tgsi: provide a way to encode memory qualifiers for SSBO

Each load/store on most hardware can specify what caching to do. Since
SSBO allows individual variables to also have separate caching modes,
allow loads/stores to have the qualifiers instead of attempting to
encode them in declarations.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

src/gallium/auxiliary/tgsi/tgsi_build.c

@@ -620,6 +620,7 @@ tgsi_default_instruction( void )
    instruction.NumSrcRegs = 1;
    instruction.Label = 0;
    instruction.Texture = 0;
+   instruction.Memory = 0;
    instruction.Padding = 0;
 
    return instruction;
@@ -766,6 +767,34 @@ tgsi_build_instruction_texture(
    return instruction_texture;
 }
 
+static struct tgsi_instruction_memory
+tgsi_default_instruction_memory( void )
+{
+   struct tgsi_instruction_memory instruction_memory;
+
+   instruction_memory.Qualifier = 0;
+   instruction_memory.Padding = 0;
+
+   return instruction_memory;
+}
+
+static struct tgsi_instruction_memory
+tgsi_build_instruction_memory(
+   unsigned qualifier,
+   struct tgsi_token *prev_token,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header )
+{
+   struct tgsi_instruction_memory instruction_memory;
+
+   instruction_memory.Qualifier = qualifier;
+   instruction_memory.Padding = 0;
+   instruction->Memory = 1;
+
+   instruction_grow( instruction, header );
+
+   return instruction_memory;
+}
 
 static struct tgsi_texture_offset
 tgsi_default_texture_offset( void )
@@ -1012,6 +1041,7 @@ tgsi_default_full_instruction( void )
    full_instruction.Predicate = tgsi_default_instruction_predicate();
    full_instruction.Label = tgsi_default_instruction_label();
    full_instruction.Texture = tgsi_default_instruction_texture();
+   full_instruction.Memory = tgsi_default_instruction_memory();
    for( i = 0;  i < TGSI_FULL_MAX_TEX_OFFSETS; i++ ) {
       full_instruction.TexOffsets[i] = tgsi_default_texture_offset();
    }
@@ -1123,6 +1153,24 @@ tgsi_build_full_instruction(
          prev_token = (struct tgsi_token *) texture_offset;
       }
    }
+
+   if (full_inst->Instruction.Memory) {
+      struct tgsi_instruction_memory *instruction_memory;
+
+      if( maxsize <= size )
+         return 0;
+      instruction_memory =
+         (struct  tgsi_instruction_memory *) &tokens[size];
+      size++;
+
+      *instruction_memory = tgsi_build_instruction_memory(
+         full_inst->Memory.Qualifier,
+         prev_token,
+         instruction,
+         header );
+      prev_token = (struct tgsi_token  *) instruction_memory;
+   }
+
    for( i = 0;  i <   full_inst->Instruction.NumDstRegs; i++ ) {
       const struct tgsi_full_dst_register *reg = &full_inst->Dst[i];
       struct tgsi_dst_register *dst_register;

src/gallium/auxiliary/tgsi/tgsi_dump.c

@@ -624,6 +624,16 @@ iter_instruction(
       }
    }
 
+   if (inst->Instruction.Memory) {
+      uint32_t qualifier = inst->Memory.Qualifier;
+      while (qualifier) {
+         int bit = ffs(qualifier) - 1;
+         qualifier &= ~(1U << bit);
+         TXT(", ");
+         ENM(bit, tgsi_memory_names);
+      }
+   }
+
    switch (inst->Instruction.Opcode) {
    case TGSI_OPCODE_IF:
    case TGSI_OPCODE_UIF:

src/gallium/auxiliary/tgsi/tgsi_parse.c

@@ -195,6 +195,10 @@ tgsi_parse_token(
          }
       }
 
+      if (inst->Instruction.Memory) {
+         next_token(ctx, &inst->Memory);
+      }
+
       assert( inst->Instruction.NumDstRegs <= TGSI_FULL_MAX_DST_REGISTERS );
 
       for (i = 0; i < inst->Instruction.NumDstRegs; i++) {

src/gallium/auxiliary/tgsi/tgsi_parse.h

@@ -91,6 +91,7 @@ struct tgsi_full_instruction
    struct tgsi_instruction_predicate   Predicate;
    struct tgsi_instruction_label       Label;
    struct tgsi_instruction_texture     Texture;
+   struct tgsi_instruction_memory      Memory;
    struct tgsi_full_dst_register       Dst[TGSI_FULL_MAX_DST_REGISTERS];
    struct tgsi_full_src_register       Src[TGSI_FULL_MAX_SRC_REGISTERS];
    struct tgsi_texture_offset          TexOffsets[TGSI_FULL_MAX_TEX_OFFSETS];

src/gallium/auxiliary/tgsi/tgsi_strings.c

@@ -208,6 +208,13 @@ const char *tgsi_immediate_type_names[4] =
    "FLT64"
 };
 
+const char *tgsi_memory_names[3] =
+{
+   "COHERENT",
+   "RESTRICT",
+   "VOLATILE",
+};
+
 
 static inline void
 tgsi_strings_check(void)

src/gallium/auxiliary/tgsi/tgsi_strings.h

@@ -60,6 +60,8 @@ extern const char *tgsi_fs_coord_pixel_center_names[2];
 
 extern const char *tgsi_immediate_type_names[4];
 
+extern const char *tgsi_memory_names[3];
+
 
 const char *
 tgsi_file_name(unsigned file);

src/gallium/auxiliary/tgsi/tgsi_text.c

@@ -1039,6 +1039,12 @@ parse_instruction(
       inst.Texture.Texture = TGSI_TEXTURE_UNKNOWN;
    }
 
+   if ((i >= TGSI_OPCODE_LOAD && i <= TGSI_OPCODE_ATOMIMAX) ||
+       i == TGSI_OPCODE_RESQ) {
+      inst.Instruction.Memory = 1;
+      inst.Memory.Qualifier = 0;
+   }
+
    /* Parse instruction operands.
     */
    for (i = 0; i < info->num_dst + info->num_src + info->is_tex; i++) {
@@ -1090,6 +1096,27 @@ parse_instruction(
    }
    inst.Texture.NumOffsets = i;
 
+   cur = ctx->cur;
+   eat_opt_white(&cur);
+   for (i = 0; inst.Instruction.Memory && *cur == ','; i++) {
+      uint j;
+      cur++;
+      eat_opt_white(&cur);
+      ctx->cur = cur;
+      for (j = 0; j < 3; j++) {
+         if (str_match_nocase_whole(&ctx->cur, tgsi_memory_names[j])) {
+            inst.Memory.Qualifier |= 1U << j;
+            break;
+         }
+      }
+      if (j == 3) {
+         report_error(ctx, "Expected memory qualifier");
+         return FALSE;
+      }
+      cur = ctx->cur;
+      eat_opt_white(&cur);
+   }
+
    cur = ctx->cur;
    eat_opt_white( &cur );
    if (info->is_branch && *cur == ':') {

src/gallium/auxiliary/tgsi/tgsi_ureg.c

@@ -60,6 +60,7 @@ union tgsi_any_token {
    struct tgsi_instruction_predicate insn_predicate;
    struct tgsi_instruction_label insn_label;
    struct tgsi_instruction_texture insn_texture;
+   struct tgsi_instruction_memory insn_memory;
    struct tgsi_texture_offset insn_texture_offset;
    struct tgsi_src_register src;
    struct tgsi_ind_register ind;
@@ -1226,6 +1227,21 @@ ureg_emit_texture_offset(struct ureg_program *ureg,
    
 }
 
+void
+ureg_emit_memory(struct ureg_program *ureg,
+                 unsigned extended_token,
+                 unsigned qualifier)
+{
+   union tgsi_any_token *out, *insn;
+
+   out = get_tokens( ureg, DOMAIN_INSN, 1 );
+   insn = retrieve_token( ureg, DOMAIN_INSN, extended_token );
+
+   insn->insn.Memory = 1;
+
+   out[0].value = 0;
+   out[0].insn_memory.Qualifier = qualifier;
+}
 
 void
 ureg_fixup_insn_size(struct ureg_program *ureg,
@@ -1378,6 +1394,42 @@ ureg_label_insn(struct ureg_program *ureg,
 }
 
 
+void
+ureg_memory_insn(struct ureg_program *ureg,
+                 unsigned opcode,
+                 const struct ureg_dst *dst,
+                 unsigned nr_dst,
+                 const struct ureg_src *src,
+                 unsigned nr_src,
+                 unsigned qualifier)
+{
+   struct ureg_emit_insn_result insn;
+   unsigned i;
+
+   insn = ureg_emit_insn(ureg,
+                         opcode,
+                         FALSE,
+                         FALSE,
+                         FALSE,
+                         TGSI_SWIZZLE_X,
+                         TGSI_SWIZZLE_Y,
+                         TGSI_SWIZZLE_Z,
+                         TGSI_SWIZZLE_W,
+                         nr_dst,
+                         nr_src);
+
+   ureg_emit_memory(ureg, insn.extended_token, qualifier);
+
+   for (i = 0; i < nr_dst; i++)
+      ureg_emit_dst(ureg, dst[i]);
+
+   for (i = 0; i < nr_src; i++)
+      ureg_emit_src(ureg, src[i]);
+
+   ureg_fixup_insn_size(ureg, insn.insn_token);
+}
+
+
 static void
 emit_decl_semantic(struct ureg_program *ureg,
                    unsigned file,

src/gallium/auxiliary/tgsi/tgsi_ureg.h

@@ -531,6 +531,14 @@ ureg_label_insn(struct ureg_program *ureg,
                 unsigned nr_src,
                 unsigned *label);
 
+void
+ureg_memory_insn(struct ureg_program *ureg,
+                 unsigned opcode,
+                 const struct ureg_dst *dst,
+                 unsigned nr_dst,
+                 const struct ureg_src *src,
+                 unsigned nr_src,
+                 unsigned qualifier);
 
 /***********************************************************************
  * Internal instruction helpers, don't call these directly:
@@ -568,6 +576,11 @@ void
 ureg_emit_texture_offset(struct ureg_program *ureg,
                          const struct tgsi_texture_offset *offset);
 
+void
+ureg_emit_memory(struct ureg_program *ureg,
+                 unsigned insn_token,
+                 unsigned qualifier);
+
 void 
 ureg_emit_dst( struct ureg_program *ureg,
                struct ureg_dst dst );

src/gallium/include/pipe/p_shader_tokens.h

@@ -572,7 +572,8 @@ struct tgsi_instruction
    unsigned Predicate  : 1;  /* BOOL */
    unsigned Label      : 1;
    unsigned Texture    : 1;
-   unsigned Padding    : 2;
+   unsigned Memory     : 1;
+   unsigned Padding    : 1;
 };
 
 /*
@@ -729,6 +730,19 @@ struct tgsi_dst_register
    unsigned Padding     : 6;
 };
 
+#define TGSI_MEMORY_COHERENT (1 << 0)
+#define TGSI_MEMORY_RESTRICT (1 << 1)
+#define TGSI_MEMORY_VOLATILE (1 << 2)
+
+/**
+ * Specifies the type of memory access to do for the LOAD/STORE instruction.
+ */
+struct tgsi_instruction_memory
+{
+   unsigned Qualifier : 3;  /* TGSI_MEMORY_ */
+   unsigned Padding   : 29;
+};
+
 
 #ifdef __cplusplus
 }