src/gallium/drivers/freedreno/a2xx/ir2_ra.c

/*
 * Copyright (C) 2018 Jonathan Marek <jonathan@marek.ca>
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 *
 * Authors:
 *    Jonathan Marek <jonathan@marek.ca>
 */

#include "ir2_private.h"

/* if an instruction has side effects, we should never kill it */
static bool has_side_effects(struct ir2_instr *instr)
{
	if (instr->type == IR2_CF)
		return true;
	else if (instr->type == IR2_FETCH)
		return false;

	switch (instr->alu.scalar_opc) {
	case PRED_SETEs ... KILLONEs:
		return true;
	default:
		break;
	}

	switch (instr->alu.vector_opc) {
	case PRED_SETE_PUSHv ... KILLNEv:
		return true;
	default:
		break;
	}

	return instr->alu.export >= 0;
}

/* mark an instruction as required, and all its sources recursively */
static void set_need_emit(struct ir2_context *ctx, struct ir2_instr *instr)
{
	struct ir2_reg *reg;

	/* don't repeat work already done */
	if (instr->need_emit)
		return;

	instr->need_emit = true;

	ir2_foreach_src(src, instr) {
		switch (src->type) {
		case IR2_SRC_SSA:
			set_need_emit(ctx, &ctx->instr[src->num]);
			break;
		case IR2_SRC_REG:
			/* slow ..  */
			reg = get_reg_src(ctx, src);
			ir2_foreach_instr(instr, ctx) {
				if (!instr->is_ssa && instr->reg == reg)
					set_need_emit(ctx, instr);
			}
		default:
			break;
		}
	}
}

/* get current bit mask of allocated components for a register */
static unsigned reg_mask(struct ir2_context *ctx, unsigned idx)
{
	return ctx->reg_state[idx/8] >> idx%8*4 & 0xf;
}

static void reg_setmask(struct ir2_context *ctx, unsigned idx, unsigned c)
{
	idx = idx * 4 + c;
	ctx->reg_state[idx/32] |= 1 << idx%32;
}

static void reg_freemask(struct ir2_context *ctx, unsigned idx, unsigned c)
{
	idx = idx * 4 + c;
	ctx->reg_state[idx/32] &= ~(1 << idx%32);
}

void ra_count_refs(struct ir2_context *ctx)
{
	struct ir2_reg *reg;

	/* mark instructions as needed
	 * need to do this because "substitutions" pass makes many movs not needed
	 */
	ir2_foreach_instr(instr, ctx) {
		if (has_side_effects(instr))
			set_need_emit(ctx, instr);
	}

	/* compute ref_counts */
	ir2_foreach_instr(instr, ctx) {
		/* kill non-needed so they can be skipped */
		if (!instr->need_emit) {
			instr->type = IR2_NONE;
			continue;
		}

		ir2_foreach_src(src, instr) {
			if (src->type == IR2_SRC_CONST)
				continue;

			reg = get_reg_src(ctx, src);
			for (int i = 0; i < src_ncomp(instr); i++)
				reg->comp[swiz_get(src->swizzle, i)].ref_count++;
		}
	}
}

void ra_reg(struct ir2_context *ctx, struct ir2_reg *reg, int force_idx,
	bool export, uint8_t export_writemask)
{
	/* for export, don't allocate anything but set component layout */
	if (export) {
		for (int i = 0; i < 4; i++)
			reg->comp[i].c = i;
		return;
	}

	unsigned idx = force_idx;

	/* TODO: allocate into the same register if theres room
	 * note: the blob doesn't do it, so verify that it is indeed better
	 * also, doing it would conflict with scalar mov insertion
	 */

	/* check if already allocated */
	for (int i = 0; i < reg->ncomp; i++) {
		if (reg->comp[i].alloc)
			return;
	}

	if (force_idx < 0) {
		for (idx = 0; idx < 64; idx++) {
			if (reg_mask(ctx, idx) == 0)
				break;
		}
	}
	assert(idx != 64); /* TODO ran out of register space.. */

	/* update max_reg value */
	ctx->info->max_reg = MAX2(ctx->info->max_reg, (int) idx);

	unsigned mask = reg_mask(ctx, idx);

	for (int i = 0; i < reg->ncomp; i++) {
		/* don't allocate never used values */
		if (reg->comp[i].ref_count == 0) {
			reg->comp[i].c = 7;
			continue;
		}

		/* TODO */
		unsigned c = 1 ? i : (ffs(~mask) - 1);
		mask |= 1 << c;
		reg->comp[i].c = c;
		reg_setmask(ctx, idx, c);
		reg->comp[i].alloc = true;
	}

	reg->idx = idx;
	ctx->live_regs[reg->idx] = reg;
}

/* reduce srcs ref_count and free if needed */
void ra_src_free(struct ir2_context *ctx, struct ir2_instr *instr)
{
	struct ir2_reg *reg;
	struct ir2_reg_component *comp;

	ir2_foreach_src(src, instr) {
		if (src->type == IR2_SRC_CONST)
			continue;

		reg = get_reg_src(ctx, src);
		/* XXX use before write case */

		for (int i = 0; i < src_ncomp(instr); i++) {
			comp = &reg->comp[swiz_get(src->swizzle, i)];
			if (!--comp->ref_count && reg->block_idx_free < 0) {
				reg_freemask(ctx, reg->idx, comp->c);
				comp->alloc = false;
			}
		}
	}
}

/* free any regs left for a block */
void ra_block_free(struct ir2_context *ctx, unsigned block)
{
	ir2_foreach_live_reg(reg, ctx) {
		if (reg->block_idx_free != block)
			continue;

		for (int i = 0; i < reg->ncomp; i++) {
			if (!reg->comp[i].alloc) /* XXX should never be true? */
				continue;

			reg_freemask(ctx, reg->idx, reg->comp[i].c);
			reg->comp[i].alloc = false;
		}
		ctx->live_regs[reg->idx] = NULL;
	}
}