#include "midgard_ops.h"
#include "util/u_math.h"
#include "util/u_memory.h"
-#include "lcra.h"
#include "midgard_quirks.h"
struct phys_reg {
/* First, we need liveness information to be computed per block */
mir_compute_liveness(ctx);
+ /* We need to force r1.w live throughout a blend shader */
+
+ if (ctx->is_blend) {
+ unsigned r1w = ~0;
+
+ mir_foreach_block(ctx, _block) {
+ midgard_block *block = (midgard_block *) _block;
+ mir_foreach_instr_in_block_rev(block, ins) {
+ if (ins->writeout)
+ r1w = ins->src[2];
+ }
+
+ if (r1w != ~0)
+ break;
+ }
+
+ mir_foreach_instr_global(ctx, ins) {
+ if (ins->dest < ctx->temp_count)
+ lcra_add_node_interference(l, ins->dest, mir_bytemask(ins), r1w, 0xF);
+ }
+ }
+
/* Now that every block has live_in/live_out computed, we can determine
* interference by walking each block linearly. Take live_out at the
* end of each block and walk the block backwards. */
- mir_foreach_block(ctx, blk) {
- uint16_t *live = mem_dup(blk->live_out, ctx->temp_count * sizeof(uint16_t));
+ mir_foreach_block(ctx, _blk) {
+ midgard_block *blk = (midgard_block *) _blk;
+ uint16_t *live = mem_dup(_blk->live_out, ctx->temp_count * sizeof(uint16_t));
mir_foreach_instr_in_block_rev(blk, ins) {
/* Mark all registers live after the instruction as
unsigned *min_alignment = calloc(sizeof(unsigned), ctx->temp_count);
mir_foreach_instr_global(ctx, ins) {
+ /* Swizzles of 32-bit sources on 64-bit instructions need to be
+ * aligned to either bottom (xy) or top (zw). More general
+ * swizzle lowering should happen prior to scheduling (TODO),
+ * but once we get RA we shouldn't disrupt this further. Align
+ * sources of 64-bit instructions. */
+
+ if (ins->type == TAG_ALU_4 && ins->alu.reg_mode == midgard_reg_mode_64) {
+ mir_foreach_src(ins, v) {
+ unsigned s = ins->src[v];
+
+ if (s < ctx->temp_count)
+ min_alignment[s] = 3;
+ }
+ }
+
+ if (ins->type == TAG_LOAD_STORE_4 && OP_HAS_ADDRESS(ins->load_store.op)) {
+ mir_foreach_src(ins, v) {
+ unsigned s = ins->src[v];
+ unsigned size = mir_srcsize(ins, v);
+
+ if (s < ctx->temp_count)
+ min_alignment[s] = (size == midgard_reg_mode_64) ? 3 : 2;
+ }
+ }
+
if (ins->dest >= SSA_FIXED_MINIMUM) continue;
/* 0 for x, 1 for xy, 2 for xyz, 3 for xyzw */
set_class(l->class, ins->src[1], REG_CLASS_LDST);
set_class(l->class, ins->src[2], REG_CLASS_LDST);
- if (OP_IS_VEC4_ONLY(ins->load_store.op))
+ if (OP_IS_VEC4_ONLY(ins->load_store.op)) {
lcra_restrict_range(l, ins->dest, 16);
+ lcra_restrict_range(l, ins->src[0], 16);
+ lcra_restrict_range(l, ins->src[1], 16);
+ lcra_restrict_range(l, ins->src[2], 16);
+ }
} else if (ins->type == TAG_TEXTURE_4) {
set_class(l->class, ins->dest, REG_CLASS_TEXW);
set_class(l->class, ins->src[0], REG_CLASS_TEXR);
set_class(l->class, ins->src[1], REG_CLASS_TEXR);
set_class(l->class, ins->src[2], REG_CLASS_TEXR);
+ set_class(l->class, ins->src[3], REG_CLASS_TEXR);
}
}
mir_foreach_instr_global(ctx, ins) {
if (!(ins->compact_branch && ins->writeout)) continue;
- if (ins->src[0] < ctx->temp_count)
- l->solutions[ins->src[0]] = 0;
+ if (ins->src[0] < ctx->temp_count) {
+ if (ins->writeout_depth)
+ l->solutions[ins->src[0]] = (16 * 1) + COMPONENT_X * 4;
+ else if (ins->writeout_stencil)
+ l->solutions[ins->src[0]] = (16 * 1) + COMPONENT_Y * 4;
+ else
+ l->solutions[ins->src[0]] = 0;
+ }
if (ins->src[1] < ctx->temp_count)
l->solutions[ins->src[1]] = (16 * 1) + COMPONENT_Z * 4;
return l;
}
+
/* Once registers have been decided via register allocation
* (allocate_registers), we need to rewrite the MIR to use registers instead of
* indices */
unsigned src2 = ins->src[1];
unsigned src3 = ins->src[2];
+ midgard_reg_mode m32 = midgard_reg_mode_32;
if (src2 != ~0) {
- struct phys_reg src = index_to_reg(ctx, l, src2, mir_srcsize(ins, 1));
+ struct phys_reg src = index_to_reg(ctx, l, src2, m32);
unsigned component = src.offset / src.size;
assert(component * src.size == src.offset);
ins->load_store.arg_1 |= midgard_ldst_reg(src.reg, component);
}
if (src3 != ~0) {
- struct phys_reg src = index_to_reg(ctx, l, src3, mir_srcsize(ins, 2));
+ struct phys_reg src = index_to_reg(ctx, l, src3, m32);
unsigned component = src.offset / src.size;
assert(component * src.size == src.offset);
ins->load_store.arg_2 |= midgard_ldst_reg(src.reg, component);
}
case TAG_TEXTURE_4: {
+ if (ins->texture.op == TEXTURE_OP_BARRIER)
+ break;
+
/* Grab RA results */
struct phys_reg dest = index_to_reg(ctx, l, ins->dest, mir_typesize(ins));
struct phys_reg coord = index_to_reg(ctx, l, ins->src[1], mir_srcsize(ins, 1));
struct phys_reg lod = index_to_reg(ctx, l, ins->src[2], mir_srcsize(ins, 2));
+ struct phys_reg offset = index_to_reg(ctx, l, ins->src[3], mir_srcsize(ins, 2));
/* First, install the texture coordinate */
ins->texture.in_reg_full = 1;
if (ins->src[2] != ~0) {
assert(!(lod.offset & 3));
midgard_tex_register_select sel = {
- .select = lod.reg,
+ .select = lod.reg & 1,
.full = 1,
.component = lod.offset / 4
};
ins->texture.bias = packed;
}
+ /* If there is an offset register, install it */
+ if (ins->src[3] != ~0) {
+ unsigned x = offset.offset / 4;
+ unsigned y = x + 1;
+ unsigned z = x + 2;
+
+ /* Check range, TODO: half-registers */
+ assert(z < 4);
+
+ ins->texture.offset =
+ (1) | /* full */
+ (offset.reg & 1) << 1 | /* select */
+ (0 << 2) | /* upper */
+ (x << 3) | /* swizzle */
+ (y << 5) | /* swizzle */
+ (z << 7); /* swizzle */
+ }
+
break;
}
if (is_special_w)
spill_slot = spill_index++;
- mir_foreach_block(ctx, block) {
+ mir_foreach_block(ctx, _block) {
+ midgard_block *block = (midgard_block *) _block;
mir_foreach_instr_in_block_safe(block, ins) {
if (ins->dest != spill_node) continue;
* work registers to back special registers; TLS
* spilling is to use memory to back work registers) */
- mir_foreach_block(ctx, block) {
+ mir_foreach_block(ctx, _block) {
+ midgard_block *block = (midgard_block *) _block;
mir_foreach_instr_in_block(block, ins) {
/* We can't rewrite the moves used to spill in the
* first place. These moves are hinted. */