ir3.c

/*
 * Copyright (c) 2012 Rob Clark <[email protected]>
 *
 * 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.
 */

#include "ir3.h"

#include <assert.h>
#include <errno.h>
#include <stdbool.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#include "util/bitscan.h"
#include "util/half_float.h"
#include "util/ralloc.h"
#include "util/u_math.h"

#include "instr-a3xx.h"
#include "ir3_shader.h"

/* 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)
{
   return rzalloc_size(shader, sz); /* TODO: don't use rzalloc */
}

struct ir3 *
ir3_create(struct ir3_compiler *compiler, struct ir3_shader_variant *v)
{
   struct ir3 *shader = rzalloc(v, struct ir3);

   shader->compiler = compiler;
   shader->type = v->type;

   list_inithead(&shader->block_list);
   list_inithead(&shader->array_list);

   return shader;
}

void
ir3_destroy(struct ir3 *shader)
{
   ralloc_free(shader);
}

static void
collect_reg_info(struct ir3_instruction *instr, struct ir3_register *reg,
                 struct ir3_info *info)
{
   struct ir3_shader_variant *v = info->data;
   unsigned repeat = instr->repeat;

   if (reg->flags & IR3_REG_IMMED) {
      /* nothing to do */
      return;
   }

   if (!(reg->flags & IR3_REG_R)) {
      repeat = 0;
   }

   unsigned components;
   int16_t max;

   if (reg->flags & IR3_REG_RELATIV) {
      components = reg->size;
      max = (reg->array.base + components - 1);
   } else {
      components = util_last_bit(reg->wrmask);
      max = (reg->num + repeat + components - 1);
   }

   if (reg->flags & IR3_REG_CONST) {
      info->max_const = MAX2(info->max_const, max >> 2);
   } else if (max < regid(48, 0)) {
      if (reg->flags & IR3_REG_HALF) {
         if (v->mergedregs) {
            /* starting w/ a6xx, half regs conflict with full regs: */
            info->max_reg = MAX2(info->max_reg, max >> 3);
         } else {
            info->max_half_reg = MAX2(info->max_half_reg, max >> 2);
         }
      } else {
         info->max_reg = MAX2(info->max_reg, max >> 2);
      }
   }
}

bool
ir3_should_double_threadsize(struct ir3_shader_variant *v, unsigned regs_count)
{
   const struct ir3_compiler *compiler = v->shader->compiler;

   /* We can't support more than compiler->branchstack_size diverging threads
    * in a wave. Thus, doubling the threadsize is only possible if we don't
    * exceed the branchstack size limit.
    */
   if (MIN2(v->branchstack, compiler->threadsize_base * 2) >
       compiler->branchstack_size) {
      return false;
   }

   switch (v->type) {
   case MESA_SHADER_COMPUTE: {
      unsigned threads_per_wg =
         v->local_size[0] * v->local_size[1] * v->local_size[2];

      /* For a5xx, if the workgroup size is greater than the maximum number
       * of threads per core with 32 threads per wave (512) then we have to
       * use the doubled threadsize because otherwise the workgroup wouldn't
       * fit. For smaller workgroup sizes, we follow the blob and use the
       * smaller threadsize.
       */
      if (compiler->gpu_id < 600) {
         return v->local_size_variable ||
                threads_per_wg >
                   compiler->threadsize_base * compiler->max_waves;
      }

      /* On a6xx, we prefer the larger threadsize unless the workgroup is
       * small enough that it would be useless. Note that because
       * threadsize_base is bumped to 64, we don't have to worry about the
       * workgroup fitting, unlike the a5xx case.
       */
      if (!v->local_size_variable) {
         if (threads_per_wg <= compiler->threadsize_base)
            return false;
      }
   }
      FALLTHROUGH;
   case MESA_SHADER_FRAGMENT: {
      /* Check that doubling the threadsize wouldn't exceed the regfile size */
      return regs_count * 2 <= compiler->reg_size_vec4;
   }

   default:
      /* On a6xx+, it's impossible to use a doubled wavesize in the geometry
       * stages - the bit doesn't exist. The blob never used it for the VS
       * on earlier gen's anyway.
       */
      return false;
   }
}

/* Get the maximum number of waves that could be used even if this shader
 * didn't use any registers.
 */
unsigned
ir3_get_reg_independent_max_waves(struct ir3_shader_variant *v,
                                  bool double_threadsize)
{
   const struct ir3_compiler *compiler = v->shader->compiler;
   unsigned max_waves = compiler->max_waves;

   /* If this is a compute shader, compute the limit based on shared size */
   if (v->type == MESA_SHADER_COMPUTE) {
      /* Shared is allocated in chunks of 1k */
      unsigned shared_per_wg = ALIGN_POT(v->shared_size, 1024);
      if (shared_per_wg > 0 && !v->local_size_variable) {
         unsigned wgs_per_core = compiler->local_mem_size / shared_per_wg;
         unsigned threads_per_wg =
            v->local_size[0] * v->local_size[1] * v->local_size[2];
         unsigned waves_per_wg =
            DIV_ROUND_UP(threads_per_wg, compiler->threadsize_base *
                                            (double_threadsize ? 2 : 1) *
                                            compiler->wave_granularity);
         max_waves = MIN2(max_waves, waves_per_wg * wgs_per_core *
                                        compiler->wave_granularity);
      }
   }

   /* Compute the limit based on branchstack */
   if (v->branchstack > 0) {
      unsigned branchstack_max_waves = compiler->branchstack_size /
                                       v->branchstack *
                                       compiler->wave_granularity;
      max_waves = MIN2(max_waves, branchstack_max_waves);
   }

   return max_waves;
}

/* Get the maximum number of waves that could be launched limited by reg size.
 */
unsigned
ir3_get_reg_dependent_max_waves(const struct ir3_compiler *compiler,
                                unsigned reg_count, bool double_threadsize)
{
   return reg_count ? (compiler->reg_size_vec4 /
                       (reg_count * (double_threadsize ? 2 : 1)) *
                       compiler->wave_granularity)
                    : compiler->max_waves;
}

void
ir3_collect_info(struct ir3_shader_variant *v)
{
   struct ir3_info *info = &v->info;
   struct ir3 *shader = v->ir;
   const struct ir3_compiler *compiler = v->shader->compiler;

   memset(info, 0, sizeof(*info));
   info->data = v;
   info->max_reg = -1;
   info->max_half_reg = -1;
   info->max_const = -1;
   info->multi_dword_ldp_stp = false;

   uint32_t instr_count = 0;
   foreach_block (block, &shader->block_list) {
      foreach_instr (instr, &block->instr_list) {
         instr_count++;
      }
   }

   v->instrlen = DIV_ROUND_UP(instr_count, compiler->instr_align);

   /* Pad out with NOPs to instrlen, including at least 4 so that cffdump
    * doesn't try to decode the following data as instructions (such as the
    * next stage's shader in turnip)
    */
   info->size = MAX2(v->instrlen * compiler->instr_align, instr_count + 4) * 8;
   info->sizedwords = info->size / 4;

   foreach_block (block, &shader->block_list) {
      int sfu_delay = 0;

      foreach_instr (instr, &block->instr_list) {

         foreach_src (reg, instr) {
            collect_reg_info(instr, reg, info);
         }

         foreach_dst (reg, instr) {
            if (is_dest_gpr(reg)) {
               collect_reg_info(instr, reg, info);
            }
         }

         if ((instr->opc == OPC_STP || instr->opc == OPC_LDP)) {
            struct ir3_register *base =
               (instr->opc == OPC_STP) ? instr->srcs[2] : instr->srcs[1];
            if (base->iim_val * type_size(instr->cat6.type) > 32) {
               info->multi_dword_ldp_stp = true;
            }
         }

         if ((instr->opc == OPC_BARY_F) && (instr->dsts[0]->flags & IR3_REG_EI))
            info->last_baryf = info->instrs_count;

         unsigned instrs_count = 1 + instr->repeat + instr->nop;
         unsigned nops_count = instr->nop;

         if (instr->opc == OPC_NOP) {
            nops_count = 1 + instr->repeat;
            info->instrs_per_cat[0] += nops_count;
         } else {
            info->instrs_per_cat[opc_cat(instr->opc)] += 1 + instr->repeat;
            info->instrs_per_cat[0] += nops_count;
         }

         if (instr->opc == OPC_MOV) {
            if (instr->cat1.src_type == instr->cat1.dst_type) {
               info->mov_count += 1 + instr->repeat;
            } else {
               info->cov_count += 1 + instr->repeat;
            }
         }

         info->instrs_count += instrs_count;
         info->nops_count += nops_count;

         if (instr->flags & IR3_INSTR_SS) {
            info->ss++;
            info->sstall += sfu_delay;
            sfu_delay = 0;
         }

         if (instr->flags & IR3_INSTR_SY)
            info->sy++;

         if (is_sfu(instr)) {
            sfu_delay = 10;
         } else {
            int n = MIN2(sfu_delay, 1 + instr->repeat + instr->nop);
            sfu_delay -= n;
         }
      }
   }

   /* TODO: for a5xx and below, is there a separate regfile for
    * half-registers?
    */
   unsigned regs_count =
      info->max_reg + 1 +
      (compiler->gpu_id >= 600 ? ((info->max_half_reg + 2) / 2) : 0);

   info->double_threadsize = ir3_should_double_threadsize(v, regs_count);
   unsigned reg_independent_max_waves =
      ir3_get_reg_independent_max_waves(v, info->double_threadsize);
   unsigned reg_dependent_max_waves = ir3_get_reg_dependent_max_waves(
      compiler, regs_count, info->double_threadsize);
   info->max_waves = MIN2(reg_independent_max_waves, reg_dependent_max_waves);
   assert(info->max_waves <= v->shader->compiler->max_waves);
}

static struct ir3_register *
reg_create(struct ir3 *shader, int num, int flags)
{
   struct ir3_register *reg = ir3_alloc(shader, sizeof(struct ir3_register));
   reg->wrmask = 1;
   reg->flags = flags;
   reg->num = num;
   return reg;
}

static void
insert_instr(struct ir3_block *block, struct ir3_instruction *instr)
{
   struct ir3 *shader = block->shader;

   instr->serialno = ++shader->instr_count;

   list_addtail(&instr->node, &block->instr_list);

   if (is_input(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
   block->serialno = ++shader->block_count;
#endif
   block->shader = shader;
   list_inithead(&block->node);
   list_inithead(&block->instr_list);
   return block;
}

void
ir3_block_add_predecessor(struct ir3_block *block, struct ir3_block *pred)
{
   array_insert(block, block->predecessors, pred);
}

void
ir3_block_add_physical_predecessor(struct ir3_block *block,
                                   struct ir3_block *pred)
{
   array_insert(block, block->physical_predecessors, pred);
}

void
ir3_block_remove_predecessor(struct ir3_block *block, struct ir3_block *pred)
{
   for (unsigned i = 0; i < block->predecessors_count; i++) {
      if (block->predecessors[i] == pred) {
         if (i < block->predecessors_count - 1) {
            block->predecessors[i] =
               block->predecessors[block->predecessors_count - 1];
         }

         block->predecessors_count--;
         return;
      }
   }
}

unsigned
ir3_block_get_pred_index(struct ir3_block *block, struct ir3_block *pred)
{
   for (unsigned i = 0; i < block->predecessors_count; i++) {
      if (block->predecessors[i] == pred) {
         return i;
      }
   }

   unreachable("ir3_block_get_pred_index() invalid predecessor");
}

static struct ir3_instruction *
instr_create(struct ir3_block *block, opc_t opc, int ndst, int nsrc)
{
   /* Add extra sources for array destinations and the address reg */
   if (1 <= opc_cat(opc))
      nsrc += 2;
   struct ir3_instruction *instr;
   unsigned sz = sizeof(*instr) + (ndst * sizeof(instr->dsts[0])) +
                 (nsrc * sizeof(instr->srcs[0]));
   char *ptr = ir3_alloc(block->shader, sz);

   instr = (struct ir3_instruction *)ptr;
   ptr += sizeof(*instr);
   instr->dsts = (struct ir3_register **)ptr;
   instr->srcs = instr->dsts + ndst;

#ifdef DEBUG
   instr->dsts_max = ndst;
   instr->srcs_max = nsrc;
#endif

   return instr;
}

struct ir3_instruction *
ir3_instr_create(struct ir3_block *block, opc_t opc, int ndst, int nsrc)
{
   struct ir3_instruction *instr = instr_create(block, opc, ndst, nsrc);
   instr->block = block;
   instr->opc = opc;
   insert_instr(block, instr);
   return instr;
}

struct ir3_instruction *
ir3_instr_clone(struct ir3_instruction *instr)
{
   struct ir3_instruction *new_instr = instr_create(
      instr->block, instr->opc, instr->dsts_count, instr->srcs_count);
   struct ir3_register **dsts, **srcs;

   dsts = new_instr->dsts;
   srcs = new_instr->srcs;
   *new_instr = *instr;
   new_instr->dsts = dsts;
   new_instr->srcs = srcs;

   insert_instr(instr->block, new_instr);

   /* clone registers: */
   new_instr->dsts_count = 0;
   new_instr->srcs_count = 0;
   foreach_dst (reg, instr) {
      struct ir3_register *new_reg =
         ir3_dst_create(new_instr, reg->num, reg->flags);
      *new_reg = *reg;
      if (new_reg->instr)
         new_reg->instr = new_instr;
   }
   foreach_src (reg, instr) {
      struct ir3_register *new_reg =
         ir3_src_create(new_instr, reg->num, reg->flags);
      *new_reg = *reg;
   }

   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)
{
   for (unsigned i = 0; i < instr->deps_count; i++) {
      if (instr->deps[i] == dep)
         return;
   }

   array_insert(instr, instr->deps, dep);
}

struct ir3_register *
ir3_src_create(struct ir3_instruction *instr, int num, int flags)
{
   struct ir3 *shader = instr->block->shader;
#ifdef DEBUG
   debug_assert(instr->srcs_count < instr->srcs_max);
#endif
   struct ir3_register *reg = reg_create(shader, num, flags);
   instr->srcs[instr->srcs_count++] = reg;
   return reg;
}

struct ir3_register *
ir3_dst_create(struct ir3_instruction *instr, int num, int flags)
{
   struct ir3 *shader = instr->block->shader;
#ifdef DEBUG
   debug_assert(instr->dsts_count < instr->dsts_max);
#endif
   struct ir3_register *reg = reg_create(shader, num, flags);
   instr->dsts[instr->dsts_count++] = reg;
   return reg;
}

struct ir3_register *
ir3_reg_clone(struct ir3 *shader, struct ir3_register *reg)
{
   struct ir3_register *new_reg = reg_create(shader, 0, 0);
   *new_reg = *reg;
   return new_reg;
}

void
ir3_reg_set_last_array(struct ir3_instruction *instr, struct ir3_register *reg,
                       struct ir3_register *last_write)
{
   assert(reg->flags & IR3_REG_ARRAY);
   struct ir3_register *new_reg = ir3_src_create(instr, 0, 0);
   *new_reg = *reg;
   new_reg->def = last_write;
   ir3_reg_tie(reg, new_reg);
}

void
ir3_instr_set_address(struct ir3_instruction *instr,
                      struct ir3_instruction *addr)
{
   if (!instr->address) {
      struct ir3 *ir = instr->block->shader;

      debug_assert(instr->block == addr->block);

      instr->address =
         ir3_src_create(instr, addr->dsts[0]->num, addr->dsts[0]->flags);
      instr->address->def = addr->dsts[0];
      debug_assert(reg_num(addr->dsts[0]) == REG_A0);
      unsigned comp = reg_comp(addr->dsts[0]);
      if (comp == 0) {
         array_insert(ir, ir->a0_users, instr);
      } else {
         debug_assert(comp == 1);
         array_insert(ir, ir->a1_users, instr);
      }
   } else {
      debug_assert(instr->address->def->instr == addr);
   }
}

void
ir3_block_clear_mark(struct ir3_block *block)
{
   foreach_instr (instr, &block->instr_list)
      instr->flags &= ~IR3_INSTR_MARK;
}

void
ir3_clear_mark(struct ir3 *ir)
{
   foreach_block (block, &ir->block_list) {
      ir3_block_clear_mark(block);
   }
}

unsigned
ir3_count_instructions(struct ir3 *ir)
{
   unsigned cnt = 1;
   foreach_block (block, &ir->block_list) {
      block->start_ip = cnt;
      foreach_instr (instr, &block->instr_list) {
         instr->ip = cnt++;
      }
      block->end_ip = cnt;
   }
   return cnt;
}

/* When counting instructions for RA, we insert extra fake instructions at the
 * beginning of each block, where values become live, and at the end where
 * values die. This prevents problems where values live-in at the beginning or
 * live-out at the end of a block from being treated as if they were
 * live-in/live-out at the first/last instruction, which would be incorrect.
 * In ir3_legalize these ip's are assumed to be actual ip's of the final
 * program, so it would be incorrect to use this everywhere.
 */

unsigned
ir3_count_instructions_ra(struct ir3 *ir)
{
   unsigned cnt = 1;
   foreach_block (block, &ir->block_list) {
      block->start_ip = cnt++;
      foreach_instr (instr, &block->instr_list) {
         instr->ip = cnt++;
      }
      block->end_ip = cnt++;
   }
   return cnt;
}

struct ir3_array *
ir3_lookup_array(struct ir3 *ir, unsigned id)
{
   foreach_array (arr, &ir->array_list)
      if (arr->id == id)
         return arr;
   return NULL;
}

void
ir3_find_ssa_uses(struct ir3 *ir, void *mem_ctx, bool falsedeps)
{
   /* We could do this in a single pass if we can assume instructions
    * are always sorted.  Which currently might not always be true.
    * (In particular after ir3_group pass, but maybe other places.)
    */
   foreach_block (block, &ir->block_list)
      foreach_instr (instr, &block->instr_list)
         instr->uses = NULL;

   foreach_block (block, &ir->block_list) {
      foreach_instr (instr, &block->instr_list) {
         foreach_ssa_src_n (src, n, instr) {
            if (__is_false_dep(instr, n) && !falsedeps)
               continue;
            if (!src->uses)
               src->uses = _mesa_pointer_set_create(mem_ctx);
            _mesa_set_add(src->uses, instr);
         }
      }
   }
}

/**
 * Set the destination type of an instruction, for example if a
 * conversion is folded in, handling the special cases where the
 * instruction's dest type or opcode needs to be fixed up.
 */
void
ir3_set_dst_type(struct ir3_instruction *instr, bool half)
{
   if (half) {
      instr->dsts[0]->flags |= IR3_REG_HALF;
   } else {
      instr->dsts[0]->flags &= ~IR3_REG_HALF;
   }

   switch (opc_cat(instr->opc)) {
   case 1: /* move instructions */
      if (half) {
         instr->cat1.dst_type = half_type(instr->cat1.dst_type);
      } else {
         instr->cat1.dst_type = full_type(instr->cat1.dst_type);
      }
      break;
   case 4:
      if (half) {
         instr->opc = cat4_half_opc(instr->opc);
      } else {
         instr->opc = cat4_full_opc(instr->opc);
      }
      break;
   case 5:
      if (half) {
         instr->cat5.type = half_type(instr->cat5.type);
      } else {
         instr->cat5.type = full_type(instr->cat5.type);
      }
      break;
   }
}

/**
 * One-time fixup for instruction src-types.  Other than cov's that
 * are folded, an instruction's src type does not change.
 */
void
ir3_fixup_src_type(struct ir3_instruction *instr)
{
   switch (opc_cat(instr->opc)) {
   case 1: /* move instructions */
      if (instr->srcs[0]->flags & IR3_REG_HALF) {
         instr->cat1.src_type = half_type(instr->cat1.src_type);
      } else {
         instr->cat1.src_type = full_type(instr->cat1.src_type);
      }
      break;
   case 3:
      if (instr->srcs[0]->flags & IR3_REG_HALF) {
         instr->opc = cat3_half_opc(instr->opc);
      } else {
         instr->opc = cat3_full_opc(instr->opc);
      }
      break;
   }
}

/**
 * Map a floating point immed to FLUT (float lookup table) value,
 * returns negative for immediates that cannot be mapped.
 */
int
ir3_flut(struct ir3_register *src_reg)
{
   static const struct {
      uint32_t f32;
      uint16_t f16;
   } flut[] = {
         { .f32 = 0x00000000, .f16 = 0x0000 },    /* 0.0 */
         { .f32 = 0x3f000000, .f16 = 0x3800 },    /* 0.5 */
         { .f32 = 0x3f800000, .f16 = 0x3c00 },    /* 1.0 */
         { .f32 = 0x40000000, .f16 = 0x4000 },    /* 2.0 */
         { .f32 = 0x402df854, .f16 = 0x4170 },    /* e */
         { .f32 = 0x40490fdb, .f16 = 0x4248 },    /* pi */
         { .f32 = 0x3ea2f983, .f16 = 0x3518 },    /* 1/pi */
         { .f32 = 0x3f317218, .f16 = 0x398c },    /* 1/log2(e) */
         { .f32 = 0x3fb8aa3b, .f16 = 0x3dc5 },    /* log2(e) */
         { .f32 = 0x3e9a209b, .f16 = 0x34d1 },    /* 1/log2(10) */
         { .f32 = 0x40549a78, .f16 = 0x42a5 },    /* log2(10) */
         { .f32 = 0x40800000, .f16 = 0x4400 },    /* 4.0 */
   };

   if (src_reg->flags & IR3_REG_HALF) {
      /* Note that half-float immeds are already lowered to 16b in nir: */
      uint32_t imm = src_reg->uim_val;
      for (unsigned i = 0; i < ARRAY_SIZE(flut); i++) {
         if (flut[i].f16 == imm) {
            return i;
         }
      }
   } else {
      uint32_t imm = src_reg->uim_val;
      for (unsigned i = 0; i < ARRAY_SIZE(flut); i++) {
         if (flut[i].f32 == imm) {
            return i;
         }
      }
   }

   return -1;
}

static unsigned
cp_flags(unsigned flags)
{
   /* only considering these flags (at least for now): */
   flags &= (IR3_REG_CONST | IR3_REG_IMMED | IR3_REG_FNEG | IR3_REG_FABS |
             IR3_REG_SNEG | IR3_REG_SABS | IR3_REG_BNOT | IR3_REG_RELATIV |
             IR3_REG_SHARED);
   return flags;
}

bool
ir3_valid_flags(struct ir3_instruction *instr, unsigned n, unsigned flags)
{
   struct ir3_compiler *compiler = instr->block->shader->compiler;
   unsigned valid_flags;

   if ((flags & IR3_REG_SHARED) && opc_cat(instr->opc) > 3)
      return false;

   flags = cp_flags(flags);

   /* If destination is indirect, then source cannot be.. at least
    * I don't think so..
    */
   if (instr->dsts_count > 0 && (instr->dsts[0]->flags & IR3_REG_RELATIV) &&
       (flags & IR3_REG_RELATIV))
      return false;

   if (flags & IR3_REG_RELATIV) {
      /* TODO need to test on earlier gens.. pretty sure the earlier
       * problem was just that we didn't check that the src was from
       * same block (since we can't propagate address register values
       * across blocks currently)
       */
      if (compiler->gpu_id < 600)
         return false;

      /* NOTE in the special try_swap_mad_two_srcs() case we can be
       * called on a src that has already had an indirect load folded
       * in, in which case ssa() returns NULL
       */
      if (instr->srcs[n]->flags & IR3_REG_SSA) {
         struct ir3_instruction *src = ssa(instr->srcs[n]);
         if (src->address->def->instr->block != instr->block)
            return false;
      }
   }

   if (is_meta(instr)) {
      /* collect and phi nodes support const/immed sources, which will be
       * turned into move instructions, but not anything else.
       */
      if (flags & ~(IR3_REG_IMMED | IR3_REG_CONST | IR3_REG_SHARED))
         return false;

      if ((flags & IR3_REG_SHARED) && !(instr->dsts[0]->flags & IR3_REG_SHARED))
         return false;

      return true;
   }

   switch (opc_cat(instr->opc)) {
   case 0: /* end, chmask */
      return flags == 0;
   case 1:
      switch (instr->opc) {
      case OPC_MOVMSK:
      case OPC_SWZ:
      case OPC_SCT:
      case OPC_GAT:
         valid_flags = IR3_REG_SHARED;
         break;
      default:
         valid_flags =
            IR3_REG_IMMED | IR3_REG_CONST | IR3_REG_RELATIV | IR3_REG_SHARED;
      }
      if (flags & ~valid_flags)
         return false;
      break;
   case 2:
      valid_flags = ir3_cat2_absneg(instr->opc) | IR3_REG_CONST |
                    IR3_REG_RELATIV | IR3_REG_IMMED | IR3_REG_SHARED;

      if (flags & ~valid_flags)
         return false;

      if (flags & (IR3_REG_CONST | IR3_REG_IMMED | IR3_REG_SHARED)) {
         unsigned m = n ^ 1;
         /* cannot deal w/ const or shared in both srcs:
          * (note that some cat2 actually only have a single src)
          */
         if (m < instr->srcs_count) {
            struct ir3_register *reg = instr->srcs[m];
            if ((flags & (IR3_REG_CONST | IR3_REG_SHARED)) &&
                (reg->flags & (IR3_REG_CONST | IR3_REG_SHARED)))
               return false;
            if ((flags & IR3_REG_IMMED) && reg->flags & (IR3_REG_IMMED))
               return false;
         }
      }
      break;
   case 3:
      valid_flags =
         ir3_cat3_absneg(instr->opc) | IR3_REG_RELATIV | IR3_REG_SHARED;

      if (instr->opc == OPC_SHLG_B16) {
         valid_flags |= IR3_REG_IMMED;
         /* shlg.b16 can be RELATIV+CONST but not CONST: */
         if (flags & IR3_REG_RELATIV)
            valid_flags |= IR3_REG_CONST;
      } else {
         valid_flags |= IR3_REG_CONST;
      }

      if (flags & ~valid_flags)
         return false;

      if (flags & (IR3_REG_CONST | IR3_REG_SHARED | IR3_REG_RELATIV)) {
         /* cannot deal w/ const/shared/relativ in 2nd src: */
         if (n == 1)
            return false;
      }

      break;
   case 4:
      /* seems like blob compiler avoids const as src.. */
      /* TODO double check if this is still the case on a4xx */
      if (flags & (IR3_REG_CONST | IR3_REG_IMMED))
         return false;
      if (flags & (IR3_REG_SABS | IR3_REG_SNEG))
         return false;
      break;
   case 5:
      /* no flags allowed */
      if (flags)
         return false;
      break;
   case 6:
      valid_flags = IR3_REG_IMMED;
      if (flags & ~valid_flags)
         return false;

      if (flags & IR3_REG_IMMED) {
         /* doesn't seem like we can have immediate src for store
          * instructions:
          *
          * TODO this restriction could also apply to load instructions,
          * but for load instructions this arg is the address (and not
          * really sure any good way to test a hard-coded immed addr src)
          */
         if (is_store(instr) && (instr->opc != OPC_STG) && (n == 1))
            return false;

         if ((instr->opc == OPC_LDL) && (n == 0))
            return false;

         if ((instr->opc == OPC_STL) && (n != 2))
            return false;

         if ((instr->opc == OPC_LDP) && (n == 0))
            return false;

         if ((instr->opc == OPC_STP) && (n != 2))
            return false;

         if (instr->opc == OPC_STLW && n == 0)
            return false;

         if (instr->opc == OPC_LDLW && n == 0)
            return false;

         /* disallow immediates in anything but the SSBO slot argument for
          * cat6 instructions:
          */
         if (is_atomic(instr->opc) && (n != 0))
            return false;

         if (is_atomic(instr->opc) && !(instr->flags & IR3_INSTR_G))
            return false;

         if (instr->opc == OPC_STG && (n == 2))
            return false;

         if (instr->opc == OPC_STG_A && (n == 4))
            return false;

         /* as with atomics, these cat6 instrs can only have an immediate
          * for SSBO/IBO slot argument
          */
         switch (instr->opc) {
         case OPC_LDIB:
         case OPC_STIB:
         case OPC_LDC:
         case OPC_RESINFO:
            if (n != 0)
               return false;
            break;
         default:
            break;
         }
      }

      break;
   }

   return true;
}