction *instr)
 		instr = instr->next;
 	}
 	printf("%u instructions\n", n);
+
+	for (n = 0; n < block->noutputs; n++) {
+		if (!block->outputs[n])
+			continue;
+		printf("out%d: ", n);
+		ir3_dump_instr_single(block->outputs[n]);
+	}
 }

src/gallium/drivers/freedreno/ir3/ir3_ra.c

@@ -62,6 +62,22 @@ struct ir3_ra_ctx {
 	bool error;
 };
 
+#define ra_debug 0
+
+#define ra_dump_list(msg, n) do { \
+		if (ra_debug) { \
+			debug_printf("-- " msg); \
+			ir3_dump_instr_list(n); \
+		} \
+	} while (0)
+
+#define ra_dump_instr(msg, n) do { \
+		if (ra_debug) { \
+			debug_printf(">> " msg); \
+			ir3_dump_instr_single(n); \
+		} \
+	} while (0)
+
 /* sorta ugly way to retrofit half-precision support.. rather than
  * passing extra param around, just OR in a high bit.  All the low
  * value arithmetic (ie. +/- offset within a contiguous vec4, etc)
@@ -92,7 +108,8 @@ static struct ir3_ra_assignment ra_calc(struct ir3_instruction *instr);
 /* check that the register exists, is a GPR and is not special (a0/p0) */
 static struct ir3_register * reg_check(struct ir3_instruction *instr, unsigned n)
 {
-	if ((n < instr->regs_count) && reg_gpr(instr->regs[n]))
+	if ((n < instr->regs_count) && reg_gpr(instr->regs[n]) &&
+			!(instr->regs[n]->flags & IR3_REG_SSA))
 		return instr->regs[n];
 	return NULL;
 }
@@ -444,9 +461,6 @@ static void ra_assign_reg(struct ir3_visitor *v,
 {
 	struct ra_assign_visitor *a = ra_assign_visitor(v);
 
-	if (is_flow(instr) && (instr->opc == OPC_KILL))
-		return;
-
 	reg->flags &= ~IR3_REG_SSA;
 	reg->num = a->num & ~REG_HALF;
 
@@ -582,7 +596,8 @@ static void ir3_instr_ra(struct ir3_ra_ctx *ctx,
 		num = regid(REG_A0, 0) | REG_HALF;
 	} else {
 		/* predicate register (p0).. etc */
-		return;
+		num = regid(REG_P0, 0);
+		debug_assert(dst->num == num);
 	}
 
 	ra_assign(ctx, instr, num);
@@ -735,6 +750,8 @@ static int block_ra(struct ir3_ra_ctx *ctx, struct ir3_block *block)
 {
 	struct ir3_instruction *n;
 
+	ra_dump_list("before:\n", block->head);
+
 	if (!block->parent) {
 		unsigned i, j;
 		int base, off = output_base(ctx);
@@ -765,14 +782,16 @@ static int block_ra(struct ir3_ra_ctx *ctx, struct ir3_block *block)
 		}
 	}
 
+	ra_dump_list("after:\n", block->head);
+
 	/* then loop over instruction list and assign registers:
 	 */
-	n = block->head;
-	while (n) {
+	for (n = block->head; n; n = n->next) {
+		ra_dump_instr("ASSIGN: ", n);
 		ir3_instr_ra(ctx, n);
 		if (ctx->error)
 			return -1;
-		n = n->next;
+		ra_dump_list("-------", block->head);
 	}
 
 	legalize(ctx, block);
@@ -784,6 +803,7 @@ int ir3_block_ra(struct ir3_block *block, enum shader_t type,
 		bool half_precision, bool frag_coord, bool frag_face,
 		bool *has_samp, int *max_bary)
 {
+	struct ir3_instruction *n;
 	struct ir3_ra_ctx ctx = {
 			.block = block,
 			.type = type,
@@ -794,6 +814,13 @@ int ir3_block_ra(struct ir3_block *block, enum shader_t type,
 	};
 	int ret;
 
+	/* mark dst registers w/ SSA flag so we can see which
+	 * have been assigned so far:
+	 */
+	for (n = block->head; n; n = n->next)
+		if (n->regs_count > 0)
+			n->regs[0]->flags |= IR3_REG_SSA;
+
 	ir3_clear_mark(block->shader);
 	ret = block_ra(&ctx, block);
 	*has_samp = ctx.has_samp;