gitlab-ci: Add ppc64el and s390x cross-build jobs
[mesa.git] / src / compiler / nir / nir_lower_clip.c
index 62540ac11d937868b78299a242c29ac4d1b44d01..0f0d2eb2c73d3e182609285274d487916a0d533a 100644 (file)
@@ -31,7 +31,7 @@
 
 /* Generates the lowering code for user-clip-planes, generating CLIPDIST
  * from UCP[n] + CLIPVERTEX or POSITION.  Additionally, an optional pass
- * for fragment shaders to insert conditional kill's based on the inter-
+ * for fragment shaders to insert conditional kills based on the inter-
  * polated CLIPDIST
  *
  * NOTE: should be run after nir_lower_outputs_to_temporaries() (or at
 
 static nir_variable *
 create_clipdist_var(nir_shader *shader, unsigned drvloc,
-                    bool output, gl_varying_slot slot)
+                    bool output, gl_varying_slot slot, unsigned array_size)
 {
    nir_variable *var = rzalloc(shader, nir_variable);
 
    var->data.driver_location = drvloc;
-   var->type = glsl_vec4_type();
    var->data.mode = output ? nir_var_shader_out : nir_var_shader_in;
    var->name = ralloc_asprintf(var, "clipdist_%d", drvloc);
    var->data.index = 0;
    var->data.location = slot;
 
+   if (array_size > 0) {
+      var->type = glsl_array_type(glsl_float_type(), array_size,
+                                  sizeof(float));
+      var->data.compact = 1;
+   } else
+      var->type = glsl_vec4_type();
+
    if (output) {
       exec_list_push_tail(&shader->outputs, &var->node);
       shader->num_outputs++; /* TODO use type_size() */
@@ -64,6 +70,28 @@ create_clipdist_var(nir_shader *shader, unsigned drvloc,
    return var;
 }
 
+static void
+create_clipdist_vars(nir_shader *shader, nir_variable **io_vars,
+                     unsigned ucp_enables, int *drvloc, bool output,
+                     bool use_clipdist_array)
+{
+   if (use_clipdist_array) {
+      io_vars[0] =
+         create_clipdist_var(shader, ++(*drvloc), true,
+                             VARYING_SLOT_CLIP_DIST0,
+                             util_last_bit(ucp_enables));
+   } else {
+      if (ucp_enables & 0x0f)
+         io_vars[0] =
+            create_clipdist_var(shader, ++(*drvloc), output,
+                                VARYING_SLOT_CLIP_DIST0, 0);
+      if (ucp_enables & 0xf0)
+         io_vars[1] =
+            create_clipdist_var(shader, ++(*drvloc), output,
+                                VARYING_SLOT_CLIP_DIST1, 0);
+   }
+}
+
 static void
 store_clipdist_output(nir_builder *b, nir_variable *out, nir_ssa_def **val)
 {
@@ -107,7 +135,7 @@ find_output_in_block(nir_block *block, unsigned drvloc)
          if ((intr->intrinsic == nir_intrinsic_store_output) &&
              nir_intrinsic_base(intr) == drvloc) {
             assert(intr->src[0].is_ssa);
-            assert(nir_src_as_const_value(intr->src[1]));
+            assert(nir_src_is_const(intr->src[1]));
             return intr->src[0].ssa;
          }
       }
@@ -146,67 +174,148 @@ find_output(nir_shader *shader, unsigned drvloc)
    return def;
 }
 
-/*
- * VS lowering
- */
+static bool
+find_clipvertex_and_position_outputs(nir_shader *shader,
+                                     nir_variable **clipvertex,
+                                     nir_variable **position)
+{
+   nir_foreach_variable(var, &shader->outputs) {
+      switch (var->data.location) {
+      case VARYING_SLOT_POS:
+         *position = var;
+         break;
+      case VARYING_SLOT_CLIP_VERTEX:
+         *clipvertex = var;
+         break;
+      case VARYING_SLOT_CLIP_DIST0:
+      case VARYING_SLOT_CLIP_DIST1:
+         /* if shader is already writing CLIPDIST, then
+          * there should be no user-clip-planes to deal
+          * with.
+          *
+          * We assume nir_remove_dead_variables has removed the clipdist
+          * variables if they're not written.
+          */
+         return false;
+      }
+   }
+
+   return *clipvertex || *position;
+}
+
+static nir_ssa_def *
+get_ucp(nir_builder *b, int plane,
+        const gl_state_index16 clipplane_state_tokens[][STATE_LENGTH])
+{
+   if (clipplane_state_tokens) {
+      char tmp[100];
+      snprintf(tmp, ARRAY_SIZE(tmp), "gl_ClipPlane%dMESA", plane);
+      nir_variable *var = nir_variable_create(b->shader,
+                                              nir_var_uniform,
+                                              glsl_vec4_type(),
+                                              tmp);
+
+      var->num_state_slots = 1;
+      var->state_slots = ralloc_array(var, nir_state_slot, 1);
+      memcpy(var->state_slots[0].tokens,
+             clipplane_state_tokens[plane],
+             sizeof(var->state_slots[0].tokens));
+      return nir_load_var(b, var);
+   } else
+      return nir_load_user_clip_plane(b, plane);
+}
+
 
 static void
-lower_clip_vs(nir_function_impl *impl, unsigned ucp_enables,
-              nir_ssa_def *cv, nir_variable **out)
+lower_clip_outputs(nir_builder *b, nir_variable *position,
+                   nir_variable *clipvertex, nir_variable **out,
+                   unsigned ucp_enables, bool use_vars,
+                   bool use_clipdist_array,
+                   const gl_state_index16 clipplane_state_tokens[][STATE_LENGTH])
 {
    nir_ssa_def *clipdist[MAX_CLIP_PLANES];
-   nir_builder b;
+   nir_ssa_def *cv;
 
-   nir_builder_init(&b, impl);
+   if (use_vars) {
+      cv = nir_load_var(b, clipvertex ? clipvertex : position);
 
-   /* NIR should ensure that, even in case of loops/if-else, there
-    * should be only a single predecessor block to end_block, which
-    * makes the perfect place to insert the clipdist calculations.
-    *
-    * NOTE: in case of early return's, these would have to be lowered
-    * to jumps to end_block predecessor in a previous pass.  Not sure
-    * if there is a good way to sanity check this, but for now the
-    * users of this pass don't support sub-routines.
-    */
-   assert(impl->end_block->predecessors->entries == 1);
-   b.cursor = nir_after_cf_list(&impl->body);
+      if (clipvertex) {
+         exec_node_remove(&clipvertex->node);
+         clipvertex->data.mode = nir_var_shader_temp;
+         exec_list_push_tail(&b->shader->globals, &clipvertex->node);
+         nir_fixup_deref_modes(b->shader);
+      }
+   } else {
+      if (clipvertex)
+         cv = find_output(b->shader, clipvertex->data.driver_location);
+      else {
+         assert(position);
+         cv = find_output(b->shader, position->data.driver_location);
+      }
+   }
 
    for (int plane = 0; plane < MAX_CLIP_PLANES; plane++) {
       if (ucp_enables & (1 << plane)) {
-         nir_ssa_def *ucp =
-            nir_load_system_value(&b, nir_intrinsic_load_user_clip_plane, plane);
+         nir_ssa_def *ucp = get_ucp(b, plane, clipplane_state_tokens);
 
          /* calculate clipdist[plane] - dot(ucp, cv): */
-         clipdist[plane] = nir_fdot4(&b, ucp, cv);
-      }
-      else {
+         clipdist[plane] = nir_fdot4(b, ucp, cv);
+      } else {
          /* 0.0 == don't-clip == disabled: */
-         clipdist[plane] = nir_imm_float(&b, 0.0);
+         clipdist[plane] = nir_imm_float(b, 0.0);
+      }
+      if (use_clipdist_array && plane < util_last_bit(ucp_enables)) {
+         assert(use_vars);
+         nir_deref_instr *deref;
+         deref = nir_build_deref_array_imm(b,
+                                           nir_build_deref_var(b, out[0]),
+                                           plane);
+         nir_store_deref(b, deref, clipdist[plane], 1);
       }
    }
 
-   if (ucp_enables & 0x0f)
-      store_clipdist_output(&b, out[0], &clipdist[0]);
-   if (ucp_enables & 0xf0)
-      store_clipdist_output(&b, out[1], &clipdist[4]);
-
-   nir_metadata_preserve(impl, nir_metadata_dominance);
+   if (!use_clipdist_array) {
+      if (use_vars) {
+         if (ucp_enables & 0x0f)
+            nir_store_var(b, out[0], nir_vec(b, clipdist, 4), 0xf);
+         if (ucp_enables & 0xf0)
+            nir_store_var(b, out[1], nir_vec(b, &clipdist[4], 4), 0xf);
+      } else {
+         if (ucp_enables & 0x0f)
+            store_clipdist_output(b, out[0], &clipdist[0]);
+         if (ucp_enables & 0xf0)
+            store_clipdist_output(b, out[1], &clipdist[4]);
+      }
+   }
 }
 
-/* ucp_enables is bitmask of enabled ucp's.  Actual ucp values are
+/*
+ * VS lowering
+ */
+
+/* ucp_enables is bitmask of enabled ucps.  Actual ucp values are
  * passed in to shader via user_clip_plane system-values
+ *
+ * If use_vars is true, the pass will use variable loads and stores instead
+ * of working with store_output intrinsics.
+ *
+ * If use_clipdist_array is true, the pass will use compact arrays for the
+ * clipdist output instead of two vec4s.
  */
-void
-nir_lower_clip_vs(nir_shader *shader, unsigned ucp_enables)
+bool
+nir_lower_clip_vs(nir_shader *shader, unsigned ucp_enables, bool use_vars,
+                  bool use_clipdist_array,
+                  const gl_state_index16 clipplane_state_tokens[][STATE_LENGTH])
 {
-   int clipvertex = -1;
-   int position = -1;
+   nir_function_impl *impl = nir_shader_get_entrypoint(shader);
+   nir_builder b;
    int maxloc = -1;
-   nir_ssa_def *cv;
+   nir_variable *position = NULL;
+   nir_variable *clipvertex = NULL;
    nir_variable *out[2] = { NULL };
 
    if (!ucp_enables)
-      return;
+      return false;
 
    /* find clipvertex/position outputs: */
    nir_foreach_variable(var, &shader->outputs) {
@@ -217,43 +326,100 @@ nir_lower_clip_vs(nir_shader *shader, unsigned ucp_enables)
        * output:
        */
       maxloc = MAX2(maxloc, loc);
+   }
 
-      switch (var->data.location) {
-      case VARYING_SLOT_POS:
-         position = loc;
+   nir_builder_init(&b, impl);
+
+   /* NIR should ensure that, even in case of loops/if-else, there
+    * should be only a single predecessor block to end_block, which
+    * makes the perfect place to insert the clipdist calculations.
+    *
+    * NOTE: in case of early returns, these would have to be lowered
+    * to jumps to end_block predecessor in a previous pass.  Not sure
+    * if there is a good way to sanity check this, but for now the
+    * users of this pass don't support sub-routines.
+    */
+   assert(impl->end_block->predecessors->entries == 1);
+   b.cursor = nir_after_cf_list(&impl->body);
+
+   /* find clipvertex/position outputs */
+   if (!find_clipvertex_and_position_outputs(shader, &clipvertex, &position))
+      return false;
+
+   /* insert CLIPDIST outputs */
+   create_clipdist_vars(shader, out, ucp_enables, &maxloc, true,
+                        use_clipdist_array);
+
+   lower_clip_outputs(&b, position, clipvertex, out, ucp_enables, use_vars,
+                      use_clipdist_array, clipplane_state_tokens);
+
+   nir_metadata_preserve(impl, nir_metadata_dominance);
+
+   return true;
+}
+
+static void
+lower_clip_in_gs_block(nir_builder *b, nir_block *block, nir_variable *position,
+                       nir_variable *clipvertex, nir_variable **out,
+                       unsigned ucp_enables, bool use_clipdist_array,
+                       const gl_state_index16 clipplane_state_tokens[][STATE_LENGTH])
+{
+   nir_foreach_instr_safe(instr, block) {
+      if (instr->type != nir_instr_type_intrinsic)
+         continue;
+
+      nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+      switch (intrin->intrinsic) {
+      case nir_intrinsic_emit_vertex_with_counter:
+      case nir_intrinsic_emit_vertex:
+         b->cursor = nir_before_instr(instr);
+         lower_clip_outputs(b, position, clipvertex, out, ucp_enables, true,
+                            use_clipdist_array, clipplane_state_tokens);
          break;
-      case VARYING_SLOT_CLIP_VERTEX:
-         clipvertex = loc;
+      default:
+         /* not interesting; skip this */
          break;
-      case VARYING_SLOT_CLIP_DIST0:
-      case VARYING_SLOT_CLIP_DIST1:
-         /* if shader is already writing CLIPDIST, then
-          * there should be no user-clip-planes to deal
-          * with.
-          */
-         return;
       }
    }
+}
 
-   if (clipvertex != -1)
-      cv = find_output(shader, clipvertex);
-   else if (position != -1)
-      cv = find_output(shader, position);
-   else
-      return;
+/*
+ * GS lowering
+ */
 
-   /* insert CLIPDIST outputs: */
-   if (ucp_enables & 0x0f)
-      out[0] =
-         create_clipdist_var(shader, ++maxloc, true, VARYING_SLOT_CLIP_DIST0);
-   if (ucp_enables & 0xf0)
-      out[1] =
-         create_clipdist_var(shader, ++maxloc, true, VARYING_SLOT_CLIP_DIST1);
+bool
+nir_lower_clip_gs(nir_shader *shader, unsigned ucp_enables,
+                  bool use_clipdist_array,
+                  const gl_state_index16 clipplane_state_tokens[][STATE_LENGTH])
+{
+   nir_function_impl *impl = nir_shader_get_entrypoint(shader);
+   nir_builder b;
+   int maxloc = -1;
+   nir_variable *position = NULL;
+   nir_variable *clipvertex = NULL;
+   nir_variable *out[2] = { NULL };
 
-   nir_foreach_function(function, shader) {
-      if (!strcmp(function->name, "main"))
-         lower_clip_vs(function->impl, ucp_enables, cv, out);
-   }
+   if (!ucp_enables)
+      return false;
+
+   /* find clipvertex/position outputs */
+   if (!find_clipvertex_and_position_outputs(shader, &clipvertex, &position))
+      return false;
+
+   /* insert CLIPDIST outputs */
+   create_clipdist_vars(shader, out, ucp_enables, &maxloc, true,
+                        use_clipdist_array);
+
+   nir_builder_init(&b, impl);
+
+   nir_foreach_block(block, impl)
+      lower_clip_in_gs_block(&b, block, position, clipvertex, out,
+                             ucp_enables, use_clipdist_array,
+                             clipplane_state_tokens);
+
+   nir_metadata_preserve(impl, nir_metadata_dominance);
+
+   return true;
 }
 
 /*
@@ -287,21 +453,24 @@ lower_clip_fs(nir_function_impl *impl, unsigned ucp_enables,
          discard->src[0] = nir_src_for_ssa(cond);
          nir_builder_instr_insert(&b, &discard->instr);
 
-         b.shader->info->fs.uses_discard = true;
+         b.shader->info.fs.uses_discard = true;
       }
    }
+
+   nir_metadata_preserve(impl, nir_metadata_dominance);
 }
 
 /* insert conditional kill based on interpolated CLIPDIST
  */
-void
-nir_lower_clip_fs(nir_shader *shader, unsigned ucp_enables)
+bool
+nir_lower_clip_fs(nir_shader *shader, unsigned ucp_enables,
+                  bool use_clipdist_array)
 {
    nir_variable *in[2];
    int maxloc = -1;
 
    if (!ucp_enables)
-      return;
+      return false;
 
    nir_foreach_variable(var, &shader->inputs) {
       int loc = var->data.driver_location;
@@ -316,18 +485,14 @@ nir_lower_clip_fs(nir_shader *shader, unsigned ucp_enables)
    /* The shader won't normally have CLIPDIST inputs, so we
     * must add our own:
     */
-   /* insert CLIPDIST outputs: */
-   if (ucp_enables & 0x0f)
-      in[0] =
-         create_clipdist_var(shader, ++maxloc, false,
-                             VARYING_SLOT_CLIP_DIST0);
-   if (ucp_enables & 0xf0)
-      in[1] =
-         create_clipdist_var(shader, ++maxloc, false,
-                             VARYING_SLOT_CLIP_DIST1);
+   /* insert CLIPDIST inputs */
+   create_clipdist_vars(shader, in, ucp_enables, &maxloc, false,
+                        use_clipdist_array);
 
    nir_foreach_function(function, shader) {
       if (!strcmp(function->name, "main"))
          lower_clip_fs(function->impl, ucp_enables, in);
    }
+
+   return true;
 }