freedreno/ir3: add NIR compiler
authorRob Clark <robclark@freedesktop.org>
Thu, 12 Mar 2015 21:18:33 +0000 (17:18 -0400)
committerRob Clark <robclark@freedesktop.org>
Sun, 5 Apr 2015 20:36:40 +0000 (16:36 -0400)
The NIR compiler frontend is an alternative to the TGSI f/e, producing
the same ir3 IR and using the same backend passes for scheduling, etc.

It is not enabled by default yet, as there are still some regressions.
To enable, use 'FD_MESA_DEBUG=nir'.  It is enough to use with, for
example, xonotic or supertuxkart.

With the NIR f/e, scalarizing and a number of other lowering steps
happen in NIR, so we don't have to do them in ir3.  Which simplifies the
f/e and allows the lowered instructions to pass through other
optimization stages.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
src/gallium/drivers/freedreno/Makefile.sources
src/gallium/drivers/freedreno/freedreno_screen.c
src/gallium/drivers/freedreno/freedreno_util.h
src/gallium/drivers/freedreno/ir3/ir3_cmdline.c
src/gallium/drivers/freedreno/ir3/ir3_compiler.h
src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c [new file with mode: 0644]
src/gallium/drivers/freedreno/ir3/ir3_shader.c

index 1f22ce9c622f5cce7b25b5525775389f8510d74a..3224efc9994d51ad6f91d383c5da6c4873e32780 100644 (file)
@@ -121,6 +121,7 @@ ir3_SOURCES := \
        ir3/instr-a3xx.h \
        ir3/ir3.c \
        ir3/ir3_compiler.c \
+       ir3/ir3_compiler_nir.c \
        ir3/ir3_compiler.h \
        ir3/ir3_cp.c \
        ir3/ir3_depth.c \
index 66fe0e571cf8013328ee7a4608c3ab13871369c1..1b89387678d1105465f731d45bd719545c3d53de 100644 (file)
@@ -71,6 +71,7 @@ static const struct debug_named_value debug_options[] = {
                {"optdump",   FD_DBG_OPTDUMP,"Dump shader DAG to .dot files"},
                {"glsl120",   FD_DBG_GLSL120,"Temporary flag to force GLSL 120 (rather than 130) on a3xx+"},
                {"nocp",      FD_DBG_NOCP,   "Disable copy-propagation"},
+               {"nir",       FD_DBG_NIR,    "Enable experimental NIR compiler"},
                DEBUG_NAMED_VALUE_END
 };
 
@@ -345,6 +346,10 @@ fd_screen_get_shader_param(struct pipe_screen *pscreen, unsigned shader,
        case PIPE_SHADER_CAP_MAX_TEX_INDIRECTIONS:
                return 16384;
        case PIPE_SHADER_CAP_MAX_CONTROL_FLOW_DEPTH:
+               /* for now, let someone else flatten if/else when using NIR: */
+               if ((fd_mesa_debug & FD_DBG_NIR) &&
+                               (is_a3xx(screen) || is_a4xx(screen)))
+                       return 0;
                return 8; /* XXX */
        case PIPE_SHADER_CAP_MAX_INPUTS:
        case PIPE_SHADER_CAP_MAX_OUTPUTS:
index b3076b9f5c64a9b74540df6d571eb6e56ebfd54a..2735ae4131537eaac2b20af76b0cb7b56e1a6f1d 100644 (file)
@@ -66,6 +66,7 @@ enum adreno_stencil_op fd_stencil_op(unsigned op);
 #define FD_DBG_OPTDUMP  0x0800
 #define FD_DBG_GLSL120  0x1000
 #define FD_DBG_NOCP     0x2000
+#define FD_DBG_NIR      0x4000
 
 extern int fd_mesa_debug;
 extern bool fd_binning_enabled;
index 81d0180933fc5945004fc0714c7f55ed2064bf8e..d0517aab8ce9a868cc69ac2004cff58fe4cd0f52 100644 (file)
@@ -215,6 +215,7 @@ static void print_usage(void)
        printf("    --saturate-t MASK - bitmask of samplers to saturate T coord\n");
        printf("    --saturate-r MASK - bitmask of samplers to saturate R coord\n");
        printf("    --nocp            - disable copy propagation\n");
+       printf("    --nir             - use NIR compiler\n");
        printf("    --help            - show this message\n");
 }
 
@@ -229,6 +230,7 @@ int main(int argc, char **argv)
        const char *info;
        void *ptr;
        size_t size;
+       int use_nir = 0;
 
        fd_mesa_debug |= FD_DBG_DISASM;
 
@@ -293,6 +295,11 @@ int main(int argc, char **argv)
                        n++;
                        continue;
                }
+               if (!strcmp(argv[n], "--nir")) {
+                       use_nir = true;
+                       n++;
+                       continue;
+               }
 
                if (!strcmp(argv[n], "--help")) {
                        print_usage();
@@ -333,8 +340,13 @@ int main(int argc, char **argv)
                break;
        }
 
-       info = "compiler";
-       ret = ir3_compile_shader(&v, toks, key, true);
+       if (use_nir) {
+               info = "NIR compiler";
+               ret = ir3_compile_shader_nir(&v, toks, key);
+       } else {
+               info = "TGSI compiler";
+               ret = ir3_compile_shader(&v, toks, key, true);
+       }
 
        if (ret) {
                reset_variant(&v, "compiler failed, trying without copy propagation!");
@@ -348,3 +360,11 @@ int main(int argc, char **argv)
        }
        dump_info(&v, info);
 }
+
+void _mesa_error_no_memory(const char *caller);
+
+void
+_mesa_error_no_memory(const char *caller)
+{
+       fprintf(stderr, "Mesa error: out of memory in %s", caller);
+}
index 6459482c007c9f596e46f73dfd586fd545ef9450..ed9637bd2d0c5cf83e7024d0cc1730e46733201e 100644 (file)
@@ -32,6 +32,9 @@
 #include "ir3_shader.h"
 
 
+int ir3_compile_shader_nir(struct ir3_shader_variant *so,
+               const struct tgsi_token *tokens, struct ir3_shader_key key);
+
 int ir3_compile_shader(struct ir3_shader_variant *so,
                const struct tgsi_token *tokens,
                struct ir3_shader_key key, bool cp);
diff --git a/src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c b/src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c
new file mode 100644 (file)
index 0000000..fdf814f
--- /dev/null
@@ -0,0 +1,1717 @@
+/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
+
+/*
+ * Copyright (C) 2015 Rob Clark <robclark@freedesktop.org>
+ *
+ * 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, sublicense,
+ * 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 NONINFRINGEMENT.  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:
+ *    Rob Clark <robclark@freedesktop.org>
+ */
+
+#include <stdarg.h>
+
+#include "pipe/p_state.h"
+#include "util/u_string.h"
+#include "util/u_memory.h"
+#include "util/u_inlines.h"
+#include "tgsi/tgsi_lowering.h"
+#include "tgsi/tgsi_strings.h"
+
+#include "nir/tgsi_to_nir.h"
+#include "glsl/shader_enums.h"
+
+#include "freedreno_util.h"
+
+#include "ir3_compiler.h"
+#include "ir3_shader.h"
+
+#include "instr-a3xx.h"
+#include "ir3.h"
+
+
+static struct ir3_instruction * create_immed(struct ir3_block *block, uint32_t val);
+
+struct ir3_compile {
+       const struct tgsi_token *tokens;
+       struct nir_shader *s;
+
+       struct ir3 *ir;
+       struct ir3_shader_variant *so;
+
+       /* bitmask of which samplers are integer: */
+       uint16_t integer_s;
+
+       struct ir3_block *block;
+
+       /* For fragment shaders, from the hw perspective the only
+        * actual input is r0.xy position register passed to bary.f.
+        * But TGSI doesn't know that, it still declares things as
+        * IN[] registers.  So we do all the input tracking normally
+        * and fix things up after compile_instructions()
+        *
+        * NOTE that frag_pos is the hardware position (possibly it
+        * is actually an index or tag or some such.. it is *not*
+        * values that can be directly used for gl_FragCoord..)
+        */
+       struct ir3_instruction *frag_pos, *frag_face, *frag_coord[4];
+
+       /* For vertex shaders, keep track of the system values sources */
+       struct ir3_instruction *vertex_id, *basevertex, *instance_id;
+
+       /* mapping from nir_register to defining instruction: */
+       struct hash_table *def_ht;
+
+       /* a common pattern for indirect addressing is to request the
+        * same address register multiple times.  To avoid generating
+        * duplicate instruction sequences (which our backend does not
+        * try to clean up, since that should be done as the NIR stage)
+        * we cache the address value generated for a given src value:
+        */
+       struct hash_table *addr_ht;
+
+       /* for calculating input/output positions/linkages: */
+       unsigned next_inloc;
+
+       /* a4xx (at least patchlevel 0) cannot seem to flat-interpolate
+        * so we need to use ldlv.u32 to load the varying directly:
+        */
+       bool flat_bypass;
+
+       /* for looking up which system value is which */
+       unsigned sysval_semantics[8];
+
+       /* list of kill instructions: */
+       struct ir3_instruction *kill[16];
+       unsigned int kill_count;
+
+       /* set if we encounter something we can't handle yet, so we
+        * can bail cleanly and fallback to TGSI compiler f/e
+        */
+       bool error;
+};
+
+
+static struct nir_shader *to_nir(const struct tgsi_token *tokens)
+{
+       struct nir_shader_compiler_options options = {
+                       .lower_fpow = true,
+                       .lower_fsat = true,
+                       .lower_scmp = true,
+                       .lower_flrp = true,
+                       .native_integers = true,
+       };
+       bool progress;
+
+       struct nir_shader *s = tgsi_to_nir(tokens, &options);
+
+       if (fd_mesa_debug & FD_DBG_OPTMSGS) {
+               debug_printf("----------------------\n");
+               nir_print_shader(s, stdout);
+               debug_printf("----------------------\n");
+       }
+
+       nir_opt_global_to_local(s);
+       nir_convert_to_ssa(s);
+       nir_lower_idiv(s);
+
+       do {
+               progress = false;
+
+               nir_lower_vars_to_ssa(s);
+               nir_lower_alu_to_scalar(s);
+
+               progress |= nir_copy_prop(s);
+               progress |= nir_opt_dce(s);
+               progress |= nir_opt_cse(s);
+               progress |= nir_opt_peephole_select(s);
+               progress |= nir_opt_algebraic(s);
+               progress |= nir_opt_constant_folding(s);
+
+       } while (progress);
+
+       nir_remove_dead_variables(s);
+       nir_validate_shader(s);
+
+       if (fd_mesa_debug & FD_DBG_OPTMSGS) {
+               debug_printf("----------------------\n");
+               nir_print_shader(s, stdout);
+               debug_printf("----------------------\n");
+       }
+
+       return s;
+}
+
+/* TODO nir doesn't lower everything for us yet, but ideally it would: */
+static const struct tgsi_token *
+lower_tgsi(const struct tgsi_token *tokens, struct ir3_shader_variant *so)
+{
+       struct tgsi_shader_info info;
+       struct tgsi_lowering_config lconfig = {
+                       .color_two_side = so->key.color_two_side,
+                       .lower_FRC = true,
+       };
+
+       switch (so->type) {
+       case SHADER_FRAGMENT:
+       case SHADER_COMPUTE:
+               lconfig.saturate_s = so->key.fsaturate_s;
+               lconfig.saturate_t = so->key.fsaturate_t;
+               lconfig.saturate_r = so->key.fsaturate_r;
+               break;
+       case SHADER_VERTEX:
+               lconfig.saturate_s = so->key.vsaturate_s;
+               lconfig.saturate_t = so->key.vsaturate_t;
+               lconfig.saturate_r = so->key.vsaturate_r;
+               break;
+       }
+
+       if (!so->shader) {
+               /* hack for standalone compiler which does not have
+                * screen/context:
+                */
+       } else if (ir3_shader_gpuid(so->shader) >= 400) {
+               /* a4xx seems to have *no* sam.p */
+               lconfig.lower_TXP = ~0;  /* lower all txp */
+       } else {
+               /* a3xx just needs to avoid sam.p for 3d tex */
+               lconfig.lower_TXP = (1 << TGSI_TEXTURE_3D);
+       }
+
+       return tgsi_transform_lowering(&lconfig, tokens, &info);
+}
+
+static struct ir3_compile *
+compile_init(struct ir3_shader_variant *so,
+               const struct tgsi_token *tokens)
+{
+       struct ir3_compile *ctx = rzalloc(NULL, struct ir3_compile);
+       const struct tgsi_token *lowered_tokens;
+
+       if (!so->shader) {
+               /* hack for standalone compiler which does not have
+                * screen/context:
+                */
+       } else if (ir3_shader_gpuid(so->shader) >= 400) {
+               /* need special handling for "flat" */
+               ctx->flat_bypass = true;
+       } else {
+               /* no special handling for "flat" */
+               ctx->flat_bypass = false;
+       }
+
+       switch (so->type) {
+       case SHADER_FRAGMENT:
+       case SHADER_COMPUTE:
+               ctx->integer_s = so->key.finteger_s;
+               break;
+       case SHADER_VERTEX:
+               ctx->integer_s = so->key.vinteger_s;
+               break;
+       }
+
+       ctx->ir = so->ir;
+       ctx->so = so;
+       ctx->next_inloc = 8;
+       ctx->def_ht = _mesa_hash_table_create(ctx,
+                       _mesa_hash_pointer, _mesa_key_pointer_equal);
+       ctx->addr_ht = _mesa_hash_table_create(ctx,
+                       _mesa_hash_pointer, _mesa_key_pointer_equal);
+
+       lowered_tokens = lower_tgsi(tokens, so);
+       if (!lowered_tokens)
+               lowered_tokens = tokens;
+       ctx->s = to_nir(lowered_tokens);
+
+       if (lowered_tokens != tokens)
+               free((void *)lowered_tokens);
+
+       so->first_immediate = ctx->s->num_uniforms;
+
+       return ctx;
+}
+
+static void
+compile_error(struct ir3_compile *ctx, const char *format, ...)
+{
+       va_list ap;
+       va_start(ap, format);
+       _debug_vprintf(format, ap);
+       va_end(ap);
+       nir_print_shader(ctx->s, stdout);
+       ctx->error = true;
+}
+
+#define compile_assert(ctx, cond) do { \
+               if (!(cond)) compile_error((ctx), "failed assert: "#cond"\n"); \
+       } while (0)
+
+static void
+compile_free(struct ir3_compile *ctx)
+{
+       ralloc_free(ctx);
+}
+
+/* allocate a n element value array (to be populated by caller) and
+ * insert in def_ht
+ */
+static struct ir3_instruction **
+__get_dst(struct ir3_compile *ctx, void *key, unsigned n)
+{
+       struct ir3_instruction **value =
+               ralloc_array(ctx->def_ht, struct ir3_instruction *, n);
+       _mesa_hash_table_insert(ctx->def_ht, key, value);
+       return value;
+}
+
+static struct ir3_instruction **
+get_dst(struct ir3_compile *ctx, nir_dest *dst, unsigned n)
+{
+       if (dst->is_ssa) {
+               return __get_dst(ctx, &dst->ssa, n);
+       } else {
+               return __get_dst(ctx, dst->reg.reg, n);
+       }
+}
+
+static struct ir3_instruction **
+get_dst_ssa(struct ir3_compile *ctx, nir_ssa_def *dst, unsigned n)
+{
+       return __get_dst(ctx, dst, n);
+}
+
+static struct ir3_instruction **
+get_src(struct ir3_compile *ctx, nir_src *src)
+{
+       struct hash_entry *entry;
+       if (src->is_ssa) {
+               entry = _mesa_hash_table_search(ctx->def_ht, src->ssa);
+       } else {
+               entry = _mesa_hash_table_search(ctx->def_ht, src->reg.reg);
+       }
+       compile_assert(ctx, entry);
+       return entry->data;
+}
+
+static struct ir3_instruction *
+create_immed(struct ir3_block *block, uint32_t val)
+{
+       struct ir3_instruction *mov;
+
+       mov = ir3_instr_create(block, 1, 0);
+       mov->cat1.src_type = TYPE_U32;
+       mov->cat1.dst_type = TYPE_U32;
+       ir3_reg_create(mov, 0, 0);
+       ir3_reg_create(mov, 0, IR3_REG_IMMED)->uim_val = val;
+
+       return mov;
+}
+
+static struct ir3_instruction *
+create_addr(struct ir3_block *block, struct ir3_instruction *src)
+{
+       struct ir3_instruction *instr, *immed;
+
+       /* TODO in at least some cases, the backend could probably be
+        * made clever enough to propagate IR3_REG_HALF..
+        */
+       instr = ir3_COV(block, src, TYPE_U32, TYPE_S16);
+       instr->regs[0]->flags |= IR3_REG_HALF;
+
+       immed = create_immed(block, 2);
+       immed->regs[0]->flags |= IR3_REG_HALF;
+
+       instr = ir3_SHL_B(block, instr, 0, immed, 0);
+       instr->regs[0]->flags |= IR3_REG_HALF;
+       instr->regs[1]->flags |= IR3_REG_HALF;
+
+       instr = ir3_MOV(block, instr, TYPE_S16);
+       instr->regs[0]->flags |= IR3_REG_ADDR | IR3_REG_HALF;
+       instr->regs[1]->flags |= IR3_REG_HALF;
+
+       return instr;
+}
+
+/* caches addr values to avoid generating multiple cov/shl/mova
+ * sequences for each use of a given NIR level src as address
+ */
+static struct ir3_instruction *
+get_addr(struct ir3_compile *ctx, struct ir3_instruction *src)
+{
+       struct ir3_instruction *addr;
+       struct hash_entry *entry;
+       entry = _mesa_hash_table_search(ctx->addr_ht, src);
+       if (entry)
+               return entry->data;
+
+       /* TODO do we need to cache per block? */
+       addr = create_addr(ctx->block, src);
+       _mesa_hash_table_insert(ctx->addr_ht, src, addr);
+
+       return addr;
+}
+
+static struct ir3_instruction *
+create_uniform(struct ir3_block *block, unsigned n)
+{
+       struct ir3_instruction *mov;
+
+       mov = ir3_instr_create(block, 1, 0);
+       /* TODO get types right? */
+       mov->cat1.src_type = TYPE_F32;
+       mov->cat1.dst_type = TYPE_F32;
+       ir3_reg_create(mov, 0, 0);
+       ir3_reg_create(mov, n, IR3_REG_CONST);
+
+       return mov;
+}
+
+static struct ir3_instruction *
+create_uniform_indirect(struct ir3_block *block, unsigned n,
+               struct ir3_instruction *address)
+{
+       struct ir3_instruction *mov;
+
+       mov = ir3_instr_create(block, 1, 0);
+       mov->cat1.src_type = TYPE_U32;
+       mov->cat1.dst_type = TYPE_U32;
+       ir3_reg_create(mov, 0, 0);
+       ir3_reg_create(mov, n, IR3_REG_CONST | IR3_REG_RELATIV);
+       mov->address = address;
+
+       return mov;
+}
+
+static struct ir3_instruction *
+create_indirect(struct ir3_block *block, struct ir3_instruction **arr,
+               unsigned arrsz, unsigned n, struct ir3_instruction *address)
+{
+       struct ir3_instruction *mov, *collect;
+       struct ir3_register *src;
+
+       collect = ir3_instr_create2(block, -1, OPC_META_FI, 1 + arrsz);
+       ir3_reg_create(collect, 0, 0);
+       for (unsigned i = 0; i < arrsz; i++)
+               ir3_reg_create(collect, 0, IR3_REG_SSA)->instr = arr[i];
+
+       mov = ir3_instr_create(block, 1, 0);
+       mov->cat1.src_type = TYPE_U32;
+       mov->cat1.dst_type = TYPE_U32;
+       ir3_reg_create(mov, 0, 0);
+       src = ir3_reg_create(mov, 0, IR3_REG_SSA | IR3_REG_RELATIV);
+       src->instr = collect;
+       src->size  = arrsz;
+       mov->address = address;
+
+       return mov;
+}
+
+static struct ir3_instruction *
+create_input(struct ir3_block *block, struct ir3_instruction *instr,
+               unsigned n)
+{
+       struct ir3_instruction *in;
+
+       in = ir3_instr_create(block, -1, OPC_META_INPUT);
+       in->inout.block = block;
+       ir3_reg_create(in, n, 0);
+       if (instr)
+               ir3_reg_create(in, 0, IR3_REG_SSA)->instr = instr;
+
+       return in;
+}
+
+static struct ir3_instruction *
+create_frag_input(struct ir3_compile *ctx, unsigned n, bool use_ldlv)
+{
+       struct ir3_block *block = ctx->block;
+       struct ir3_instruction *instr;
+       struct ir3_instruction *inloc = create_immed(block, n);
+
+       if (use_ldlv) {
+               instr = ir3_LDLV(block, inloc, 0, create_immed(block, 1), 0);
+               instr->cat6.type = TYPE_U32;
+               instr->cat6.iim_val = 1;
+       } else {
+               instr = ir3_BARY_F(block, inloc, 0, ctx->frag_pos, 0);
+               instr->regs[2]->wrmask = 0x3;
+       }
+
+       return instr;
+}
+
+static struct ir3_instruction *
+create_frag_coord(struct ir3_compile *ctx, unsigned comp)
+{
+       struct ir3_block *block = ctx->block;
+       struct ir3_instruction *instr;
+
+       compile_assert(ctx, !ctx->frag_coord[comp]);
+
+       ctx->frag_coord[comp] = create_input(ctx->block, NULL, 0);
+
+       switch (comp) {
+       case 0: /* .x */
+       case 1: /* .y */
+               /* for frag_coord, we get unsigned values.. we need
+                * to subtract (integer) 8 and divide by 16 (right-
+                * shift by 4) then convert to float:
+                *
+                *    add.s tmp, src, -8
+                *    shr.b tmp, tmp, 4
+                *    mov.u32f32 dst, tmp
+                *
+                */
+               instr = ir3_ADD_S(block, ctx->frag_coord[comp], 0,
+                               create_immed(block, -8), 0);
+               instr = ir3_SHR_B(block, instr, 0,
+                               create_immed(block, 4), 0);
+               instr = ir3_COV(block, instr, TYPE_U32, TYPE_F32);
+
+               return instr;
+       case 2: /* .z */
+       case 3: /* .w */
+       default:
+               /* seems that we can use these as-is: */
+               return ctx->frag_coord[comp];
+       }
+}
+
+static struct ir3_instruction *
+create_frag_face(struct ir3_compile *ctx, unsigned comp)
+{
+       struct ir3_block *block = ctx->block;
+       struct ir3_instruction *instr;
+
+       switch (comp) {
+       case 0: /* .x */
+               compile_assert(ctx, !ctx->frag_face);
+
+               ctx->frag_face = create_input(block, NULL, 0);
+
+               /* for faceness, we always get -1 or 0 (int).. but TGSI expects
+                * positive vs negative float.. and piglit further seems to
+                * expect -1.0 or 1.0:
+                *
+                *    mul.s tmp, hr0.x, 2
+                *    add.s tmp, tmp, 1
+                *    mov.s32f32, dst, tmp
+                *
+                */
+               instr = ir3_MUL_S(block, ctx->frag_face, 0,
+                               create_immed(block, 2), 0);
+               instr = ir3_ADD_S(block, instr, 0,
+                               create_immed(block, 1), 0);
+               instr = ir3_COV(block, instr, TYPE_S32, TYPE_F32);
+
+               return instr;
+       case 1: /* .y */
+       case 2: /* .z */
+               return create_immed(block, fui(0.0));
+       default:
+       case 3: /* .w */
+               return create_immed(block, fui(1.0));
+       }
+}
+
+/*
+ * Adreno uses uint rather than having dedicated bool type,
+ * which (potentially) requires some conversion, in particular
+ * when using output of an bool instr to int input, or visa
+ * versa.
+ *
+ *         | Adreno  |  NIR  |
+ *  -------+---------+-------+-
+ *   true  |    1    |  ~0   |
+ *   false |    0    |   0   |
+ *
+ * To convert from an adreno bool (uint) to nir, use:
+ *
+ *    absneg.s dst, (neg)src
+ *
+ * To convert back in the other direction:
+ *
+ *    absneg.s dst, (abs)arc
+ *
+ * The CP step can clean up the absneg.s that cancel each other
+ * out, and with a slight bit of extra cleverness (to recognize
+ * the instructions which produce either a 0 or 1) can eliminate
+ * the absneg.s's completely when an instruction that wants
+ * 0/1 consumes the result.  For example, when a nir 'bcsel'
+ * consumes the result of 'feq'.  So we should be able to get by
+ * without a boolean resolve step, and without incuring any
+ * extra penalty in instruction count.
+ */
+
+/* NIR bool -> native (adreno): */
+static struct ir3_instruction *
+ir3_b2n(struct ir3_block *block, struct ir3_instruction *instr)
+{
+       return ir3_ABSNEG_S(block, instr, IR3_REG_SABS);
+}
+
+/* native (adreno) -> NIR bool: */
+static struct ir3_instruction *
+ir3_n2b(struct ir3_block *block, struct ir3_instruction *instr)
+{
+       return ir3_ABSNEG_S(block, instr, IR3_REG_SNEG);
+}
+
+/*
+ * alu/sfu instructions:
+ */
+
+static void
+emit_alu(struct ir3_compile *ctx, nir_alu_instr *alu)
+{
+       const nir_op_info *info = &nir_op_infos[alu->op];
+       struct ir3_instruction **dst, *src[info->num_inputs];
+       struct ir3_block *b = ctx->block;
+
+       dst = get_dst(ctx, &alu->dest.dest, MAX2(info->output_size, 1));
+
+       /* Vectors are special in that they have non-scalarized writemasks,
+        * and just take the first swizzle channel for each argument in
+        * order into each writemask channel.
+        */
+       if ((alu->op == nir_op_vec2) ||
+                       (alu->op == nir_op_vec3) ||
+                       (alu->op == nir_op_vec4)) {
+
+               for (int i = 0; i < info->num_inputs; i++) {
+                       nir_alu_src *asrc = &alu->src[i];
+
+                       compile_assert(ctx, !asrc->abs);
+                       compile_assert(ctx, !asrc->negate);
+
+                       src[i] = get_src(ctx, &asrc->src)[asrc->swizzle[0]];
+                       dst[i] = ir3_MOV(b, src[i], TYPE_U32);
+               }
+
+               return;
+       }
+
+       /* General case: We can just grab the one used channel per src. */
+       for (int i = 0; i < info->num_inputs; i++) {
+               unsigned chan = ffs(alu->dest.write_mask) - 1;
+               nir_alu_src *asrc = &alu->src[i];
+
+               compile_assert(ctx, !asrc->abs);
+               compile_assert(ctx, !asrc->negate);
+
+               src[i] = get_src(ctx, &asrc->src)[asrc->swizzle[chan]];
+       }
+
+       switch (alu->op) {
+       case nir_op_f2i:
+               dst[0] = ir3_COV(b, src[0], TYPE_F32, TYPE_S32);
+               break;
+       case nir_op_f2u:
+               dst[0] = ir3_COV(b, src[0], TYPE_F32, TYPE_U32);
+               break;
+       case nir_op_i2f:
+               dst[0] = ir3_COV(b, src[0], TYPE_S32, TYPE_F32);
+               break;
+       case nir_op_u2f:
+               dst[0] = ir3_COV(b, src[0], TYPE_U32, TYPE_F32);
+               break;
+       case nir_op_imov:
+               dst[0] = ir3_MOV(b, src[0], TYPE_S32);
+               break;
+       case nir_op_f2b:
+               dst[0] = ir3_CMPS_F(b, src[0], 0, create_immed(b, fui(0.0)), 0);
+               dst[0]->cat2.condition = IR3_COND_NE;
+               dst[0] = ir3_n2b(b, dst[0]);
+               break;
+       case nir_op_b2f:
+               dst[0] = ir3_COV(b, ir3_b2n(b, src[0]), TYPE_U32, TYPE_F32);
+               break;
+       case nir_op_b2i:
+               dst[0] = ir3_b2n(b, src[0]);
+               break;
+       case nir_op_i2b:
+               dst[0] = ir3_CMPS_S(b, src[0], 0, create_immed(b, 0), 0);
+               dst[0]->cat2.condition = IR3_COND_NE;
+               dst[0] = ir3_n2b(b, dst[0]);
+               break;
+
+       case nir_op_fneg:
+               dst[0] = ir3_ABSNEG_F(b, src[0], IR3_REG_FNEG);
+               break;
+       case nir_op_fabs:
+               dst[0] = ir3_ABSNEG_F(b, src[0], IR3_REG_FABS);
+               break;
+       case nir_op_fmax:
+               dst[0] = ir3_MAX_F(b, src[0], 0, src[1], 0);
+               break;
+       case nir_op_fmin:
+               dst[0] = ir3_MIN_F(b, src[0], 0, src[1], 0);
+               break;
+       case nir_op_fmul:
+               dst[0] = ir3_MUL_F(b, src[0], 0, src[1], 0);
+               break;
+       case nir_op_fadd:
+               dst[0] = ir3_ADD_F(b, src[0], 0, src[1], 0);
+               break;
+       case nir_op_fsub:
+               dst[0] = ir3_ADD_F(b, src[0], 0, src[1], IR3_REG_FNEG);
+               break;
+       case nir_op_ffma:
+               dst[0] = ir3_MAD_F32(b, src[0], 0, src[1], 0, src[2], 0);
+               break;
+       case nir_op_fddx:
+               dst[0] = ir3_DSX(b, src[0], 0);
+               dst[0]->cat5.type = TYPE_F32;
+               break;
+       case nir_op_fddy:
+               dst[0] = ir3_DSY(b, src[0], 0);
+               dst[0]->cat5.type = TYPE_F32;
+               break;
+               break;
+       case nir_op_flt:
+               dst[0] = ir3_CMPS_F(b, src[0], 0, src[1], 0);
+               dst[0]->cat2.condition = IR3_COND_LT;
+               dst[0] = ir3_n2b(b, dst[0]);
+               break;
+       case nir_op_fge:
+               dst[0] = ir3_CMPS_F(b, src[0], 0, src[1], 0);
+               dst[0]->cat2.condition = IR3_COND_GE;
+               dst[0] = ir3_n2b(b, dst[0]);
+               break;
+       case nir_op_feq:
+               dst[0] = ir3_CMPS_F(b, src[0], 0, src[1], 0);
+               dst[0]->cat2.condition = IR3_COND_EQ;
+               dst[0] = ir3_n2b(b, dst[0]);
+               break;
+       case nir_op_fne:
+               dst[0] = ir3_CMPS_F(b, src[0], 0, src[1], 0);
+               dst[0]->cat2.condition = IR3_COND_NE;
+               dst[0] = ir3_n2b(b, dst[0]);
+               break;
+       case nir_op_fceil:
+               dst[0] = ir3_CEIL_F(b, src[0], 0);
+               break;
+       case nir_op_ffloor:
+               dst[0] = ir3_FLOOR_F(b, src[0], 0);
+               break;
+       case nir_op_ftrunc:
+               dst[0] = ir3_TRUNC_F(b, src[0], 0);
+               break;
+       case nir_op_fround_even:
+               dst[0] = ir3_RNDNE_F(b, src[0], 0);
+               break;
+       case nir_op_fsign:
+               dst[0] = ir3_SIGN_F(b, src[0], 0);
+               break;
+
+       case nir_op_fsin:
+               dst[0] = ir3_SIN(b, src[0], 0);
+               break;
+       case nir_op_fcos:
+               dst[0] = ir3_COS(b, src[0], 0);
+               break;
+       case nir_op_frsq:
+               dst[0] = ir3_RSQ(b, src[0], 0);
+               break;
+       case nir_op_frcp:
+               dst[0] = ir3_RCP(b, src[0], 0);
+               break;
+       case nir_op_flog2:
+               dst[0] = ir3_LOG2(b, src[0], 0);
+               break;
+       case nir_op_fexp2:
+               dst[0] = ir3_EXP2(b, src[0], 0);
+               break;
+       case nir_op_fsqrt:
+               dst[0] = ir3_SQRT(b, src[0], 0);
+               break;
+
+       case nir_op_iabs:
+               dst[0] = ir3_ABSNEG_S(b, src[0], IR3_REG_SABS);
+               break;
+       case nir_op_iadd:
+               dst[0] = ir3_ADD_U(b, src[0], 0, src[1], 0);
+               break;
+       case nir_op_iand:
+               dst[0] = ir3_AND_B(b, src[0], 0, src[1], 0);
+               break;
+       case nir_op_imax:
+               dst[0] = ir3_MAX_S(b, src[0], 0, src[1], 0);
+               break;
+       case nir_op_imin:
+               dst[0] = ir3_MIN_S(b, src[0], 0, src[1], 0);
+               break;
+       case nir_op_imul:
+               /*
+                * dst = (al * bl) + (ah * bl << 16) + (al * bh << 16)
+                *   mull.u tmp0, a, b           ; mul low, i.e. al * bl
+                *   madsh.m16 tmp1, a, b, tmp0  ; mul-add shift high mix, i.e. ah * bl << 16
+                *   madsh.m16 dst, b, a, tmp1   ; i.e. al * bh << 16
+                */
+               dst[0] = ir3_MADSH_M16(b, src[1], 0, src[0], 0,
+                                       ir3_MADSH_M16(b, src[0], 0, src[1], 0,
+                                               ir3_MULL_U(b, src[0], 0, src[1], 0), 0), 0);
+               break;
+       case nir_op_ineg:
+               dst[0] = ir3_ABSNEG_S(b, src[0], IR3_REG_SNEG);
+               break;
+       case nir_op_inot:
+               dst[0] = ir3_NOT_B(b, src[0], 0);
+               break;
+       case nir_op_ior:
+               dst[0] = ir3_OR_B(b, src[0], 0, src[1], 0);
+               break;
+       case nir_op_ishl:
+               dst[0] = ir3_SHL_B(b, src[0], 0, src[1], 0);
+               break;
+       case nir_op_ishr:
+               dst[0] = ir3_ASHR_B(b, src[0], 0, src[1], 0);
+               break;
+       case nir_op_isign: {
+               /* maybe this would be sane to lower in nir.. */
+               struct ir3_instruction *neg, *pos;
+
+               neg = ir3_CMPS_S(b, src[0], 0, create_immed(b, 0), 0);
+               neg->cat2.condition = IR3_COND_LT;
+
+               pos = ir3_CMPS_S(b, src[0], 0, create_immed(b, 0), 0);
+               pos->cat2.condition = IR3_COND_GT;
+
+               dst[0] = ir3_SUB_U(b, pos, 0, neg, 0);
+
+               break;
+       }
+       case nir_op_isub:
+               dst[0] = ir3_SUB_U(b, src[0], 0, src[1], 0);
+               break;
+       case nir_op_ixor:
+               dst[0] = ir3_XOR_B(b, src[0], 0, src[1], 0);
+               break;
+       case nir_op_ushr:
+               dst[0] = ir3_SHR_B(b, src[0], 0, src[1], 0);
+               break;
+       case nir_op_ilt:
+               dst[0] = ir3_CMPS_S(b, src[0], 0, src[1], 0);
+               dst[0]->cat2.condition = IR3_COND_LT;
+               dst[0] = ir3_n2b(b, dst[0]);
+               break;
+       case nir_op_ige:
+               dst[0] = ir3_CMPS_S(b, src[0], 0, src[1], 0);
+               dst[0]->cat2.condition = IR3_COND_GE;
+               dst[0] = ir3_n2b(b, dst[0]);
+               break;
+       case nir_op_ieq:
+               dst[0] = ir3_CMPS_S(b, src[0], 0, src[1], 0);
+               dst[0]->cat2.condition = IR3_COND_EQ;
+               dst[0] = ir3_n2b(b, dst[0]);
+               break;
+       case nir_op_ine:
+               dst[0] = ir3_CMPS_S(b, src[0], 0, src[1], 0);
+               dst[0]->cat2.condition = IR3_COND_NE;
+               dst[0] = ir3_n2b(b, dst[0]);
+               break;
+       case nir_op_ult:
+               dst[0] = ir3_CMPS_U(b, src[0], 0, src[1], 0);
+               dst[0]->cat2.condition = IR3_COND_LT;
+               dst[0] = ir3_n2b(b, dst[0]);
+               break;
+       case nir_op_uge:
+               dst[0] = ir3_CMPS_U(b, src[0], 0, src[1], 0);
+               dst[0]->cat2.condition = IR3_COND_GE;
+               dst[0] = ir3_n2b(b, dst[0]);
+               break;
+
+       case nir_op_bcsel:
+               dst[0] = ir3_SEL_B32(b, src[1], 0, ir3_b2n(b, src[0]), 0, src[2], 0);
+               break;
+
+       default:
+               compile_error(ctx, "Unhandled ALU op: %s\n",
+                               nir_op_infos[alu->op].name);
+               break;
+       }
+}
+
+static void
+emit_intrinisic(struct ir3_compile *ctx, nir_intrinsic_instr *intr)
+{
+       const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
+       struct ir3_instruction **dst, **src;
+       struct ir3_block *b = ctx->block;
+       unsigned idx = intr->const_index[0];
+
+       if (info->has_dest) {
+               dst = get_dst(ctx, &intr->dest, intr->num_components);
+       }
+
+       switch (intr->intrinsic) {
+       case nir_intrinsic_load_uniform:
+               compile_assert(ctx, intr->const_index[1] == 1);
+               for (int i = 0; i < intr->num_components; i++) {
+                       unsigned n = idx * 4 + i;
+                       dst[i] = create_uniform(b, n);
+               }
+               break;
+       case nir_intrinsic_load_uniform_indirect:
+               compile_assert(ctx, intr->const_index[1] == 1);
+               src = get_src(ctx, &intr->src[0]);
+               for (int i = 0; i < intr->num_components; i++) {
+                       unsigned n = idx * 4 + i;
+                       dst[i] = create_uniform_indirect(b, n,
+                                       get_addr(ctx, src[0]));
+               }
+               break;
+       case nir_intrinsic_load_input:
+               compile_assert(ctx, intr->const_index[1] == 1);
+               for (int i = 0; i < intr->num_components; i++) {
+                       unsigned n = idx * 4 + i;
+                       dst[i] = b->inputs[n];
+               }
+               break;
+       case nir_intrinsic_load_input_indirect:
+               compile_assert(ctx, intr->const_index[1] == 1);
+               src = get_src(ctx, &intr->src[0]);
+               for (int i = 0; i < intr->num_components; i++) {
+                       unsigned n = idx * 4 + i;
+                       dst[i] = create_indirect(b, b->inputs, b->ninputs, n,
+                                       get_addr(ctx, src[i]));
+               }
+               break;
+       case nir_intrinsic_store_output:
+               compile_assert(ctx, intr->const_index[1] == 1);
+               src = get_src(ctx, &intr->src[0]);
+               for (int i = 0; i < intr->num_components; i++) {
+                       unsigned n = idx * 4 + i;
+                       b->outputs[n] = src[i];
+               }
+               break;
+       case nir_intrinsic_discard_if:
+       case nir_intrinsic_discard: {
+               struct ir3_instruction *cond, *kill;
+
+               if (intr->intrinsic == nir_intrinsic_discard_if) {
+                       /* conditional discard: */
+                       src = get_src(ctx, &intr->src[0]);
+                       cond = ir3_b2n(b, src[0]);
+               } else {
+                       /* unconditional discard: */
+                       cond = create_immed(b, 1);
+               }
+
+               cond = ir3_CMPS_S(b, cond, 0, create_immed(b, 0), 0);
+               cond->cat2.condition = IR3_COND_NE;
+
+               /* condition always goes in predicate register: */
+               cond->regs[0]->num = regid(REG_P0, 0);
+
+               kill = ir3_KILL(b, cond, 0);
+
+               ctx->kill[ctx->kill_count++] = kill;
+               ctx->so->has_kill = true;
+
+               break;
+       }
+       default:
+               compile_error(ctx, "Unhandled intrinsic type: %s\n",
+                               nir_intrinsic_infos[intr->intrinsic].name);
+               break;
+       }
+}
+
+static void
+emit_load_const(struct ir3_compile *ctx, nir_load_const_instr *instr)
+{
+       struct ir3_instruction **dst = get_dst_ssa(ctx, &instr->def,
+                       instr->def.num_components);
+       for (int i = 0; i < instr->def.num_components; i++)
+               dst[i] = create_immed(ctx->block, instr->value.u[i]);
+}
+
+static void
+emit_undef(struct ir3_compile *ctx, nir_ssa_undef_instr *undef)
+{
+       struct ir3_instruction **dst = get_dst_ssa(ctx, &undef->def,
+                       undef->def.num_components);
+       /* backend doesn't want undefined instructions, so just plug
+        * in 0.0..
+        */
+       for (int i = 0; i < undef->def.num_components; i++)
+               dst[i] = create_immed(ctx->block, fui(0.0));
+}
+
+/*
+ * texture fetch/sample instructions:
+ */
+
+static void
+emit_tex(struct ir3_compile *ctx, nir_tex_instr *tex)
+{
+       struct ir3_block *b = ctx->block;
+       struct ir3_instruction **dst, *src0, *src1, *sam;
+       struct ir3_instruction **coord, *lod, *compare, *proj, **off, **ddx, **ddy;
+       struct ir3_register *reg;
+       bool has_bias = false, has_lod = false, has_proj = false, has_off = false;
+       unsigned i, coords, flags = 0;
+       opc_t opc;
+
+       /* TODO: might just be one component for gathers? */
+       dst = get_dst(ctx, &tex->dest, 4);
+
+       for (unsigned i = 0; i < tex->num_srcs; i++) {
+               switch (tex->src[i].src_type) {
+               case nir_tex_src_coord:
+                       coord = get_src(ctx, &tex->src[i].src);
+                       break;
+               case nir_tex_src_bias:
+                       lod = get_src(ctx, &tex->src[i].src)[0];
+                       has_bias = true;
+                       break;
+               case nir_tex_src_lod:
+                       lod = get_src(ctx, &tex->src[i].src)[0];
+                       has_lod = true;
+                       break;
+               case nir_tex_src_comparitor: /* shadow comparator */
+                       compare = get_src(ctx, &tex->src[i].src)[0];
+                       break;
+               case nir_tex_src_projector:
+                       proj = get_src(ctx, &tex->src[i].src)[0];
+                       has_proj = true;
+                       break;
+               case nir_tex_src_offset:
+                       off = get_src(ctx, &tex->src[i].src);
+                       has_off = true;
+                       break;
+               case nir_tex_src_ddx:
+                       ddx = get_src(ctx, &tex->src[i].src);
+                       break;
+               case nir_tex_src_ddy:
+                       ddy = get_src(ctx, &tex->src[i].src);
+                       break;
+               default:
+                       compile_error(ctx, "Unhandled NIR tex serc type: %d\n",
+                                       tex->src[i].src_type);
+                       return;
+               }
+       }
+
+       /*
+        * lay out the first argument in the proper order:
+        *  - actual coordinates first
+        *  - shadow reference
+        *  - array index
+        *  - projection w
+        *  - starting at offset 4, dpdx.xy, dpdy.xy
+        *
+        * bias/lod go into the second arg
+        */
+
+       src0 = ir3_instr_create2(b, -1, OPC_META_FI, 12);
+       ir3_reg_create(src0, 0, 0);
+
+       coords = tex->coord_components;
+       if (tex->is_array)       /* array idx goes after shadow ref */
+               coords--;
+
+       /* insert tex coords: */
+       for (i = 0; i < coords; i++)
+               ir3_reg_create(src0, 0, IR3_REG_SSA)->instr = coord[i];
+
+       if (coords == 1) {
+               /* hw doesn't do 1d, so we treat it as 2d with
+                * height of 1, and patch up the y coord.
+                * TODO: y coord should be (int)0 in some cases..
+                */
+               ir3_reg_create(src0, 0, IR3_REG_SSA)->instr =
+                               create_immed(b, fui(0.5));
+       }
+
+       if (tex->is_shadow) {
+               ir3_reg_create(src0, 0, IR3_REG_SSA)->instr = compare;
+               flags |= IR3_INSTR_S;
+       }
+
+       if (tex->is_array) {
+               ir3_reg_create(src0, 0, IR3_REG_SSA)->instr = coord[coords];
+               flags |= IR3_INSTR_A;
+       }
+
+       if (has_proj) {
+               ir3_reg_create(src0, 0, IR3_REG_SSA)->instr = proj;
+               flags |= IR3_INSTR_P;
+       }
+
+       /* pad to 4, then ddx/ddy: */
+       if (tex->op == nir_texop_txd) {
+               while (src0->regs_count < 5) {
+                       ir3_reg_create(src0, 0, IR3_REG_SSA)->instr =
+                                       create_immed(b, fui(0.0));
+               }
+               for (i = 0; i < coords; i++) {
+                       ir3_reg_create(src0, 0, IR3_REG_SSA)->instr = ddx[i];
+               }
+               if (coords < 2) {
+                       ir3_reg_create(src0, 0, IR3_REG_SSA)->instr =
+                                       create_immed(b, fui(0.0));
+               }
+               for (i = 0; i < coords; i++) {
+                       ir3_reg_create(src0, 0, IR3_REG_SSA)->instr = ddy[i];
+               }
+               if (coords < 2) {
+                       ir3_reg_create(src0, 0, IR3_REG_SSA)->instr =
+                                       create_immed(b, fui(0.0));
+               }
+       }
+
+       /*
+        * second argument (if applicable):
+        *  - offsets
+        *  - lod
+        *  - bias
+        */
+       if (has_off | has_lod | has_bias) {
+               src1 = ir3_instr_create2(b, -1, OPC_META_FI, 5);
+               ir3_reg_create(src1, 0, 0);
+
+               if (has_off) {
+                       for (i = 0; i < coords; i++) {
+                               ir3_reg_create(src0, 0, IR3_REG_SSA)->instr = off[i];
+                       }
+                       if (coords < 2) {
+                               ir3_reg_create(src0, 0, IR3_REG_SSA)->instr =
+                                               create_immed(b, fui(0.0));
+                       }
+                       flags |= IR3_INSTR_O;
+               }
+
+               if (has_lod | has_bias) {
+                       ir3_reg_create(src1, 0, IR3_REG_SSA)->instr = lod;
+               }
+       } else {
+               src1 = NULL;
+       }
+
+       switch (tex->op) {
+       case nir_texop_tex:      opc = OPC_SAM;      break;
+       case nir_texop_txb:      opc = OPC_SAMB;     break;
+       case nir_texop_txl:      opc = OPC_SAML;     break;
+       case nir_texop_txd:      opc = OPC_SAMGQ;    break;
+       case nir_texop_txf:      opc = OPC_ISAML;    break;
+       case nir_texop_txf_ms:
+       case nir_texop_txs:
+       case nir_texop_lod:
+       case nir_texop_tg4:
+       case nir_texop_query_levels:
+               compile_error(ctx, "Unhandled NIR tex type: %d\n", tex->op);
+               return;
+       }
+
+       sam = ir3_instr_create(b, 5, opc);
+       sam->flags |= flags;
+       ir3_reg_create(sam, 0, 0)->wrmask = 0xf;  // TODO proper wrmask??
+       reg = ir3_reg_create(sam, 0, IR3_REG_SSA);
+       reg->wrmask = (1 << (src0->regs_count - 1)) - 1;
+       reg->instr = src0;
+       if (src1) {
+               reg = ir3_reg_create(sam, 0, IR3_REG_SSA);
+               reg->instr = src1;
+               reg->wrmask = (1 << (src1->regs_count - 1)) - 1;
+       }
+       sam->cat5.samp = tex->sampler_index;
+       sam->cat5.tex  = tex->sampler_index;
+
+       switch (tex->dest_type) {
+       case nir_type_invalid:
+       case nir_type_float:
+               sam->cat5.type = TYPE_F32;
+               break;
+       case nir_type_int:
+               sam->cat5.type = TYPE_S32;
+               break;
+       case nir_type_unsigned:
+       case nir_type_bool:
+               sam->cat5.type = TYPE_U32;
+       }
+
+       // TODO maybe split this out into a helper, for other cases that
+       // write multiple?
+       struct ir3_instruction *prev = NULL;
+       for (int i = 0; i < 4; i++) {
+               struct ir3_instruction *split =
+                               ir3_instr_create(b, -1, OPC_META_FO);
+               ir3_reg_create(split, 0, IR3_REG_SSA);
+               ir3_reg_create(split, 0, IR3_REG_SSA)->instr = sam;
+               split->fo.off = i;
+
+               if (prev) {
+                       split->cp.left = prev;
+                       split->cp.left_cnt++;
+                       prev->cp.right = split;
+                       prev->cp.right_cnt++;
+               }
+               prev = split;
+
+               dst[i] = split;
+       }
+}
+
+
+static void
+emit_instr(struct ir3_compile *ctx, nir_instr *instr)
+{
+       switch (instr->type) {
+       case nir_instr_type_alu:
+               emit_alu(ctx, nir_instr_as_alu(instr));
+               break;
+       case nir_instr_type_intrinsic:
+               emit_intrinisic(ctx, nir_instr_as_intrinsic(instr));
+               break;
+       case nir_instr_type_load_const:
+               emit_load_const(ctx, nir_instr_as_load_const(instr));
+               break;
+       case nir_instr_type_ssa_undef:
+               emit_undef(ctx, nir_instr_as_ssa_undef(instr));
+               break;
+       case nir_instr_type_tex:
+               emit_tex(ctx, nir_instr_as_tex(instr));
+               break;
+
+       case nir_instr_type_call:
+       case nir_instr_type_jump:
+       case nir_instr_type_phi:
+       case nir_instr_type_parallel_copy:
+               compile_error(ctx, "Unhandled NIR instruction type: %d\n", instr->type);
+               break;
+       }
+}
+
+static void
+emit_block(struct ir3_compile *ctx, nir_block *block)
+{
+       nir_foreach_instr(block, instr) {
+               emit_instr(ctx, instr);
+               if (ctx->error)
+                       return;
+       }
+}
+
+static void
+emit_function(struct ir3_compile *ctx, nir_function_impl *impl)
+{
+       foreach_list_typed(nir_cf_node, node, node, &impl->body) {
+               switch (node->type) {
+               case nir_cf_node_block:
+                       emit_block(ctx, nir_cf_node_as_block(node));
+                       break;
+               case nir_cf_node_if:
+               case nir_cf_node_loop:
+               case nir_cf_node_function:
+                       compile_error(ctx, "TODO\n");
+                       break;
+               }
+               if (ctx->error)
+                       return;
+       }
+}
+
+static void
+setup_input(struct ir3_compile *ctx, nir_variable *in)
+{
+       struct ir3_shader_variant *so = ctx->so;
+       unsigned array_len = MAX2(glsl_get_length(in->type), 1);
+       unsigned ncomp = glsl_get_components(in->type);
+       /* XXX: map loc slots to semantics */
+       unsigned semantic_name = in->data.location;
+       unsigned semantic_index = in->data.index;
+       unsigned n = in->data.driver_location;
+
+       DBG("; in: %u:%u, len=%ux%u, loc=%u\n",
+                       semantic_name, semantic_index, array_len,
+                       ncomp, n);
+
+       so->inputs[n].semantic =
+                       ir3_semantic_name(semantic_name, semantic_index);
+       so->inputs[n].compmask = (1 << ncomp) - 1;
+       so->inputs[n].inloc = ctx->next_inloc;
+       so->inputs[n].interpolate = 0;
+       so->inputs_count = MAX2(so->inputs_count, n + 1);
+
+       /* the fdN_program_emit() code expects tgsi consts here, so map
+        * things back to tgsi for now:
+        */
+       switch (in->data.interpolation) {
+       case INTERP_QUALIFIER_FLAT:
+               so->inputs[n].interpolate = TGSI_INTERPOLATE_CONSTANT;
+               break;
+       case INTERP_QUALIFIER_NOPERSPECTIVE:
+               so->inputs[n].interpolate = TGSI_INTERPOLATE_LINEAR;
+               break;
+       case INTERP_QUALIFIER_SMOOTH:
+               so->inputs[n].interpolate = TGSI_INTERPOLATE_PERSPECTIVE;
+               break;
+       }
+
+       for (int i = 0; i < ncomp; i++) {
+               struct ir3_instruction *instr = NULL;
+               unsigned idx = (n * 4) + i;
+
+               if (ctx->so->type == SHADER_FRAGMENT) {
+                       if (semantic_name == TGSI_SEMANTIC_POSITION) {
+                               so->inputs[n].bary = false;
+                               so->frag_coord = true;
+                               instr = create_frag_coord(ctx, i);
+                       } else if (semantic_name == TGSI_SEMANTIC_FACE) {
+                               so->inputs[n].bary = false;
+                               so->frag_face = true;
+                               instr = create_frag_face(ctx, i);
+                       } else {
+                               bool use_ldlv = false;
+
+                               /* with NIR, we need to infer TGSI_INTERPOLATE_COLOR
+                                * from the semantic name:
+                                */
+                               if (semantic_name == TGSI_SEMANTIC_COLOR)
+                                       so->inputs[n].interpolate = TGSI_INTERPOLATE_COLOR;
+
+                               if (ctx->flat_bypass) {
+                                       /* with NIR, we need to infer TGSI_INTERPOLATE_COLOR
+                                        * from the semantic name:
+                                        */
+                                       switch (so->inputs[n].interpolate) {
+                                       case TGSI_INTERPOLATE_COLOR:
+                                               if (!ctx->so->key.rasterflat)
+                                                       break;
+                                               /* fallthrough */
+                                       case TGSI_INTERPOLATE_CONSTANT:
+                                               use_ldlv = true;
+                                               break;
+                                       }
+                               }
+
+                               so->inputs[n].bary = true;
+
+                               instr = create_frag_input(ctx, idx, use_ldlv);
+                       }
+               } else {
+                       instr = create_input(ctx->block, NULL, idx);
+               }
+
+               ctx->block->inputs[idx] = instr;
+       }
+
+       if (so->inputs[n].bary || (ctx->so->type == SHADER_VERTEX)) {
+               ctx->next_inloc += ncomp;
+               so->total_in += ncomp;
+       }
+}
+
+static void
+setup_output(struct ir3_compile *ctx, nir_variable *out)
+{
+       struct ir3_shader_variant *so = ctx->so;
+       unsigned array_len = MAX2(glsl_get_length(out->type), 1);
+       unsigned ncomp = glsl_get_components(out->type);
+       /* XXX: map loc slots to semantics */
+       unsigned semantic_name = out->data.location;
+       unsigned semantic_index = out->data.index;
+       unsigned n = out->data.driver_location;
+       unsigned comp = 0;
+
+       DBG("; out: %u:%u, len=%ux%u, loc=%u\n",
+                       semantic_name, semantic_index, array_len,
+                       ncomp, n);
+
+       if (ctx->so->type == SHADER_VERTEX) {
+               switch (semantic_name) {
+               case TGSI_SEMANTIC_POSITION:
+                       so->writes_pos = true;
+                       break;
+               case TGSI_SEMANTIC_PSIZE:
+                       so->writes_psize = true;
+                       break;
+               case TGSI_SEMANTIC_COLOR:
+               case TGSI_SEMANTIC_BCOLOR:
+               case TGSI_SEMANTIC_GENERIC:
+               case TGSI_SEMANTIC_FOG:
+               case TGSI_SEMANTIC_TEXCOORD:
+                       break;
+               default:
+                       compile_error(ctx, "unknown VS semantic name: %s\n",
+                                       tgsi_semantic_names[semantic_name]);
+               }
+       } else {
+               switch (semantic_name) {
+               case TGSI_SEMANTIC_POSITION:
+                       comp = 2;  /* tgsi will write to .z component */
+                       so->writes_pos = true;
+                       break;
+               case TGSI_SEMANTIC_COLOR:
+                       break;
+               default:
+                       compile_error(ctx, "unknown FS semantic name: %s\n",
+                                       tgsi_semantic_names[semantic_name]);
+               }
+       }
+
+       compile_assert(ctx, n < ARRAY_SIZE(so->outputs));
+
+       so->outputs[n].semantic =
+                       ir3_semantic_name(semantic_name, semantic_index);
+       so->outputs[n].regid = regid(n, comp);
+       so->outputs_count = MAX2(so->outputs_count, n + 1);
+
+       for (int i = 0; i < ncomp; i++) {
+               unsigned idx = (n * 4) + i;
+
+               ctx->block->outputs[idx] = create_immed(ctx->block, fui(0.0));
+       }
+}
+
+static void
+emit_instructions(struct ir3_compile *ctx)
+{
+       unsigned ninputs  = exec_list_length(&ctx->s->inputs) * 4;
+       unsigned noutputs = exec_list_length(&ctx->s->outputs) * 4;
+
+       /* we need to allocate big enough outputs array so that
+        * we can stuff the kill's at the end:
+        */
+       if (ctx->so->type == SHADER_FRAGMENT)
+               noutputs += ARRAY_SIZE(ctx->kill);
+
+       ctx->block = ir3_block_create(ctx->ir, 0, ninputs, noutputs);
+
+       if (ctx->so->type == SHADER_FRAGMENT)
+               ctx->block->noutputs -= ARRAY_SIZE(ctx->kill);
+
+
+       /* for fragment shader, we have a single input register (usually
+        * r0.xy) which is used as the base for bary.f varying fetch instrs:
+        */
+       if (ctx->so->type == SHADER_FRAGMENT) {
+               // TODO maybe a helper for fi since we need it a few places..
+               struct ir3_instruction *instr;
+               instr = ir3_instr_create(ctx->block, -1, OPC_META_FI);
+               ir3_reg_create(instr, 0, 0);
+               ir3_reg_create(instr, 0, IR3_REG_SSA);    /* r0.x */
+               ir3_reg_create(instr, 0, IR3_REG_SSA);    /* r0.y */
+               ctx->frag_pos = instr;
+       }
+
+       /* Setup inputs: */
+       foreach_list_typed(nir_variable, var, node, &ctx->s->inputs) {
+               setup_input(ctx, var);
+               if (ctx->error)
+                       return;
+       }
+
+       /* Setup outputs: */
+       foreach_list_typed(nir_variable, var, node, &ctx->s->outputs) {
+               setup_output(ctx, var);
+               if (ctx->error)
+                       return;
+       }
+
+       /* Find the main function and emit the body: */
+       nir_foreach_overload(ctx->s, overload) {
+               compile_assert(ctx, strcmp(overload->function->name, "main") == 0);
+               compile_assert(ctx, overload->impl);
+               emit_function(ctx, overload->impl);
+               if (ctx->error)
+                       return;
+       }
+}
+
+/* from NIR perspective, we actually have inputs.  But most of the "inputs"
+ * for a fragment shader are just bary.f instructions.  The *actual* inputs
+ * from the hw perspective are the frag_pos and optionally frag_coord and
+ * frag_face.
+ */
+static void
+fixup_frag_inputs(struct ir3_compile *ctx)
+{
+       struct ir3_shader_variant *so = ctx->so;
+       struct ir3_block *block = ctx->block;
+       struct ir3_instruction **inputs;
+       struct ir3_instruction *instr;
+       int n, regid = 0;
+
+       block->ninputs = 0;
+
+       n  = 4;  /* always have frag_pos */
+       n += COND(so->frag_face, 4);
+       n += COND(so->frag_coord, 4);
+
+       inputs = ir3_alloc(ctx->ir, n * (sizeof(struct ir3_instruction *)));
+
+       if (so->frag_face) {
+               /* this ultimately gets assigned to hr0.x so doesn't conflict
+                * with frag_coord/frag_pos..
+                */
+               inputs[block->ninputs++] = ctx->frag_face;
+               ctx->frag_face->regs[0]->num = 0;
+
+               /* remaining channels not used, but let's avoid confusing
+                * other parts that expect inputs to come in groups of vec4
+                */
+               inputs[block->ninputs++] = NULL;
+               inputs[block->ninputs++] = NULL;
+               inputs[block->ninputs++] = NULL;
+       }
+
+       /* since we don't know where to set the regid for frag_coord,
+        * we have to use r0.x for it.  But we don't want to *always*
+        * use r1.x for frag_pos as that could increase the register
+        * footprint on simple shaders:
+        */
+       if (so->frag_coord) {
+               ctx->frag_coord[0]->regs[0]->num = regid++;
+               ctx->frag_coord[1]->regs[0]->num = regid++;
+               ctx->frag_coord[2]->regs[0]->num = regid++;
+               ctx->frag_coord[3]->regs[0]->num = regid++;
+
+               inputs[block->ninputs++] = ctx->frag_coord[0];
+               inputs[block->ninputs++] = ctx->frag_coord[1];
+               inputs[block->ninputs++] = ctx->frag_coord[2];
+               inputs[block->ninputs++] = ctx->frag_coord[3];
+       }
+
+       /* we always have frag_pos: */
+       so->pos_regid = regid;
+
+       /* r0.x */
+       instr = create_input(block, NULL, block->ninputs);
+       instr->regs[0]->num = regid++;
+       inputs[block->ninputs++] = instr;
+       ctx->frag_pos->regs[1]->instr = instr;
+
+       /* r0.y */
+       instr = create_input(block, NULL, block->ninputs);
+       instr->regs[0]->num = regid++;
+       inputs[block->ninputs++] = instr;
+       ctx->frag_pos->regs[2]->instr = instr;
+
+       block->inputs = inputs;
+}
+
+static void
+compile_dump(struct ir3_compile *ctx)
+{
+       const char *name = (ctx->so->type == SHADER_VERTEX) ? "vert" : "frag";
+       static unsigned n = 0;
+       char fname[16];
+       FILE *f;
+       snprintf(fname, sizeof(fname), "%s-%04u.dot", name, n++);
+       f = fopen(fname, "w");
+       if (!f)
+               return;
+       ir3_block_depth(ctx->block);
+       ir3_dump(ctx->ir, name, ctx->block, f);
+       fclose(f);
+}
+
+int
+ir3_compile_shader_nir(struct ir3_shader_variant *so,
+               const struct tgsi_token *tokens, struct ir3_shader_key key)
+{
+       struct ir3_compile *ctx;
+       struct ir3_block *block;
+       struct ir3_instruction **inputs;
+       unsigned i, j, actual_in;
+       int ret = 0, max_bary;
+
+       assert(!so->ir);
+
+       so->ir = ir3_create();
+
+       assert(so->ir);
+
+       ctx = compile_init(so, tokens);
+       if (!ctx) {
+               DBG("INIT failed!");
+               ret = -1;
+               goto out;
+       }
+
+       emit_instructions(ctx);
+
+       if (ctx->error) {
+               DBG("EMIT failed!");
+               ret = -1;
+               goto out;
+       }
+
+       block = ctx->block;
+       so->ir->block = block;
+
+       /* keep track of the inputs from TGSI perspective.. */
+       inputs = block->inputs;
+
+       /* but fixup actual inputs for frag shader: */
+       if (so->type == SHADER_FRAGMENT)
+               fixup_frag_inputs(ctx);
+
+       /* at this point, for binning pass, throw away unneeded outputs: */
+       if (key.binning_pass) {
+               for (i = 0, j = 0; i < so->outputs_count; i++) {
+                       unsigned name = sem2name(so->outputs[i].semantic);
+                       unsigned idx = sem2idx(so->outputs[i].semantic);
+
+                       /* throw away everything but first position/psize */
+                       if ((idx == 0) && ((name == TGSI_SEMANTIC_POSITION) ||
+                                       (name == TGSI_SEMANTIC_PSIZE))) {
+                               if (i != j) {
+                                       so->outputs[j] = so->outputs[i];
+                                       block->outputs[(j*4)+0] = block->outputs[(i*4)+0];
+                                       block->outputs[(j*4)+1] = block->outputs[(i*4)+1];
+                                       block->outputs[(j*4)+2] = block->outputs[(i*4)+2];
+                                       block->outputs[(j*4)+3] = block->outputs[(i*4)+3];
+                               }
+                               j++;
+                       }
+               }
+               so->outputs_count = j;
+               block->noutputs = j * 4;
+       }
+
+       /* if we want half-precision outputs, mark the output registers
+        * as half:
+        */
+       if (key.half_precision) {
+               for (i = 0; i < block->noutputs; i++) {
+                       if (!block->outputs[i])
+                               continue;
+                       block->outputs[i]->regs[0]->flags |= IR3_REG_HALF;
+               }
+       }
+
+       /* at this point, we want the kill's in the outputs array too,
+        * so that they get scheduled (since they have no dst).. we've
+        * already ensured that the array is big enough in push_block():
+        */
+       if (so->type == SHADER_FRAGMENT) {
+               for (i = 0; i < ctx->kill_count; i++)
+                       block->outputs[block->noutputs++] = ctx->kill[i];
+       }
+
+       if (fd_mesa_debug & FD_DBG_OPTDUMP)
+               compile_dump(ctx);
+
+       if (fd_mesa_debug & FD_DBG_OPTMSGS) {
+               printf("BEFORE CP:\n");
+               ir3_dump_instr_list(block->head);
+       }
+
+       ir3_block_depth(block);
+
+       ir3_block_cp(block);
+
+       if (fd_mesa_debug & FD_DBG_OPTMSGS) {
+               printf("BEFORE GROUPING:\n");
+               ir3_dump_instr_list(block->head);
+       }
+
+       /* Group left/right neighbors, inserting mov's where needed to
+        * solve conflicts:
+        */
+       ir3_block_group(block);
+
+       if (fd_mesa_debug & FD_DBG_OPTDUMP)
+               compile_dump(ctx);
+
+       ir3_block_depth(block);
+
+       if (fd_mesa_debug & FD_DBG_OPTMSGS) {
+               printf("AFTER DEPTH:\n");
+               ir3_dump_instr_list(block->head);
+       }
+
+       ret = ir3_block_sched(block);
+       if (ret) {
+               DBG("SCHED failed!");
+               goto out;
+       }
+
+       if (fd_mesa_debug & FD_DBG_OPTMSGS) {
+               printf("AFTER SCHED:\n");
+               ir3_dump_instr_list(block->head);
+       }
+
+       ret = ir3_block_ra(block, so->type, so->frag_coord, so->frag_face);
+       if (ret) {
+               DBG("RA failed!");
+               goto out;
+       }
+
+       if (fd_mesa_debug & FD_DBG_OPTMSGS) {
+               printf("AFTER RA:\n");
+               ir3_dump_instr_list(block->head);
+       }
+
+       ir3_block_legalize(block, &so->has_samp, &max_bary);
+
+       /* fixup input/outputs: */
+       for (i = 0; i < so->outputs_count; i++) {
+               so->outputs[i].regid = block->outputs[i*4]->regs[0]->num;
+               /* preserve hack for depth output.. tgsi writes depth to .z,
+                * but what we give the hw is the scalar register:
+                */
+               if ((so->type == SHADER_FRAGMENT) &&
+                       (sem2name(so->outputs[i].semantic) == TGSI_SEMANTIC_POSITION))
+                       so->outputs[i].regid += 2;
+       }
+
+       /* Note that some or all channels of an input may be unused: */
+       actual_in = 0;
+       for (i = 0; i < so->inputs_count; i++) {
+               unsigned j, regid = ~0, compmask = 0;
+               so->inputs[i].ncomp = 0;
+               for (j = 0; j < 4; j++) {
+                       struct ir3_instruction *in = inputs[(i*4) + j];
+                       if (in) {
+                               compmask |= (1 << j);
+                               regid = in->regs[0]->num - j;
+                               actual_in++;
+                               so->inputs[i].ncomp++;
+                       }
+               }
+               so->inputs[i].regid = regid;
+               so->inputs[i].compmask = compmask;
+       }
+
+       /* fragment shader always gets full vec4's even if it doesn't
+        * fetch all components, but vertex shader we need to update
+        * with the actual number of components fetch, otherwise thing
+        * will hang due to mismaptch between VFD_DECODE's and
+        * TOTALATTRTOVS
+        */
+       if (so->type == SHADER_VERTEX)
+               so->total_in = actual_in;
+       else
+               so->total_in = align(max_bary + 1, 4);
+
+out:
+       if (ret) {
+               ir3_destroy(so->ir);
+               so->ir = NULL;
+       }
+       compile_free(ctx);
+
+       return ret;
+}
index 0cf357e17d8200ee49f1177a778ade79021971d9..9bf4e64c7f1f42650306ff003c12aedb66ca3d67 100644 (file)
@@ -177,10 +177,20 @@ create_variant(struct ir3_shader *shader, struct ir3_shader_key key)
                tgsi_dump(tokens, 0);
        }
 
-       ret = ir3_compile_shader(v, tokens, key, true);
+       if (fd_mesa_debug & FD_DBG_NIR) {
+               ret = ir3_compile_shader_nir(v, tokens, key);
+               if (ret)
+                       reset_variant(v, "NIR compiler failed, fallback to TGSI!");
+       } else {
+               ret = -1;
+       }
+
        if (ret) {
-               reset_variant(v, "new compiler failed, trying without copy propagation!");
-               ret = ir3_compile_shader(v, tokens, key, false);
+               ret = ir3_compile_shader(v, tokens, key, true);
+               if (ret) {
+                       reset_variant(v, "new compiler failed, trying without copy propagation!");
+                       ret = ir3_compile_shader(v, tokens, key, false);
+               }
        }
 
        if (ret) {