nir/lower_doubles: Handle fdiv and fsub directly

Reviewed-by: Eric Anholt <eric@anholt.net>

diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 6805d067a80b8f78ce7fc3120f31fc295239a04d..99fa4baa16a3a6c5243505458a0ed0c909a0cc6e 100644 (file)
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -2476,7 +2476,9 @@ typedef enum {
    nir_lower_dfract = (1 << 6),
    nir_lower_dround_even = (1 << 7),
    nir_lower_dmod = (1 << 8),
-   nir_lower_fp64_full_software = (1 << 9),
+   nir_lower_dsub = (1 << 9),
+   nir_lower_ddiv = (1 << 10),
+   nir_lower_fp64_full_software = (1 << 11),
 } nir_lower_doubles_options;
 
 typedef struct nir_shader_compiler_options {

diff --git a/src/compiler/nir/nir_lower_double_ops.c b/src/compiler/nir/nir_lower_double_ops.c
index 211c9359558f728d317f6274ea80f55b7a7404af..45ac155de51c622e7fd50d688acda1688c4cc94e 100644 (file)
--- a/src/compiler/nir/nir_lower_double_ops.c
+++ b/src/compiler/nir/nir_lower_double_ops.c
@@ -600,6 +600,8 @@ nir_lower_doubles_op_to_options_mask(nir_op opcode)
    case nir_op_ffract:        return nir_lower_dfract;
    case nir_op_fround_even:   return nir_lower_dround_even;
    case nir_op_fmod:          return nir_lower_dmod;
+   case nir_op_fsub:          return nir_lower_dsub;
+   case nir_op_fdiv:          return nir_lower_ddiv;
    default:                   return 0;
    }
 }
@@ -673,10 +675,21 @@ lower_doubles_instr(nir_builder *b, nir_instr *instr, void *_data)
    case nir_op_fround_even:
       return lower_round_even(b, src);
 
+   case nir_op_fdiv:
+   case nir_op_fsub:
    case nir_op_fmod: {
       nir_ssa_def *src1 = nir_mov_alu(b, alu->src[1],
                                       alu->dest.dest.ssa.num_components);
-      return lower_mod(b, src, src1);
+      switch (alu->op) {
+      case nir_op_fdiv:
+         return nir_fmul(b, src, nir_frcp(b, src1));
+      case nir_op_fsub:
+         return nir_fadd(b, src, nir_fneg(b, src1));
+      case nir_op_fmod:
+         return lower_mod(b, src, src1);
+      default:
+         unreachable("unhandled opcode");
+      }
    }
    default:
       unreachable("unhandled opcode");