#include "midgard_ops.h"
#include "util/register_allocate.h"
#include "util/u_math.h"
+#include "util/u_memory.h"
/* For work registers, we can subdivide in various ways. So we create
* classes for the various sizes and conflict accordingly, keeping in
*/
#define WORK_STRIDE 10
-#define SHADOW_R27 17
+
+/* We have overlapping register classes for special registers, handled via
+ * shadows */
+
+#define SHADOW_R0 17
+#define SHADOW_R28 18
+#define SHADOW_R29 19
/* Prepacked masks/swizzles for virtual register types */
static unsigned reg_type_to_mask[WORK_STRIDE] = {
0x1, 0x1 << 1, 0x1 << 2, 0x1 << 3 /* x */
};
-static unsigned reg_type_to_swizzle[WORK_STRIDE] = {
- SWIZZLE(COMPONENT_X, COMPONENT_Y, COMPONENT_Z, COMPONENT_W),
-
- SWIZZLE(COMPONENT_X, COMPONENT_Y, COMPONENT_Z, COMPONENT_W),
- SWIZZLE(COMPONENT_Y, COMPONENT_Z, COMPONENT_W, COMPONENT_W),
-
- SWIZZLE(COMPONENT_X, COMPONENT_Y, COMPONENT_Z, COMPONENT_W),
- SWIZZLE(COMPONENT_Y, COMPONENT_Z, COMPONENT_Z, COMPONENT_W),
- SWIZZLE(COMPONENT_Z, COMPONENT_W, COMPONENT_Z, COMPONENT_W),
-
- SWIZZLE(COMPONENT_X, COMPONENT_Y, COMPONENT_Z, COMPONENT_W),
- SWIZZLE(COMPONENT_Y, COMPONENT_Y, COMPONENT_Z, COMPONENT_W),
- SWIZZLE(COMPONENT_Z, COMPONENT_Y, COMPONENT_Z, COMPONENT_W),
- SWIZZLE(COMPONENT_W, COMPONENT_Y, COMPONENT_Z, COMPONENT_W),
-};
-
struct phys_reg {
+ /* Physical register: 0-31 */
unsigned reg;
- unsigned mask;
- unsigned swizzle;
-};
-/* Given the mask/swizzle of both the register and the original source,
- * compose to find the actual mask/swizzle to give the hardware */
+ /* Byte offset into the physical register: 0-15 */
+ unsigned offset;
-static unsigned
-compose_writemask(unsigned mask, struct phys_reg reg)
-{
- /* Note: the reg mask is guaranteed to be contiguous. So we shift
- * into the X place, compose via a simple AND, and shift back */
+ /* Number of bytes in a component of this register */
+ unsigned size;
+};
- unsigned shift = __builtin_ctz(reg.mask);
- return ((reg.mask >> shift) & mask) << shift;
-}
+/* Shift each component up by reg_offset and shift all components horizontally
+ * by dst_offset. TODO: vec8+ */
-static unsigned
-compose_swizzle(unsigned swizzle, unsigned mask,
- struct phys_reg reg, struct phys_reg dst)
+static void
+offset_swizzle(unsigned *swizzle, unsigned reg_offset, unsigned srcsize, unsigned dst_offset, unsigned dstsize)
{
- unsigned out = pan_compose_swizzle(swizzle, reg.swizzle);
-
- /* Based on the register mask, we need to adjust over. E.g if we're
- * writing to yz, a base swizzle of xy__ becomes _xy_. Save the
- * original first component (x). But to prevent duplicate shifting
- * (only applies to ALU -- mask param is set to xyzw out on L/S to
- * prevent changes), we have to account for the shift inherent to the
- * original writemask */
+ unsigned out[MIR_VEC_COMPONENTS];
- unsigned rep = out & 0x3;
- unsigned shift = __builtin_ctz(dst.mask) - __builtin_ctz(mask);
- unsigned shifted = out << (2*shift);
+ signed reg_comp = reg_offset / srcsize;
+ signed dst_comp = dst_offset / dstsize;
- /* ..but we fill in the gaps so it appears to replicate */
+ assert(reg_comp * srcsize == reg_offset);
+ assert(dst_comp * dstsize == dst_offset);
- for (unsigned s = 0; s < shift; ++s)
- shifted |= rep << (2*s);
+ for (signed c = 0; c < MIR_VEC_COMPONENTS; ++c) {
+ signed comp = MAX2(c - dst_comp, 0);
+ out[c] = MIN2(swizzle[comp] + reg_comp, 4 - 1);
+ }
- return shifted;
+ memcpy(swizzle, out, sizeof(out));
}
/* Helper to return the default phys_reg for a given register */
{
struct phys_reg r = {
.reg = reg,
- .mask = 0xF, /* xyzw */
- .swizzle = 0xE4 /* xyzw */
+ .offset = 0,
+ .size = 4
};
return r;
* register corresponds to */
static struct phys_reg
-index_to_reg(compiler_context *ctx, struct ra_graph *g, int reg)
+index_to_reg(compiler_context *ctx, struct ra_graph *g, unsigned reg, midgard_reg_mode size)
{
/* Check for special cases */
- if (reg >= SSA_FIXED_MINIMUM)
+ if (reg == ~0)
+ return default_phys_reg(REGISTER_UNUSED);
+ else if (reg >= SSA_FIXED_MINIMUM)
return default_phys_reg(SSA_REG_FROM_FIXED(reg));
- else if ((reg < 0) || !g)
+ else if (!g)
return default_phys_reg(REGISTER_UNUSED);
/* Special cases aside, we pick the underlying register */
/* Apply shadow registers */
- if (phys == SHADOW_R27)
- phys = 27;
+ if (phys >= SHADOW_R28 && phys <= SHADOW_R29)
+ phys += 28 - SHADOW_R28;
+ else if (phys == SHADOW_R0)
+ phys = 0;
+
+ unsigned bytes = mir_bytes_for_mode(size);
struct phys_reg r = {
.reg = phys,
- .mask = reg_type_to_mask[type],
- .swizzle = reg_type_to_swizzle[type]
+ .offset = __builtin_ctz(reg_type_to_mask[type]) * bytes,
+ .size = bytes
};
/* Report that we actually use this register, and return it */
* work registers, although it is also used to create the register set for
* special register allocation */
+static void
+add_shadow_conflicts (struct ra_regs *regs, unsigned base, unsigned shadow, unsigned shadow_count)
+{
+ for (unsigned a = 0; a < WORK_STRIDE; ++a) {
+ unsigned reg_a = (WORK_STRIDE * base) + a;
+
+ for (unsigned b = 0; b < shadow_count; ++b) {
+ unsigned reg_b = (WORK_STRIDE * shadow) + b;
+
+ ra_add_reg_conflict(regs, reg_a, reg_b);
+ ra_add_reg_conflict(regs, reg_b, reg_a);
+ }
+ }
+}
+
static struct ra_regs *
create_register_set(unsigned work_count, unsigned *classes)
{
/* First, initialize the RA */
struct ra_regs *regs = ra_alloc_reg_set(NULL, virtual_count, true);
- for (unsigned c = 0; c < NR_REG_CLASSES; ++c) {
+ for (unsigned c = 0; c < (NR_REG_CLASSES - 1); ++c) {
int work_vec4 = ra_alloc_reg_class(regs);
int work_vec3 = ra_alloc_reg_class(regs);
int work_vec2 = ra_alloc_reg_class(regs);
/* Special register classes have other register counts */
unsigned count =
- (c == REG_CLASS_WORK) ? work_count :
- (c == REG_CLASS_LDST27) ? 1 : 2;
+ (c == REG_CLASS_WORK) ? work_count : 2;
- /* We arbitraily pick r17 (RA unused) as the shadow for r27 */
unsigned first_reg =
(c == REG_CLASS_LDST) ? 26 :
- (c == REG_CLASS_LDST27) ? SHADOW_R27 :
- (c == REG_CLASS_TEX) ? 28 : 0;
+ (c == REG_CLASS_TEXR) ? 28 :
+ (c == REG_CLASS_TEXW) ? SHADOW_R28 :
+ 0;
/* Add the full set of work registers */
for (unsigned i = first_reg; i < (first_reg + count); ++i) {
}
}
+ int fragc = ra_alloc_reg_class(regs);
- /* All of the r27 registers in in LDST conflict with all of the
- * registers in LD27 (pseudo/shadow register) */
-
- for (unsigned a = 0; a < WORK_STRIDE; ++a) {
- unsigned reg_a = (WORK_STRIDE * 27) + a;
+ classes[4*REG_CLASS_FRAGC + 0] = fragc;
+ classes[4*REG_CLASS_FRAGC + 1] = fragc;
+ classes[4*REG_CLASS_FRAGC + 2] = fragc;
+ classes[4*REG_CLASS_FRAGC + 3] = fragc;
+ ra_class_add_reg(regs, fragc, WORK_STRIDE * SHADOW_R0);
- for (unsigned b = 0; b < WORK_STRIDE; ++b) {
- unsigned reg_b = (WORK_STRIDE * SHADOW_R27) + b;
-
- ra_add_reg_conflict(regs, reg_a, reg_b);
- ra_add_reg_conflict(regs, reg_b, reg_a);
- }
- }
+ /* We have duplicate classes */
+ add_shadow_conflicts(regs, 0, SHADOW_R0, 1);
+ add_shadow_conflicts(regs, 28, SHADOW_R28, WORK_STRIDE);
+ add_shadow_conflicts(regs, 29, SHADOW_R29, WORK_STRIDE);
/* We're done setting up */
ra_set_finalize(regs, NULL);
set_class(unsigned *classes, unsigned node, unsigned class)
{
/* Check that we're even a node */
- if ((node < 0) || (node >= SSA_FIXED_MINIMUM))
+ if (node >= SSA_FIXED_MINIMUM)
return;
/* First 4 are work, next 4 are load/store.. */
if (class == current_class)
return;
-
- if ((current_class == REG_CLASS_LDST27) && (class == REG_CLASS_LDST))
- return;
-
- /* If we're changing, we must not have already assigned a special class
- */
-
- bool compat = current_class == REG_CLASS_WORK;
- compat |= (current_class == REG_CLASS_LDST) && (class == REG_CLASS_LDST27);
-
- assert(compat);
+ /* If we're changing, we haven't assigned a special class */
+ assert(current_class == REG_CLASS_WORK);
classes[node] &= 0x3;
classes[node] |= (class << 2);
static void
force_vec4(unsigned *classes, unsigned node)
{
- if ((node < 0) || (node >= SSA_FIXED_MINIMUM))
+ if (node >= SSA_FIXED_MINIMUM)
return;
/* Force vec4 = 3 */
check_read_class(unsigned *classes, unsigned tag, unsigned node)
{
/* Non-nodes are implicitly ok */
- if ((node < 0) || (node >= SSA_FIXED_MINIMUM))
+ if (node >= SSA_FIXED_MINIMUM)
return true;
unsigned current_class = classes[node] >> 2;
switch (current_class) {
case REG_CLASS_LDST:
- case REG_CLASS_LDST27:
return (tag == TAG_LOAD_STORE_4);
- default:
+ case REG_CLASS_TEXR:
+ return (tag == TAG_TEXTURE_4);
+ case REG_CLASS_TEXW:
return (tag != TAG_LOAD_STORE_4);
+ case REG_CLASS_WORK:
+ return IS_ALU(tag);
+ default:
+ unreachable("Invalid class");
+ }
+}
+
+static bool
+check_write_class(unsigned *classes, unsigned tag, unsigned node)
+{
+ /* Non-nodes are implicitly ok */
+ if (node >= SSA_FIXED_MINIMUM)
+ return true;
+
+ unsigned current_class = classes[node] >> 2;
+
+ switch (current_class) {
+ case REG_CLASS_TEXR:
+ return true;
+ case REG_CLASS_TEXW:
+ return (tag == TAG_TEXTURE_4);
+ case REG_CLASS_LDST:
+ case REG_CLASS_WORK:
+ return IS_ALU(tag) || (tag == TAG_LOAD_STORE_4);
+ default:
+ unreachable("Invalid class");
}
}
static void
mark_node_class (unsigned *bitfield, unsigned node)
{
- if ((node >= 0) && (node < SSA_FIXED_MINIMUM))
+ if (node < SSA_FIXED_MINIMUM)
BITSET_SET(bitfield, node);
}
-static midgard_instruction *
-mir_find_last_write(compiler_context *ctx, unsigned i)
-{
- midgard_instruction *last_write = NULL;
-
- mir_foreach_instr_global(ctx, ins) {
- if (ins->compact_branch) continue;
-
- if (ins->ssa_args.dest == i)
- last_write = ins;
- }
-
- return last_write;
-}
-
void
mir_lower_special_reads(compiler_context *ctx)
{
size_t sz = BITSET_WORDS(ctx->temp_count) * sizeof(BITSET_WORD);
- /* Bitfields for the various types of registers we could have */
+ /* Bitfields for the various types of registers we could have. aluw can
+ * be written by either ALU or load/store */
unsigned *alur = calloc(sz, 1);
+ unsigned *aluw = calloc(sz, 1);
+ unsigned *brar = calloc(sz, 1);
unsigned *ldst = calloc(sz, 1);
unsigned *texr = calloc(sz, 1);
unsigned *texw = calloc(sz, 1);
/* Pass #1 is analysis, a linear scan to fill out the bitfields */
mir_foreach_instr_global(ctx, ins) {
- if (ins->compact_branch) continue;
-
switch (ins->type) {
case TAG_ALU_4:
- mark_node_class(alur, ins->ssa_args.src0);
- mark_node_class(alur, ins->ssa_args.src1);
+ mark_node_class(aluw, ins->dest);
+ mark_node_class(alur, ins->src[0]);
+ mark_node_class(alur, ins->src[1]);
+ mark_node_class(alur, ins->src[2]);
+
+ if (ins->compact_branch && ins->writeout)
+ mark_node_class(brar, ins->src[0]);
+
break;
+
case TAG_LOAD_STORE_4:
- mark_node_class(ldst, ins->ssa_args.src0);
- mark_node_class(ldst, ins->ssa_args.src1);
+ mark_node_class(aluw, ins->dest);
+ mark_node_class(ldst, ins->src[0]);
+ mark_node_class(ldst, ins->src[1]);
+ mark_node_class(ldst, ins->src[2]);
break;
+
case TAG_TEXTURE_4:
- mark_node_class(texr, ins->ssa_args.src0);
- mark_node_class(texr, ins->ssa_args.src1);
- mark_node_class(texw, ins->ssa_args.dest);
+ mark_node_class(texr, ins->src[0]);
+ mark_node_class(texr, ins->src[1]);
+ mark_node_class(texr, ins->src[2]);
+ mark_node_class(texw, ins->dest);
break;
}
}
for (unsigned i = 0; i < ctx->temp_count; ++i) {
bool is_alur = BITSET_TEST(alur, i);
+ bool is_aluw = BITSET_TEST(aluw, i);
+ bool is_brar = BITSET_TEST(brar, i);
bool is_ldst = BITSET_TEST(ldst, i);
bool is_texr = BITSET_TEST(texr, i);
bool is_texw = BITSET_TEST(texw, i);
bool collision =
(is_alur && (is_ldst || is_texr)) ||
(is_ldst && (is_alur || is_texr || is_texw)) ||
- (is_texr && (is_alur || is_ldst)) ||
- (is_texw && (is_ldst));
+ (is_texr && (is_alur || is_ldst || is_texw)) ||
+ (is_texw && (is_aluw || is_ldst || is_texr)) ||
+ (is_brar && is_texw);
if (!collision)
continue;
/* Use the index as-is as the work copy. Emit copies for
* special uses */
- if (is_ldst) {
+ unsigned classes[] = { TAG_LOAD_STORE_4, TAG_TEXTURE_4, TAG_TEXTURE_4, TAG_ALU_4};
+ bool collisions[] = { is_ldst, is_texr, is_texw && is_aluw, is_brar };
+
+ for (unsigned j = 0; j < ARRAY_SIZE(collisions); ++j) {
+ if (!collisions[j]) continue;
+
+ /* When the hazard is from reading, we move and rewrite
+ * sources (typical case). When it's from writing, we
+ * flip the move and rewrite destinations (obscure,
+ * only from control flow -- impossible in SSA) */
+
+ bool hazard_write = (j == 2);
+
unsigned idx = spill_idx++;
- midgard_instruction m = v_mov(i, blank_alu_src, idx);
- midgard_instruction *use = mir_next_op(mir_find_last_write(ctx, i));
- assert(use);
- mir_insert_instruction_before(use, m);
- /* Rewrite to use */
- mir_rewrite_index_src_tag(ctx, i, idx, TAG_LOAD_STORE_4);
+ midgard_instruction m = hazard_write ?
+ v_mov(idx, i) : v_mov(i, idx);
+
+ /* Insert move before each read/write, depending on the
+ * hazard we're trying to account for */
+
+ mir_foreach_instr_global_safe(ctx, pre_use) {
+ if (pre_use->type != classes[j])
+ continue;
+
+ if (hazard_write) {
+ if (pre_use->dest != i)
+ continue;
+ } else {
+ if (!mir_has_arg(pre_use, i))
+ continue;
+ }
+
+ if (hazard_write) {
+ midgard_instruction *use = mir_next_op(pre_use);
+ assert(use);
+ mir_insert_instruction_before(ctx, use, m);
+ mir_rewrite_index_dst_single(pre_use, i, idx);
+ } else {
+ idx = spill_idx++;
+ m = v_mov(i, idx);
+ m.mask = mir_from_bytemask(mir_bytemask_of_read_components(pre_use, i), midgard_reg_mode_32);
+ mir_insert_instruction_before(ctx, pre_use, m);
+ mir_rewrite_index_src_single(pre_use, i, idx);
+ }
+ }
}
}
free(alur);
+ free(aluw);
+ free(brar);
free(ldst);
free(texr);
free(texw);
}
+static void
+mir_compute_interference(
+ compiler_context *ctx,
+ struct ra_graph *g)
+{
+ /* First, we need liveness information to be computed per block */
+ mir_compute_liveness(ctx);
+
+ /* 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_instr_in_block_rev(blk, ins) {
+ /* Mark all registers live after the instruction as
+ * interfering with the destination */
+
+ unsigned dest = ins->dest;
+
+ if (dest < ctx->temp_count) {
+ for (unsigned i = 0; i < ctx->temp_count; ++i)
+ if (live[i])
+ ra_add_node_interference(g, dest, i);
+ }
+
+ /* Update live_in */
+ mir_liveness_ins_update(live, ins, ctx->temp_count);
+ }
+
+ free(live);
+ }
+}
+
/* This routine performs the actual register allocation. It should be succeeded
* by install_registers */
unsigned *found_class = calloc(sizeof(unsigned), ctx->temp_count);
mir_foreach_instr_global(ctx, ins) {
- if (ins->compact_branch) continue;
- if (ins->ssa_args.dest < 0) continue;
- if (ins->ssa_args.dest >= SSA_FIXED_MINIMUM) continue;
+ if (ins->dest >= SSA_FIXED_MINIMUM) continue;
/* 0 for x, 1 for xy, 2 for xyz, 3 for xyzw */
int class = util_logbase2(ins->mask);
/* Use the largest class if there's ambiguity, this
* handles partial writes */
- int dest = ins->ssa_args.dest;
+ int dest = ins->dest;
found_class[dest] = MAX2(found_class[dest], class);
}
* nodes (TODO) */
mir_foreach_instr_global(ctx, ins) {
- if (ins->compact_branch) continue;
-
/* Check if this operation imposes any classes */
if (ins->type == TAG_LOAD_STORE_4) {
- bool force_r27 = OP_IS_R27_ONLY(ins->load_store.op);
- unsigned class = force_r27 ? REG_CLASS_LDST27 : REG_CLASS_LDST;
+ bool force_vec4_only = OP_IS_VEC4_ONLY(ins->load_store.op);
- set_class(found_class, ins->ssa_args.src0, class);
- set_class(found_class, ins->ssa_args.src1, class);
+ set_class(found_class, ins->src[0], REG_CLASS_LDST);
+ set_class(found_class, ins->src[1], REG_CLASS_LDST);
+ set_class(found_class, ins->src[2], REG_CLASS_LDST);
- if (force_r27) {
- force_vec4(found_class, ins->ssa_args.dest);
- force_vec4(found_class, ins->ssa_args.src0);
- force_vec4(found_class, ins->ssa_args.src1);
+ if (force_vec4_only) {
+ force_vec4(found_class, ins->dest);
+ force_vec4(found_class, ins->src[0]);
+ force_vec4(found_class, ins->src[1]);
+ force_vec4(found_class, ins->src[2]);
}
+ } else if (ins->type == TAG_TEXTURE_4) {
+ set_class(found_class, ins->dest, REG_CLASS_TEXW);
+ set_class(found_class, ins->src[0], REG_CLASS_TEXR);
+ set_class(found_class, ins->src[1], REG_CLASS_TEXR);
+ set_class(found_class, ins->src[2], REG_CLASS_TEXR);
}
}
/* Check that the semantics of the class are respected */
mir_foreach_instr_global(ctx, ins) {
- if (ins->compact_branch) continue;
+ assert(check_write_class(found_class, ins->type, ins->dest));
+ assert(check_read_class(found_class, ins->type, ins->src[0]));
+ assert(check_read_class(found_class, ins->type, ins->src[1]));
+ assert(check_read_class(found_class, ins->type, ins->src[2]));
+ }
- /* Non-load-store cannot read load/store */
- assert(check_read_class(found_class, ins->type, ins->ssa_args.src0));
- assert(check_read_class(found_class, ins->type, ins->ssa_args.src1));
+ /* Mark writeout to r0 */
+ mir_foreach_instr_global(ctx, ins) {
+ if (ins->compact_branch && ins->writeout)
+ set_class(found_class, ins->src[0], REG_CLASS_FRAGC);
}
for (unsigned i = 0; i < ctx->temp_count; ++i) {
ra_set_node_class(g, i, classes[class]);
}
- /* Determine liveness */
-
- int *live_start = malloc(nodes * sizeof(int));
- int *live_end = malloc(nodes * sizeof(int));
-
- /* Initialize as non-existent */
-
- for (int i = 0; i < nodes; ++i) {
- live_start[i] = live_end[i] = -1;
- }
-
- int d = 0;
-
- mir_foreach_block(ctx, block) {
- mir_foreach_instr_in_block(block, ins) {
- if (ins->compact_branch) continue;
-
- if (ins->ssa_args.dest < SSA_FIXED_MINIMUM) {
- /* If this destination is not yet live, it is
- * now since we just wrote it */
-
- int dest = ins->ssa_args.dest;
-
- if (dest >= 0 && live_start[dest] == -1)
- live_start[dest] = d;
- }
-
- /* Since we just used a source, the source might be
- * dead now. Scan the rest of the block for
- * invocations, and if there are none, the source dies
- * */
-
- int sources[2] = {
- ins->ssa_args.src0, ins->ssa_args.src1
- };
-
- for (int src = 0; src < 2; ++src) {
- int s = sources[src];
-
- if (ins->ssa_args.inline_constant && src == 1)
- continue;
-
- if (s < 0) continue;
-
- if (s >= SSA_FIXED_MINIMUM) continue;
-
- if (!mir_is_live_after(ctx, block, ins, s)) {
- live_end[s] = d;
- }
- }
-
- ++d;
- }
- }
-
- /* If a node still hasn't been killed, kill it now */
-
- for (int i = 0; i < nodes; ++i) {
- /* live_start == -1 most likely indicates a pinned output */
-
- if (live_end[i] == -1)
- live_end[i] = d;
- }
-
- /* Setup interference between nodes that are live at the same time */
-
- for (int i = 0; i < nodes; ++i) {
- for (int j = i + 1; j < nodes; ++j) {
- bool j_overlaps_i = live_start[j] < live_end[i];
- bool i_overlaps_j = live_end[j] < live_start[i];
-
- if (i_overlaps_j || j_overlaps_i)
- ra_add_node_interference(g, i, j);
- }
- }
-
- /* Cleanup */
- free(live_start);
- free(live_end);
+ mir_compute_interference(ctx, g);
if (!ra_allocate(g)) {
*spilled = true;
struct ra_graph *g,
midgard_instruction *ins)
{
- ssa_args args = ins->ssa_args;
-
switch (ins->type) {
- case TAG_ALU_4: {
- int adjusted_src = args.inline_constant ? -1 : args.src1;
- struct phys_reg src1 = index_to_reg(ctx, g, args.src0);
- struct phys_reg src2 = index_to_reg(ctx, g, adjusted_src);
- struct phys_reg dest = index_to_reg(ctx, g, args.dest);
+ case TAG_ALU_4:
+ case TAG_ALU_8:
+ case TAG_ALU_12:
+ case TAG_ALU_16: {
+ if (ins->compact_branch)
+ return;
+
+ struct phys_reg src1 = index_to_reg(ctx, g, ins->src[0], mir_srcsize(ins, 0));
+ struct phys_reg src2 = index_to_reg(ctx, g, ins->src[1], mir_srcsize(ins, 1));
+ struct phys_reg dest = index_to_reg(ctx, g, ins->dest, mir_typesize(ins));
- unsigned uncomposed_mask = ins->mask;
- ins->mask = compose_writemask(uncomposed_mask, dest);
+ mir_set_bytemask(ins, mir_bytemask(ins) << dest.offset);
- /* Adjust the dest mask if necessary. Mostly this is a no-op
- * but it matters for dot products */
- dest.mask = effective_writemask(&ins->alu, ins->mask);
+ unsigned dest_offset =
+ GET_CHANNEL_COUNT(alu_opcode_props[ins->alu.op].props) ? 0 :
+ dest.offset;
- midgard_vector_alu_src mod1 =
- vector_alu_from_unsigned(ins->alu.src1);
- mod1.swizzle = compose_swizzle(mod1.swizzle, uncomposed_mask, src1, dest);
- ins->alu.src1 = vector_alu_srco_unsigned(mod1);
+ offset_swizzle(ins->swizzle[0], src1.offset, src1.size, dest_offset, dest.size);
ins->registers.src1_reg = src1.reg;
- ins->registers.src2_imm = args.inline_constant;
+ ins->registers.src2_imm = ins->has_inline_constant;
- if (args.inline_constant) {
+ if (ins->has_inline_constant) {
/* Encode inline 16-bit constant. See disassembler for
* where the algorithm is from */
} else {
midgard_vector_alu_src mod2 =
vector_alu_from_unsigned(ins->alu.src2);
- mod2.swizzle = compose_swizzle(
- mod2.swizzle, uncomposed_mask, src2, dest);
+ offset_swizzle(ins->swizzle[1], src2.offset, src2.size, dest_offset, dest.size);
ins->alu.src2 = vector_alu_srco_unsigned(mod2);
ins->registers.src2_reg = src2.reg;
}
case TAG_LOAD_STORE_4: {
- bool fixed = args.src0 >= SSA_FIXED_MINIMUM;
+ /* Which physical register we read off depends on
+ * whether we are loading or storing -- think about the
+ * logical dataflow */
- if (OP_IS_STORE_R26(ins->load_store.op) && fixed) {
- ins->load_store.reg = SSA_REG_FROM_FIXED(args.src0);
- } else if (OP_IS_STORE_VARY(ins->load_store.op)) {
- struct phys_reg src = index_to_reg(ctx, g, args.src0);
+ bool encodes_src = OP_IS_STORE(ins->load_store.op);
+
+ if (encodes_src) {
+ struct phys_reg src = index_to_reg(ctx, g, ins->src[0], mir_srcsize(ins, 0));
assert(src.reg == 26 || src.reg == 27);
ins->load_store.reg = src.reg - 26;
+ offset_swizzle(ins->swizzle[0], src.offset, src.size, 0, 4);
+ } else {
+ struct phys_reg dst = index_to_reg(ctx, g, ins->dest, mir_typesize(ins));
- /* TODO: swizzle/mask */
- } else {
- /* Which physical register we read off depends on
- * whether we are loading or storing -- think about the
- * logical dataflow */
+ ins->load_store.reg = dst.reg;
+ offset_swizzle(ins->swizzle[0], 0, 4, dst.offset, dst.size);
+ mir_set_bytemask(ins, mir_bytemask(ins) << dst.offset);
+ }
+
+ /* We also follow up by actual arguments */
- unsigned r = OP_IS_STORE(ins->load_store.op) ?
- args.src0 : args.dest;
- struct phys_reg src = index_to_reg(ctx, g, r);
+ unsigned src2 = ins->src[1];
+ unsigned src3 = ins->src[2];
- ins->load_store.reg = src.reg;
+ if (src2 != ~0) {
+ struct phys_reg src = index_to_reg(ctx, g, src2, mir_srcsize(ins, 1));
+ unsigned component = src.offset / src.size;
+ assert(component * src.size == src.offset);
+ ins->load_store.arg_1 |= midgard_ldst_reg(src.reg, component);
+ }
- ins->load_store.swizzle = compose_swizzle(
- ins->load_store.swizzle, 0xF,
- default_phys_reg(0), src);
+ if (src3 != ~0) {
+ struct phys_reg src = index_to_reg(ctx, g, src3, mir_srcsize(ins, 2));
+ unsigned component = src.offset / src.size;
+ assert(component * src.size == src.offset);
+ ins->load_store.arg_2 |= midgard_ldst_reg(src.reg, component);
+ }
+
+ break;
+ }
- ins->mask = compose_writemask(
- ins->mask, src);
+ case TAG_TEXTURE_4: {
+ /* Grab RA results */
+ struct phys_reg dest = index_to_reg(ctx, g, ins->dest, mir_typesize(ins));
+ struct phys_reg coord = index_to_reg(ctx, g, ins->src[1], mir_srcsize(ins, 1));
+ struct phys_reg lod = index_to_reg(ctx, g, ins->src[2], mir_srcsize(ins, 2));
+
+ assert(dest.reg == 28 || dest.reg == 29);
+ assert(coord.reg == 28 || coord.reg == 29);
+
+ /* First, install the texture coordinate */
+ ins->texture.in_reg_full = 1;
+ ins->texture.in_reg_upper = 0;
+ ins->texture.in_reg_select = coord.reg - 28;
+ offset_swizzle(ins->swizzle[1], coord.offset, coord.size, 0, 4);
+
+ /* Next, install the destination */
+ ins->texture.out_full = 1;
+ ins->texture.out_upper = 0;
+ ins->texture.out_reg_select = dest.reg - 28;
+ offset_swizzle(ins->swizzle[0], 0, 4, dest.offset, dest.size);
+ mir_set_bytemask(ins, mir_bytemask(ins) << dest.offset);
+
+ /* If there is a register LOD/bias, use it */
+ if (ins->src[2] != ~0) {
+ assert(!(lod.offset & 3));
+ midgard_tex_register_select sel = {
+ .select = lod.reg,
+ .full = 1,
+ .component = lod.offset / 4
+ };
+
+ uint8_t packed;
+ memcpy(&packed, &sel, sizeof(packed));
+ ins->texture.bias = packed;
}
break;
void
install_registers(compiler_context *ctx, struct ra_graph *g)
{
- mir_foreach_block(ctx, block) {
- mir_foreach_instr_in_block(block, ins) {
- if (ins->compact_branch) continue;
- install_registers_instr(ctx, g, ins);
- }
- }
-
+ mir_foreach_instr_global(ctx, ins)
+ install_registers_instr(ctx, g, ins);
}
projects / mesa.git / blobdiff