r600/sfn: rework getting a vector and uniforms from the value pool

Signed-off-by: Gert Wollny <[email protected]>
Reviewed-by: Reviewed-by: Dave Airlie <[email protected]>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5085>

11 files changed, 151 insertions, 87 deletions

diff --git a/src/gallium/drivers/r600/sfn/sfn_emitinstruction.cpp b/src/gallium/drivers/r600/sfn/sfn_emitinstruction.cpp
index c9c8662a669..a5ace9a164d 100644
--- a/src/gallium/drivers/r600/sfn/sfn_emitinstruction.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_emitinstruction.cpp
@@ -98,10 +98,15 @@ PValue EmitInstruction::from_nir_with_fetch_constant(const nir_src& src, unsigne
    return m_proc.from_nir_with_fetch_constant(src, component);
 }
 
-GPRVector *EmitInstruction::vec_from_nir_with_fetch_constant(const nir_src& src, unsigned mask,
-                                            const GPRVector::Swizzle& swizzle)
+GPRVector EmitInstruction::vec_from_nir_with_fetch_constant(const nir_src& src, unsigned mask,
+                                                            const GPRVector::Swizzle& swizzle, bool match)
 {
-   return m_proc.vec_from_nir_with_fetch_constant(src, mask, swizzle);
+   return m_proc.vec_from_nir_with_fetch_constant(src, mask, swizzle, match);
+}
+
+void EmitInstruction::add_uniform(unsigned index, const PValue &value)
+{
+   m_proc.add_uniform(index, value);
 }
 
 void EmitInstruction::load_uniform(const nir_alu_src& src)
@@ -153,6 +158,24 @@ enum chip_class EmitInstruction::get_chip_class(void) const
    return m_proc.get_chip_class();
 }
 
+PValue EmitInstruction::literal(uint32_t value)
+{
+   return m_proc.literal(value);
+}
+
+GPRVector EmitInstruction::vec_from_nir(const nir_dest& dst, int num_components)
+{
+   return m_proc.vec_from_nir(dst, num_components);
+}
+
+bool EmitInstruction::inject_register(unsigned sel, unsigned swizzle,
+                                      const PValue& reg, bool map)
+{
+   return m_proc.inject_register(sel, swizzle, reg, map);
+}
+
+
+
 const std::set<AluModifiers> EmitInstruction::empty = {};
 const std::set<AluModifiers> EmitInstruction::write = {alu_write};
 const std::set<AluModifiers> EmitInstruction::last_write = {alu_write, alu_last_instr};
diff --git a/src/gallium/drivers/r600/sfn/sfn_emitinstruction.h b/src/gallium/drivers/r600/sfn/sfn_emitinstruction.h
index 961c0612994..a4470cb9f23 100644
--- a/src/gallium/drivers/r600/sfn/sfn_emitinstruction.h
+++ b/src/gallium/drivers/r600/sfn/sfn_emitinstruction.h
@@ -77,14 +77,22 @@ protected:
                          const std::set<AluModifiers>& m_flags);
 
    PValue from_nir_with_fetch_constant(const nir_src& src, unsigned component);
-   GPRVector *vec_from_nir_with_fetch_constant(const nir_src& src, unsigned mask,
-                                               const GPRVector::Swizzle& swizzle);
+   GPRVector vec_from_nir_with_fetch_constant(const nir_src& src, unsigned mask,
+                                              const GPRVector::Swizzle& swizzle, bool match = false);
 
+   void add_uniform(unsigned index, const PValue &value);
    void load_uniform(const nir_alu_src& src);
    const nir_variable *get_deref_location(const nir_src& v) const;
 
    enum chip_class get_chip_class(void) const;
 
+   PValue literal(uint32_t value);
+
+   GPRVector vec_from_nir(const nir_dest& dst, int num_components);
+
+   bool inject_register(unsigned sel, unsigned swizzle,
+                        const PValue& reg, bool map);
+
 private:
 
    ShaderFromNirProcessor& m_proc;
diff --git a/src/gallium/drivers/r600/sfn/sfn_emitssboinstruction.cpp b/src/gallium/drivers/r600/sfn/sfn_emitssboinstruction.cpp
index 302e8ef3086..5d022e9ab6d 100644
--- a/src/gallium/drivers/r600/sfn/sfn_emitssboinstruction.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_emitssboinstruction.cpp
@@ -242,17 +242,18 @@ bool EmitSSBOInstruction::emit_store_ssbo(const nir_intrinsic_instr* instr)
    emit_instruction(ir);
 #else
 
-   PValue value(from_nir_with_fetch_constant(instr->src[0], 0));
-   GPRVector out_vec({value, value, value, value});
+   auto values = vec_from_nir_with_fetch_constant(instr->src[0],
+         (1 << instr->src[0].ssa->num_components) - 1, {0,1,2,3}, true);
+
    emit_instruction(new RatInstruction(cf_mem_rat, RatInstruction::STORE_TYPED,
-                                       out_vec, addr_vec, 0, rat_id, 1,
+                                       values, addr_vec, 0, rat_id, 1,
                                        1, 0, false));
    for (int i = 1; i < instr->src[0].ssa->num_components; ++i) {
-      emit_instruction(new AluInstruction(op1_mov, out_vec.reg_i(0), from_nir(instr->src[0], i), write));
+      emit_instruction(new AluInstruction(op1_mov, values.reg_i(0), from_nir(instr->src[0], i), write));
       emit_instruction(new AluInstruction(op2_add_int, addr_vec.reg_i(0),
                                           {addr_vec.reg_i(0), Value::one_i}, last_write));
       emit_instruction(new RatInstruction(cf_mem_rat, RatInstruction::STORE_TYPED,
-                                          out_vec, addr_vec, 0, rat_id, 1,
+                                          values, addr_vec, 0, rat_id, 1,
                                           1, 0, false));
    }
 #endif
diff --git a/src/gallium/drivers/r600/sfn/sfn_emittexinstruction.cpp b/src/gallium/drivers/r600/sfn/sfn_emittexinstruction.cpp
index 4e43becc8be..b3c3b06b450 100644
--- a/src/gallium/drivers/r600/sfn/sfn_emittexinstruction.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_emittexinstruction.cpp
@@ -658,15 +658,19 @@ bool EmitTexInstruction::emit_tex_tg4(nir_tex_instr* instr, TexInputs& src)
          for (unsigned i = 0; i < instr->coord_components; ++i)
             swizzle[i] = i;
 
-         std::unique_ptr<GPRVector> ofs(vec_from_nir_with_fetch_constant(*src.offset,
-                                                                         ( 1 << instr->coord_components) -1,
-                                                                         swizzle));
+         int noffsets = instr->coord_components;
+         if (instr->is_array)
+            --noffsets;
+
+         auto ofs = vec_from_nir_with_fetch_constant(*src.offset,
+                                                     ( 1 << noffsets) - 1,
+                                                     swizzle);
          GPRVector dummy(0, {7,7,7,7});
          tex_op = (tex_op == TexInstruction::gather4_c) ?
                      TexInstruction::gather4_c_o : TexInstruction::gather4_o;
 
          auto set_ofs = new TexInstruction(TexInstruction::set_offsets, dummy,
-                                           *ofs, sampler.id,
+                                           ofs, sampler.id,
                                            sampler.id + R600_MAX_CONST_BUFFERS, src.sampler_offset);
          set_ofs->set_dest_swizzle({7,7,7,7});
          emit_instruction(set_ofs);
@@ -838,29 +842,25 @@ bool EmitTexInstruction::get_inputs(const nir_tex_instr& instr, TexInputs &src)
          break;
 
       case nir_tex_src_coord: {
-         std::unique_ptr<GPRVector> coord(vec_from_nir_with_fetch_constant(instr.src[i].src,
-                                                                           (1 << instr.coord_components) - 1,
-                                                                           {0,1,2,3}));
-         src.coord = *coord;
-
+         src.coord = vec_from_nir_with_fetch_constant(instr.src[i].src,
+                                                      (1 << instr.coord_components) - 1,
+         {0,1,2,3});
       } break;
       case nir_tex_src_comparator:
          src.comperator = from_nir(instr.src[i], 0);
          break;
       case nir_tex_src_ddx: {
          sfn_log << SfnLog::tex << "Get DDX ";
-         std::unique_ptr<GPRVector> coord(vec_from_nir_with_fetch_constant(instr.src[i].src,
-                                                                           (1 << grad_components) - 1,
-                                                                           swizzle_from_mask(grad_components)));
-         src.ddx = *coord;
+         src.ddx = vec_from_nir_with_fetch_constant(instr.src[i].src,
+                                                    (1 << grad_components) - 1,
+                                                    swizzle_from_comps(grad_components));
          sfn_log << SfnLog::tex << src.ddx << "\n";
       } break;
       case nir_tex_src_ddy:{
          sfn_log << SfnLog::tex << "Get DDY ";
-         std::unique_ptr<GPRVector> coord(vec_from_nir_with_fetch_constant(instr.src[i].src,
-                                                                           (1 << grad_components) - 1,
-                                                                           swizzle_from_mask(grad_components)));
-         src.ddy = *coord;
+         src.ddy = vec_from_nir_with_fetch_constant(instr.src[i].src,
+                                                    (1 << grad_components) - 1,
+                                                    swizzle_from_comps(grad_components));
          sfn_log << SfnLog::tex << src.ddy << "\n";
       }  break;
       case nir_tex_src_lod:
diff --git a/src/gallium/drivers/r600/sfn/sfn_shader_base.cpp b/src/gallium/drivers/r600/sfn/sfn_shader_base.cpp
index 4d1d1c15242..3354e712876 100644
--- a/src/gallium/drivers/r600/sfn/sfn_shader_base.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_shader_base.cpp
@@ -592,9 +592,8 @@ bool ShaderFromNirProcessor::emit_store_scratch(nir_intrinsic_instr* instr)
 {
    PValue address = from_nir(instr->src[1], 0, 0);
 
-   std::unique_ptr<GPRVector> vec(vec_from_nir_with_fetch_constant(instr->src[0], (1 << instr->num_components) - 1,
-                                  swizzle_from_mask(instr->num_components)));
-   GPRVector value(*vec);
+   auto value = vec_from_nir_with_fetch_constant(instr->src[0], (1 << instr->num_components) - 1,
+         swizzle_from_comps(instr->num_components));
 
    int writemask = nir_intrinsic_write_mask(instr);
    int align = nir_intrinsic_align_mul(instr);
@@ -629,21 +628,44 @@ bool ShaderFromNirProcessor::emit_load_scratch(nir_intrinsic_instr* instr)
    return true;
 }
 
-GPRVector *ShaderFromNirProcessor::vec_from_nir_with_fetch_constant(const nir_src& src,
-                                                                    UNUSED unsigned mask,
-                                                                    const GPRVector::Swizzle& swizzle)
+GPRVector ShaderFromNirProcessor::vec_from_nir_with_fetch_constant(const nir_src& src,
+                                                                   unsigned mask,
+                                                                   const GPRVector::Swizzle& swizzle,
+                                                                   bool match)
 {
-   GPRVector *result = nullptr;
-   int sel = lookup_register_index(src);
-   if (sel >= 0 && from_nir(src, 0)->type() == Value::gpr &&
-       from_nir(src, 0)->chan() == 0) {
-      /* If the x-channel is really an x-channel register then we are pretty
-       * save that the value come like we need them */
-      result = new GPRVector(from_nir(src, 0)->sel(), swizzle);
-   } else {
+   bool use_same = true;
+   GPRVector::Values v;
+
+   for (int i = 0; i < 4 && use_same; ++i)  {
+      if ((1 << i) & mask) {
+         if (swizzle[i] < 4) {
+            v[i] = from_nir(src, swizzle[i]);
+            assert(v[i]);
+            if (v[i]->type() != Value::gpr)
+               use_same = false;
+            if (match && (v[i]->chan() != swizzle[i]))
+                use_same = false;
+         }
+      }
+   }
+
+   if (use_same) {
+      int i = 0;
+      while (!v[i] && i < 4) ++i;
+      assert(i < 4);
+
+      unsigned sel = v[i]->sel();
+      for (i = 0; i < 4 && use_same; ++i) {
+         if (!v[i])
+            v[i] = PValue(new GPRValue(sel, swizzle[i]));
+         else
+            use_same &= v[i]->sel() == sel;
+      }
+   }
+
+   if (!use_same) {
       AluInstruction *ir = nullptr;
       int sel = allocate_temp_register();
-      GPRVector::Values v;
       for (int i = 0; i < 4; ++i) {
          v[i] = PValue(new GPRValue(sel, swizzle[i]));
          if (swizzle[i] < 4 && (mask & (1 << i))) {
@@ -654,10 +676,8 @@ GPRVector *ShaderFromNirProcessor::vec_from_nir_with_fetch_constant(const nir_sr
       }
       if (ir)
          ir->set_flag(alu_last_instr);
-
-      result = new GPRVector(v);
    }
-   return result;
+   return GPRVector(v);;
 }
 
 bool ShaderFromNirProcessor::emit_load_ubo(nir_intrinsic_instr* instr)
diff --git a/src/gallium/drivers/r600/sfn/sfn_shader_base.h b/src/gallium/drivers/r600/sfn/sfn_shader_base.h
index dd0cd58dba5..4d799bf8f6e 100644
--- a/src/gallium/drivers/r600/sfn/sfn_shader_base.h
+++ b/src/gallium/drivers/r600/sfn/sfn_shader_base.h
@@ -62,8 +62,8 @@ public:
    void emit_instruction(Instruction *ir);
 
    PValue from_nir_with_fetch_constant(const nir_src& src, unsigned component);
-   GPRVector *vec_from_nir_with_fetch_constant(const nir_src& src, unsigned mask,
-                                               const GPRVector::Swizzle& swizzle);
+   GPRVector vec_from_nir_with_fetch_constant(const nir_src& src, unsigned mask,
+                                              const GPRVector::Swizzle& swizzle, bool match = false);
 
    bool emit_instruction(EAluOp opcode, PValue dest,
                          std::vector<PValue> src0,
@@ -79,7 +79,7 @@ public:
 
    r600_shader& sh_info() {return m_sh_info;}
    void add_param_output_reg(int loc, const GPRVector *gpr);
-   void set_output(unsigned pos, PValue var);
+   void set_output(unsigned pos, int sel);
    const GPRVector *output_register(unsigned location) const;
    void evaluate_spi_sid(r600_shader_io &io);
 
@@ -176,7 +176,7 @@ private:
    std::set<nir_variable*> m_arrays;
 
    std::map<unsigned, PValue> m_inputs;
-   std::map<unsigned, PValue> m_outputs;
+   std::map<unsigned, int> m_outputs;
 
    std::map<unsigned, nir_variable*> m_var_derefs;
    std::map<const nir_variable *, nir_variable_mode> m_var_mode;
diff --git a/src/gallium/drivers/r600/sfn/sfn_shader_fragment.cpp b/src/gallium/drivers/r600/sfn/sfn_shader_fragment.cpp
index 6a40b8066ef..eb37721d8c5 100644
--- a/src/gallium/drivers/r600/sfn/sfn_shader_fragment.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_shader_fragment.cpp
@@ -664,17 +664,26 @@ bool FragmentShaderFromNir::emit_export_pixel(const nir_variable *out_var, nir_i
 
    std::array<uint32_t,4> swizzle;
    unsigned writemask = nir_intrinsic_write_mask(instr);
-   if (out_var->data.location != FRAG_RESULT_STENCIL) {
+   switch (out_var->data.location) {
+   case FRAG_RESULT_STENCIL:
+      writemask = 2;
+      swizzle = {7,0,7,7};
+      break;
+   case FRAG_RESULT_SAMPLE_MASK:
+      writemask = 4;
+      swizzle = {7,7,0,7};
+      break;
+   default:
+      std::cerr << "Swizzle = ";
       for (int i = 0; i < 4; ++i) {
          swizzle[i] = (i < instr->num_components) ? i : 7;
+         std::cerr << swizzle[i] << ", ";
       }
-   } else {
-      swizzle = {7,0,7,7};
    }
 
-   GPRVector *value = vec_from_nir_with_fetch_constant(instr->src[1], writemask, swizzle);
+   auto value = vec_from_nir_with_fetch_constant(instr->src[1], writemask, swizzle);
 
-   set_output(out_var->data.driver_location, PValue(value));
+   set_output(out_var->data.driver_location, value.sel());
 
    if (out_var->data.location == FRAG_RESULT_COLOR ||
        (out_var->data.location >= FRAG_RESULT_DATA0 &&
@@ -688,7 +697,7 @@ bool FragmentShaderFromNir::emit_export_pixel(const nir_variable *out_var, nir_i
             continue;
          }
 
-         m_last_pixel_export = new ExportInstruction(location, *value, ExportInstruction::et_pixel);
+         m_last_pixel_export = new ExportInstruction(location, value, ExportInstruction::et_pixel);
 
          if (sh_info().ps_export_highest < location)
             sh_info().ps_export_highest = location;
@@ -701,16 +710,14 @@ bool FragmentShaderFromNir::emit_export_pixel(const nir_variable *out_var, nir_i
          emit_export_instruction(m_last_pixel_export);
          ++m_max_counted_color_exports;
       };
-   } else if (out_var->data.location == FRAG_RESULT_DEPTH) {
-      m_depth_exports++;
-      emit_export_instruction(new ExportInstruction(61, *value, ExportInstruction::et_pixel));
-   } else if (out_var->data.location == FRAG_RESULT_STENCIL) {
+   } else if (out_var->data.location == FRAG_RESULT_DEPTH ||
+              out_var->data.location == FRAG_RESULT_STENCIL ||
+              out_var->data.location == FRAG_RESULT_SAMPLE_MASK) {
       m_depth_exports++;
-      emit_export_instruction(new ExportInstruction(61, *value, ExportInstruction::et_pixel));
+      emit_export_instruction(new ExportInstruction(61, value, ExportInstruction::et_pixel));
    } else {
       return false;
    }
-
    return true;
 }
 
diff --git a/src/gallium/drivers/r600/sfn/sfn_shader_geometry.cpp b/src/gallium/drivers/r600/sfn/sfn_shader_geometry.cpp
index 35c22c7299f..48f02780ab2 100644
--- a/src/gallium/drivers/r600/sfn/sfn_shader_geometry.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_shader_geometry.cpp
@@ -57,18 +57,15 @@ bool GeometryShaderFromNir::do_emit_load_deref(UNUSED const nir_variable *in_var
 
 bool GeometryShaderFromNir::do_emit_store_deref(const nir_variable *out_var, nir_intrinsic_instr* instr)
 {
-   uint32_t write_mask =  (1 << instr->num_components) - 1;
-   GPRVector::Swizzle swz = swizzle_from_mask(instr->num_components);
-   std::unique_ptr<GPRVector> vec(vec_from_nir_with_fetch_constant(instr->src[1], write_mask, swz));
+   uint32_t write_mask = nir_intrinsic_write_mask(instr);
+   GPRVector::Swizzle swz = swizzle_from_mask(write_mask);
+   auto out_value = vec_from_nir_with_fetch_constant(instr->src[1], write_mask, swz, true);
 
-   GPRVector out_value  = *vec;
-
-   sh_info().output[out_var->data.driver_location].write_mask =
-         (1 << instr->num_components) - 1;
+   sh_info().output[out_var->data.driver_location].write_mask = write_mask;
 
    auto ir = new MemRingOutIntruction(cf_mem_ring, mem_write_ind, out_value,
                                       4 * out_var->data.driver_location,
-                                      4, m_export_base);
+                                      instr->num_components, m_export_base);
    emit_instruction(ir);
 
    return true;
diff --git a/src/gallium/drivers/r600/sfn/sfn_shader_tcs.cpp b/src/gallium/drivers/r600/sfn/sfn_shader_tcs.cpp
index 8e959cf27d2..866396badde 100644
--- a/src/gallium/drivers/r600/sfn/sfn_shader_tcs.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_shader_tcs.cpp
@@ -113,9 +113,9 @@ bool TcsShaderFromNir::store_tess_factor(nir_intrinsic_instr* instr)
 {
    const GPRVector::Swizzle& swizzle = (instr->src[0].ssa->num_components == 4) ?
             GPRVector::Swizzle({0, 1, 2, 3}) : GPRVector::Swizzle({0, 1, 7, 7});
-   std::unique_ptr<GPRVector> val(vec_from_nir_with_fetch_constant(instr->src[0],
-                                  0xf, swizzle));
-   emit_instruction(new GDSStoreTessFactor(*val));
+   auto val = vec_from_nir_with_fetch_constant(instr->src[0],
+         (1 << instr->src[0].ssa->num_components) - 1, swizzle);
+   emit_instruction(new GDSStoreTessFactor(val));
    return true;
 }
 
diff --git a/src/gallium/drivers/r600/sfn/sfn_value_gpr.h b/src/gallium/drivers/r600/sfn/sfn_value_gpr.h
index 92b126c989b..589e8821744 100644
--- a/src/gallium/drivers/r600/sfn/sfn_value_gpr.h
+++ b/src/gallium/drivers/r600/sfn/sfn_value_gpr.h
@@ -177,7 +177,7 @@ inline size_t GPRArrayValue::array_size() const
    return m_array->size();
 }
 
-inline GPRVector::Swizzle swizzle_from_mask(unsigned ncomp)
+inline GPRVector::Swizzle swizzle_from_comps(unsigned ncomp)
 {
    GPRVector::Swizzle swz = {0,1,2,3};
    for (int i = ncomp; i < 4; ++i)
@@ -185,6 +185,14 @@ inline GPRVector::Swizzle swizzle_from_mask(unsigned ncomp)
    return swz;
 }
 
+inline GPRVector::Swizzle swizzle_from_mask(unsigned mask)
+{
+   GPRVector::Swizzle swz;
+   for (int i = 0; i < 4; ++i)
+      swz[i] =  ((1 << i) & mask) ? i : 7;
+   return swz;
+}
+
 
 }
 
diff --git a/src/gallium/drivers/r600/sfn/sfn_vertexstageexport.cpp b/src/gallium/drivers/r600/sfn/sfn_vertexstageexport.cpp
index 013603edbb6..3de3418c7a4 100644
--- a/src/gallium/drivers/r600/sfn/sfn_vertexstageexport.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_vertexstageexport.cpp
@@ -159,8 +159,8 @@ bool VertexStageExportForFS::emit_varying_pos(const nir_variable *out_var, nir_i
 
    m_proc.sh_info().output[out_var->data.driver_location].write_mask = write_mask;
 
-   GPRVector *value = m_proc.vec_from_nir_with_fetch_constant(instr->src[1], write_mask, swizzle);
-   m_proc.set_output(out_var->data.driver_location, PValue(value));
+   GPRVector value = m_proc.vec_from_nir_with_fetch_constant(instr->src[1], write_mask, swizzle);
+   m_proc.set_output(out_var->data.driver_location, value.sel());
 
    int export_slot = 0;
 
@@ -168,8 +168,8 @@ bool VertexStageExportForFS::emit_varying_pos(const nir_variable *out_var, nir_i
    case VARYING_SLOT_EDGE: {
       m_proc.sh_info().vs_out_misc_write = 1;
       m_proc.sh_info().vs_out_edgeflag = 1;
-      m_proc.emit_instruction(op1_mov, value->reg_i(1), {value->reg_i(1)}, {alu_write, alu_dst_clamp, alu_last_instr});
-      m_proc.emit_instruction(op1_flt_to_int, value->reg_i(1), {value->reg_i(1)}, {alu_write, alu_last_instr});
+      m_proc.emit_instruction(op1_mov, value.reg_i(1), {value.reg_i(1)}, {alu_write, alu_dst_clamp, alu_last_instr});
+      m_proc.emit_instruction(op1_flt_to_int, value.reg_i(1), {value.reg_i(1)}, {alu_write, alu_last_instr});
       m_proc.sh_info().output[out_var->data.driver_location].write_mask = 0xf;
    }
       /* fallthrough */
@@ -189,7 +189,7 @@ bool VertexStageExportForFS::emit_varying_pos(const nir_variable *out_var, nir_i
       return false;
    }
 
-   m_last_pos_export = new ExportInstruction(export_slot, *value, ExportInstruction::et_pos);
+   m_last_pos_export = new ExportInstruction(export_slot, value, ExportInstruction::et_pos);
    m_proc.emit_export_instruction(m_last_pos_export);
    m_proc.add_param_output_reg(out_var->data.driver_location, m_last_pos_export->gpr_ptr());
    return true;
@@ -207,16 +207,16 @@ bool VertexStageExportForFS::emit_varying_param(const nir_variable *out_var, nir
 
    m_proc.sh_info().output[out_var->data.driver_location].write_mask = write_mask;
 
-   GPRVector *value = m_proc.vec_from_nir_with_fetch_constant(instr->src[1], write_mask, swizzle);
-   m_proc.sh_info().output[out_var->data.driver_location].gpr = value->sel();
+   GPRVector value = m_proc.vec_from_nir_with_fetch_constant(instr->src[1], write_mask, swizzle);
+   m_proc.sh_info().output[out_var->data.driver_location].gpr = value.sel();
 
    /* This should use the registers!! */
-   m_proc.set_output(out_var->data.driver_location, PValue(value));
+   m_proc.set_output(out_var->data.driver_location, value.sel());
 
    auto param_loc = m_param_map.find(out_var->data.location);
    assert(param_loc != m_param_map.end());
 
-   m_last_param_export = new ExportInstruction(param_loc->second, *value, ExportInstruction::et_param);
+   m_last_param_export = new ExportInstruction(param_loc->second, value, ExportInstruction::et_param);
    m_proc.emit_export_instruction(m_last_param_export);
    m_proc.add_param_output_reg(out_var->data.driver_location, m_last_param_export->gpr_ptr());
    return true;
@@ -227,7 +227,7 @@ bool VertexStageExportForFS::emit_clip_vertices(const nir_variable *out_var, nir
    m_proc.sh_info().cc_dist_mask = 0xff;
    m_proc.sh_info().clip_dist_write = 0xff;
 
-   std::unique_ptr<GPRVector> clip_vertex(m_proc.vec_from_nir_with_fetch_constant(instr->src[1], 0xf, {0,1,2,3}));
+   GPRVector clip_vertex = m_proc.vec_from_nir_with_fetch_constant(instr->src[1], 0xf, {0,1,2,3});
 
    for (int i = 0; i < 4; ++i)
       m_proc.sh_info().output[out_var->data.driver_location].write_mask |= 1 << i;
@@ -239,7 +239,7 @@ bool VertexStageExportForFS::emit_clip_vertices(const nir_variable *out_var, nir
       int ochan = i & 3;
       AluInstruction *ir = nullptr;
       for (int j = 0; j < 4; j++) {
-         ir = new AluInstruction(op2_dot4_ieee, clip_dist[oreg].reg_i(j), clip_vertex->reg_i(j),
+         ir = new AluInstruction(op2_dot4_ieee, clip_dist[oreg].reg_i(j), clip_vertex.reg_i(j),
                                  PValue(new UniformValue(512 + i, j, R600_BUFFER_INFO_CONST_BUFFER)),
                                  (j == ochan) ? EmitInstruction::write : EmitInstruction::empty);
          m_proc.emit_instruction(ir);
@@ -419,10 +419,10 @@ bool VertexStageExportForGS::store_deref(const nir_variable *out_var, nir_intrin
 
    uint32_t write_mask =  (1 << instr->num_components) - 1;
 
-   std::unique_ptr<GPRVector> value(m_proc.vec_from_nir_with_fetch_constant(instr->src[1], write_mask,
-                                    swizzle_from_mask(instr->num_components)));
+   GPRVector value = m_proc.vec_from_nir_with_fetch_constant(instr->src[1], write_mask,
+         swizzle_from_comps(instr->num_components));
 
-   auto ir = new MemRingOutIntruction(cf_mem_ring, mem_write, *value,
+   auto ir = new MemRingOutIntruction(cf_mem_ring, mem_write, value,
                                       ring_offset >> 2, 4, PValue());
    m_proc.emit_export_instruction(ir);