etnaviv: add alternative NIR compiler

enable with ETNA_MESA_DEBUG=nir

Signed-off-by: Jonathan Marek <[email protected]>
Reviewed-by: Philipp Zabel <[email protected]>

12 files changed, 2388 insertions, 37 deletions

diff --git a/src/gallium/drivers/etnaviv/Makefile.sources b/src/gallium/drivers/etnaviv/Makefile.sources
index 01e7e49a38a..1d0269acffb 100644
--- a/src/gallium/drivers/etnaviv/Makefile.sources
+++ b/src/gallium/drivers/etnaviv/Makefile.sources
@@ -18,6 +18,8 @@ C_SOURCES :=  \
 	etnaviv_clear_blit.h \
 	etnaviv_compiler.c \
 	etnaviv_compiler.h \
+	etnaviv_compiler_nir.c \
+	etnaviv_compiler_nir_emit.h \
 	etnaviv_context.c \
 	etnaviv_context.h \
 	etnaviv_debug.h \
diff --git a/src/gallium/drivers/etnaviv/etnaviv_compiler.c b/src/gallium/drivers/etnaviv/etnaviv_compiler.c
index 474e3d23405..8214d4f5770 100644
--- a/src/gallium/drivers/etnaviv/etnaviv_compiler.c
+++ b/src/gallium/drivers/etnaviv/etnaviv_compiler.c
@@ -2286,6 +2286,9 @@ copy_uniform_state_to_shader(struct etna_compile *c, struct etna_shader_variant
 bool
 etna_compile_shader(struct etna_shader_variant *v)
 {
+   if (DBG_ENABLED(ETNA_DBG_NIR))
+      return etna_compile_shader_nir(v);
+
    /* Create scratch space that may be too large to fit on stack
     */
    bool ret;
@@ -2449,11 +2452,12 @@ etna_compile_shader(struct etna_shader_variant *v)
    etna_compile_fill_in_labels(c);
 
    /* fill in output structure */
-   v->processor = c->info.processor;
+   v->stage = c->info.processor == PIPE_SHADER_FRAGMENT ? MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX;
    v->code_size = c->inst_ptr * 4;
    v->code = mem_dup(c->code, c->inst_ptr * 16);
    v->num_loops = c->num_loops;
    v->num_temps = c->next_free_native;
+   v->vs_id_in_reg = -1;
    v->vs_pos_out_reg = -1;
    v->vs_pointsize_out_reg = -1;
    v->ps_color_out_reg = -1;
@@ -2483,7 +2487,7 @@ extern const char *tgsi_swizzle_names[];
 void
 etna_dump_shader(const struct etna_shader_variant *shader)
 {
-   if (shader->processor == PIPE_SHADER_VERTEX)
+   if (shader->stage == MESA_SHADER_VERTEX)
       printf("VERT\n");
    else
       printf("FRAG\n");
@@ -2502,22 +2506,42 @@ etna_dump_shader(const struct etna_shader_variant *shader)
              shader->uniforms.imm_data[idx],
              shader->uniforms.imm_contents[idx]);
    }
-   printf("inputs:\n");
-   for (int idx = 0; idx < shader->infile.num_reg; ++idx) {
-      printf(" [%i] name=%s index=%i comps=%i\n", shader->infile.reg[idx].reg,
-             tgsi_semantic_names[shader->infile.reg[idx].semantic.Name],
-             shader->infile.reg[idx].semantic.Index,
-             shader->infile.reg[idx].num_components);
-   }
-   printf("outputs:\n");
-   for (int idx = 0; idx < shader->outfile.num_reg; ++idx) {
-      printf(" [%i] name=%s index=%i comps=%i\n", shader->outfile.reg[idx].reg,
-             tgsi_semantic_names[shader->outfile.reg[idx].semantic.Name],
-             shader->outfile.reg[idx].semantic.Index,
-             shader->outfile.reg[idx].num_components);
+
+   if (DBG_ENABLED(ETNA_DBG_NIR)) {
+      printf("inputs:\n");
+      for (int idx = 0; idx < shader->infile.num_reg; ++idx) {
+         printf(" [%i] name=%s comps=%i\n", shader->infile.reg[idx].reg,
+                (shader->stage == MESA_SHADER_VERTEX) ?
+                  gl_vert_attrib_name(shader->infile.reg[idx].slot) :
+                  gl_varying_slot_name(shader->infile.reg[idx].slot),
+                shader->infile.reg[idx].num_components);
+      }
+      printf("outputs:\n");
+      for (int idx = 0; idx < shader->outfile.num_reg; ++idx) {
+         printf(" [%i] name=%s comps=%i\n", shader->outfile.reg[idx].reg,
+                (shader->stage == MESA_SHADER_VERTEX) ?
+                  gl_varying_slot_name(shader->outfile.reg[idx].slot) :
+                  gl_frag_result_name(shader->outfile.reg[idx].slot),
+                shader->outfile.reg[idx].num_components);
+      }
+   } else {
+      printf("inputs:\n");
+      for (int idx = 0; idx < shader->infile.num_reg; ++idx) {
+         printf(" [%i] name=%s index=%i comps=%i\n", shader->infile.reg[idx].reg,
+                tgsi_semantic_names[shader->infile.reg[idx].semantic.Name],
+                shader->infile.reg[idx].semantic.Index,
+                shader->infile.reg[idx].num_components);
+      }
+      printf("outputs:\n");
+      for (int idx = 0; idx < shader->outfile.num_reg; ++idx) {
+         printf(" [%i] name=%s index=%i comps=%i\n", shader->outfile.reg[idx].reg,
+                tgsi_semantic_names[shader->outfile.reg[idx].semantic.Name],
+                shader->outfile.reg[idx].semantic.Index,
+                shader->outfile.reg[idx].num_components);
+      }
    }
    printf("special:\n");
-   if (shader->processor == PIPE_SHADER_VERTEX) {
+   if (shader->stage == MESA_SHADER_VERTEX) {
       printf("  vs_pos_out_reg=%i\n", shader->vs_pos_out_reg);
       printf("  vs_pointsize_out_reg=%i\n", shader->vs_pointsize_out_reg);
       printf("  vs_load_balancing=0x%08x\n", shader->vs_load_balancing);
@@ -2531,6 +2555,9 @@ etna_dump_shader(const struct etna_shader_variant *shader)
 void
 etna_destroy_shader(struct etna_shader_variant *shader)
 {
+   if (DBG_ENABLED(ETNA_DBG_NIR))
+      return etna_destroy_shader_nir(shader);
+
    assert(shader);
 
    FREE(shader->code);
@@ -2554,6 +2581,9 @@ bool
 etna_link_shader(struct etna_shader_link_info *info,
                  const struct etna_shader_variant *vs, const struct etna_shader_variant *fs)
 {
+   if (DBG_ENABLED(ETNA_DBG_NIR))
+      return etna_link_shader_nir(info, vs, fs);
+
    int comp_ofs = 0;
    /* For each fragment input we need to find the associated vertex shader
     * output, which can be found by matching on semantic name and index. A
diff --git a/src/gallium/drivers/etnaviv/etnaviv_compiler.h b/src/gallium/drivers/etnaviv/etnaviv_compiler.h
index 48b1b218750..b7feeb6d49b 100644
--- a/src/gallium/drivers/etnaviv/etnaviv_compiler.h
+++ b/src/gallium/drivers/etnaviv/etnaviv_compiler.h
@@ -32,6 +32,7 @@
 #include "etnaviv_shader.h"
 #include "pipe/p_compiler.h"
 #include "pipe/p_shader_tokens.h"
+#include "compiler/shader_enums.h"
 
 /* XXX some of these are pretty arbitrary limits, may be better to switch
  * to dynamic allocation at some point.
@@ -47,6 +48,7 @@
 struct etna_shader_inout {
    int reg; /* native register */
    struct tgsi_declaration_semantic semantic; /* tgsi semantic name and index */
+   int slot; /* nir: gl_varying_slot or gl_vert_attrib */
    int num_components;
 };
 
@@ -59,7 +61,7 @@ struct etna_shader_io_file {
 struct etna_shader_variant {
    uint32_t id; /* for debug */
 
-   uint processor; /* TGSI_PROCESSOR_... */
+   gl_shader_stage stage;
    uint32_t code_size; /* code size in uint32 words */
    uint32_t *code;
    unsigned num_loops;
@@ -77,12 +79,13 @@ struct etna_shader_variant {
    /* outputs (for linking) */
    struct etna_shader_io_file outfile;
 
-   /* index into outputs (for linking) */
+   /* index into outputs (for linking) - only for TGSI compiler */
    int output_count_per_semantic[TGSI_SEMANTIC_COUNT];
    struct etna_shader_inout * *output_per_semantic_list; /* list of pointers to outputs */
    struct etna_shader_inout **output_per_semantic[TGSI_SEMANTIC_COUNT];
 
-   /* special outputs (vs only) */
+   /* special inputs/outputs (vs only) */
+   int vs_id_in_reg; /* vertexid+instanceid input */
    int vs_pos_out_reg; /* VS position output */
    int vs_pointsize_out_reg; /* VS point size output */
    uint32_t vs_load_balancing;
@@ -134,4 +137,20 @@ etna_link_shader(struct etna_shader_link_info *info,
 void
 etna_destroy_shader(struct etna_shader_variant *shader);
 
+/* NIR compiler */
+
+bool
+etna_compile_shader_nir(struct etna_shader_variant *shader);
+
+void
+etna_dump_shader_nir(const struct etna_shader_variant *shader);
+
+bool
+etna_link_shader_nir(struct etna_shader_link_info *info,
+                     const struct etna_shader_variant *vs,
+                     const struct etna_shader_variant *fs);
+
+void
+etna_destroy_shader_nir(struct etna_shader_variant *shader);
+
 #endif
diff --git a/src/gallium/drivers/etnaviv/etnaviv_compiler_nir.c b/src/gallium/drivers/etnaviv/etnaviv_compiler_nir.c
new file mode 100644
index 00000000000..44e0c1a29fb
--- /dev/null
+++ b/src/gallium/drivers/etnaviv/etnaviv_compiler_nir.c
@@ -0,0 +1,853 @@
+/*
+ * Copyright (c) 2012-2019 Etnaviv Project
+ * Copyright (c) 2019 Zodiac Inflight Innovations
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sub license,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ *
+ * Authors:
+ *    Jonathan Marek <[email protected]>
+ *    Wladimir J. van der Laan <[email protected]>
+ */
+
+#include "etnaviv_compiler.h"
+#include "etnaviv_asm.h"
+#include "etnaviv_context.h"
+#include "etnaviv_debug.h"
+#include "etnaviv_disasm.h"
+#include "etnaviv_uniforms.h"
+#include "etnaviv_util.h"
+
+#include <math.h>
+#include "util/u_memory.h"
+#include "util/register_allocate.h"
+#include "compiler/nir/nir_builder.h"
+#include "compiler/nir/nir_worklist.h"
+
+#include "util/u_half.h"
+
+struct etna_compile {
+   nir_shader *nir;
+#define is_fs(c) ((c)->nir->info.stage == MESA_SHADER_FRAGMENT)
+   const struct etna_specs *specs;
+   struct etna_shader_variant *variant;
+
+   /* register assigned to each output, indexed by driver_location */
+   unsigned output_reg[ETNA_NUM_INPUTS];
+
+   /* block # to instr index */
+   unsigned *block_ptr;
+
+   /* Code generation */
+   int inst_ptr; /* current instruction pointer */
+   struct etna_inst code[ETNA_MAX_INSTRUCTIONS * ETNA_INST_SIZE];
+
+   /* There was an error during compilation */
+   bool error;
+};
+
+#define compile_error(ctx, args...) ({ \
+   printf(args); \
+   ctx->error = true; \
+   assert(0); \
+})
+
+/* io related lowering
+ * run after lower_int_to_float because it adds i2f/f2i ops
+ */
+static void
+etna_lower_io(nir_shader *shader, struct etna_shader_variant *v)
+{
+   bool rb_swap = shader->info.stage == MESA_SHADER_FRAGMENT && v->key.frag_rb_swap;
+
+   unsigned color_location = 0;
+   nir_foreach_variable(var, &shader->outputs) {
+      switch (var->data.location) {
+      case FRAG_RESULT_COLOR:
+      case FRAG_RESULT_DATA0:
+         color_location = var->data.driver_location;
+         break;
+      }
+   }
+
+   nir_foreach_function(function, shader) {
+      nir_builder b;
+      nir_builder_init(&b, function->impl);
+
+      nir_foreach_block(block, function->impl) {
+         nir_foreach_instr_safe(instr, block) {
+            if (instr->type == nir_instr_type_intrinsic) {
+               nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+
+               switch (intr->intrinsic) {
+               case nir_intrinsic_load_front_face: {
+                  /* front face inverted (run after int_to_float, so invert as float) */
+                  b.cursor = nir_after_instr(instr);
+
+                  nir_ssa_def *ssa = nir_seq(&b, &intr->dest.ssa, nir_imm_float(&b, 0.0));
+                  nir_ssa_def_rewrite_uses_after(&intr->dest.ssa,
+                                                 nir_src_for_ssa(ssa),
+                                                 ssa->parent_instr);
+               } break;
+               case nir_intrinsic_store_output: {
+                  if (!rb_swap || nir_intrinsic_base(intr) != color_location)
+                     break;
+                  b.cursor = nir_before_instr(instr);
+
+                  nir_ssa_def *ssa = nir_mov(&b, intr->src[0].ssa);
+                  nir_alu_instr *alu = nir_instr_as_alu(ssa->parent_instr);
+                  alu->src[0].swizzle[0] = 2;
+                  alu->src[0].swizzle[2] = 0;
+                  nir_instr_rewrite_src(instr, &intr->src[0], nir_src_for_ssa(ssa));
+               } break;
+               case nir_intrinsic_load_instance_id: {
+                  b.cursor = nir_after_instr(instr);
+                  nir_ssa_def *ssa = nir_i2f32(&b, &intr->dest.ssa);
+                  nir_ssa_def_rewrite_uses_after(&intr->dest.ssa,
+                                                 nir_src_for_ssa(ssa),
+                                                 ssa->parent_instr);
+               } break;
+               case nir_intrinsic_load_uniform: {
+                  /* multiply by 16 and convert to int */
+                  b.cursor = nir_before_instr(instr);
+                  nir_ssa_def *ssa = nir_f2u32(&b, nir_fmul(&b, intr->src[0].ssa,
+                                                                nir_imm_float(&b, 16.0f)));
+                  nir_instr_rewrite_src(instr, &intr->src[0], nir_src_for_ssa(ssa));
+               } break;
+               default:
+                  break;
+               }
+            }
+
+            if (instr->type != nir_instr_type_tex)
+               continue;
+
+            nir_tex_instr *tex = nir_instr_as_tex(instr);
+            nir_src *coord = NULL;
+            nir_src *lod_bias = NULL;
+            unsigned lod_bias_idx;
+
+            assert(tex->sampler_index == tex->texture_index);
+
+            for (unsigned i = 0; i < tex->num_srcs; i++) {
+               switch (tex->src[i].src_type) {
+               case nir_tex_src_coord:
+                  coord = &tex->src[i].src;
+                  break;
+               case nir_tex_src_bias:
+               case nir_tex_src_lod:
+                  assert(!lod_bias);
+                  lod_bias = &tex->src[i].src;
+                  lod_bias_idx = i;
+                  break;
+               default:
+                  assert(0);
+                  break;
+               }
+            }
+
+            if (tex->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
+               /* use a dummy load_uniform here to represent texcoord scale */
+               b.cursor = nir_before_instr(instr);
+               nir_intrinsic_instr *load =
+                  nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_uniform);
+               nir_intrinsic_set_base(load, ~tex->sampler_index);
+               load->num_components = 2;
+               load->src[0] = nir_src_for_ssa(nir_imm_float(&b, 0.0f));
+               nir_ssa_dest_init(&load->instr, &load->dest, 2, 32, NULL);
+               nir_intrinsic_set_type(load, nir_type_float);
+
+               nir_builder_instr_insert(&b, &load->instr);
+
+               nir_ssa_def *new_coord = nir_fmul(&b, coord->ssa, &load->dest.ssa);
+               nir_instr_rewrite_src(&tex->instr, coord, nir_src_for_ssa(new_coord));
+            }
+
+            /* pre HALTI5 needs texture sources in a single source */
+
+            if (!lod_bias || v->shader->specs->halti >= 5)
+               continue;
+
+            assert(coord && lod_bias && tex->coord_components < 4);
+
+            nir_alu_instr *vec = nir_alu_instr_create(shader, nir_op_vec4);
+            for (unsigned i = 0; i < tex->coord_components; i++) {
+               vec->src[i].src = nir_src_for_ssa(coord->ssa);
+               vec->src[i].swizzle[0] = i;
+            }
+            for (unsigned i = tex->coord_components; i < 4; i++)
+               vec->src[i].src = nir_src_for_ssa(lod_bias->ssa);
+
+            vec->dest.write_mask = 0xf;
+            nir_ssa_dest_init(&vec->instr, &vec->dest.dest, 4, 32, NULL);
+
+            nir_tex_instr_remove_src(tex, lod_bias_idx);
+            nir_instr_rewrite_src(&tex->instr, coord, nir_src_for_ssa(&vec->dest.dest.ssa));
+            tex->coord_components = 4;
+
+            nir_instr_insert_before(&tex->instr, &vec->instr);
+         }
+      }
+   }
+}
+
+static void
+etna_lower_alu_to_scalar(nir_shader *shader, const struct etna_specs *specs)
+{
+   BITSET_DECLARE(scalar_ops, nir_num_opcodes);
+   BITSET_ZERO(scalar_ops);
+
+   BITSET_SET(scalar_ops, nir_op_frsq);
+   BITSET_SET(scalar_ops, nir_op_frcp);
+   BITSET_SET(scalar_ops, nir_op_flog2);
+   BITSET_SET(scalar_ops, nir_op_fexp2);
+   BITSET_SET(scalar_ops, nir_op_fsqrt);
+   BITSET_SET(scalar_ops, nir_op_fcos);
+   BITSET_SET(scalar_ops, nir_op_fsin);
+   BITSET_SET(scalar_ops, nir_op_fdiv);
+
+   if (!specs->has_halti2_instructions)
+      BITSET_SET(scalar_ops, nir_op_fdot2);
+
+   nir_lower_alu_to_scalar(shader, scalar_ops);
+}
+
+static void
+etna_lower_alu_impl(nir_function_impl *impl, struct etna_compile *c)
+{
+   nir_shader *shader = impl->function->shader;
+
+   nir_builder b;
+   nir_builder_init(&b, impl);
+
+   /* in a seperate loop so we can apply the multiple-uniform logic to the new fmul */
+   nir_foreach_block(block, impl) {
+      nir_foreach_instr_safe(instr, block) {
+         if (instr->type != nir_instr_type_alu)
+            continue;
+
+         nir_alu_instr *alu = nir_instr_as_alu(instr);
+         /* multiply sin/cos src by constant
+          * TODO: do this earlier (but it breaks const_prop opt)
+          */
+         if (alu->op == nir_op_fsin || alu->op == nir_op_fcos) {
+            b.cursor = nir_before_instr(instr);
+
+            nir_ssa_def *imm = c->specs->has_new_transcendentals ?
+               nir_imm_float(&b, 1.0 / M_PI) :
+               nir_imm_float(&b, 2.0 / M_PI);
+
+            nir_instr_rewrite_src(instr, &alu->src[0].src,
+               nir_src_for_ssa(nir_fmul(&b, alu->src[0].src.ssa, imm)));
+         }
+
+         /* change transcendental ops to vec2 and insert vec1 mul for the result
+          * TODO: do this earlier (but it breaks with optimizations)
+          */
+         if (c->specs->has_new_transcendentals && (
+             alu->op == nir_op_fdiv || alu->op == nir_op_flog2 ||
+             alu->op == nir_op_fsin || alu->op == nir_op_fcos)) {
+            nir_ssa_def *ssa = &alu->dest.dest.ssa;
+
+            assert(ssa->num_components == 1);
+
+            nir_alu_instr *mul = nir_alu_instr_create(shader, nir_op_fmul);
+            mul->src[0].src = mul->src[1].src = nir_src_for_ssa(ssa);
+            mul->src[1].swizzle[0] = 1;
+
+            mul->dest.write_mask = 1;
+            nir_ssa_dest_init(&mul->instr, &mul->dest.dest, 1, 32, NULL);
+
+            ssa->num_components = 2;
+
+            mul->dest.saturate = alu->dest.saturate;
+            alu->dest.saturate = 0;
+
+            nir_instr_insert_after(instr, &mul->instr);
+
+            nir_ssa_def_rewrite_uses_after(ssa, nir_src_for_ssa(&mul->dest.dest.ssa), &mul->instr);
+         }
+      }
+   }
+}
+
+static void etna_lower_alu(nir_shader *shader, struct etna_compile *c)
+{
+   nir_foreach_function(function, shader) {
+      if (function->impl)
+         etna_lower_alu_impl(function->impl, c);
+   }
+}
+
+static void
+emit_inst(struct etna_compile *c, struct etna_inst *inst)
+{
+   c->code[c->inst_ptr++] = *inst;
+}
+
+/* to map nir srcs should to etna_inst srcs */
+enum {
+   SRC_0_1_2 = (0 << 0) | (1 << 2) | (2 << 4),
+   SRC_0_1_X = (0 << 0) | (1 << 2) | (3 << 4),
+   SRC_0_X_X = (0 << 0) | (3 << 2) | (3 << 4),
+   SRC_0_X_1 = (0 << 0) | (3 << 2) | (1 << 4),
+   SRC_0_1_0 = (0 << 0) | (1 << 2) | (0 << 4),
+   SRC_X_X_0 = (3 << 0) | (3 << 2) | (0 << 4),
+   SRC_0_X_0 = (0 << 0) | (3 << 2) | (0 << 4),
+};
+
+/* info to translate a nir op to etna_inst */
+struct etna_op_info {
+   uint8_t opcode; /* INST_OPCODE_ */
+   uint8_t src; /* SRC_ enum  */
+   uint8_t cond; /* INST_CONDITION_ */
+   uint8_t type; /* INST_TYPE_ */
+};
+
+static const struct etna_op_info etna_ops[] = {
+   [0 ... nir_num_opcodes - 1] = {0xff},
+#undef TRUE
+#undef FALSE
+#define OPCT(nir, op, src, cond, type) [nir_op_##nir] = { \
+   INST_OPCODE_##op, \
+   SRC_##src, \
+   INST_CONDITION_##cond, \
+   INST_TYPE_##type \
+}
+#define OPC(nir, op, src, cond) OPCT(nir, op, src, cond, F32)
+#define OP(nir, op, src) OPC(nir, op, src, TRUE)
+   OP(mov, MOV, X_X_0), OP(fneg, MOV, X_X_0), OP(fabs, MOV, X_X_0), OP(fsat, MOV, X_X_0),
+   OP(fmul, MUL, 0_1_X), OP(fadd, ADD, 0_X_1), OP(ffma, MAD, 0_1_2),
+   OP(fdot2, DP2, 0_1_X), OP(fdot3, DP3, 0_1_X), OP(fdot4, DP4, 0_1_X),
+   OPC(fmin, SELECT, 0_1_0, GT), OPC(fmax, SELECT, 0_1_0, LT),
+   OP(ffract, FRC, X_X_0), OP(frcp, RCP, X_X_0), OP(frsq, RSQ, X_X_0),
+   OP(fsqrt, SQRT, X_X_0), OP(fsin, SIN, X_X_0), OP(fcos, COS, X_X_0),
+   OP(fsign, SIGN, X_X_0), OP(ffloor, FLOOR, X_X_0), OP(fceil, CEIL, X_X_0),
+   OP(flog2, LOG, X_X_0), OP(fexp2, EXP, X_X_0),
+   OPC(seq, SET, 0_1_X, EQ), OPC(sne, SET, 0_1_X, NE), OPC(sge, SET, 0_1_X, GE), OPC(slt, SET, 0_1_X, LT),
+   OPC(fcsel, SELECT, 0_1_2, NZ),
+   OP(fdiv, DIV, 0_1_X),
+   OP(fddx, DSX, 0_X_0), OP(fddy, DSY, 0_X_0),
+
+   /* integer opcodes */
+   OPCT(i2f32, I2F, 0_X_X, TRUE, S32),
+   OPCT(f2u32, F2I, 0_X_X, TRUE, U32),
+};
+
+static void
+etna_emit_block_start(struct etna_compile *c, unsigned block)
+{
+   c->block_ptr[block] = c->inst_ptr;
+}
+
+static void
+etna_emit_alu(struct etna_compile *c, nir_op op, struct etna_inst_dst dst,
+              struct etna_inst_src src[3], bool saturate)
+{
+   struct etna_op_info ei = etna_ops[op];
+
+   assert(ei.opcode != 0xff);
+
+   struct etna_inst inst = {
+      .opcode = ei.opcode,
+      .type = ei.type,
+      .cond = ei.cond,
+      .dst = dst,
+      .sat = saturate,
+   };
+
+   switch (op) {
+   case nir_op_fdiv:
+   case nir_op_flog2:
+   case nir_op_fsin:
+   case nir_op_fcos:
+      if (c->specs->has_new_transcendentals)
+         inst.tex.amode = 1;
+      /* fall through */
+   case nir_op_frsq:
+   case nir_op_frcp:
+   case nir_op_fexp2:
+   case nir_op_fsqrt:
+   case nir_op_i2f32:
+   case nir_op_f2u32:
+      /* for these instructions we want src to be in x component
+       * note: on HALTI2+ i2f/f2u are not scalar but we only use them this way currently
+       */
+      src[0].swiz = inst_swiz_compose(src[0].swiz,
+                                      INST_SWIZ_BROADCAST(ffs(inst.dst.write_mask)-1));
+   default:
+      break;
+   }
+
+   for (unsigned j = 0; j < 3; j++) {
+      unsigned i = ((ei.src >> j*2) & 3);
+      if (i < 3)
+         inst.src[j] = src[i];
+   }
+
+   emit_inst(c, &inst);
+}
+
+static void
+etna_emit_tex(struct etna_compile *c, nir_texop op, unsigned texid, unsigned dst_swiz,
+              struct etna_inst_dst dst, struct etna_inst_src coord,
+              struct etna_inst_src lod_bias)
+{
+   struct etna_inst inst = {
+      .dst = dst,
+      .tex.id = texid + (is_fs(c) ? 0 : c->specs->vertex_sampler_offset),
+      .tex.swiz = dst_swiz,
+      .src[0] = coord,
+   };
+
+   if (lod_bias.use)
+      inst.src[1] = lod_bias;
+
+   switch (op) {
+   case nir_texop_tex: inst.opcode = INST_OPCODE_TEXLD; break;
+   case nir_texop_txb: inst.opcode = INST_OPCODE_TEXLDB; break;
+   case nir_texop_txl: inst.opcode = INST_OPCODE_TEXLDL; break;
+   default:
+      assert(0);
+   }
+
+   emit_inst(c, &inst);
+}
+
+static void
+etna_emit_jump(struct etna_compile *c, unsigned block, struct etna_inst_src condition)
+{
+   if (!condition.use) {
+      emit_inst(c, &(struct etna_inst) {.opcode = INST_OPCODE_BRANCH, .imm = block });
+      return;
+   }
+
+   struct etna_inst inst = {
+      .opcode = INST_OPCODE_BRANCH,
+      .cond = INST_CONDITION_NOT,
+      .type = INST_TYPE_U32,
+      .src[0] = condition,
+      .imm = block,
+   };
+   inst.src[0].swiz = INST_SWIZ_BROADCAST(inst.src[0].swiz & 3);
+   emit_inst(c, &inst);
+}
+
+static void
+etna_emit_discard(struct etna_compile *c, struct etna_inst_src condition)
+{
+   if (!condition.use) {
+      emit_inst(c, &(struct etna_inst) { .opcode = INST_OPCODE_TEXKILL });
+      return;
+   }
+
+   struct etna_inst inst = {
+      .opcode = INST_OPCODE_TEXKILL,
+      .cond = INST_CONDITION_GZ,
+      .src[0] = condition,
+   };
+   inst.src[0].swiz = INST_SWIZ_BROADCAST(inst.src[0].swiz & 3);
+   emit_inst(c, &inst);
+}
+
+static void
+etna_emit_output(struct etna_compile *c, unsigned index, struct etna_inst_src src)
+{
+   c->output_reg[index] = src.reg;
+}
+
+static void
+etna_emit_load_ubo(struct etna_compile *c, struct etna_inst_dst dst,
+                   struct etna_inst_src src, struct etna_inst_src base)
+{
+   emit_inst(c, &(struct etna_inst) {
+      .opcode = INST_OPCODE_LOAD,
+      .type = INST_TYPE_U32,
+      .dst = dst,
+      .src[0] = src,
+      .src[1] = base,
+   });
+}
+
+#define OPT(nir, pass, ...) ({                             \
+   bool this_progress = false;                             \
+   NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__);      \
+   this_progress;                                          \
+})
+#define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
+
+static void
+etna_optimize_loop(nir_shader *s)
+{
+   bool progress;
+   do {
+      progress = false;
+
+      OPT_V(s, nir_lower_vars_to_ssa);
+      progress |= OPT(s, nir_opt_copy_prop_vars);
+      progress |= OPT(s, nir_copy_prop);
+      progress |= OPT(s, nir_opt_dce);
+      progress |= OPT(s, nir_opt_cse);
+      progress |= OPT(s, nir_opt_peephole_select, 16, true, true);
+      progress |= OPT(s, nir_opt_intrinsics);
+      progress |= OPT(s, nir_opt_algebraic);
+      progress |= OPT(s, nir_opt_constant_folding);
+      progress |= OPT(s, nir_opt_dead_cf);
+      if (OPT(s, nir_opt_trivial_continues)) {
+         progress = true;
+         /* If nir_opt_trivial_continues makes progress, then we need to clean
+          * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
+          * to make progress.
+          */
+         OPT(s, nir_copy_prop);
+         OPT(s, nir_opt_dce);
+      }
+      progress |= OPT(s, nir_opt_loop_unroll, nir_var_all);
+      progress |= OPT(s, nir_opt_if, false);
+      progress |= OPT(s, nir_opt_remove_phis);
+      progress |= OPT(s, nir_opt_undef);
+   }
+   while (progress);
+}
+
+static int
+etna_glsl_type_size(const struct glsl_type *type, bool bindless)
+{
+   return glsl_count_attribute_slots(type, false);
+}
+
+static void
+copy_uniform_state_to_shader(struct etna_shader_variant *sobj, uint64_t *consts, unsigned count)
+{
+   struct etna_shader_uniform_info *uinfo = &sobj->uniforms;
+
+   uinfo->imm_count = count * 4;
+   uinfo->imm_data = MALLOC(uinfo->imm_count * sizeof(*uinfo->imm_data));
+   uinfo->imm_contents = MALLOC(uinfo->imm_count * sizeof(*uinfo->imm_contents));
+
+   for (unsigned i = 0; i < uinfo->imm_count; i++) {
+      uinfo->imm_data[i] = consts[i];
+      uinfo->imm_contents[i] = consts[i] >> 32;
+   }
+
+   etna_set_shader_uniforms_dirty_flags(sobj);
+}
+
+#include "etnaviv_compiler_nir_emit.h"
+
+bool
+etna_compile_shader_nir(struct etna_shader_variant *v)
+{
+   if (unlikely(!v))
+      return false;
+
+   struct etna_compile *c = CALLOC_STRUCT(etna_compile);
+   if (!c)
+      return false;
+
+   c->variant = v;
+   c->specs = v->shader->specs;
+   c->nir = nir_shader_clone(NULL, v->shader->nir);
+
+   nir_shader *s = c->nir;
+   const struct etna_specs *specs = c->specs;
+
+   v->stage = s->info.stage;
+   v->num_loops = 0; /* TODO */
+   v->vs_id_in_reg = -1;
+   v->vs_pos_out_reg = -1;
+   v->vs_pointsize_out_reg = -1;
+   v->ps_color_out_reg = 0; /* 0 for shader that doesn't write fragcolor.. */
+   v->ps_depth_out_reg = -1;
+
+   /* setup input linking */
+   struct etna_shader_io_file *sf = &v->infile;
+   if (s->info.stage == MESA_SHADER_VERTEX) {
+      nir_foreach_variable(var, &s->inputs) {
+         unsigned idx = var->data.driver_location;
+         sf->reg[idx].reg = idx;
+         sf->reg[idx].slot = var->data.location;
+         sf->reg[idx].num_components = 4; /* TODO */
+         sf->num_reg = MAX2(sf->num_reg, idx+1);
+      }
+   } else {
+      unsigned count = 0;
+      nir_foreach_variable(var, &s->inputs) {
+         unsigned idx = var->data.driver_location;
+         sf->reg[idx].reg = idx + 1;
+         sf->reg[idx].slot = var->data.location;
+         sf->reg[idx].num_components = 4; /* TODO */
+         sf->num_reg = MAX2(sf->num_reg, idx+1);
+         count++;
+      }
+      assert(sf->num_reg == count);
+   }
+
+   NIR_PASS_V(s, nir_lower_io, nir_var_all, etna_glsl_type_size,
+            (nir_lower_io_options)0);
+
+   OPT_V(s, nir_lower_regs_to_ssa);
+   OPT_V(s, nir_lower_vars_to_ssa);
+   OPT_V(s, nir_lower_indirect_derefs, nir_var_all);
+   OPT_V(s, nir_lower_tex, &(struct nir_lower_tex_options) { .lower_txp = ~0u });
+   OPT_V(s, etna_lower_alu_to_scalar, specs);
+
+   etna_optimize_loop(s);
+
+   /* use opt_algebraic between int_to_float and boot_to_float because
+    * int_to_float emits ftrunc, and ftrunc lowering generates bool ops
+    */
+   OPT_V(s, nir_lower_int_to_float);
+   OPT_V(s, nir_opt_algebraic);
+   OPT_V(s, nir_lower_bool_to_float);
+
+   /* after int to float because insert i2f for instance_id */
+   OPT_V(s, etna_lower_io, v);
+
+   etna_optimize_loop(s);
+
+   if (DBG_ENABLED(ETNA_DBG_DUMP_SHADERS))
+      nir_print_shader(s, stdout);
+
+   while( OPT(s, nir_opt_vectorize) );
+   OPT_V(s, etna_lower_alu_to_scalar, specs);
+
+   NIR_PASS_V(s, nir_remove_dead_variables, nir_var_function_temp);
+   NIR_PASS_V(s, nir_opt_algebraic_late);
+
+   NIR_PASS_V(s, nir_move_vec_src_uses_to_dest);
+   NIR_PASS_V(s, nir_copy_prop);
+   NIR_PASS_V(s, nir_lower_to_source_mods, ~nir_lower_int_source_mods);
+   /* need copy prop after uses_to_dest, and before src mods: see
+    * dEQP-GLES2.functional.shaders.random.all_features.fragment.95
+    */
+
+   NIR_PASS_V(s, nir_opt_dce);
+
+   NIR_PASS_V(s, etna_lower_alu, c);
+
+   if (DBG_ENABLED(ETNA_DBG_DUMP_SHADERS))
+      nir_print_shader(s, stdout);
+
+   uint64_t consts[ETNA_MAX_IMM] = {};
+
+   unsigned block_ptr[nir_shader_get_entrypoint(s)->num_blocks];
+   c->block_ptr = block_ptr;
+   struct emit_options options = {
+      .max_temps = ETNA_MAX_TEMPS,
+      .max_consts = ETNA_MAX_IMM / 4,
+      .id_reg = sf->num_reg,
+      .single_const_src = c->specs->halti < 5,
+      .etna_new_transcendentals = c->specs->has_new_transcendentals,
+      .user = c,
+      .consts = consts,
+   };
+
+   unsigned num_consts;
+   bool ok = emit_shader(c->nir, &options, &v->num_temps, &num_consts);
+   assert(ok);
+
+   /* empty shader, emit NOP */
+   if (!c->inst_ptr)
+      emit_inst(c, &(struct etna_inst) { .opcode = INST_OPCODE_NOP });
+
+   /* assemble instructions, fixing up labels */
+   uint32_t *code = MALLOC(c->inst_ptr * 16 + 1024);
+   for (unsigned i = 0; i < c->inst_ptr; i++) {
+      struct etna_inst *inst = &c->code[i];
+      if (inst->opcode == INST_OPCODE_BRANCH)
+         inst->imm = block_ptr[inst->imm];
+
+      inst->halti5 = specs->halti >= 5;
+      etna_assemble(&code[i * 4], inst);
+   }
+
+   v->code_size = c->inst_ptr * 4;
+   v->code = code;
+   v->needs_icache = c->inst_ptr > specs->max_instructions;
+
+   copy_uniform_state_to_shader(v, consts, num_consts);
+
+   if (s->info.stage == MESA_SHADER_FRAGMENT) {
+      v->input_count_unk8 = 31; /* XXX what is this */
+
+      nir_foreach_variable(var, &s->outputs) {
+         unsigned reg = c->output_reg[var->data.driver_location];
+         switch (var->data.location) {
+         case FRAG_RESULT_COLOR:
+         case FRAG_RESULT_DATA0: /* DATA0 is used by gallium shaders for color */
+            v->ps_color_out_reg = reg;
+            break;
+         case FRAG_RESULT_DEPTH:
+            v->ps_depth_out_reg = reg;
+            break;
+         default:
+            compile_error(c, "Unsupported fs output %s\n", gl_frag_result_name(var->data.location));
+         }
+      }
+      assert(v->ps_depth_out_reg <= 0);
+      v->outfile.num_reg = 0;
+      ralloc_free(c->nir);
+      FREE(c);
+      return true;
+   }
+
+   v->input_count_unk8 = DIV_ROUND_UP(v->infile.num_reg + 4, 16); /* XXX what is this */
+
+   sf = &v->outfile;
+   sf->num_reg = 0;
+   nir_foreach_variable(var, &s->outputs) {
+      unsigned native = c->output_reg[var->data.driver_location];
+
+      if (var->data.location == VARYING_SLOT_POS) {
+         v->vs_pos_out_reg = native;
+         continue;
+      }
+
+      if (var->data.location == VARYING_SLOT_PSIZ) {
+         v->vs_pointsize_out_reg = native;
+         continue;
+      }
+
+      sf->reg[sf->num_reg].reg = native;
+      sf->reg[sf->num_reg].slot = var->data.location;
+      sf->reg[sf->num_reg].num_components = 4; /* TODO */
+      sf->num_reg++;
+   }
+
+   /* fill in "mystery meat" load balancing value. This value determines how
+    * work is scheduled between VS and PS
+    * in the unified shader architecture. More precisely, it is determined from
+    * the number of VS outputs, as well as chip-specific
+    * vertex output buffer size, vertex cache size, and the number of shader
+    * cores.
+    *
+    * XXX this is a conservative estimate, the "optimal" value is only known for
+    * sure at link time because some
+    * outputs may be unused and thus unmapped. Then again, in the general use
+    * case with GLSL the vertex and fragment
+    * shaders are linked already before submitting to Gallium, thus all outputs
+    * are used.
+    *
+    * note: TGSI compiler counts all outputs (including position and pointsize), here
+    * v->outfile.num_reg only counts varyings, +1 to compensate for the position output
+    * TODO: might have a problem that we don't count pointsize when it is used
+    */
+
+   int half_out = v->outfile.num_reg / 2 + 1;
+   assert(half_out);
+
+   uint32_t b = ((20480 / (specs->vertex_output_buffer_size -
+                           2 * half_out * specs->vertex_cache_size)) +
+                 9) /
+                10;
+   uint32_t a = (b + 256 / (specs->shader_core_count * half_out)) / 2;
+   v->vs_load_balancing = VIVS_VS_LOAD_BALANCING_A(MIN2(a, 255)) |
+                             VIVS_VS_LOAD_BALANCING_B(MIN2(b, 255)) |
+                             VIVS_VS_LOAD_BALANCING_C(0x3f) |
+                             VIVS_VS_LOAD_BALANCING_D(0x0f);
+
+   ralloc_free(c->nir);
+   FREE(c);
+   return true;
+}
+
+void
+etna_destroy_shader_nir(struct etna_shader_variant *shader)
+{
+   assert(shader);
+
+   FREE(shader->code);
+   FREE(shader->uniforms.imm_data);
+   FREE(shader->uniforms.imm_contents);
+   FREE(shader);
+}
+
+static const struct etna_shader_inout *
+etna_shader_vs_lookup(const struct etna_shader_variant *sobj,
+                      const struct etna_shader_inout *in)
+{
+   for (int i = 0; i < sobj->outfile.num_reg; i++)
+      if (sobj->outfile.reg[i].slot == in->slot)
+         return &sobj->outfile.reg[i];
+
+   return NULL;
+}
+
+bool
+etna_link_shader_nir(struct etna_shader_link_info *info,
+                     const struct etna_shader_variant *vs,
+                     const struct etna_shader_variant *fs)
+{
+   int comp_ofs = 0;
+   /* For each fragment input we need to find the associated vertex shader
+    * output, which can be found by matching on semantic name and index. A
+    * binary search could be used because the vs outputs are sorted by their
+    * semantic index and grouped by semantic type by fill_in_vs_outputs.
+    */
+   assert(fs->infile.num_reg < ETNA_NUM_INPUTS);
+   info->pcoord_varying_comp_ofs = -1;
+
+   for (int idx = 0; idx < fs->infile.num_reg; ++idx) {
+      const struct etna_shader_inout *fsio = &fs->infile.reg[idx];
+      const struct etna_shader_inout *vsio = etna_shader_vs_lookup(vs, fsio);
+      struct etna_varying *varying;
+      bool interpolate_always = true;
+
+      assert(fsio->reg > 0 && fsio->reg <= ARRAY_SIZE(info->varyings));
+
+      if (fsio->reg > info->num_varyings)
+         info->num_varyings = fsio->reg;
+
+      varying = &info->varyings[fsio->reg - 1];
+      varying->num_components = fsio->num_components;
+
+      if (!interpolate_always) /* colors affected by flat shading */
+         varying->pa_attributes = 0x200;
+      else /* texture coord or other bypasses flat shading */
+         varying->pa_attributes = 0x2f1;
+
+      varying->use[0] = VARYING_COMPONENT_USE_UNUSED;
+      varying->use[1] = VARYING_COMPONENT_USE_UNUSED;
+      varying->use[2] = VARYING_COMPONENT_USE_UNUSED;
+      varying->use[3] = VARYING_COMPONENT_USE_UNUSED;
+
+      /* point coord is an input to the PS without matching VS output,
+       * so it gets a varying slot without being assigned a VS register.
+       */
+      if (fsio->slot == VARYING_SLOT_PNTC) {
+         varying->use[0] = VARYING_COMPONENT_USE_POINTCOORD_X;
+         varying->use[1] = VARYING_COMPONENT_USE_POINTCOORD_Y;
+
+         info->pcoord_varying_comp_ofs = comp_ofs;
+      } else {
+         if (vsio == NULL) { /* not found -- link error */
+            BUG("Semantic value not found in vertex shader outputs\n");
+            return true;
+         }
+         varying->reg = vsio->reg;
+      }
+
+      comp_ofs += varying->num_components;
+   }
+
+   assert(info->num_varyings == fs->infile.num_reg);
+
+   return false;
+}
diff --git a/src/gallium/drivers/etnaviv/etnaviv_compiler_nir_emit.h b/src/gallium/drivers/etnaviv/etnaviv_compiler_nir_emit.h
new file mode 100644
index 00000000000..0d30fe54dfd
--- /dev/null
+++ b/src/gallium/drivers/etnaviv/etnaviv_compiler_nir_emit.h
@@ -0,0 +1,1396 @@
+/*
+ * Copyright (c) 2019 Zodiac Inflight Innovations
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sub license,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ *
+ * Authors:
+ *    Jonathan Marek <[email protected]>
+ */
+
+#include "etnaviv_asm.h"
+#include "etnaviv_context.h"
+
+#include "compiler/nir/nir.h"
+#include "compiler/nir/nir_builder.h"
+#include "compiler/nir/nir_worklist.h"
+#include "util/register_allocate.h"
+
+struct emit_options {
+   unsigned max_temps; /* max # of vec4 registers */
+   unsigned max_consts; /* max # of vec4 consts */
+   unsigned id_reg; /* register with vertex/instance id */
+   bool single_const_src : 1; /* limited to 1 vec4 const src */
+   bool etna_new_transcendentals : 1;
+   void *user;
+   uint64_t *consts;
+};
+
+#define ALU_SWIZ(s) INST_SWIZ((s)->swizzle[0], (s)->swizzle[1], (s)->swizzle[2], (s)->swizzle[3])
+#define SRC_DISABLE ((hw_src){})
+#define SRC_CONST(idx, s) ((hw_src){.use=1, .rgroup = INST_RGROUP_UNIFORM_0, .reg=idx, .swiz=s})
+#define SRC_REG(idx, s) ((hw_src){.use=1, .rgroup = INST_RGROUP_TEMP, .reg=idx, .swiz=s})
+
+#define option(name) (state->options->name)
+#define emit(type, args...) etna_emit_##type(state->options->user, args)
+
+typedef struct etna_inst_dst hw_dst;
+typedef struct etna_inst_src hw_src;
+
+enum {
+   BYPASS_DST = 1,
+   BYPASS_SRC = 2,
+};
+
+struct state {
+   const struct emit_options *options;
+   unsigned const_count;
+
+   nir_shader *shader;
+   nir_function_impl *impl;
+
+   /* ra state */
+   struct ra_graph *g;
+   struct ra_regs *regs;
+   unsigned *live_map;
+   unsigned num_nodes;
+};
+
+static inline hw_src
+src_swizzle(hw_src src, unsigned swizzle)
+{
+   src.swiz = inst_swiz_compose(src.swiz, swizzle);
+   return src;
+}
+
+static inline bool is_sysval(nir_instr *instr)
+{
+   if (instr->type != nir_instr_type_intrinsic)
+      return false;
+
+   nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+   return intr->intrinsic == nir_intrinsic_load_front_face ||
+          intr->intrinsic == nir_intrinsic_load_frag_coord;
+}
+
+/* constants are represented as 64-bit ints
+ * 32-bit for the value and 32-bit for the type (imm, uniform, etc)
+ */
+
+#define CONST_VAL(a, b) (nir_const_value) {.u64 = (uint64_t)(a) << 32 | (uint64_t)(b)}
+#define CONST(x) CONST_VAL(ETNA_IMMEDIATE_CONSTANT, x)
+#define UNIFORM(x) CONST_VAL(ETNA_IMMEDIATE_UNIFORM, x)
+#define UNIFORM_BASE(x) CONST_VAL(ETNA_IMMEDIATE_UBO0_ADDR, x)
+#define TEXSCALE(x, i) CONST_VAL(ETNA_IMMEDIATE_TEXRECT_SCALE_X + (i), x)
+
+static int
+const_add(uint64_t *c, uint64_t value)
+{
+   for (unsigned i = 0; i < 4; i++) {
+      if (c[i] == value || !c[i]) {
+         c[i] = value;
+         return i;
+      }
+   }
+   return -1;
+}
+
+static hw_src
+const_src(struct state *state, nir_const_value *value, unsigned num_components)
+{
+   unsigned i;
+   int swiz = -1;
+   for (i = 0; swiz < 0; i++) {
+      uint64_t *a = &option(consts)[i*4];
+      uint64_t save[4];
+      memcpy(save, a, sizeof(save));
+      swiz = 0;
+      for (unsigned j = 0; j < num_components; j++) {
+         int c = const_add(a, value[j].u64);
+         if (c < 0) {
+            memcpy(a, save, sizeof(save));
+            swiz = -1;
+            break;
+         }
+         swiz |= c << j * 2;
+      }
+   }
+
+   assert(i <= option(max_consts));
+   state->const_count = MAX2(state->const_count, i);
+
+   return SRC_CONST(i - 1, swiz);
+}
+
+struct ssa_reg {
+   uint8_t idx;
+   uint8_t src_swizzle;
+   uint8_t dst_swizzle;
+   uint8_t write_mask;
+};
+
+/* Swizzles and write masks can be used to layer virtual non-interfering
+ * registers on top of the real VEC4 registers. For example, the virtual
+ * VEC3_XYZ register and the virtual SCALAR_W register that use the same
+ * physical VEC4 base register do not interfere.
+ */
+enum {
+   REG_CLASS_VIRT_SCALAR,
+   REG_CLASS_VIRT_VEC2,
+   REG_CLASS_VIRT_VEC3,
+   REG_CLASS_VEC4,
+   /* special vec2 class for fast transcendentals, limited to XY or ZW */
+   REG_CLASS_VIRT_VEC2T,
+   NUM_REG_CLASSES,
+} reg_class;
+
+enum {
+   REG_TYPE_VEC4,
+   REG_TYPE_VIRT_VEC3_XYZ,
+   REG_TYPE_VIRT_VEC3_XYW,
+   REG_TYPE_VIRT_VEC3_XZW,
+   REG_TYPE_VIRT_VEC3_YZW,
+   REG_TYPE_VIRT_VEC2_XY,
+   REG_TYPE_VIRT_VEC2_XZ,
+   REG_TYPE_VIRT_VEC2_XW,
+   REG_TYPE_VIRT_VEC2_YZ,
+   REG_TYPE_VIRT_VEC2_YW,
+   REG_TYPE_VIRT_VEC2_ZW,
+   REG_TYPE_VIRT_SCALAR_X,
+   REG_TYPE_VIRT_SCALAR_Y,
+   REG_TYPE_VIRT_SCALAR_Z,
+   REG_TYPE_VIRT_SCALAR_W,
+   REG_TYPE_VIRT_VEC2T_XY,
+   REG_TYPE_VIRT_VEC2T_ZW,
+   NUM_REG_TYPES,
+} reg_type;
+
+/* writemask when used as dest */
+static const uint8_t
+reg_writemask[NUM_REG_TYPES] = {
+   [REG_TYPE_VEC4] = 0xf,
+   [REG_TYPE_VIRT_SCALAR_X] = 0x1,
+   [REG_TYPE_VIRT_SCALAR_Y] = 0x2,
+   [REG_TYPE_VIRT_VEC2_XY] = 0x3,
+   [REG_TYPE_VIRT_VEC2T_XY] = 0x3,
+   [REG_TYPE_VIRT_SCALAR_Z] = 0x4,
+   [REG_TYPE_VIRT_VEC2_XZ] = 0x5,
+   [REG_TYPE_VIRT_VEC2_YZ] = 0x6,
+   [REG_TYPE_VIRT_VEC3_XYZ] = 0x7,
+   [REG_TYPE_VIRT_SCALAR_W] = 0x8,
+   [REG_TYPE_VIRT_VEC2_XW] = 0x9,
+   [REG_TYPE_VIRT_VEC2_YW] = 0xa,
+   [REG_TYPE_VIRT_VEC3_XYW] = 0xb,
+   [REG_TYPE_VIRT_VEC2_ZW] = 0xc,
+   [REG_TYPE_VIRT_VEC2T_ZW] = 0xc,
+   [REG_TYPE_VIRT_VEC3_XZW] = 0xd,
+   [REG_TYPE_VIRT_VEC3_YZW] = 0xe,
+};
+
+/* how to swizzle when used as a src */
+static const uint8_t
+reg_swiz[NUM_REG_TYPES] = {
+   [REG_TYPE_VEC4] = INST_SWIZ_IDENTITY,
+   [REG_TYPE_VIRT_SCALAR_X] = INST_SWIZ_IDENTITY,
+   [REG_TYPE_VIRT_SCALAR_Y] = SWIZZLE(Y, Y, Y, Y),
+   [REG_TYPE_VIRT_VEC2_XY] = INST_SWIZ_IDENTITY,
+   [REG_TYPE_VIRT_VEC2T_XY] = INST_SWIZ_IDENTITY,
+   [REG_TYPE_VIRT_SCALAR_Z] = SWIZZLE(Z, Z, Z, Z),
+   [REG_TYPE_VIRT_VEC2_XZ] = SWIZZLE(X, Z, X, Z),
+   [REG_TYPE_VIRT_VEC2_YZ] = SWIZZLE(Y, Z, Y, Z),
+   [REG_TYPE_VIRT_VEC3_XYZ] = INST_SWIZ_IDENTITY,
+   [REG_TYPE_VIRT_SCALAR_W] = SWIZZLE(W, W, W, W),
+   [REG_TYPE_VIRT_VEC2_XW] = SWIZZLE(X, W, X, W),
+   [REG_TYPE_VIRT_VEC2_YW] = SWIZZLE(Y, W, Y, W),
+   [REG_TYPE_VIRT_VEC3_XYW] = SWIZZLE(X, Y, W, X),
+   [REG_TYPE_VIRT_VEC2_ZW] = SWIZZLE(Z, W, Z, W),
+   [REG_TYPE_VIRT_VEC2T_ZW] = SWIZZLE(Z, W, Z, W),
+   [REG_TYPE_VIRT_VEC3_XZW] = SWIZZLE(X, Z, W, X),
+   [REG_TYPE_VIRT_VEC3_YZW] = SWIZZLE(Y, Z, W, X),
+};
+
+/* how to swizzle when used as a dest */
+static const uint8_t
+reg_dst_swiz[NUM_REG_TYPES] = {
+   [REG_TYPE_VEC4] = INST_SWIZ_IDENTITY,
+   [REG_TYPE_VIRT_SCALAR_X] = INST_SWIZ_IDENTITY,
+   [REG_TYPE_VIRT_SCALAR_Y] = SWIZZLE(X, X, X, X),
+   [REG_TYPE_VIRT_VEC2_XY] = INST_SWIZ_IDENTITY,
+   [REG_TYPE_VIRT_VEC2T_XY] = INST_SWIZ_IDENTITY,
+   [REG_TYPE_VIRT_SCALAR_Z] = SWIZZLE(X, X, X, X),
+   [REG_TYPE_VIRT_VEC2_XZ] = SWIZZLE(X, X, Y, Y),
+   [REG_TYPE_VIRT_VEC2_YZ] = SWIZZLE(X, X, Y, Y),
+   [REG_TYPE_VIRT_VEC3_XYZ] = INST_SWIZ_IDENTITY,
+   [REG_TYPE_VIRT_SCALAR_W] = SWIZZLE(X, X, X, X),
+   [REG_TYPE_VIRT_VEC2_XW] = SWIZZLE(X, X, Y, Y),
+   [REG_TYPE_VIRT_VEC2_YW] = SWIZZLE(X, X, Y, Y),
+   [REG_TYPE_VIRT_VEC3_XYW] = SWIZZLE(X, Y, Z, Z),
+   [REG_TYPE_VIRT_VEC2_ZW] = SWIZZLE(X, X, X, Y),
+   [REG_TYPE_VIRT_VEC2T_ZW] = SWIZZLE(X, X, X, Y),
+   [REG_TYPE_VIRT_VEC3_XZW] = SWIZZLE(X, Y, Y, Z),
+   [REG_TYPE_VIRT_VEC3_YZW] = SWIZZLE(X, X, Y, Z),
+};
+
+static inline int reg_get_type(int virt_reg)
+{
+   return virt_reg % NUM_REG_TYPES;
+}
+
+static inline int reg_get_base(struct state *state, int virt_reg)
+{
+   /* offset by 1 to avoid reserved position register */
+   if (state->shader->info.stage == MESA_SHADER_FRAGMENT)
+      return virt_reg / NUM_REG_TYPES + 1;
+   return virt_reg / NUM_REG_TYPES;
+}
+
+static inline int reg_get_class(int virt_reg)
+{
+   switch (reg_get_type(virt_reg)) {
+   case REG_TYPE_VEC4:
+      return REG_CLASS_VEC4;
+   case REG_TYPE_VIRT_VEC3_XYZ:
+   case REG_TYPE_VIRT_VEC3_XYW:
+   case REG_TYPE_VIRT_VEC3_XZW:
+   case REG_TYPE_VIRT_VEC3_YZW:
+      return REG_CLASS_VIRT_VEC3;
+   case REG_TYPE_VIRT_VEC2_XY:
+   case REG_TYPE_VIRT_VEC2_XZ:
+   case REG_TYPE_VIRT_VEC2_XW:
+   case REG_TYPE_VIRT_VEC2_YZ:
+   case REG_TYPE_VIRT_VEC2_YW:
+   case REG_TYPE_VIRT_VEC2_ZW:
+      return REG_CLASS_VIRT_VEC2;
+   case REG_TYPE_VIRT_SCALAR_X:
+   case REG_TYPE_VIRT_SCALAR_Y:
+   case REG_TYPE_VIRT_SCALAR_Z:
+   case REG_TYPE_VIRT_SCALAR_W:
+      return REG_CLASS_VIRT_SCALAR;
+   case REG_TYPE_VIRT_VEC2T_XY:
+   case REG_TYPE_VIRT_VEC2T_ZW:
+      return REG_CLASS_VIRT_VEC2T;
+   }
+
+   assert(false);
+   return 0;
+}
+
+/* get unique ssa/reg index for nir_src */
+static unsigned
+src_index(nir_function_impl *impl, nir_src *src)
+{
+   return src->is_ssa ? src->ssa->index : (src->reg.reg->index + impl->ssa_alloc);
+}
+
+/* get unique ssa/reg index for nir_dest */
+static unsigned
+dest_index(nir_function_impl *impl, nir_dest *dest)
+{
+   return dest->is_ssa ? dest->ssa.index : (dest->reg.reg->index + impl->ssa_alloc);
+}
+
+/* nir_src to allocated register */
+static hw_src
+ra_src(struct state *state, nir_src *src)
+{
+   unsigned reg = ra_get_node_reg(state->g, state->live_map[src_index(state->impl, src)]);
+   return SRC_REG(reg_get_base(state, reg), reg_swiz[reg_get_type(reg)]);
+}
+
+static hw_src
+get_src(struct state *state, nir_src *src)
+{
+   if (!src->is_ssa)
+      return ra_src(state, src);
+
+   nir_instr *instr = src->ssa->parent_instr;
+
+   if (instr->pass_flags & BYPASS_SRC) {
+      assert(instr->type == nir_instr_type_alu);
+      nir_alu_instr *alu = nir_instr_as_alu(instr);
+      assert(alu->op == nir_op_mov);
+      return src_swizzle(get_src(state, &alu->src[0].src), ALU_SWIZ(&alu->src[0]));
+   }
+
+   switch (instr->type) {
+   case nir_instr_type_load_const:
+      return const_src(state, nir_instr_as_load_const(instr)->value, src->ssa->num_components);
+   case nir_instr_type_intrinsic: {
+      nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+      switch (intr->intrinsic) {
+      case nir_intrinsic_load_input:
+      case nir_intrinsic_load_instance_id:
+      case nir_intrinsic_load_uniform:
+         return ra_src(state, src);
+      case nir_intrinsic_load_front_face:
+         return (hw_src) { .use = 1, .rgroup = INST_RGROUP_INTERNAL };
+      case nir_intrinsic_load_frag_coord:
+         return SRC_REG(0, INST_SWIZ_IDENTITY);
+      default:
+         assert(0);
+         break;
+      }
+   } break;
+   case nir_instr_type_alu:
+   case nir_instr_type_tex:
+      return ra_src(state, src);
+   case nir_instr_type_ssa_undef: {
+      /* return zero to deal with broken Blur demo */
+      nir_const_value value = CONST(0);
+      return src_swizzle(const_src(state, &value, 1), SWIZZLE(X,X,X,X));
+   }
+   default:
+      assert(0);
+      break;
+   }
+
+   return SRC_DISABLE;
+}
+
+static void
+update_swiz_mask(nir_alu_instr *alu, nir_dest *dest, unsigned *swiz, unsigned *mask)
+{
+   if (!swiz)
+      return;
+
+   bool is_vec = dest != NULL;
+   unsigned swizzle = 0, write_mask = 0;
+   for (unsigned i = 0; i < 4; i++) {
+      /* channel not written */
+      if (!(alu->dest.write_mask & (1 << i)))
+         continue;
+      /* src is different (only check for vecN) */
+      if (is_vec && alu->src[i].src.ssa != &dest->ssa)
+         continue;
+
+      unsigned src_swiz = is_vec ? alu->src[i].swizzle[0] : alu->src[0].swizzle[i];
+      swizzle |= (*swiz >> src_swiz * 2 & 3) << i * 2;
+      /* this channel isn't written through this chain */
+      if (*mask & (1 << src_swiz))
+         write_mask |= 1 << i;
+   }
+   *swiz = swizzle;
+   *mask = write_mask;
+}
+
+static bool
+vec_dest_has_swizzle(nir_alu_instr *vec, nir_ssa_def *ssa)
+{
+   for (unsigned i = 0; i < 4; i++) {
+      if (!(vec->dest.write_mask & (1 << i)) || vec->src[i].src.ssa != ssa)
+         continue;
+
+      if (vec->src[i].swizzle[0] != i)
+         return true;
+   }
+
+   /* don't deal with possible bypassed vec/mov chain */
+   nir_foreach_use(use_src, ssa) {
+      nir_instr *instr = use_src->parent_instr;
+      if (instr->type != nir_instr_type_alu)
+         continue;
+
+      nir_alu_instr *alu = nir_instr_as_alu(instr);
+
+      switch (alu->op) {
+      case nir_op_mov:
+      case nir_op_vec2:
+      case nir_op_vec3:
+      case nir_op_vec4:
+         return true;
+      default:
+         break;
+      }
+   }
+   return false;
+}
+
+static nir_dest *
+real_dest(nir_dest *dest, unsigned *swiz, unsigned *mask)
+{
+   if (!dest || !dest->is_ssa)
+      return dest;
+
+   bool can_bypass_src = !list_length(&dest->ssa.if_uses);
+   nir_instr *p_instr = dest->ssa.parent_instr;
+
+   /* if used by a vecN, the "real" destination becomes the vecN destination
+    * lower_alu guarantees that values used by a vecN are only used by that vecN
+    * we can apply the same logic to movs in a some cases too
+    */
+   nir_foreach_use(use_src, &dest->ssa) {
+      nir_instr *instr = use_src->parent_instr;
+
+      /* src bypass check: for now only deal with tex src mov case
+       * note: for alu don't bypass mov for multiple uniform sources
+       */
+      switch (instr->type) {
+      case nir_instr_type_tex:
+         if (p_instr->type == nir_instr_type_alu &&
+             nir_instr_as_alu(p_instr)->op == nir_op_mov) {
+            break;
+         }
+      default:
+         can_bypass_src = false;
+         break;
+      }
+
+      if (instr->type != nir_instr_type_alu)
+         continue;
+
+      nir_alu_instr *alu = nir_instr_as_alu(instr);
+
+      switch (alu->op) {
+      case nir_op_vec2:
+      case nir_op_vec3:
+      case nir_op_vec4:
+         assert(list_length(&dest->ssa.if_uses) == 0);
+         nir_foreach_use(use_src, &dest->ssa)
+            assert(use_src->parent_instr == instr);
+
+         update_swiz_mask(alu, dest, swiz, mask);
+         break;
+      case nir_op_mov: {
+         switch (dest->ssa.parent_instr->type) {
+         case nir_instr_type_alu:
+         case nir_instr_type_tex:
+            break;
+         default:
+            continue;
+         }
+         if (list_length(&dest->ssa.if_uses) || list_length(&dest->ssa.uses) > 1)
+            continue;
+
+         update_swiz_mask(alu, NULL, swiz, mask);
+         break;
+      };
+      default:
+         continue;
+      }
+
+      assert(!(instr->pass_flags & BYPASS_SRC));
+      instr->pass_flags |= BYPASS_DST;
+      return real_dest(&alu->dest.dest, swiz, mask);
+   }
+
+   if (can_bypass_src && !(p_instr->pass_flags & BYPASS_DST)) {
+      p_instr->pass_flags |= BYPASS_SRC;
+      return NULL;
+   }
+
+   return dest;
+}
+
+/* get allocated dest register for nir_dest
+ * *p_swiz tells how the components need to be placed into register
+ */
+static hw_dst
+ra_dest(struct state *state, nir_dest *dest, unsigned *p_swiz)
+{
+   unsigned swiz = INST_SWIZ_IDENTITY, mask = 0xf;
+   dest = real_dest(dest, &swiz, &mask);
+
+   unsigned r = ra_get_node_reg(state->g, state->live_map[dest_index(state->impl, dest)]);
+   unsigned t = reg_get_type(r);
+
+   *p_swiz = inst_swiz_compose(swiz, reg_dst_swiz[t]);
+
+   return (hw_dst) {
+      .use = 1,
+      .reg = reg_get_base(state, r),
+      .write_mask = inst_write_mask_compose(mask, reg_writemask[t]),
+   };
+}
+
+/* if instruction dest needs a register, return nir_dest for it */
+static nir_dest *
+dest_for_instr(nir_instr *instr)
+{
+   nir_dest *dest = NULL;
+
+   switch (instr->type) {
+   case nir_instr_type_alu:
+      dest = &nir_instr_as_alu(instr)->dest.dest;
+      break;
+   case nir_instr_type_tex:
+      dest =&nir_instr_as_tex(instr)->dest;
+      break;
+   case nir_instr_type_intrinsic: {
+      nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+      if (intr->intrinsic == nir_intrinsic_load_uniform ||
+          intr->intrinsic == nir_intrinsic_load_input ||
+          intr->intrinsic == nir_intrinsic_load_instance_id)
+         dest = &intr->dest;
+   }
+   default:
+      break;
+   }
+   return real_dest(dest, NULL, NULL);
+}
+
+struct live_def {
+   nir_instr *instr;
+   nir_dest *dest; /* cached dest_for_instr */
+   unsigned live_start, live_end; /* live range */
+};
+
+static void
+range_include(struct live_def *def, unsigned index)
+{
+   if (def->live_start > index)
+      def->live_start = index;
+   if (def->live_end < index)
+      def->live_end = index;
+}
+
+struct live_defs_state {
+   unsigned num_defs;
+   unsigned bitset_words;
+
+   nir_function_impl *impl;
+   nir_block *block; /* current block pointer */
+   unsigned index; /* current live index */
+
+   struct live_def *defs;
+   unsigned *live_map; /* to map ssa/reg index into defs array */
+
+   nir_block_worklist worklist;
+};
+
+static bool
+init_liveness_block(nir_block *block,
+                    struct live_defs_state *state)
+{
+   block->live_in = reralloc(block, block->live_in, BITSET_WORD,
+                             state->bitset_words);
+   memset(block->live_in, 0, state->bitset_words * sizeof(BITSET_WORD));
+
+   block->live_out = reralloc(block, block->live_out, BITSET_WORD,
+                              state->bitset_words);
+   memset(block->live_out, 0, state->bitset_words * sizeof(BITSET_WORD));
+
+   nir_block_worklist_push_head(&state->worklist, block);
+
+   return true;
+}
+
+static bool
+set_src_live(nir_src *src, void *void_state)
+{
+   struct live_defs_state *state = void_state;
+
+   if (src->is_ssa) {
+      nir_instr *instr = src->ssa->parent_instr;
+
+      if (is_sysval(instr))
+         return true;
+
+      switch (instr->type) {
+      case nir_instr_type_load_const:
+      case nir_instr_type_ssa_undef:
+         return true;
+      case nir_instr_type_alu: {
+         /* alu op bypass */
+         nir_alu_instr *alu = nir_instr_as_alu(instr);
+         if (instr->pass_flags & BYPASS_SRC) {
+            for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++)
+               set_src_live(&alu->src[i].src, state);
+            return true;
+         }
+      } break;
+      default:
+         break;
+      }
+   }
+
+   unsigned i = state->live_map[src_index(state->impl, src)];
+   assert(i != ~0u);
+
+   BITSET_SET(state->block->live_in, i);
+   range_include(&state->defs[i], state->index);
+
+   return true;
+}
+
+static bool
+propagate_across_edge(nir_block *pred, nir_block *succ,
+                      struct live_defs_state *state)
+{
+   BITSET_WORD progress = 0;
+   for (unsigned i = 0; i < state->bitset_words; ++i) {
+      progress |= succ->live_in[i] & ~pred->live_out[i];
+      pred->live_out[i] |= succ->live_in[i];
+   }
+   return progress != 0;
+}
+
+static unsigned
+live_defs(nir_function_impl *impl, struct live_def *defs, unsigned *live_map)
+{
+   struct live_defs_state state;
+   unsigned block_live_index[impl->num_blocks + 1];
+
+   state.impl = impl;
+   state.defs = defs;
+   state.live_map = live_map;
+
+   state.num_defs = 0;
+   nir_foreach_block(block, impl) {
+      block_live_index[block->index] = state.num_defs;
+      nir_foreach_instr(instr, block) {
+         nir_dest *dest = dest_for_instr(instr);
+         if (!dest)
+            continue;
+
+         unsigned idx = dest_index(impl, dest);
+         /* register is already in defs */
+         if (live_map[idx] != ~0u)
+            continue;
+
+         defs[state.num_defs] = (struct live_def) {instr, dest, state.num_defs, 0};
+
+         /* input live from the start */
+         if (instr->type == nir_instr_type_intrinsic) {
+            nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+            if (intr->intrinsic == nir_intrinsic_load_input ||
+                intr->intrinsic == nir_intrinsic_load_instance_id)
+               defs[state.num_defs].live_start = 0;
+         }
+
+         live_map[idx] = state.num_defs;
+         state.num_defs++;
+      }
+   }
+   block_live_index[impl->num_blocks] = state.num_defs;
+
+   nir_block_worklist_init(&state.worklist, impl->num_blocks, NULL);
+
+   /* We now know how many unique ssa definitions we have and we can go
+    * ahead and allocate live_in and live_out sets and add all of the
+    * blocks to the worklist.
+    */
+   state.bitset_words = BITSET_WORDS(state.num_defs);
+   nir_foreach_block(block, impl) {
+      init_liveness_block(block, &state);
+   }
+
+   /* We're now ready to work through the worklist and update the liveness
+    * sets of each of the blocks.  By the time we get to this point, every
+    * block in the function implementation has been pushed onto the
+    * worklist in reverse order.  As long as we keep the worklist
+    * up-to-date as we go, everything will get covered.
+    */
+   while (!nir_block_worklist_is_empty(&state.worklist)) {
+      /* We pop them off in the reverse order we pushed them on.  This way
+       * the first walk of the instructions is backwards so we only walk
+       * once in the case of no control flow.
+       */
+      nir_block *block = nir_block_worklist_pop_head(&state.worklist);
+      state.block = block;
+
+      memcpy(block->live_in, block->live_out,
+             state.bitset_words * sizeof(BITSET_WORD));
+
+      state.index = block_live_index[block->index + 1];
+
+      nir_if *following_if = nir_block_get_following_if(block);
+      if (following_if)
+         set_src_live(&following_if->condition, &state);
+
+      nir_foreach_instr_reverse(instr, block) {
+         /* when we come across the next "live" instruction, decrement index */
+         if (state.index && instr == defs[state.index - 1].instr) {
+            state.index--;
+            /* the only source of writes to registers is phis:
+             * we don't expect any partial write_mask alus
+             * so clearing live_in here is OK
+             */
+            BITSET_CLEAR(block->live_in, state.index);
+         }
+
+         /* don't set_src_live for not-emitted instructions */
+         if (instr->pass_flags)
+            continue;
+
+         unsigned index = state.index;
+
+         /* output live till the end */
+         if (instr->type == nir_instr_type_intrinsic) {
+            nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+            if (intr->intrinsic == nir_intrinsic_store_output)
+               state.index = ~0u;
+         }
+
+         nir_foreach_src(instr, set_src_live, &state);
+
+         state.index = index;
+      }
+      assert(state.index == block_live_index[block->index]);
+
+      /* Walk over all of the predecessors of the current block updating
+       * their live in with the live out of this one.  If anything has
+       * changed, add the predecessor to the work list so that we ensure
+       * that the new information is used.
+       */
+      set_foreach(block->predecessors, entry) {
+         nir_block *pred = (nir_block *)entry->key;
+         if (propagate_across_edge(pred, block, &state))
+            nir_block_worklist_push_tail(&state.worklist, pred);
+      }
+   }
+
+   nir_block_worklist_fini(&state.worklist);
+
+   /* apply live_in/live_out to ranges */
+
+   nir_foreach_block(block, impl) {
+      BITSET_WORD tmp;
+      int i;
+
+      BITSET_FOREACH_SET(i, tmp, block->live_in, state.num_defs)
+         range_include(&state.defs[i], block_live_index[block->index]);
+
+      BITSET_FOREACH_SET(i, tmp, block->live_out, state.num_defs)
+         range_include(&state.defs[i], block_live_index[block->index + 1]);
+   }
+
+   return state.num_defs;
+}
+
+/* precomputed by register_allocate  */
+static unsigned int *q_values[] = {
+   (unsigned int[]) { 1, 2, 3, 4, 2 },
+   (unsigned int[]) { 3, 5, 6, 6, 5 },
+   (unsigned int[]) { 3, 4, 4, 4, 4 },
+   (unsigned int[]) { 1, 1, 1, 1, 1 },
+   (unsigned int[]) { 1, 2, 2, 2, 1 },
+};
+
+static void
+ra_assign(struct state *state, nir_shader *shader)
+{
+   struct ra_regs *regs = ra_alloc_reg_set(NULL, option(max_temps) *
+                  NUM_REG_TYPES, false);
+
+   /* classes always be created from index 0, so equal to the class enum
+    * which represents a register with (c+1) components
+    */
+   for (int c = 0; c < NUM_REG_CLASSES; c++)
+      ra_alloc_reg_class(regs);
+   /* add each register of each class */
+   for (int r = 0; r < NUM_REG_TYPES * option(max_temps); r++)
+      ra_class_add_reg(regs, reg_get_class(r), r);
+   /* set conflicts */
+   for (int r = 0; r < option(max_temps); r++) {
+      for (int i = 0; i < NUM_REG_TYPES; i++) {
+         for (int j = 0; j < i; j++) {
+            if (reg_writemask[i] & reg_writemask[j]) {
+               ra_add_reg_conflict(regs, NUM_REG_TYPES * r + i,
+                                         NUM_REG_TYPES * r + j);
+            }
+         }
+      }
+   }
+   ra_set_finalize(regs, q_values);
+
+   nir_function_impl *impl = nir_shader_get_entrypoint(shader);
+
+   /* liveness and interference */
+
+   nir_index_blocks(impl);
+   nir_index_ssa_defs(impl);
+   nir_foreach_block(block, impl) {
+      nir_foreach_instr(instr, block)
+         instr->pass_flags = 0;
+   }
+
+   /* this gives an approximation/upper limit on how many nodes are needed
+    * (some ssa values do not represent an allocated register)
+    */
+   unsigned max_nodes = impl->ssa_alloc + impl->reg_alloc;
+   unsigned *live_map = ralloc_array(NULL, unsigned, max_nodes);
+   memset(live_map, 0xff, sizeof(unsigned) * max_nodes);
+   struct live_def *defs = rzalloc_array(NULL, struct live_def, max_nodes);
+
+   unsigned num_nodes = live_defs(impl, defs, live_map);
+   struct ra_graph *g = ra_alloc_interference_graph(regs, num_nodes);
+
+   /* set classes from num_components */
+   for (unsigned i = 0; i < num_nodes; i++) {
+      nir_instr *instr = defs[i].instr;
+      nir_dest *dest = defs[i].dest;
+
+      ra_set_node_class(g, i, nir_dest_num_components(*dest) - 1);
+
+      if (instr->type == nir_instr_type_alu && option(etna_new_transcendentals)) {
+         switch (nir_instr_as_alu(instr)->op) {
+         case nir_op_fdiv:
+         case nir_op_flog2:
+         case nir_op_fsin:
+         case nir_op_fcos:
+            assert(dest->is_ssa);
+            ra_set_node_class(g, i, REG_CLASS_VIRT_VEC2T);
+         default:
+            break;
+         }
+      }
+   }
+
+   nir_foreach_block(block, impl) {
+      nir_foreach_instr(instr, block) {
+         if (instr->type != nir_instr_type_intrinsic)
+            continue;
+
+         nir_dest *dest = dest_for_instr(instr);
+         nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+         unsigned reg;
+
+         switch (intr->intrinsic) {
+         case nir_intrinsic_store_output: {
+            /* don't want output to be swizzled
+             * TODO: better would be to set the type to X/XY/XYZ/XYZW
+             */
+            ra_set_node_class(g, live_map[src_index(impl, &intr->src[0])], REG_CLASS_VEC4);
+         } continue;
+         case nir_intrinsic_load_input:
+            reg = nir_intrinsic_base(intr) * NUM_REG_TYPES + (unsigned[]) {
+               REG_TYPE_VIRT_SCALAR_X,
+               REG_TYPE_VIRT_VEC2_XY,
+               REG_TYPE_VIRT_VEC3_XYZ,
+               REG_TYPE_VEC4,
+            }[nir_dest_num_components(*dest) - 1];
+            break;
+         case nir_intrinsic_load_instance_id:
+            reg = option(id_reg) * NUM_REG_TYPES + REG_TYPE_VIRT_SCALAR_Y;
+            break;
+         default:
+            continue;
+         }
+
+         ra_set_node_reg(g, live_map[dest_index(impl, dest)], reg);
+      }
+   }
+
+   /* add interference for intersecting live ranges */
+   for (unsigned i = 0; i < num_nodes; i++) {
+      assert(defs[i].live_start < defs[i].live_end);
+      for (unsigned j = 0; j < i; j++) {
+         if (defs[i].live_start >= defs[j].live_end || defs[j].live_start >= defs[i].live_end)
+            continue;
+         ra_add_node_interference(g, i, j);
+      }
+   }
+
+   ralloc_free(defs);
+
+   /* Allocate registers */
+   bool ok = ra_allocate(g);
+   assert(ok);
+
+   state->g = g;
+   state->regs = regs;
+   state->live_map = live_map;
+   state->num_nodes = num_nodes;
+}
+
+static unsigned
+ra_finish(struct state *state)
+{
+   /* TODO: better way to get number of registers used? */
+   unsigned j = 0;
+   for (unsigned i = 0; i < state->num_nodes; i++) {
+      j = MAX2(j, reg_get_base(state, ra_get_node_reg(state->g, i)) + 1);
+   }
+
+   ralloc_free(state->g);
+   ralloc_free(state->regs);
+   ralloc_free(state->live_map);
+
+   return j;
+}
+
+static void
+emit_alu(struct state *state, nir_alu_instr * alu)
+{
+   const nir_op_info *info = &nir_op_infos[alu->op];
+
+   /* marked as dead instruction (vecN and other bypassed instr) */
+   if (alu->instr.pass_flags)
+      return;
+
+   assert(!(alu->op >= nir_op_vec2 && alu->op <= nir_op_vec4));
+
+   unsigned dst_swiz;
+   hw_dst dst = ra_dest(state, &alu->dest.dest, &dst_swiz);
+
+   /* compose alu write_mask with RA write mask */
+   if (!alu->dest.dest.is_ssa)
+      dst.write_mask = inst_write_mask_compose(alu->dest.write_mask, dst.write_mask);
+
+   switch (alu->op) {
+   case nir_op_fdot2:
+   case nir_op_fdot3:
+   case nir_op_fdot4:
+      /* not per-component - don't compose dst_swiz */
+      dst_swiz = INST_SWIZ_IDENTITY;
+      break;
+   default:
+      break;
+   }
+
+   hw_src srcs[3];
+
+   for (int i = 0; i < info->num_inputs; i++) {
+      nir_alu_src *asrc = &alu->src[i];
+      hw_src src;
+
+      src = src_swizzle(get_src(state, &asrc->src), ALU_SWIZ(asrc));
+      src = src_swizzle(src, dst_swiz);
+
+      if (src.rgroup != INST_RGROUP_IMMEDIATE) {
+         src.neg = asrc->negate || (alu->op == nir_op_fneg);
+         src.abs = asrc->abs || (alu->op == nir_op_fabs);
+      } else {
+         assert(!asrc->negate && alu->op != nir_op_fneg);
+         assert(!asrc->abs && alu->op != nir_op_fabs);
+      }
+
+      srcs[i] = src;
+   }
+
+   emit(alu, alu->op, dst, srcs, alu->dest.saturate || (alu->op == nir_op_fsat));
+}
+
+static void
+emit_tex(struct state *state, nir_tex_instr * tex)
+{
+   unsigned dst_swiz;
+   hw_dst dst = ra_dest(state, &tex->dest, &dst_swiz);
+   nir_src *coord = NULL;
+   nir_src *lod_bias = NULL;
+
+   for (unsigned i = 0; i < tex->num_srcs; i++) {
+      switch (tex->src[i].src_type) {
+      case nir_tex_src_coord:
+         coord = &tex->src[i].src;
+         break;
+      case nir_tex_src_bias:
+      case nir_tex_src_lod:
+         assert(!lod_bias);
+         lod_bias = &tex->src[i].src;
+         break;
+      default:
+         assert(0);
+         break;
+      }
+   }
+
+   emit(tex, tex->op, tex->sampler_index, dst_swiz, dst, get_src(state, coord),
+        lod_bias ? get_src(state, lod_bias) : SRC_DISABLE);
+}
+
+static void
+emit_intrinsic(struct state *state, nir_intrinsic_instr * intr)
+{
+   switch (intr->intrinsic) {
+   case nir_intrinsic_store_output:
+      emit(output, nir_intrinsic_base(intr), get_src(state, &intr->src[0]));
+      break;
+   case nir_intrinsic_discard_if:
+      emit(discard, get_src(state, &intr->src[0]));
+   break;
+   case nir_intrinsic_discard:
+      emit(discard, SRC_DISABLE);
+      break;
+   case nir_intrinsic_load_uniform: {
+      unsigned dst_swiz;
+      hw_dst dst = ra_dest(state, &intr->dest, &dst_swiz);
+      /* TODO: might have a problem with dst_swiz .. */
+      emit(load_ubo, dst, get_src(state, &intr->src[0]), const_src(state, &UNIFORM_BASE(nir_intrinsic_base(intr) * 16), 1));
+   } break;
+   case nir_intrinsic_load_front_face:
+   case nir_intrinsic_load_frag_coord:
+      assert(intr->dest.is_ssa); /* TODO - lower phis could cause this */
+      break;
+   case nir_intrinsic_load_input:
+   case nir_intrinsic_load_instance_id:
+      break;
+   default:
+      assert(0);
+   }
+}
+
+static void
+emit_instr(struct state *state, nir_instr * instr)
+{
+   switch (instr->type) {
+   case nir_instr_type_alu:
+      emit_alu(state, nir_instr_as_alu(instr));
+      break;
+   case nir_instr_type_tex:
+      emit_tex(state, nir_instr_as_tex(instr));
+      break;
+   case nir_instr_type_intrinsic:
+      emit_intrinsic(state, nir_instr_as_intrinsic(instr));
+      break;
+   case nir_instr_type_jump:
+      assert(nir_instr_is_last(instr));
+   case nir_instr_type_load_const:
+   case nir_instr_type_ssa_undef:
+      break;
+   default:
+      assert(0);
+      break;
+   }
+}
+
+static void
+emit_block(struct state *state, nir_block * block)
+{
+   emit(block_start, block->index);
+
+   nir_foreach_instr(instr, block)
+      emit_instr(state, instr);
+
+   /* succs->index < block->index is for the loop case  */
+   nir_block *succs = block->successors[0];
+   if (nir_block_ends_in_jump(block) || succs->index < block->index)
+      emit(jump, succs->index, SRC_DISABLE);
+}
+
+static void
+emit_cf_list(struct state *state, struct exec_list *list);
+
+static void
+emit_if(struct state *state, nir_if * nif)
+{
+   emit(jump, nir_if_first_else_block(nif)->index, get_src(state, &nif->condition));
+   emit_cf_list(state, &nif->then_list);
+
+   /* jump at end of then_list to skip else_list
+    * not needed if then_list already ends with a jump or else_list is empty
+    */
+   if (!nir_block_ends_in_jump(nir_if_last_then_block(nif)) &&
+       !nir_cf_list_is_empty_block(&nif->else_list))
+      emit(jump, nir_if_last_else_block(nif)->successors[0]->index, SRC_DISABLE);
+
+   emit_cf_list(state, &nif->else_list);
+}
+
+static void
+emit_cf_list(struct state *state, struct exec_list *list)
+{
+   foreach_list_typed(nir_cf_node, node, node, list) {
+      switch (node->type) {
+      case nir_cf_node_block:
+         emit_block(state, nir_cf_node_as_block(node));
+         break;
+      case nir_cf_node_if:
+         emit_if(state, nir_cf_node_as_if(node));
+         break;
+      case nir_cf_node_loop:
+         emit_cf_list(state, &nir_cf_node_as_loop(node)->body);
+         break;
+      default:
+         assert(0);
+         break;
+      }
+   }
+}
+
+/* based on nir_lower_vec_to_movs */
+static unsigned
+insert_vec_mov(nir_alu_instr *vec, unsigned start_idx, nir_shader *shader)
+{
+   assert(start_idx < nir_op_infos[vec->op].num_inputs);
+   unsigned write_mask = (1u << start_idx);
+
+   nir_alu_instr *mov = nir_alu_instr_create(shader, nir_op_mov);
+   nir_alu_src_copy(&mov->src[0], &vec->src[start_idx], mov);
+
+   mov->src[0].swizzle[0] = vec->src[start_idx].swizzle[0];
+   mov->src[0].negate = vec->src[start_idx].negate;
+   mov->src[0].abs = vec->src[start_idx].abs;
+
+   unsigned num_components = 1;
+
+   for (unsigned i = start_idx + 1; i < 4; i++) {
+      if (!(vec->dest.write_mask & (1 << i)))
+         continue;
+
+      if (nir_srcs_equal(vec->src[i].src, vec->src[start_idx].src) &&
+          vec->src[i].negate == vec->src[start_idx].negate &&
+          vec->src[i].abs == vec->src[start_idx].abs) {
+         write_mask |= (1 << i);
+         mov->src[0].swizzle[num_components] = vec->src[i].swizzle[0];
+         num_components++;
+      }
+   }
+
+   mov->dest.write_mask = (1 << num_components) - 1;
+   nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components, 32, NULL);
+
+   /* replace vec srcs with inserted mov */
+   for (unsigned i = 0, j = 0; i < 4; i++) {
+      if (!(write_mask & (1 << i)))
+         continue;
+
+      nir_instr_rewrite_src(&vec->instr, &vec->src[i].src, nir_src_for_ssa(&mov->dest.dest.ssa));
+      vec->src[i].swizzle[0] = j++;
+   }
+
+   nir_instr_insert_before(&vec->instr, &mov->instr);
+
+   return write_mask;
+}
+
+/*
+ * for vecN instructions:
+ * -merge constant sources into a single src
+ * -insert movs (nir_lower_vec_to_movs equivalent)
+ * for non-vecN instructions:
+ * -try to merge constants as single constant
+ * -insert movs for multiple constants (pre-HALTI5)
+ */
+static void
+lower_alu(struct state *state, nir_alu_instr *alu)
+{
+   const nir_op_info *info = &nir_op_infos[alu->op];
+
+   nir_builder b;
+   nir_builder_init(&b, state->impl);
+   b.cursor = nir_before_instr(&alu->instr);
+
+   switch (alu->op) {
+   case nir_op_vec2:
+   case nir_op_vec3:
+   case nir_op_vec4: {
+      nir_const_value value[4];
+      unsigned num_components = 0;
+
+      for (unsigned i = 0; i < info->num_inputs; i++) {
+         nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
+         if (cv)
+            value[num_components++] = cv[alu->src[i].swizzle[0]];
+      }
+
+      if (num_components <= 1) /* nothing to do */
+         break;
+
+      nir_ssa_def *def = nir_build_imm(&b, num_components, 32, value);
+
+      if (num_components == info->num_inputs) {
+         nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(def));
+         nir_instr_remove(&alu->instr);
+         return;
+      }
+
+      for (unsigned i = 0, j = 0; i < info->num_inputs; i++) {
+         nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
+         if (!cv)
+            continue;
+
+         nir_instr_rewrite_src(&alu->instr, &alu->src[i].src, nir_src_for_ssa(def));
+         alu->src[i].swizzle[0] = j++;
+      }
+   } break;
+   default: {
+      if (!option(single_const_src))
+         return;
+
+      /* pre-GC7000L can only have 1 uniform src per instruction */
+      nir_const_value value[4] = {};
+      uint8_t swizzle[4][4] = {};
+      unsigned swiz_max = 0, num_const = 0;
+
+      for (unsigned i = 0; i < info->num_inputs; i++) {
+         nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
+         if (!cv)
+            continue;
+
+         unsigned num_components = info->input_sizes[i] ?: alu->dest.dest.ssa.num_components;
+         for (unsigned j = 0; j < num_components; j++) {
+            int idx = const_add(&value[0].u64, cv[alu->src[i].swizzle[j]].u64);
+            swizzle[i][j] = idx;
+            swiz_max = MAX2(swiz_max, (unsigned) idx);
+         }
+         num_const++;
+      }
+
+      /* nothing to do */
+      if (num_const <= 1)
+         return;
+
+      /* resolve with single combined const src */
+      if (swiz_max < 4) {
+         nir_ssa_def *def = nir_build_imm(&b, swiz_max + 1, 32, value);
+
+         for (unsigned i = 0; i < info->num_inputs; i++) {
+            nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
+            if (!cv)
+               continue;
+
+            nir_instr_rewrite_src(&alu->instr, &alu->src[i].src, nir_src_for_ssa(def));
+
+            for (unsigned j = 0; j < 4; j++)
+               alu->src[i].swizzle[j] = swizzle[i][j];
+         }
+         return;
+      }
+
+      /* resolve with movs */
+      num_const = 0;
+      for (unsigned i = 0; i < info->num_inputs; i++) {
+         nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
+         if (!cv)
+            continue;
+
+         num_const++;
+         if (num_const == 1)
+            continue;
+
+         nir_ssa_def *mov = nir_mov(&b, alu->src[i].src.ssa);
+         nir_instr_rewrite_src(&alu->instr, &alu->src[i].src, nir_src_for_ssa(mov));
+      }
+   } return;
+   }
+
+   unsigned finished_write_mask = 0;
+   for (unsigned i = 0; i < 4; i++) {
+      if (!(alu->dest.write_mask & (1 << i)))
+            continue;
+
+      nir_ssa_def *ssa = alu->src[i].src.ssa;
+
+      /* check that vecN instruction is only user of this */
+      bool need_mov = list_length(&ssa->if_uses) != 0;
+      nir_foreach_use(use_src, ssa) {
+         if (use_src->parent_instr != &alu->instr)
+            need_mov = true;
+      }
+
+      nir_instr *instr = ssa->parent_instr;
+      switch (instr->type) {
+      case nir_instr_type_alu:
+      case nir_instr_type_tex:
+         break;
+      case nir_instr_type_intrinsic:
+         if (nir_instr_as_intrinsic(instr)->intrinsic == nir_intrinsic_load_input) {
+            need_mov = vec_dest_has_swizzle(alu, &nir_instr_as_intrinsic(instr)->dest.ssa);
+            break;
+         }
+      default:
+         need_mov = true;
+      }
+
+      if (need_mov && !(finished_write_mask & (1 << i)))
+         finished_write_mask |= insert_vec_mov(alu, i, state->shader);
+   }
+}
+
+static bool
+emit_shader(nir_shader *shader, const struct emit_options *options,
+            unsigned *num_temps, unsigned *num_consts)
+{
+   struct state state = {
+      .options = options,
+      .shader = shader,
+      .impl = nir_shader_get_entrypoint(shader),
+   };
+
+   nir_builder b;
+   nir_builder_init(&b, state.impl);
+
+   /* convert non-dynamic uniform loads to constants, etc */
+   nir_foreach_block(block, state.impl) {
+      nir_foreach_instr_safe(instr, block) {
+         switch(instr->type) {
+         case nir_instr_type_alu:
+            /* deals with vecN and const srcs */
+            lower_alu(&state, nir_instr_as_alu(instr));
+            break;
+         case nir_instr_type_load_const: {
+            nir_load_const_instr *load_const = nir_instr_as_load_const(instr);
+            for (unsigned  i = 0; i < load_const->def.num_components; i++)
+               load_const->value[i] = CONST(load_const->value[i].u32);
+         } break;
+         case nir_instr_type_intrinsic: {
+            nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+            if (intr->intrinsic != nir_intrinsic_load_uniform)
+               break;
+            nir_const_value *off = nir_src_as_const_value(intr->src[0]);
+            if (!off || off[0].u64 >> 32 != ETNA_IMMEDIATE_CONSTANT)
+               break;
+
+            unsigned base = nir_intrinsic_base(intr) + off[0].u32 / 16;
+            nir_const_value value[4];
+
+            for (unsigned i = 0; i < intr->dest.ssa.num_components; i++) {
+               if (nir_intrinsic_base(intr) < 0)
+                  value[i] = TEXSCALE(~nir_intrinsic_base(intr), i);
+               else
+                  value[i] = UNIFORM(base * 4 + i);
+            }
+
+            b.cursor = nir_after_instr(instr);
+            nir_ssa_def *def = nir_build_imm(&b, intr->dest.ssa.num_components, 32, value);
+
+            nir_ssa_def_rewrite_uses(&intr->dest.ssa, nir_src_for_ssa(def));
+            nir_instr_remove(instr);
+         } break;
+         default:
+            break;
+         }
+      }
+   }
+
+   /* add mov for any store output using sysval/const  */
+   nir_foreach_block(block, state.impl) {
+      nir_foreach_instr_safe(instr, block) {
+         if (instr->type != nir_instr_type_intrinsic)
+            continue;
+
+         nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+
+         switch (intr->intrinsic) {
+         case nir_intrinsic_store_output: {
+            nir_src *src = &intr->src[0];
+            if (nir_src_is_const(*src) || is_sysval(src->ssa->parent_instr)) {
+               b.cursor = nir_before_instr(instr);
+               nir_instr_rewrite_src(instr, src, nir_src_for_ssa(nir_mov(&b, src->ssa)));
+            }
+         } break;
+         default:
+            break;
+         }
+      }
+   }
+
+   /* call directly to avoid validation (load_const don't pass validation at this point) */
+   nir_convert_from_ssa(shader, true);
+   nir_opt_dce(shader);
+
+   ra_assign(&state, shader);
+
+   emit_cf_list(&state, &nir_shader_get_entrypoint(shader)->body);
+
+   *num_temps = ra_finish(&state);
+   *num_consts = state.const_count;
+   return true;
+}
diff --git a/src/gallium/drivers/etnaviv/etnaviv_debug.h b/src/gallium/drivers/etnaviv/etnaviv_debug.h
index 4051e95dd5f..7676e5adce1 100644
--- a/src/gallium/drivers/etnaviv/etnaviv_debug.h
+++ b/src/gallium/drivers/etnaviv/etnaviv_debug.h
@@ -53,6 +53,7 @@
 #define ETNA_DBG_DRAW_STALL      0x400000 /* Stall FE/PE after every draw op */
 #define ETNA_DBG_SHADERDB        0x800000 /* dump program compile information */
 #define ETNA_DBG_NO_SINGLEBUF    0x1000000 /* disable single buffer feature */
+#define ETNA_DBG_NIR             0x2000000 /* use new NIR compiler */
 
 extern int etna_mesa_debug; /* set in etna_screen.c from ETNA_DEBUG */
 
diff --git a/src/gallium/drivers/etnaviv/etnaviv_screen.c b/src/gallium/drivers/etnaviv/etnaviv_screen.c
index c0b39ec600c..88ab1b02249 100644
--- a/src/gallium/drivers/etnaviv/etnaviv_screen.c
+++ b/src/gallium/drivers/etnaviv/etnaviv_screen.c
@@ -72,6 +72,7 @@ static const struct debug_named_value debug_options[] = {
    {"draw_stall",     ETNA_DBG_DRAW_STALL, "Stall FE/PE after each rendered primitive"},
    {"shaderdb",       ETNA_DBG_SHADERDB, "Enable shaderdb output"},
    {"no_singlebuffer",ETNA_DBG_NO_SINGLEBUF, "Disable single buffer feature"},
+   {"nir",            ETNA_DBG_NIR, "use new NIR compiler"},
    DEBUG_NAMED_VALUE_END
 };
 
@@ -154,6 +155,11 @@ etna_screen_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
       return 1;
    case PIPE_CAP_NATIVE_FENCE_FD:
       return screen->drm_version >= ETNA_DRM_VERSION_FENCE_FD;
+   case PIPE_CAP_TGSI_FS_POSITION_IS_SYSVAL:
+   case PIPE_CAP_TGSI_FS_FACE_IS_INTEGER_SYSVAL: /* note: not integer */
+      return DBG_ENABLED(ETNA_DBG_NIR);
+   case PIPE_CAP_TGSI_FS_POINT_IS_SYSVAL:
+      return 0;
 
    /* Memory */
    case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT:
@@ -322,7 +328,7 @@ etna_screen_get_shader_param(struct pipe_screen *pscreen,
                 ? screen->specs.fragment_sampler_count
                 : screen->specs.vertex_sampler_count;
    case PIPE_SHADER_CAP_PREFERRED_IR:
-      return PIPE_SHADER_IR_TGSI;
+      return DBG_ENABLED(ETNA_DBG_NIR) ? PIPE_SHADER_IR_NIR : PIPE_SHADER_IR_TGSI;
    case PIPE_SHADER_CAP_MAX_CONST_BUFFER_SIZE:
       return 4096;
    case PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED:
@@ -743,6 +749,13 @@ etna_screen_bo_from_handle(struct pipe_screen *pscreen,
    return bo;
 }
 
+static const void *
+etna_get_compiler_options(struct pipe_screen *pscreen,
+                          enum pipe_shader_ir ir, unsigned shader)
+{
+   return &etna_screen(pscreen)->options;
+}
+
 struct pipe_screen *
 etna_screen_create(struct etna_device *dev, struct etna_gpu *gpu,
                    struct renderonly *ro)
@@ -845,6 +858,26 @@ etna_screen_create(struct etna_device *dev, struct etna_gpu *gpu,
    if (!etna_get_specs(screen))
       goto fail;
 
+   screen->options = (nir_shader_compiler_options) {
+      .lower_fpow = true,
+      .lower_sub = true,
+      .lower_ftrunc = true,
+      .fuse_ffma = true,
+      .lower_bitops = true,
+      .lower_all_io_to_temps = true,
+      .vertex_id_zero_based = true,
+      .lower_flrp32 = true,
+      .lower_fmod = true,
+      .lower_vector_cmp = true,
+      .lower_fdph = true,
+      .lower_fdiv = true, /* !screen->specs.has_new_transcendentals */
+      .lower_fsign = !screen->specs.has_sign_floor_ceil,
+      .lower_ffloor = !screen->specs.has_sign_floor_ceil,
+      .lower_fceil = !screen->specs.has_sign_floor_ceil,
+      .lower_fsqrt = !screen->specs.has_sin_cos_sqrt,
+      .lower_sincos = !screen->specs.has_sin_cos_sqrt,
+   };
+
    /* apply debug options that disable individual features */
    if (DBG_ENABLED(ETNA_DBG_NO_EARLY_Z))
       screen->features[viv_chipFeatures] |= chipFeatures_NO_EARLY_Z;
@@ -861,6 +894,7 @@ etna_screen_create(struct etna_device *dev, struct etna_gpu *gpu,
    pscreen->get_param = etna_screen_get_param;
    pscreen->get_paramf = etna_screen_get_paramf;
    pscreen->get_shader_param = etna_screen_get_shader_param;
+   pscreen->get_compiler_options = etna_get_compiler_options;
 
    pscreen->get_name = etna_screen_get_name;
    pscreen->get_vendor = etna_screen_get_vendor;
diff --git a/src/gallium/drivers/etnaviv/etnaviv_screen.h b/src/gallium/drivers/etnaviv/etnaviv_screen.h
index 4e850d4b7ee..99e2cc20ac7 100644
--- a/src/gallium/drivers/etnaviv/etnaviv_screen.h
+++ b/src/gallium/drivers/etnaviv/etnaviv_screen.h
@@ -38,6 +38,7 @@
 #include "util/slab.h"
 #include "util/u_dynarray.h"
 #include "util/u_helpers.h"
+#include "compiler/nir/nir.h"
 
 struct etna_bo;
 
@@ -87,6 +88,8 @@ struct etna_screen {
    /* set of resources used by currently-unsubmitted renders */
    mtx_t lock;
    struct set *used_resources;
+
+   nir_shader_compiler_options options;
 };
 
 static inline struct etna_screen *
diff --git a/src/gallium/drivers/etnaviv/etnaviv_shader.c b/src/gallium/drivers/etnaviv/etnaviv_shader.c
index 479c88bb44d..f0cde53d023 100644
--- a/src/gallium/drivers/etnaviv/etnaviv_shader.c
+++ b/src/gallium/drivers/etnaviv/etnaviv_shader.c
@@ -33,6 +33,7 @@
 #include "etnaviv_util.h"
 
 #include "tgsi/tgsi_parse.h"
+#include "nir/tgsi_to_nir.h"
 #include "util/u_math.h"
 #include "util/u_memory.h"
 
@@ -49,7 +50,7 @@ static bool etna_icache_upload_shader(struct etna_context *ctx, struct etna_shad
    etna_bo_cpu_prep(v->bo, DRM_ETNA_PREP_WRITE);
    memcpy(buf, v->code, v->code_size*4);
    etna_bo_cpu_fini(v->bo);
-   DBG("Uploaded %s of %u words to bo %p", v->processor == PIPE_SHADER_FRAGMENT ? "fs":"vs", v->code_size, v->bo);
+   DBG("Uploaded %s of %u words to bo %p", v->stage == MESA_SHADER_FRAGMENT ? "fs":"vs", v->code_size, v->bo);
    return true;
 }
 
@@ -67,8 +68,8 @@ etna_link_shaders(struct etna_context *ctx, struct compiled_shader_state *cs,
 {
    struct etna_shader_link_info link = { };
 
-   assert(vs->processor == PIPE_SHADER_VERTEX);
-   assert(fs->processor == PIPE_SHADER_FRAGMENT);
+   assert(vs->stage == MESA_SHADER_VERTEX);
+   assert(fs->stage == MESA_SHADER_FRAGMENT);
 
 #ifdef DEBUG
    if (DBG_ENABLED(ETNA_DBG_DUMP_SHADERS)) {
@@ -275,10 +276,10 @@ etna_shader_update_vs_inputs(struct compiled_shader_state *cs,
 static inline const char *
 etna_shader_stage(struct etna_shader_variant *shader)
 {
-   switch (shader->processor) {
-   case PIPE_SHADER_VERTEX:     return "VERT";
-   case PIPE_SHADER_FRAGMENT:   return "FRAG";
-   case PIPE_SHADER_COMPUTE:    return "CL";
+   switch (shader->stage) {
+   case MESA_SHADER_VERTEX:     return "VERT";
+   case MESA_SHADER_FRAGMENT:   return "FRAG";
+   case MESA_SHADER_COMPUTE:    return "CL";
    default:
       unreachable("invalid type");
       return NULL;
@@ -372,7 +373,14 @@ etna_create_shader_state(struct pipe_context *pctx,
    static uint32_t id;
    shader->id = id++;
    shader->specs = &ctx->specs;
-   shader->tokens = tgsi_dup_tokens(pss->tokens);
+
+   if (DBG_ENABLED(ETNA_DBG_NIR))
+      shader->nir = (pss->type == PIPE_SHADER_IR_NIR) ? pss->ir.nir :
+                     tgsi_to_nir(pss->tokens, pctx->screen);
+   else
+      shader->tokens = tgsi_dup_tokens(pss->tokens);
+
+
 
    if (etna_mesa_debug & ETNA_DBG_SHADERDB) {
       /* if shader-db run, create a standard variant immediately
@@ -401,7 +409,7 @@ etna_delete_shader_state(struct pipe_context *pctx, void *ss)
       etna_destroy_shader(t);
    }
 
-   FREE(shader->tokens);
+   ralloc_free(shader->nir);
    FREE(shader);
 }
 
diff --git a/src/gallium/drivers/etnaviv/etnaviv_shader.h b/src/gallium/drivers/etnaviv/etnaviv_shader.h
index 121d5815ba0..3c5b6e65d94 100644
--- a/src/gallium/drivers/etnaviv/etnaviv_shader.h
+++ b/src/gallium/drivers/etnaviv/etnaviv_shader.h
@@ -31,6 +31,7 @@
 
 struct etna_context;
 struct etna_shader_variant;
+struct nir_shader;
 
 struct etna_shader_key
 {
@@ -56,14 +57,15 @@ etna_shader_key_equal(struct etna_shader_key *a, struct etna_shader_key *b)
 }
 
 struct etna_shader {
-    /* shader id (for debug): */
-    uint32_t id;
-    uint32_t variant_count;
+   /* shader id (for debug): */
+   uint32_t id;
+   uint32_t variant_count;
 
-    struct tgsi_token *tokens;
-    const struct etna_specs *specs;
+   struct tgsi_token *tokens;
+   struct nir_shader *nir;
+   const struct etna_specs *specs;
 
-    struct etna_shader_variant *variants;
+   struct etna_shader_variant *variants;
 };
 
 bool
diff --git a/src/gallium/drivers/etnaviv/meson.build b/src/gallium/drivers/etnaviv/meson.build
index c53eab16540..5b26ca69a75 100644
--- a/src/gallium/drivers/etnaviv/meson.build
+++ b/src/gallium/drivers/etnaviv/meson.build
@@ -37,6 +37,8 @@ files_etnaviv = files(
   'etnaviv_clear_blit.h',
   'etnaviv_compiler.c',
   'etnaviv_compiler.h',
+  'etnaviv_compiler_nir.c',
+  'etnaviv_compiler_nir_emit.h',
   'etnaviv_context.c',
   'etnaviv_context.h',
   'etnaviv_debug.h',
@@ -97,7 +99,7 @@ libetnaviv = static_library(
     inc_include, inc_src, inc_gallium, inc_gallium_aux, inc_etnaviv,
   ],
   link_with: libetnaviv_drm,
-  dependencies : dep_libdrm,
+  dependencies : [dep_libdrm, idep_nir_headers],
 )
 
 etnaviv_compiler = executable(
@@ -115,4 +117,5 @@ etnaviv_compiler = executable(
 driver_etnaviv = declare_dependency(
   compile_args : '-DGALLIUM_ETNAVIV',
   link_with : [libetnaviv, libetnavivdrm],
+  dependencies : idep_nir,
 )
diff --git a/src/gallium/winsys/etnaviv/drm/meson.build b/src/gallium/winsys/etnaviv/drm/meson.build
index f931fd39e5b..f5491941755 100644
--- a/src/gallium/winsys/etnaviv/drm/meson.build
+++ b/src/gallium/winsys/etnaviv/drm/meson.build
@@ -26,5 +26,5 @@ libetnavivdrm = static_library(
     inc_etnaviv,
   ],
   link_with: libetnaviv_drm,
-  dependencies : [dep_libdrm],
+  dependencies : [dep_libdrm, idep_nir_headers],
 )