st/glsl_to_tgsi: make use of driver_cache_blob with the disk cache

driver_cache_blob was introduced with the i965 disk cache, it allows
us to simplify the cache a little and possibly offers some minor
speed improvements since we load the GLSL metadata and TGSI from
disk in one pass.

Using driver_cache_blob should also make it straight forward to
implement binary support for ARB_get_program_binary in gallium.

Reviewed-by: Marek Olšák <[email protected]>

3 files changed, 109 insertions, 230 deletions

diff --git a/src/mesa/main/mtypes.h b/src/mesa/main/mtypes.h
index 9d704c1c158..96775824ea6 100644
--- a/src/mesa/main/mtypes.h
+++ b/src/mesa/main/mtypes.h
@@ -2887,13 +2887,6 @@ struct gl_shader_program_data
    unsigned NumUniformDataSlots;
    union gl_constant_value *UniformDataSlots;
 
-   /* TODO: This used by Gallium drivers to skip the cache on tgsi fallback.
-    * All structures (gl_program, uniform storage, etc) will get recreated
-    * even though we have already loaded them from cache. We should instead
-    * switch to storing the GLSL metadata and TGSI IR in a single cache item
-    * like the i965 driver does with NIR.
-    */
-   bool skip_cache;
    GLboolean Validated;
 
    /** List of all active resources after linking. */
diff --git a/src/mesa/state_tracker/st_program.h b/src/mesa/state_tracker/st_program.h
index 6049fba517b..0e6c8e00c6e 100644
--- a/src/mesa/state_tracker/st_program.h
+++ b/src/mesa/state_tracker/st_program.h
@@ -150,9 +150,6 @@ struct st_fragment_program
    struct gl_shader_program *shader_program;
 
    struct st_fp_variant *variants;
-
-   /** SHA1 hash of linked tgsi shader program, used for on-disk cache */
-   unsigned char sha1[20];
 };
 
 
diff --git a/src/mesa/state_tracker/st_shader_cache.c b/src/mesa/state_tracker/st_shader_cache.c
index 2e9b7cc6708..a5e33133b3c 100644
--- a/src/mesa/state_tracker/st_shader_cache.c
+++ b/src/mesa/state_tracker/st_shader_cache.c
@@ -40,14 +40,15 @@ write_stream_out_to_cache(struct blob *blob,
 
 static void
 write_tgsi_to_cache(struct blob *blob, struct pipe_shader_state *tgsi,
-                    struct st_context *st, unsigned char *sha1,
-                    unsigned num_tokens)
+                    struct gl_program *prog, unsigned num_tokens)
 {
    blob_write_uint32(blob, num_tokens);
    blob_write_bytes(blob, tgsi->tokens,
                     num_tokens * sizeof(struct tgsi_token));
 
-   disk_cache_put(st->ctx->Cache, sha1, blob->data, blob->size, NULL);
+   prog->driver_cache_blob = ralloc_size(NULL, blob->size);
+   memcpy(prog->driver_cache_blob, blob->data, blob->size);
+   prog->driver_cache_blob_size = blob->size;
 }
 
 /**
@@ -68,14 +69,12 @@ st_store_tgsi_in_disk_cache(struct st_context *st, struct gl_program *prog,
    if (memcmp(prog->sh.data->sha1, zero, sizeof(prog->sh.data->sha1)) == 0)
       return;
 
-   unsigned char *sha1;
    struct blob blob;
    blob_init(&blob);
 
    switch (prog->info.stage) {
    case MESA_SHADER_VERTEX: {
       struct st_vertex_program *stvp = (struct st_vertex_program *) prog;
-      sha1 = stvp->sha1;
 
       blob_write_uint32(&blob, stvp->num_inputs);
       blob_write_bytes(&blob, stvp->index_to_input,
@@ -84,31 +83,24 @@ st_store_tgsi_in_disk_cache(struct st_context *st, struct gl_program *prog,
                        sizeof(stvp->result_to_output));
 
       write_stream_out_to_cache(&blob, &stvp->tgsi);
-      write_tgsi_to_cache(&blob, &stvp->tgsi, st, sha1, num_tokens);
+      write_tgsi_to_cache(&blob, &stvp->tgsi, prog, num_tokens);
       break;
    }
    case MESA_SHADER_TESS_CTRL:
    case MESA_SHADER_TESS_EVAL:
    case MESA_SHADER_GEOMETRY: {
-      struct st_common_program *p = st_common_program(prog);
-      sha1 = p->sha1;
-
       write_stream_out_to_cache(&blob, out_state);
-      write_tgsi_to_cache(&blob, out_state, st, sha1, num_tokens);
+      write_tgsi_to_cache(&blob, out_state, prog, num_tokens);
       break;
    }
    case MESA_SHADER_FRAGMENT: {
       struct st_fragment_program *stfp = (struct st_fragment_program *) prog;
-      sha1 = stfp->sha1;
 
-      write_tgsi_to_cache(&blob, &stfp->tgsi, st, sha1, num_tokens);
+      write_tgsi_to_cache(&blob, &stfp->tgsi, prog, num_tokens);
       break;
    }
    case MESA_SHADER_COMPUTE: {
-      struct st_compute_program *stcp = (struct st_compute_program *) prog;
-      sha1 = stcp->sha1;
-
-      write_tgsi_to_cache(&blob, out_state, st, sha1, num_tokens);
+      write_tgsi_to_cache(&blob, out_state, prog, num_tokens);
       break;
    }
    default:
@@ -116,10 +108,8 @@ st_store_tgsi_in_disk_cache(struct st_context *st, struct gl_program *prog,
    }
 
    if (st->ctx->_Shader->Flags & GLSL_CACHE_INFO) {
-      char sha1_buf[41];
-      _mesa_sha1_format(sha1_buf, sha1);
-      fprintf(stderr, "putting %s tgsi_tokens in cache: %s\n",
-              _mesa_shader_stage_to_string(prog->info.stage), sha1_buf);
+      fprintf(stderr, "putting %s tgsi_tokens in cache\n",
+              _mesa_shader_stage_to_string(prog->info.stage));
    }
 
    blob_finish(&blob);
@@ -150,254 +140,153 @@ st_load_tgsi_from_disk_cache(struct gl_context *ctx,
    if (!ctx->Cache)
       return false;
 
-   unsigned char *stage_sha1[MESA_SHADER_STAGES];
-   char sha1_buf[41];
-
-   /* Compute and store sha1 for each stage. These will be reused by the
-    * cache store pass if we fail to find the cached tgsi.
-    */
-   for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
-      if (prog->_LinkedShaders[i] == NULL)
-         continue;
-
-      char *buf = ralloc_strdup(NULL, "tgsi_tokens ");
-      _mesa_sha1_format(sha1_buf,
-                        prog->_LinkedShaders[i]->Program->sh.data->sha1);
-      ralloc_strcat(&buf, sha1_buf);
-
-      struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
-      switch (glprog->info.stage) {
-      case MESA_SHADER_VERTEX: {
-         struct st_vertex_program *stvp = (struct st_vertex_program *) glprog;
-         stage_sha1[i] = stvp->sha1;
-         ralloc_strcat(&buf, " vs");
-         disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
-         break;
-      }
-      case MESA_SHADER_TESS_CTRL: {
-         struct st_common_program *stcp = st_common_program(glprog);
-         stage_sha1[i] = stcp->sha1;
-         ralloc_strcat(&buf, " tcs");
-         disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
-         break;
-      }
-      case MESA_SHADER_TESS_EVAL: {
-         struct st_common_program *step = st_common_program(glprog);
-         stage_sha1[i] = step->sha1;
-         ralloc_strcat(&buf, " tes");
-         disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
-         break;
-      }
-      case MESA_SHADER_GEOMETRY: {
-         struct st_common_program *stgp = st_common_program(glprog);
-         stage_sha1[i] = stgp->sha1;
-         ralloc_strcat(&buf, " gs");
-         disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
-         break;
-      }
-      case MESA_SHADER_FRAGMENT: {
-         struct st_fragment_program *stfp =
-            (struct st_fragment_program *) glprog;
-         stage_sha1[i] = stfp->sha1;
-         ralloc_strcat(&buf, " fs");
-         disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
-         break;
-      }
-      case MESA_SHADER_COMPUTE: {
-         struct st_compute_program *stcp =
-            (struct st_compute_program *) glprog;
-         stage_sha1[i] = stcp->sha1;
-         ralloc_strcat(&buf, " cs");
-         disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
-         break;
-      }
-      default:
-         unreachable("Unsupported stage");
-      }
-
-      ralloc_free(buf);
-   }
-
-   /* Now that we have created the sha1 keys that will be used for writting to
-    * the tgsi cache fallback to the regular glsl to tgsi path if we didn't
-    * load the GLSL IR from cache. We do this as glsl to tgsi can alter things
-    * such as gl_program_parameter_list which holds things like uniforms.
+   /* If we didn't load the GLSL metadata from cache then we could not have
+    * loaded the tgsi either.
     */
    if (prog->data->LinkStatus != linking_skipped)
       return false;
 
-   uint8_t *buffer = NULL;
-   if (ctx->_Shader->Flags & GLSL_CACHE_FALLBACK) {
-      goto fallback_recompile;
-   }
-
    struct st_context *st = st_context(ctx);
    for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
       if (prog->_LinkedShaders[i] == NULL)
          continue;
 
-      unsigned char *sha1 = stage_sha1[i];
-      size_t size;
-      buffer = (uint8_t *) disk_cache_get(ctx->Cache, sha1, &size);
-      if (buffer) {
-         struct blob_reader blob_reader;
-         blob_reader_init(&blob_reader, buffer, size);
-
-         struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
-         switch (glprog->info.stage) {
-         case MESA_SHADER_VERTEX: {
-            struct st_vertex_program *stvp =
-               (struct st_vertex_program *) glprog;
-
-            st_release_vp_variants(st, stvp);
-
-            stvp->num_inputs = blob_read_uint32(&blob_reader);
-            blob_copy_bytes(&blob_reader, (uint8_t *) stvp->index_to_input,
-                            sizeof(stvp->index_to_input));
-            blob_copy_bytes(&blob_reader, (uint8_t *) stvp->result_to_output,
-                            sizeof(stvp->result_to_output));
-
-            read_stream_out_from_cache(&blob_reader, &stvp->tgsi);
-            read_tgsi_from_cache(&blob_reader, &stvp->tgsi.tokens);
+      struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
 
-            if (st->vp == stvp)
-               st->dirty |= ST_NEW_VERTEX_PROGRAM(st, stvp);
+      size_t size = glprog->driver_cache_blob_size;
+      uint8_t *buffer = (uint8_t *) glprog->driver_cache_blob;
 
-            break;
-         }
-         case MESA_SHADER_TESS_CTRL: {
-            struct st_common_program *sttcp = st_common_program(glprog);
+      struct blob_reader blob_reader;
+      blob_reader_init(&blob_reader, buffer, size);
 
-            st_release_basic_variants(st, sttcp->Base.Target,
-                                      &sttcp->variants, &sttcp->tgsi);
+      switch (glprog->info.stage) {
+      case MESA_SHADER_VERTEX: {
+         struct st_vertex_program *stvp = (struct st_vertex_program *) glprog;
 
-            read_stream_out_from_cache(&blob_reader, &sttcp->tgsi);
-            read_tgsi_from_cache(&blob_reader, &sttcp->tgsi.tokens);
+         st_release_vp_variants(st, stvp);
 
-            if (st->tcp == sttcp)
-               st->dirty |= sttcp->affected_states;
+         stvp->num_inputs = blob_read_uint32(&blob_reader);
+         blob_copy_bytes(&blob_reader, (uint8_t *) stvp->index_to_input,
+                         sizeof(stvp->index_to_input));
+         blob_copy_bytes(&blob_reader, (uint8_t *) stvp->result_to_output,
+                         sizeof(stvp->result_to_output));
 
-            break;
-         }
-         case MESA_SHADER_TESS_EVAL: {
-            struct st_common_program *sttep = st_common_program(glprog);
+         read_stream_out_from_cache(&blob_reader, &stvp->tgsi);
+         read_tgsi_from_cache(&blob_reader, &stvp->tgsi.tokens);
 
-            st_release_basic_variants(st, sttep->Base.Target,
-                                      &sttep->variants, &sttep->tgsi);
+         if (st->vp == stvp)
+            st->dirty |= ST_NEW_VERTEX_PROGRAM(st, stvp);
 
-            read_stream_out_from_cache(&blob_reader, &sttep->tgsi);
-            read_tgsi_from_cache(&blob_reader, &sttep->tgsi.tokens);
+         break;
+      }
+      case MESA_SHADER_TESS_CTRL: {
+         struct st_common_program *sttcp = st_common_program(glprog);
 
-            if (st->tep == sttep)
-               st->dirty |= sttep->affected_states;
+         st_release_basic_variants(st, sttcp->Base.Target,
+                                   &sttcp->variants, &sttcp->tgsi);
 
-            break;
-         }
-         case MESA_SHADER_GEOMETRY: {
-            struct st_common_program *stgp = st_common_program(glprog);
+         read_stream_out_from_cache(&blob_reader, &sttcp->tgsi);
+         read_tgsi_from_cache(&blob_reader, &sttcp->tgsi.tokens);
 
-            st_release_basic_variants(st, stgp->Base.Target, &stgp->variants,
-                                      &stgp->tgsi);
+         if (st->tcp == sttcp)
+            st->dirty |= sttcp->affected_states;
 
-            read_stream_out_from_cache(&blob_reader, &stgp->tgsi);
-            read_tgsi_from_cache(&blob_reader, &stgp->tgsi.tokens);
+         break;
+      }
+      case MESA_SHADER_TESS_EVAL: {
+         struct st_common_program *sttep = st_common_program(glprog);
 
-            if (st->gp == stgp)
-               st->dirty |= stgp->affected_states;
+         st_release_basic_variants(st, sttep->Base.Target,
+                                   &sttep->variants, &sttep->tgsi);
 
-            break;
-         }
-         case MESA_SHADER_FRAGMENT: {
-            struct st_fragment_program *stfp =
-               (struct st_fragment_program *) glprog;
+         read_stream_out_from_cache(&blob_reader, &sttep->tgsi);
+         read_tgsi_from_cache(&blob_reader, &sttep->tgsi.tokens);
 
-            st_release_fp_variants(st, stfp);
+         if (st->tep == sttep)
+            st->dirty |= sttep->affected_states;
 
-            read_tgsi_from_cache(&blob_reader, &stfp->tgsi.tokens);
+         break;
+      }
+      case MESA_SHADER_GEOMETRY: {
+         struct st_common_program *stgp = st_common_program(glprog);
 
-            if (st->fp == stfp)
-               st->dirty |= stfp->affected_states;
+         st_release_basic_variants(st, stgp->Base.Target, &stgp->variants,
+                                   &stgp->tgsi);
 
-            break;
-         }
-         case MESA_SHADER_COMPUTE: {
-            struct st_compute_program *stcp =
-               (struct st_compute_program *) glprog;
+         read_stream_out_from_cache(&blob_reader, &stgp->tgsi);
+         read_tgsi_from_cache(&blob_reader, &stgp->tgsi.tokens);
 
-            st_release_cp_variants(st, stcp);
+         if (st->gp == stgp)
+            st->dirty |= stgp->affected_states;
 
-            read_tgsi_from_cache(&blob_reader,
-                                 (const struct tgsi_token**) &stcp->tgsi.prog);
+         break;
+      }
+      case MESA_SHADER_FRAGMENT: {
+         struct st_fragment_program *stfp =
+            (struct st_fragment_program *) glprog;
 
-            stcp->tgsi.req_local_mem = stcp->Base.info.cs.shared_size;
-            stcp->tgsi.req_private_mem = 0;
-            stcp->tgsi.req_input_mem = 0;
+         st_release_fp_variants(st, stfp);
 
-            if (st->cp == stcp)
-                st->dirty |= stcp->affected_states;
+         read_tgsi_from_cache(&blob_reader, &stfp->tgsi.tokens);
 
-            break;
-         }
-         default:
-            unreachable("Unsupported stage");
-         }
+         if (st->fp == stfp)
+            st->dirty |= stfp->affected_states;
 
-         if (blob_reader.current != blob_reader.end || blob_reader.overrun) {
-            /* Something very bad has gone wrong discard the item from the
-             * cache and rebuild/link from source.
-             */
-            assert(!"Invalid TGSI shader disk cache item!");
+         break;
+      }
+      case MESA_SHADER_COMPUTE: {
+         struct st_compute_program *stcp =
+            (struct st_compute_program *) glprog;
 
-            if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
-               fprintf(stderr, "Error reading program from cache (invalid "
-                       "TGSI cache item)\n");
-            }
+         st_release_cp_variants(st, stcp);
 
-            disk_cache_remove(ctx->Cache, sha1);
+         read_tgsi_from_cache(&blob_reader,
+                              (const struct tgsi_token**) &stcp->tgsi.prog);
 
-            goto fallback_recompile;
-         }
+         stcp->tgsi.req_local_mem = stcp->Base.info.cs.shared_size;
+         stcp->tgsi.req_private_mem = 0;
+         stcp->tgsi.req_input_mem = 0;
 
-         if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
-            _mesa_sha1_format(sha1_buf, sha1);
-            fprintf(stderr, "%s tgsi_tokens retrieved from cache: %s\n",
-                    _mesa_shader_stage_to_string(i), sha1_buf);
-         }
+         if (st->cp == stcp)
+            st->dirty |= stcp->affected_states;
 
-         st_set_prog_affected_state_flags(glprog);
-         _mesa_associate_uniform_storage(ctx, prog, glprog, false);
+         break;
+      }
+      default:
+         unreachable("Unsupported stage");
+      }
 
-         /* Create Gallium shaders now instead of on demand. */
-         if (ST_DEBUG & DEBUG_PRECOMPILE ||
-             st->shader_has_one_variant[glprog->info.stage])
-            st_precompile_shader_variant(st, glprog);
+      /* Make sure we don't try to read more data than we wrote. This should
+       * never happen in release builds but its useful to have this check to
+       * catch development bugs.
+       */
+      if (blob_reader.current != blob_reader.end || blob_reader.overrun) {
+         assert(!"Invalid TGSI shader disk cache item!");
 
-         free(buffer);
-      } else {
-         /* Failed to find a matching cached shader so fallback to recompile.
-          */
          if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
-            fprintf(stderr, "TGSI cache item not found.\n");
+            fprintf(stderr, "Error reading program from cache (invalid "
+                    "TGSI cache item)\n");
          }
-
-         goto fallback_recompile;
       }
-   }
 
-   return true;
+      if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
+         fprintf(stderr, "%s tgsi_tokens retrieved from cache\n",
+                 _mesa_shader_stage_to_string(i));
+      }
 
-fallback_recompile:
-   free(buffer);
+      st_set_prog_affected_state_flags(glprog);
+      _mesa_associate_uniform_storage(ctx, prog, glprog, false);
 
-   if (ctx->_Shader->Flags & GLSL_CACHE_INFO)
-      fprintf(stderr, "TGSI cache falling back to recompile.\n");
+      /* Create Gallium shaders now instead of on demand. */
+      if (ST_DEBUG & DEBUG_PRECOMPILE ||
+          st->shader_has_one_variant[glprog->info.stage])
+         st_precompile_shader_variant(st, glprog);
 
-   for (unsigned i = 0; i < prog->NumShaders; i++) {
-      _mesa_glsl_compile_shader(ctx, prog->Shaders[i], false, false, true);
+      /* We don't need the cached blob anymore so free it */
+      ralloc_free(glprog->driver_cache_blob);
+      glprog->driver_cache_blob = NULL;
+      glprog->driver_cache_blob_size = 0;
    }
 
-   prog->data->skip_cache = true;
-   _mesa_glsl_link_shader(ctx, prog);
-
    return true;
 }