i965/fs: use SSA values directly
authorConnor Abbott <cwabbott0@gmail.com>
Wed, 24 Jun 2015 19:28:47 +0000 (12:28 -0700)
committerConnor Abbott <cwabbott0@gmail.com>
Tue, 30 Jun 2015 18:18:27 +0000 (11:18 -0700)
Before, we would use registers, but set a magical "parent_instr" field
to indicate that it was actually purely an SSA value (i.e., it wasn't
involved in any phi nodes). Instead, just use SSA values directly, which
lets us get rid of the hack and reduces memory usage since we're not
allocating a nir_register for every value. It also makes our handling of
load_const more consistent compared to the other instructions.

Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
src/mesa/drivers/dri/i965/brw_fs.h
src/mesa/drivers/dri/i965/brw_fs_nir.cpp
src/mesa/drivers/dri/i965/brw_fs_visitor.cpp
src/mesa/drivers/dri/i965/brw_nir.c
src/mesa/drivers/dri/i965/brw_nir_analyze_boolean_resolves.c

index d08d438..8170f2a 100644 (file)
@@ -249,6 +249,8 @@ public:
    void nir_emit_block(nir_block *block);
    void nir_emit_instr(nir_instr *instr);
    void nir_emit_alu(const brw::fs_builder &bld, nir_alu_instr *instr);
+   void nir_emit_undef(const brw::fs_builder &bld,
+                       nir_ssa_undef_instr *instr);
    void nir_emit_intrinsic(const brw::fs_builder &bld,
                            nir_intrinsic_instr *instr);
    void nir_emit_texture(const brw::fs_builder &bld,
@@ -345,6 +347,7 @@ public:
    unsigned max_grf;
 
    fs_reg *nir_locals;
+   fs_reg *nir_ssa_values;
    fs_reg *nir_globals;
    fs_reg nir_inputs;
    fs_reg nir_outputs;
index 59081ea..166586f 100644 (file)
@@ -366,6 +366,9 @@ fs_visitor::nir_emit_impl(nir_function_impl *impl)
       nir_locals[reg->index] = bld.vgrf(BRW_REGISTER_TYPE_F, size);
    }
 
+   nir_ssa_values = reralloc(mem_ctx, nir_ssa_values, fs_reg,
+                             impl->ssa_alloc);
+
    nir_emit_cf_list(&impl->body);
 }
 
@@ -464,6 +467,10 @@ fs_visitor::nir_emit_instr(nir_instr *instr)
        */
       break;
 
+   case nir_instr_type_ssa_undef:
+      nir_emit_undef(abld, nir_instr_as_ssa_undef(instr));
+      break;
+
    case nir_instr_type_jump:
       nir_emit_jump(abld, nir_instr_as_jump(instr));
       break;
@@ -495,17 +502,12 @@ bool
 fs_visitor::optimize_frontfacing_ternary(nir_alu_instr *instr,
                                          const fs_reg &result)
 {
-   if (instr->src[0].src.is_ssa ||
-       !instr->src[0].src.reg.reg ||
-       !instr->src[0].src.reg.reg->parent_instr)
-      return false;
-
-   if (instr->src[0].src.reg.reg->parent_instr->type !=
-       nir_instr_type_intrinsic)
+   if (!instr->src[0].src.is_ssa ||
+       instr->src[0].src.ssa->parent_instr->type != nir_instr_type_intrinsic)
       return false;
 
    nir_intrinsic_instr *src0 =
-      nir_instr_as_intrinsic(instr->src[0].src.reg.reg->parent_instr);
+      nir_instr_as_intrinsic(instr->src[0].src.ssa->parent_instr);
 
    if (src0->intrinsic != nir_intrinsic_load_front_face)
       return false;
@@ -1146,6 +1148,13 @@ fs_visitor::nir_emit_alu(const fs_builder &bld, nir_alu_instr *instr)
    }
 }
 
+void
+fs_visitor::nir_emit_undef(const fs_builder &bld, nir_ssa_undef_instr *instr)
+{
+   nir_ssa_values[instr->def.index] = bld.vgrf(BRW_REGISTER_TYPE_D,
+                                               instr->def.num_components);
+}
+
 static fs_reg
 fs_reg_for_nir_reg(fs_visitor *v, nir_register *nir_reg,
                    unsigned base_offset, nir_src *indirect)
@@ -1171,30 +1180,39 @@ fs_reg_for_nir_reg(fs_visitor *v, nir_register *nir_reg,
 fs_reg
 fs_visitor::get_nir_src(nir_src src)
 {
+   fs_reg reg;
    if (src.is_ssa) {
-      assert(src.ssa->parent_instr->type == nir_instr_type_load_const);
-      nir_load_const_instr *load = nir_instr_as_load_const(src.ssa->parent_instr);
-      fs_reg reg = bld.vgrf(BRW_REGISTER_TYPE_D, src.ssa->num_components);
+      if (src.ssa->parent_instr->type == nir_instr_type_load_const) {
+         nir_load_const_instr *load =
+            nir_instr_as_load_const(src.ssa->parent_instr);
+         reg = bld.vgrf(BRW_REGISTER_TYPE_D, src.ssa->num_components);
 
-      for (unsigned i = 0; i < src.ssa->num_components; ++i)
-         bld.MOV(offset(reg, i), fs_reg(load->value.i[i]));
-
-      return reg;
+         for (unsigned i = 0; i < src.ssa->num_components; ++i)
+            bld.MOV(offset(reg, i), fs_reg(load->value.i[i]));
+      } else {
+         reg = nir_ssa_values[src.ssa->index];
+      }
    } else {
-      fs_reg reg = fs_reg_for_nir_reg(this, src.reg.reg, src.reg.base_offset,
-                                      src.reg.indirect);
-
-      /* to avoid floating-point denorm flushing problems, set the type by
-       * default to D - instructions that need floating point semantics will set
-       * this to F if they need to
-       */
-      return retype(reg, BRW_REGISTER_TYPE_D);
+      reg = fs_reg_for_nir_reg(this, src.reg.reg, src.reg.base_offset,
+                               src.reg.indirect);
    }
+
+   /* to avoid floating-point denorm flushing problems, set the type by
+    * default to D - instructions that need floating point semantics will set
+    * this to F if they need to
+    */
+   return retype(reg, BRW_REGISTER_TYPE_D);
 }
 
 fs_reg
 fs_visitor::get_nir_dest(nir_dest dest)
 {
+   if (dest.is_ssa) {
+      nir_ssa_values[dest.ssa.index] = bld.vgrf(BRW_REGISTER_TYPE_F,
+                                                dest.ssa.num_components);
+      return nir_ssa_values[dest.ssa.index];
+   }
+
    return fs_reg_for_nir_reg(this, dest.reg.reg, dest.reg.base_offset,
                              dest.reg.indirect);
 }
index 34bf32d..395af73 100644 (file)
@@ -2026,6 +2026,7 @@ fs_visitor::fs_visitor(const struct brw_compiler *compiler, void *log_data,
    this->no16_msg = NULL;
 
    this->nir_locals = NULL;
+   this->nir_ssa_values = NULL;
    this->nir_globals = NULL;
 
    memset(&this->payload, 0, sizeof(this->payload));
index d87e783..3e154c1 100644 (file)
@@ -156,7 +156,7 @@ brw_create_nir(struct brw_context *brw,
       nir_print_shader(nir, stderr);
    }
 
-   nir_convert_from_ssa(nir, false);
+   nir_convert_from_ssa(nir, true);
    nir_validate_shader(nir);
 
    /* This is the last pass we run before we start emitting stuff.  It
index f0b018c..9eb0ed9 100644 (file)
@@ -43,8 +43,8 @@
 static uint8_t
 get_resolve_status_for_src(nir_src *src)
 {
-   nir_instr *src_instr = nir_src_get_parent_instr(src);
-   if (src_instr) {
+   if (src->is_ssa) {
+      nir_instr *src_instr = src->ssa->parent_instr;
       uint8_t resolve_status = src_instr->pass_flags & BRW_NIR_BOOLEAN_MASK;
 
       /* If the source instruction needs resolve, then from the perspective
@@ -66,8 +66,8 @@ get_resolve_status_for_src(nir_src *src)
 static bool
 src_mark_needs_resolve(nir_src *src, void *void_state)
 {
-   nir_instr *src_instr = nir_src_get_parent_instr(src);
-   if (src_instr) {
+   if (src->is_ssa) {
+      nir_instr *src_instr = src->ssa->parent_instr;
       uint8_t resolve_status = src_instr->pass_flags & BRW_NIR_BOOLEAN_MASK;
 
       /* If the source instruction is unresolved, then mark it as needing
@@ -172,11 +172,11 @@ analyze_boolean_resolves_block(nir_block *block, void *void_state)
             resolve_status = BRW_NIR_NON_BOOLEAN;
          }
 
-         /* If the destination is SSA-like, go ahead allow unresolved booleans.
+         /* If the destination is SSA, go ahead allow unresolved booleans.
           * If the destination register doesn't have a well-defined parent_instr
           * we need to resolve immediately.
           */
-         if (alu->dest.dest.reg.reg->parent_instr == NULL &&
+         if (!alu->dest.dest.is_ssa &&
              resolve_status == BRW_NIR_BOOLEAN_UNRESOLVED) {
             resolve_status = BRW_NIR_BOOLEAN_NEEDS_RESOLVE;
          }