nir: remove bool lowering from lower_int_to_float

Removes the bool_to_float logic from the int_to_float pass, so that both
can be used separately. By having separate passes we have better validation
and it makes it possible to use with the lower_ftrunc option (int lowering
generates ftrunc, but lower_ftrunc generates bools, ftrunc lowering should
probably be reworked). For now we always expect lower_bool to come after
lower_int.

Also fixes f2i32 to become ftrunc and adds u2f/f2u cases.

Signed-off-by: Jonathan Marek <[email protected]>
Reviewed-by: Jason Ekstrand <[email protected]>

3 files changed, 45 insertions, 71 deletions

diff --git a/src/compiler/nir/nir_lower_int_to_float.c b/src/compiler/nir/nir_lower_int_to_float.c
index e3503dcc8f8..f082b663302 100644
--- a/src/compiler/nir/nir_lower_int_to_float.c
+++ b/src/compiler/nir/nir_lower_int_to_float.c
@@ -34,22 +34,20 @@ assert_ssa_def_is_not_int(nir_ssa_def *def, void *arg)
 }
 
 static bool
-rewrite_bool_ssa_def_to_float(nir_ssa_def *def, void *_progress)
-{
-   bool *progress = _progress;
-   if (def->bit_size == 1) {
-      def->bit_size = 32;
-      *progress = true;
-   }
-   return true;
-}
-
-static bool
 lower_alu_instr(nir_builder *b, nir_alu_instr *alu)
 {
-   const nir_op_info *op_info = &nir_op_infos[alu->op];
+   const nir_op_info *info = &nir_op_infos[alu->op];
+
+   bool is_bool_only = alu->dest.dest.ssa.bit_size == 1;
+   for (unsigned i = 0; i < info->num_inputs; i++) {
+      if (alu->src[i].src.ssa->bit_size != 1)
+         is_bool_only = false;
+   }
 
-   b->cursor = nir_before_instr(&alu->instr);
+   if (is_bool_only) {
+      /* avoid lowering integers ops are used for booleans (ieq,ine,etc) */
+      return false;
+   }
 
    /* Replacement SSA value */
    nir_ssa_def *rep = NULL;
@@ -58,69 +56,51 @@ lower_alu_instr(nir_builder *b, nir_alu_instr *alu)
    case nir_op_vec2:
    case nir_op_vec3:
    case nir_op_vec4:
-      /* These we expect to have integers or booleans but the opcode doesn't change */
+   case nir_op_bcsel:
+      /* These we expect to have integers but the opcode doesn't change */
       break;
 
-   case nir_op_b2f32: alu->op = nir_op_mov; break;
-   case nir_op_b2i32: alu->op = nir_op_mov; break;
+   case nir_op_b2i32: alu->op = nir_op_b2f32; break;
    case nir_op_i2f32: alu->op = nir_op_mov; break;
-   case nir_op_f2i32: alu->op = nir_op_mov; break;
-   case nir_op_f2i1:
-   case nir_op_f2b1:
-   case nir_op_i2b1:
-      rep = nir_sne(b, nir_ssa_for_alu_src(b, alu, 0),
-                       nir_imm_float(b, 0));
-      break;
-
-   case nir_op_flt: alu->op = nir_op_slt; break;
-   case nir_op_fge: alu->op = nir_op_sge; break;
-   case nir_op_feq: alu->op = nir_op_seq; break;
-   case nir_op_fne: alu->op = nir_op_sne; break;
-   case nir_op_ilt: alu->op = nir_op_slt; break;
-   case nir_op_ige: alu->op = nir_op_sge; break;
-   case nir_op_ieq: alu->op = nir_op_seq; break;
-   case nir_op_ine: alu->op = nir_op_sne; break;
-   case nir_op_ult: alu->op = nir_op_slt; break;
-   case nir_op_uge: alu->op = nir_op_sge; break;
-   case nir_op_ineg: alu->op = nir_op_fneg; break;
-
-   case nir_op_ball_fequal2:  alu->op = nir_op_fall_equal2; break;
-   case nir_op_ball_fequal3:  alu->op = nir_op_fall_equal3; break;
-   case nir_op_ball_fequal4:  alu->op = nir_op_fall_equal4; break;
-   case nir_op_bany_fnequal2: alu->op = nir_op_fany_nequal2; break;
-   case nir_op_bany_fnequal3: alu->op = nir_op_fany_nequal3; break;
-   case nir_op_bany_fnequal4: alu->op = nir_op_fany_nequal4; break;
-   case nir_op_ball_iequal2:  alu->op = nir_op_fall_equal2; break;
-   case nir_op_ball_iequal3:  alu->op = nir_op_fall_equal3; break;
-   case nir_op_ball_iequal4:  alu->op = nir_op_fall_equal4; break;
-   case nir_op_bany_inequal2: alu->op = nir_op_fany_nequal2; break;
-   case nir_op_bany_inequal3: alu->op = nir_op_fany_nequal3; break;
-   case nir_op_bany_inequal4: alu->op = nir_op_fany_nequal4; break;
-
-   case nir_op_bcsel: alu->op = nir_op_fcsel; break;
+   case nir_op_u2f32: alu->op = nir_op_mov; break;
+   case nir_op_f2i32: alu->op = nir_op_ftrunc; break;
+   case nir_op_f2u32: alu->op = nir_op_ffloor; break;
+   case nir_op_i2b1: alu->op = nir_op_f2b1; break;
+
+   case nir_op_ilt: alu->op = nir_op_flt; break;
+   case nir_op_ige: alu->op = nir_op_fge; break;
+   case nir_op_ieq: alu->op = nir_op_feq; break;
+   case nir_op_ine: alu->op = nir_op_fne; break;
+   case nir_op_ult: alu->op = nir_op_flt; break;
+   case nir_op_uge: alu->op = nir_op_fge; break;
 
    case nir_op_iadd: alu->op = nir_op_fadd; break;
    case nir_op_isub: alu->op = nir_op_fsub; break;
    case nir_op_imul: alu->op = nir_op_fmul; break;
-   case nir_op_iand: alu->op = nir_op_fmul; break;
-   case nir_op_ixor: alu->op = nir_op_sne; break;
-   case nir_op_ior: alu->op = nir_op_fmax; break;
-
    case nir_op_idiv:
       rep = nir_ftrunc(b, nir_fdiv(b,
                                    nir_ssa_for_alu_src(b, alu, 0),
                                    nir_ssa_for_alu_src(b, alu, 1)));
       break;
+   case nir_op_iabs: alu->op = nir_op_fabs; break;
+   case nir_op_ineg: alu->op = nir_op_fneg; break;
+   case nir_op_imax: alu->op = nir_op_fmax; break;
+   case nir_op_imin: alu->op = nir_op_fmin; break;
 
-   case nir_op_inot:
-      rep = nir_seq(b, nir_ssa_for_alu_src(b, alu, 0),
-                       nir_imm_float(b, 0));
-      break;
+   case nir_op_ball_iequal2:  alu->op = nir_op_ball_fequal2; break;
+   case nir_op_ball_iequal3:  alu->op = nir_op_ball_fequal3; break;
+   case nir_op_ball_iequal4:  alu->op = nir_op_ball_fequal4; break;
+   case nir_op_bany_inequal2: alu->op = nir_op_bany_fnequal2; break;
+   case nir_op_bany_inequal3: alu->op = nir_op_bany_fnequal3; break;
+   case nir_op_bany_inequal4: alu->op = nir_op_bany_fnequal4; break;
 
    default:
-      assert(alu->dest.dest.ssa.bit_size > 1);
-      for (unsigned i = 0; i < op_info->num_inputs; i++)
-         assert(alu->src[i].src.ssa->bit_size > 1);
+      assert(nir_alu_type_get_base_type(info->output_type) != nir_type_int &&
+             nir_alu_type_get_base_type(info->output_type) != nir_type_uint);
+      for (unsigned i = 0; i < info->num_inputs; i++) {
+         assert(nir_alu_type_get_base_type(info->input_types[i]) != nir_type_int &&
+                nir_alu_type_get_base_type(info->input_types[i]) != nir_type_uint);
+      }
       return false;
    }
 
@@ -128,9 +108,6 @@ lower_alu_instr(nir_builder *b, nir_alu_instr *alu)
       /* We've emitted a replacement instruction */
       nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(rep));
       nir_instr_remove(&alu->instr);
-   } else {
-      if (alu->dest.dest.ssa.bit_size == 1)
-         alu->dest.dest.ssa.bit_size = 32;
    }
 
    return true;
@@ -161,11 +138,7 @@ nir_lower_int_to_float_impl(nir_function_impl *impl)
 
          case nir_instr_type_load_const: {
             nir_load_const_instr *load = nir_instr_as_load_const(instr);
-            if (load->def.bit_size == 1) {
-               for (unsigned i = 0; i < load->def.num_components; i++)
-                  load->value[i].f32 = load->value[i].b ? 1.0 : 0.0;
-               load->def.bit_size = 32;
-            } else if (BITSET_TEST(int_types, load->def.index)) {
+            if (load->def.bit_size != 1 && BITSET_TEST(int_types, load->def.index)) {
                for (unsigned i = 0; i < load->def.num_components; i++)
                   load->value[i].f32 = load->value[i].i32;
             }
@@ -176,8 +149,6 @@ nir_lower_int_to_float_impl(nir_function_impl *impl)
          case nir_instr_type_ssa_undef:
          case nir_instr_type_phi:
          case nir_instr_type_tex:
-            nir_foreach_ssa_def(instr, rewrite_bool_ssa_def_to_float,
-                                &progress);
             break;
 
          default:
diff --git a/src/gallium/drivers/freedreno/a2xx/ir2_nir.c b/src/gallium/drivers/freedreno/a2xx/ir2_nir.c
index b9f372a3fd0..5cfb831b6e2 100644
--- a/src/gallium/drivers/freedreno/a2xx/ir2_nir.c
+++ b/src/gallium/drivers/freedreno/a2xx/ir2_nir.c
@@ -1073,6 +1073,7 @@ ir2_nir_compile(struct ir2_context *ctx, bool binning)
 	OPT_V(ctx->nir, nir_opt_move_comparisons);
 
 	OPT_V(ctx->nir, nir_lower_bool_to_float);
+	OPT_V(ctx->nir, nir_lower_int_to_float);
 
 	/* lower to scalar instructions that can only be scalar on a2xx */
 	OPT_V(ctx->nir, ir2_nir_lower_scalar);
diff --git a/src/gallium/drivers/lima/lima_program.c b/src/gallium/drivers/lima/lima_program.c
index e01c487cfd3..38776154edb 100644
--- a/src/gallium/drivers/lima/lima_program.c
+++ b/src/gallium/drivers/lima/lima_program.c
@@ -117,6 +117,7 @@ lima_program_optimize_vs_nir(struct nir_shader *s)
    } while (progress);
 
    NIR_PASS_V(s, nir_lower_int_to_float);
+   NIR_PASS_V(s, nir_lower_bool_to_float);
    NIR_PASS_V(s, nir_copy_prop);
    NIR_PASS_V(s, nir_lower_locals_to_regs);
    NIR_PASS_V(s, nir_convert_from_ssa, true);
@@ -155,6 +156,7 @@ lima_program_optimize_fs_nir(struct nir_shader *s)
    } while (progress);
 
    NIR_PASS_V(s, nir_lower_int_to_float);
+   NIR_PASS_V(s, nir_lower_bool_to_float);
 
    /* Lower modifiers */
    NIR_PASS_V(s, nir_lower_to_source_mods, nir_lower_all_source_mods);