lima/ppir: implement gl_FragDepth support

Mali4x0 supports writing depth and stencil from fragment shader
and we've been using it quite a while for depth/stencil buffer reload.

The missing part was specifying output register for depth/stencil.
To figure it out, I changed reload shader to use register $4 as output
and poked RSW bits (or rather consecutive 4 bit groups) until tests
that rely on reload started to pass again.

It turns out that register number for gl_FragDepth/gl_FragStencil is in
rsw->depth_test and register number for gl_FragColor is in
rsw->multi_sample and it's repeated 4 times for some reason (likely for
MSAA?)

With this knowledge we now can modify ppir compiler to support multiple
store_output intrinsics.

To do that just add destination SSA for store_output to the registers
list for regalloc and mark them explicitly as output. Since it's never
read in shader we have to take care about it in liveness analysis -
basically just mark it alive from the time when it's written to the end
of the block. If it's live only in the last instruction, mark it as
live_internal, so regalloc doesn't clobber it.

Then just let regalloc do its job, and then copy register number to the
shader state and program it in RSW.

The tricky part is gl_FragStencil, since it resides in the same register
as gl_FragDepth and with the current design of the compiler it's hard to
merge them. However gl_FragStencil doesn't seem to be part of GL2
or GLES2, so we can just leave it not implemented.

Also we need to take care of stop bit for instructions - now we can't
just set it in every instruction that stores output, since there may be
several outputs. So if there's any store_output instructions in the
block just mark that block has a stop, and set stop bit in the last
instruction in the block. The only exception is discard - we always need
to set stop bit in discard instruction.

Reviewed-by: Andreas Baierl <ichgeh@imkreisrum.de>
Reviewed-by: Erico Nunes <nunes.erico@gmail.com>
Signed-off-by: Vasily Khoruzhick <anarsoul@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13830>

11 files changed, 144 insertions, 37 deletions

diff --git a/src/gallium/drivers/lima/ir/pp/codegen.c b/src/gallium/drivers/lima/ir/pp/codegen.c
index 5a5ee14c78e..0760e19c1b9 100644
--- a/src/gallium/drivers/lima/ir/pp/codegen.c
+++ b/src/gallium/drivers/lima/ir/pp/codegen.c
@@ -773,7 +773,7 @@ static int encode_instr(ppir_instr *instr, void *code, void *last_code)
    size = align_to_word(size) + 1;
 
    ctrl->count = size;
-   if (instr->is_end)
+   if (instr->stop)
       ctrl->stop = true;
 
    if (last_code) {
@@ -818,6 +818,11 @@ bool ppir_codegen_prog(ppir_compiler *comp)
          instr->encode_size = get_instr_encode_size(instr);
          size += instr->encode_size;
       }
+      /* Set stop flag for the last instruction if block has stop flag */
+      if (block->stop) {
+         ppir_instr *instr = list_last_entry(&block->instr_list, ppir_instr, list);
+         instr->stop = true;
+      }
    }
 
    uint32_t *prog = rzalloc_size(comp->prog, size * sizeof(uint32_t));
diff --git a/src/gallium/drivers/lima/ir/pp/instr.c b/src/gallium/drivers/lima/ir/pp/instr.c
index 8e1bc95158d..fc64de8812f 100644
--- a/src/gallium/drivers/lima/ir/pp/instr.c
+++ b/src/gallium/drivers/lima/ir/pp/instr.c
@@ -284,7 +284,7 @@ void ppir_instr_print_list(ppir_compiler *comp)
    list_for_each_entry(ppir_block, block, &comp->block_list, list) {
       printf("-------block %3d-------\n", block->index);
       list_for_each_entry(ppir_instr, instr, &block->instr_list, list) {
-         printf("%c%03d: ", instr->is_end ? '*' : ' ', instr->index);
+         printf("%c%03d: ", instr->stop ? '*' : ' ', instr->index);
          for (int i = 0; i < PPIR_INSTR_SLOT_NUM; i++) {
             ppir_node *node = instr->slots[i];
             if (node)
diff --git a/src/gallium/drivers/lima/ir/pp/liveness.c b/src/gallium/drivers/lima/ir/pp/liveness.c
index 1799a53b165..8f642e2d05c 100644
--- a/src/gallium/drivers/lima/ir/pp/liveness.c
+++ b/src/gallium/drivers/lima/ir/pp/liveness.c
@@ -121,7 +121,7 @@ ppir_liveness_instr_srcs(ppir_compiler *comp, ppir_instr *instr)
 /* Update the liveness information of the instruction by removing its
  * dests from the live_in set. */
 static void
-ppir_liveness_instr_dest(ppir_compiler *comp, ppir_instr *instr)
+ppir_liveness_instr_dest(ppir_compiler *comp, ppir_instr *instr, ppir_instr *last)
 {
    for (int i = PPIR_INSTR_SLOT_NUM-1; i >= 0; i--) {
       ppir_node *node = instr->slots[i];
@@ -146,9 +146,18 @@ ppir_liveness_instr_dest(ppir_compiler *comp, ppir_instr *instr)
       unsigned int index = reg->regalloc_index;
       bool live = BITSET_TEST(instr->live_set, index);
 
+      /* If it's an out reg, it's alive till the end of the block, so add it
+       * to live_set of the last instruction */
+      if (!live && reg->out_reg && (instr != last)) {
+         BITSET_SET(last->live_set, index);
+         BITSET_CLEAR(instr->live_set, index);
+         continue;
+      }
+
       /* If a register is written but wasn't read in a later instruction, it is
-       * either dead code or a bug. For now, assign an interference to it to
-       * ensure it doesn't get assigned a live register and overwrites it. */
+       * either an output register in last instruction, dead code or a bug.
+       * For now, assign an interference to it to ensure it doesn't get assigned
+       * a live register and overwrites it. */
       if (!live) {
          BITSET_SET(instr->live_internal, index);
          continue;
@@ -230,7 +239,7 @@ ppir_liveness_compute_live_sets(ppir_compiler *comp)
                                     instr->live_mask, next_instr->live_mask);
          }
 
-         ppir_liveness_instr_dest(comp, instr);
+         ppir_liveness_instr_dest(comp, instr, last);
          ppir_liveness_instr_srcs(comp, instr);
 
          cont |= !ppir_liveness_set_equal(comp,
diff --git a/src/gallium/drivers/lima/ir/pp/nir.c b/src/gallium/drivers/lima/ir/pp/nir.c
index 6f61a0986fb..fc8030fc745 100644
--- a/src/gallium/drivers/lima/ir/pp/nir.c
+++ b/src/gallium/drivers/lima/ir/pp/nir.c
@@ -345,6 +345,18 @@ static bool ppir_emit_intrinsic(ppir_block *block, nir_instr *ni)
        * back to inserting a mov at the end.
        * If the source node will only be able to output to pipeline
        * registers, fall back to the mov as well. */
+      assert(nir_src_is_const(instr->src[1]) &&
+             "lima doesn't support indirect outputs");
+
+      nir_io_semantics io = nir_intrinsic_io_semantics(instr);
+      unsigned offset = nir_src_as_uint(instr->src[1]);
+      unsigned slot = io.location + offset;
+      ppir_output_type out_type = ppir_nir_output_to_ppir(slot);
+      if (out_type == ppir_output_invalid) {
+         ppir_debug("Unsupported output type: %d\n", slot);
+         return false;
+      }
+
       if (!block->comp->uses_discard && instr->src->is_ssa) {
          node = block->comp->var_nodes[instr->src->ssa->index];
          switch (node->op) {
@@ -352,9 +364,12 @@ static bool ppir_emit_intrinsic(ppir_block *block, nir_instr *ni)
          case ppir_op_load_texture:
          case ppir_op_const:
             break;
-         default:
-            node->is_end = 1;
+         default: {
+            ppir_dest *dest = ppir_node_get_dest(node);
+            dest->ssa.out_type = out_type;
+            node->is_out = 1;
             return true;
+            }
          }
       }
 
@@ -367,6 +382,7 @@ static bool ppir_emit_intrinsic(ppir_block *block, nir_instr *ni)
       dest->ssa.num_components = instr->num_components;
       dest->ssa.index = 0;
       dest->write_mask = u_bit_consecutive(0, instr->num_components);
+      dest->ssa.out_type = out_type;
 
       alu_node->num_src = 1;
 
@@ -376,7 +392,7 @@ static bool ppir_emit_intrinsic(ppir_block *block, nir_instr *ni)
       ppir_node_add_src(block->comp, &alu_node->node, alu_node->src, instr->src,
                         u_bit_consecutive(0, instr->num_components));
 
-      alu_node->node.is_end = 1;
+      alu_node->node.is_out = 1;
 
       list_addtail(&alu_node->node.list, &block->node_list);
       return true;
@@ -798,6 +814,7 @@ static ppir_compiler *ppir_compiler_create(void *prog, unsigned num_reg, unsigne
    comp->var_nodes = (ppir_node **)(comp + 1);
    comp->reg_base = num_ssa;
    comp->prog = prog;
+
    return comp;
 }
 
@@ -833,7 +850,7 @@ static void ppir_add_ordering_deps(ppir_compiler *comp)
          if (prev_node && ppir_node_is_root(node) && node->op != ppir_op_const) {
             ppir_node_add_dep(prev_node, node, ppir_dep_sequence);
          }
-         if (node->is_end ||
+         if (node->is_out ||
              node->op == ppir_op_discard ||
              node->op == ppir_op_store_temp ||
              node->op == ppir_op_branch) {
@@ -930,18 +947,11 @@ bool ppir_compile_nir(struct lima_fs_compiled_shader *prog, struct nir_shader *n
       }
    }
 
-   /* Validate outputs, we support only gl_FragColor */
-   nir_foreach_shader_out_variable(var, nir) {
-      switch (var->data.location) {
-      case FRAG_RESULT_COLOR:
-      case FRAG_RESULT_DATA0:
-         break;
-      default:
-         ppir_error("unsupported output type\n");
-         goto err_out0;
-         break;
-      }
-   }
+   comp->out_type_to_reg = rzalloc_size(comp, sizeof(int) * ppir_output_num);
+
+   /* -1 means reg is not written by the shader */
+   for (int i = 0; i < ppir_output_num; i++)
+      comp->out_type_to_reg[i] = -1;
 
    foreach_list_typed(nir_register, reg, node, &func->registers) {
       ppir_reg *r = rzalloc(comp, ppir_reg);
diff --git a/src/gallium/drivers/lima/ir/pp/node.c b/src/gallium/drivers/lima/ir/pp/node.c
index 99d025e2c05..cc51448bfb3 100644
--- a/src/gallium/drivers/lima/ir/pp/node.c
+++ b/src/gallium/drivers/lima/ir/pp/node.c
@@ -618,9 +618,9 @@ static ppir_node *ppir_node_insert_mov_local(ppir_node *node)
    ppir_node_add_dep(move, node, ppir_dep_src);
    list_addtail(&move->list, &node->list);
 
-   if (node->is_end) {
-      node->is_end = false;
-      move->is_end = true;
+   if (node->is_out) {
+      node->is_out = false;
+      move->is_out = true;
    }
 
    return move;
diff --git a/src/gallium/drivers/lima/ir/pp/node_to_instr.c b/src/gallium/drivers/lima/ir/pp/node_to_instr.c
index 0d5d37f4927..ff8d735d9a7 100644
--- a/src/gallium/drivers/lima/ir/pp/node_to_instr.c
+++ b/src/gallium/drivers/lima/ir/pp/node_to_instr.c
@@ -203,7 +203,7 @@ static bool ppir_do_one_node_to_instr(ppir_block *block, ppir_node *node)
    case ppir_node_type_discard:
       if (!create_new_instr(block, node))
          return false;
-      node->instr->is_end = true;
+      block->stop = true;
       break;
    case ppir_node_type_branch:
       if (!create_new_instr(block, node))
@@ -276,8 +276,13 @@ static bool ppir_do_node_to_instr(ppir_block *block, ppir_node *root)
          if (!ppir_do_one_node_to_instr(block, node))
             return false;
 
-      if (node->is_end)
-         node->instr->is_end = true;
+      /* The node writes output register. We can't stop at this exact
+       * instruction because there may be another node that writes another
+       * output, so set stop flag for the block. We will set stop flag on
+       * the last instruction of the block during codegen
+       */
+      if (node->is_out)
+         block->stop = true;
 
       ppir_node_foreach_pred(node, dep) {
          ppir_node *pred = dep->pred;
diff --git a/src/gallium/drivers/lima/ir/pp/ppir.h b/src/gallium/drivers/lima/ir/pp/ppir.h
index 74a508e504e..f434786f4cf 100644
--- a/src/gallium/drivers/lima/ir/pp/ppir.h
+++ b/src/gallium/drivers/lima/ir/pp/ppir.h
@@ -161,7 +161,7 @@ typedef struct ppir_node {
    struct ppir_instr *instr;
    int instr_pos;
    struct ppir_block *block;
-   bool is_end;
+   bool is_out;
    bool succ_different_block;
 
    /* for scheduler */
@@ -179,9 +179,42 @@ typedef enum {
    ppir_pipeline_reg_discard, /* varying load */
 } ppir_pipeline;
 
+typedef enum {
+   ppir_output_color,
+   ppir_output_depth,
+   ppir_output_num,
+   ppir_output_invalid = -1,
+} ppir_output_type;
+
+static inline const char *ppir_output_type_to_str(ppir_output_type type)
+{
+   switch (type) {
+   case ppir_output_color:
+      return "OUTPUT_COLOR";
+   case ppir_output_depth:
+      return "OUTPUT_DEPTH";
+   default:
+      return "INVALID";
+   }
+}
+
+static inline ppir_output_type ppir_nir_output_to_ppir(gl_frag_result res)
+{
+   switch (res) {
+   case FRAG_RESULT_COLOR:
+   case FRAG_RESULT_DATA0:
+      return ppir_output_color;
+   case FRAG_RESULT_DEPTH:
+      return ppir_output_depth;
+   default:
+      return ppir_output_invalid;
+   }
+}
+
 typedef struct ppir_reg {
    struct list_head list;
    int index;
+   ppir_output_type out_type;
    int regalloc_index;
    int num_components;
 
@@ -191,6 +224,7 @@ typedef struct ppir_reg {
    bool is_head;
    bool spilled;
    bool undef;
+   bool out_reg;
 } ppir_reg;
 
 typedef enum {
@@ -316,7 +350,7 @@ typedef struct ppir_instr {
 
    ppir_node *slots[PPIR_INSTR_SLOT_NUM];
    ppir_const constant[2];
-   bool is_end;
+   bool stop;
 
    /* for scheduler */
    struct list_head succ_list;
@@ -340,6 +374,7 @@ typedef struct ppir_block {
    struct list_head list;
    struct list_head node_list;
    struct list_head instr_list;
+   bool stop;
 
    struct ppir_block *successors[2];
 
@@ -370,6 +405,7 @@ typedef struct ppir_compiler {
    struct hash_table_u64 *blocks;
    int cur_index;
    int cur_instr_index;
+   int *out_type_to_reg;
 
    struct list_head reg_list;
    int reg_num;
diff --git a/src/gallium/drivers/lima/ir/pp/regalloc.c b/src/gallium/drivers/lima/ir/pp/regalloc.c
index 3ea136b5660..37b54b5a4bf 100644
--- a/src/gallium/drivers/lima/ir/pp/regalloc.c
+++ b/src/gallium/drivers/lima/ir/pp/regalloc.c
@@ -82,9 +82,6 @@ static void ppir_regalloc_update_reglist_ssa(ppir_compiler *comp)
 {
    list_for_each_entry(ppir_block, block, &comp->block_list, list) {
       list_for_each_entry(ppir_node, node, &block->node_list, list) {
-         if (node->is_end)
-            continue;
-
          if (!node->instr || node->op == ppir_op_const)
             continue;
 
@@ -94,6 +91,8 @@ static void ppir_regalloc_update_reglist_ssa(ppir_compiler *comp)
 
             if (dest->type == ppir_target_ssa) {
                reg = &dest->ssa;
+               if (node->is_out)
+                  reg->out_reg = true;
                list_addtail(&reg->list, &comp->reg_list);
                comp->reg_num++;
             }
@@ -133,6 +132,14 @@ static void ppir_regalloc_print_result(ppir_compiler *comp)
       }
    }
    printf("--------------------------\n");
+
+   printf("======ppir output regs======\n");
+   for (int i = 0; i < ppir_output_num; i++) {
+      if (comp->out_type_to_reg[i] != -1)
+         printf("%s: $%d\n", ppir_output_type_to_str(i),
+                (int)comp->out_type_to_reg[i]);
+   }
+   printf("--------------------------\n");
 }
 
 static bool create_new_instr_after(ppir_block *block, ppir_instr *ref,
@@ -578,6 +585,11 @@ static bool ppir_regalloc_prog_try(ppir_compiler *comp, bool *spilled)
    n = 0;
    list_for_each_entry(ppir_reg, reg, &comp->reg_list, list) {
       reg->index = ra_get_node_reg(g, n++);
+      if (reg->out_reg) {
+         /* We need actual reg number, we don't have swizzle for output regs */
+         assert(!(reg->index & 0x3) && "ppir: output regs don't have swizzle");
+         comp->out_type_to_reg[reg->out_type] = reg->index / 4;
+      }
    }
 
    ralloc_free(g);
@@ -604,8 +616,11 @@ bool ppir_regalloc_prog(ppir_compiler *comp)
    ppir_regalloc_update_reglist_ssa(comp);
 
    /* No registers? Probably shader consists of discard instruction */
-   if (list_is_empty(&comp->reg_list))
+   if (list_is_empty(&comp->reg_list)) {
+      comp->prog->state.frag_color_reg = 0;
+      comp->prog->state.frag_depth_reg = -1;
       return true;
+   }
 
    /* this will most likely succeed in the first
     * try, except for very complicated shaders */
@@ -613,5 +628,10 @@ bool ppir_regalloc_prog(ppir_compiler *comp)
       if (!spilled)
          return false;
 
+   comp->prog->state.frag_color_reg =
+      comp->out_type_to_reg[ppir_output_color];
+   comp->prog->state.frag_depth_reg =
+      comp->out_type_to_reg[ppir_output_depth];
+
    return true;
 }
diff --git a/src/gallium/drivers/lima/lima_context.h b/src/gallium/drivers/lima/lima_context.h
index e871d0b952d..bea1ddeb477 100644
--- a/src/gallium/drivers/lima/lima_context.h
+++ b/src/gallium/drivers/lima/lima_context.h
@@ -49,6 +49,8 @@ struct lima_fs_compiled_shader {
    struct {
       int shader_size;
       int stack_size;
+      int frag_color_reg;
+      int frag_depth_reg;
       bool uses_discard;
    } state;
 };
diff --git a/src/gallium/drivers/lima/lima_draw.c b/src/gallium/drivers/lima/lima_draw.c
index 5c9f03b1d1c..889f0192179 100644
--- a/src/gallium/drivers/lima/lima_draw.c
+++ b/src/gallium/drivers/lima/lima_draw.c
@@ -677,6 +677,12 @@ lima_pack_render_state(struct lima_context *ctx, const struct pipe_draw_info *in
    if (!rst->depth_clip_far || ctx->viewport.far == 1.0f)
       render->depth_test |= 0x20; /* don't clip depth far */
 
+   if (fs->state.frag_depth_reg != -1) {
+      render->depth_test |= (fs->state.frag_depth_reg << 6);
+      /* Shader writes depth */
+      render->depth_test |= 0x801;
+   }
+
    ushort far, near;
 
    near = float_to_ushort(ctx->viewport.near);
@@ -729,6 +735,12 @@ lima_pack_render_state(struct lima_context *ctx, const struct pipe_draw_info *in
    if (ctx->framebuffer.base.samples)
       render->multi_sample |= 0x68;
 
+   /* Set gl_FragColor register, need to specify it 4 times */
+   render->multi_sample |= (fs->state.frag_color_reg << 28) |
+                           (fs->state.frag_color_reg << 24) |
+                           (fs->state.frag_color_reg << 20) |
+                           (fs->state.frag_color_reg << 16);
+
    /* alpha test */
    if (ctx->zsa->base.alpha_enabled) {
       render->multi_sample |= ctx->zsa->base.alpha_func;
@@ -755,7 +767,8 @@ lima_pack_render_state(struct lima_context *ctx, const struct pipe_draw_info *in
       render->aux1 |= 0x00002000;
 
    if (fs->state.uses_discard ||
-       ctx->zsa->base.alpha_enabled) {
+       ctx->zsa->base.alpha_enabled ||
+       fs->state.frag_depth_reg != -1) {
       early_z = false;
       pixel_kill = false;
    }
diff --git a/src/gallium/drivers/lima/lima_parser.c b/src/gallium/drivers/lima/lima_parser.c
index 3e69daee1b2..bcacd290aff 100644
--- a/src/gallium/drivers/lima/lima_parser.c
+++ b/src/gallium/drivers/lima/lima_parser.c
@@ -525,7 +525,7 @@ parse_rsw(FILE *fp, uint32_t *value, int i, uint32_t *helper)
          fprintf(fp, ": ignore depth clip near");
       if ((*value & 0x00000020) == 0x00000020)
          fprintf(fp, ", ignore depth clip far");
-      fprintf(fp, ", unknown bits 6-9: 0x%08x", *value & 0x000003c0);
+      fprintf(fp, ", register for gl_FragDepth: $%d", (*value & 0x000003c0) >> 6);
       fprintf(fp, ", unknown bits 13-15: 0x%08x */\n", *value & 0x00000e000);
       break;
    case 4: /* DEPTH RANGE */
@@ -594,7 +594,14 @@ parse_rsw(FILE *fp, uint32_t *value, int i, uint32_t *helper)
          fprintf(fp, " */\n");
       else
          fprintf(fp, ", UNKNOWN\n");
-      fprintf(fp, "\t\t\t\t\t\t/* %s(2)", render_state_infos[i].info);
+
+      fprintf(fp, "\t\t\t\t\t\t/* %s(3)", render_state_infos[i].info);
+      fprintf(fp, ", register for gl_FragColor: $%d $%d $%d $%d */\n",
+              (*value & 0xf0000000) >> 28,
+              (*value & 0x0f000000) >> 24,
+              (*value & 0x00f00000) >> 20,
+              (*value & 0x000f0000) >> 16);
+      fprintf(fp, "\t\t\t\t\t\t/* %s(3)", render_state_infos[i].info);
       fprintf(fp, ": alpha_test_func: %d (%s) */\n",
               (*value & 0x00000007),
               lima_get_compare_func_string((*value & 0x00000007))); /* alpha_test_func */