#include <stdbool.h>
#include <errno.h>
+#include "util/ralloc.h"
+
#include "freedreno_util.h"
#include "instr-a3xx.h"
-#define CHUNK_SZ 1020
-
-struct ir3_heap_chunk {
- struct ir3_heap_chunk *next;
- uint32_t heap[CHUNK_SZ];
-};
-
-static void grow_heap(struct ir3 *shader)
-{
- struct ir3_heap_chunk *chunk = calloc(1, sizeof(*chunk));
- chunk->next = shader->chunk;
- shader->chunk = chunk;
- shader->heap_idx = 0;
-}
-
/* simple allocator to carve allocations out of an up-front allocated heap,
* so that we can free everything easily in one shot.
*/
void * ir3_alloc(struct ir3 *shader, int sz)
{
- void *ptr;
-
- sz = align(sz, 4) / 4;
-
- if ((shader->heap_idx + sz) > CHUNK_SZ)
- grow_heap(shader);
-
- ptr = &shader->chunk->heap[shader->heap_idx];
- shader->heap_idx += sz;
-
- return ptr;
+ return rzalloc_size(shader, sz); /* TODO: don't use rzalloc */
}
struct ir3 * ir3_create(struct ir3_compiler *compiler,
unsigned nin, unsigned nout)
{
- struct ir3 *shader = calloc(1, sizeof(struct ir3));
-
- grow_heap(shader);
+ struct ir3 *shader = rzalloc(compiler, struct ir3);
shader->compiler = compiler;
shader->ninputs = nin;
void ir3_destroy(struct ir3 *shader)
{
- while (shader->chunk) {
- struct ir3_heap_chunk *chunk = shader->chunk;
- shader->chunk = chunk->next;
- free(chunk);
- }
- free(shader->indirects);
- free(shader->predicates);
- free(shader->baryfs);
- free(shader);
+ ralloc_free(shader);
}
#define iassert(cond) do { \
info->max_const = MAX2(info->max_const, max);
} else if (val.num == 63) {
/* ignore writes to dummy register r63.x */
- } else if ((max != REG_A0) && (max != REG_P0)) {
+ } else if (max < 48) {
if (reg->flags & IR3_REG_HALF) {
info->max_half_reg = MAX2(info->max_half_reg, max);
} else {
{
instr_cat0_t *cat0 = ptr;
- if (info->gpu_id >= 400) {
+ if (info->gpu_id >= 500) {
+ cat0->a5xx.immed = instr->cat0.immed;
+ } else if (info->gpu_id >= 400) {
cat0->a4xx.immed = instr->cat0.immed;
} else {
cat0->a3xx.immed = instr->cat0.immed;
cat2->dst = reg(dst, info, instr->repeat,
IR3_REG_R | IR3_REG_EI | IR3_REG_HALF);
cat2->repeat = instr->repeat;
+ cat2->sat = !!(instr->flags & IR3_INSTR_SAT);
cat2->ss = !!(instr->flags & IR3_INSTR_SS);
cat2->ul = !!(instr->flags & IR3_INSTR_UL);
cat2->dst_half = !!((src1->flags ^ dst->flags) & IR3_REG_HALF);
cat3->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
cat3->repeat = instr->repeat;
+ cat3->sat = !!(instr->flags & IR3_INSTR_SAT);
cat3->ss = !!(instr->flags & IR3_INSTR_SS);
cat3->ul = !!(instr->flags & IR3_INSTR_UL);
cat3->dst_half = !!((src_flags ^ dst->flags) & IR3_REG_HALF);
cat4->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
cat4->repeat = instr->repeat;
+ cat4->sat = !!(instr->flags & IR3_INSTR_SAT);
cat4->ss = !!(instr->flags & IR3_INSTR_SS);
cat4->ul = !!(instr->flags & IR3_INSTR_UL);
cat4->dst_half = !!((src->flags ^ dst->flags) & IR3_REG_HALF);
iassert(!((dst->flags ^ type_flags(instr->cat5.type)) & IR3_REG_HALF));
+ assume(src1 || !src2);
+ assume(src2 || !src3);
+
if (src1) {
cat5->full = ! (src1->flags & IR3_REG_HALF);
cat5->src1 = reg(src1, info, instr->repeat, IR3_REG_HALF);
}
-
if (instr->flags & IR3_INSTR_S2EN) {
if (src2) {
iassert(!((src1->flags ^ src2->flags) & IR3_REG_HALF));
struct ir3_register *dst, *src1, *src2;
instr_cat6_t *cat6 = ptr;
+ cat6->type = instr->cat6.type;
+ cat6->opc = instr->opc;
+ cat6->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
+ cat6->sync = !!(instr->flags & IR3_INSTR_SY);
+ cat6->g = !!(instr->flags & IR3_INSTR_G);
+ cat6->opc_cat = 6;
+
/* the "dst" for a store instruction is (from the perspective
* of data flow in the shader, ie. register use/def, etc) in
* fact a register that is read by the instruction, rather
src2 = (instr->regs_count >= 3) ? instr->regs[2] : NULL;
}
-
/* TODO we need a more comprehensive list about which instructions
* can be encoded which way. Or possibly use IR3_INSTR_0 flag to
* indicate to use the src_off encoding even if offset is zero
* (but then what to do about dst_off?)
*/
- if (instr->cat6.src_offset || (instr->opc == OPC_LDG)) {
+ if (is_atomic(instr->opc)) {
+ instr_cat6ldgb_t *ldgb = ptr;
+
+ /* maybe these two bits both determine the instruction encoding? */
+ cat6->src_off = false;
+
+ ldgb->d = instr->cat6.d - 1;
+ ldgb->typed = instr->cat6.typed;
+ ldgb->type_size = instr->cat6.iim_val - 1;
+
+ ldgb->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
+
+ if (ldgb->g) {
+ struct ir3_register *src3 = instr->regs[3];
+ struct ir3_register *src4 = instr->regs[4];
+
+ /* first src is src_ssbo: */
+ iassert(src1->flags & IR3_REG_IMMED);
+ ldgb->src_ssbo = src1->uim_val;
+
+ ldgb->src1 = reg(src2, info, instr->repeat, IR3_REG_IMMED);
+ ldgb->src1_im = !!(src2->flags & IR3_REG_IMMED);
+ ldgb->src2 = reg(src3, info, instr->repeat, IR3_REG_IMMED);
+ ldgb->src2_im = !!(src3->flags & IR3_REG_IMMED);
+
+ ldgb->src3 = reg(src4, info, instr->repeat, 0);
+ ldgb->pad0 = 0x1;
+ ldgb->pad3 = 0x1;
+ } else {
+ ldgb->src1 = reg(src1, info, instr->repeat, IR3_REG_IMMED);
+ ldgb->src1_im = !!(src1->flags & IR3_REG_IMMED);
+ ldgb->src2 = reg(src2, info, instr->repeat, IR3_REG_IMMED);
+ ldgb->src2_im = !!(src2->flags & IR3_REG_IMMED);
+ ldgb->pad0 = 0x1;
+ ldgb->pad3 = 0x0;
+ }
+
+ return 0;
+ } else if (instr->opc == OPC_LDGB) {
+ struct ir3_register *src3 = instr->regs[3];
+ instr_cat6ldgb_t *ldgb = ptr;
+
+ /* maybe these two bits both determine the instruction encoding? */
+ cat6->src_off = false;
+
+ ldgb->d = instr->cat6.d - 1;
+ ldgb->typed = instr->cat6.typed;
+ ldgb->type_size = instr->cat6.iim_val - 1;
+
+ ldgb->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
+
+ /* first src is src_ssbo: */
+ iassert(src1->flags & IR3_REG_IMMED);
+ ldgb->src_ssbo = src1->uim_val;
+
+ /* then next two are src1/src2: */
+ ldgb->src1 = reg(src2, info, instr->repeat, IR3_REG_IMMED);
+ ldgb->src1_im = !!(src2->flags & IR3_REG_IMMED);
+ ldgb->src2 = reg(src3, info, instr->repeat, IR3_REG_IMMED);
+ ldgb->src2_im = !!(src3->flags & IR3_REG_IMMED);
+
+ ldgb->pad0 = 0x0;
+ ldgb->pad3 = 0x1;
+
+ return 0;
+ } else if (instr->opc == OPC_RESINFO) {
+ instr_cat6ldgb_t *ldgb = ptr;
+
+ ldgb->d = instr->cat6.d - 1;
+
+ ldgb->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
+
+ /* first src is src_ssbo: */
+ iassert(src1->flags & IR3_REG_IMMED);
+ ldgb->src_ssbo = src1->uim_val;
+
+ return 0;
+ } else if ((instr->opc == OPC_STGB) || (instr->opc == OPC_STIB)) {
+ struct ir3_register *src3 = instr->regs[4];
+ instr_cat6stgb_t *stgb = ptr;
+
+ /* maybe these two bits both determine the instruction encoding? */
+ cat6->src_off = true;
+ stgb->pad3 = 0x2;
+
+ stgb->d = instr->cat6.d - 1;
+ stgb->typed = instr->cat6.typed;
+ stgb->type_size = instr->cat6.iim_val - 1;
+
+ /* first src is dst_ssbo: */
+ iassert(dst->flags & IR3_REG_IMMED);
+ stgb->dst_ssbo = dst->uim_val;
+
+ /* then src1/src2/src3: */
+ stgb->src1 = reg(src1, info, instr->repeat, 0);
+ stgb->src2 = reg(src2, info, instr->repeat, IR3_REG_IMMED);
+ stgb->src2_im = !!(src2->flags & IR3_REG_IMMED);
+ stgb->src3 = reg(src3, info, instr->repeat, IR3_REG_IMMED);
+ stgb->src3_im = !!(src3->flags & IR3_REG_IMMED);
+
+ return 0;
+ } else if (instr->cat6.src_offset || (instr->opc == OPC_LDG) ||
+ (instr->opc == OPC_LDL)) {
instr_cat6a_t *cat6a = ptr;
cat6->src_off = true;
}
}
- if (instr->cat6.dst_offset || (instr->opc == OPC_STG)) {
+ if (instr->cat6.dst_offset || (instr->opc == OPC_STG) ||
+ (instr->opc == OPC_STL)) {
instr_cat6c_t *cat6c = ptr;
cat6->dst_off = true;
cat6c->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
cat6d->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
}
- cat6->type = instr->cat6.type;
- cat6->opc = instr->opc;
- cat6->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
- cat6->sync = !!(instr->flags & IR3_INSTR_SY);
- cat6->g = !!(instr->flags & IR3_INSTR_G);
- cat6->opc_cat = 6;
+ return 0;
+}
+
+static int emit_cat7(struct ir3_instruction *instr, void *ptr,
+ struct ir3_info *info)
+{
+ instr_cat7_t *cat7 = ptr;
+
+ cat7->ss = !!(instr->flags & IR3_INSTR_SS);
+ cat7->w = instr->cat7.w;
+ cat7->r = instr->cat7.r;
+ cat7->l = instr->cat7.l;
+ cat7->g = instr->cat7.g;
+ cat7->opc = instr->opc;
+ cat7->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
+ cat7->sync = !!(instr->flags & IR3_INSTR_SY);
+ cat7->opc_cat = 7;
return 0;
}
static int (*emit[])(struct ir3_instruction *instr, void *ptr,
struct ir3_info *info) = {
emit_cat0, emit_cat1, emit_cat2, emit_cat3, emit_cat4, emit_cat5, emit_cat6,
+ emit_cat7,
};
void * ir3_assemble(struct ir3 *shader, struct ir3_info *info,
info->max_const = -1;
info->instrs_count = 0;
info->sizedwords = 0;
+ info->ss = info->sy = 0;
list_for_each_entry (struct ir3_block, block, &shader->block_list, node) {
list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node) {
}
}
- /* need a integer number of instruction "groups" (sets of 16
+ /* need an integer number of instruction "groups" (sets of 16
* instructions on a4xx or sets of 4 instructions on a3xx),
* so pad out w/ NOPs if needed: (NOTE each instruction is 64bits)
*/
goto fail;
info->instrs_count += 1 + instr->repeat;
dwords += 2;
+
+ if (instr->flags & IR3_INSTR_SS)
+ info->ss++;
+
+ if (instr->flags & IR3_INSTR_SY)
+ info->sy++;
}
}
{
struct ir3 *shader = block->shader;
#ifdef DEBUG
- static uint32_t serialno = 0;
- instr->serialno = ++serialno;
+ instr->serialno = ++shader->instr_count;
#endif
list_addtail(&instr->node, &block->instr_list);
if (is_input(instr))
- array_insert(shader->baryfs, instr);
+ array_insert(shader, shader->baryfs, instr);
}
struct ir3_block * ir3_block_create(struct ir3 *shader)
{
struct ir3_block *block = ir3_alloc(shader, sizeof(*block));
#ifdef DEBUG
- static uint32_t serialno = 0;
- block->serialno = ++serialno;
+ block->serialno = ++shader->block_count;
#endif
block->shader = shader;
list_inithead(&block->node);
return new_instr;
}
+/* Add a false dependency to instruction, to ensure it is scheduled first: */
+void ir3_instr_add_dep(struct ir3_instruction *instr, struct ir3_instruction *dep)
+{
+ array_insert(instr, instr->deps, dep);
+}
+
struct ir3_register * ir3_reg_create(struct ir3_instruction *instr,
int num, int flags)
{
if (instr->address != addr) {
struct ir3 *ir = instr->block->shader;
instr->address = addr;
- array_insert(ir->indirects, instr);
+ array_insert(ir, ir->indirects, instr);
}
}
projects / mesa.git / blobdiff