intel/fs: Fix nir_op_b2[fi] with 64-bit result on Gen8 LP and Gen9 LP

Several of the Atom GPUs have additional restrictions on alignment when
moving < 64-bit source to a 64-bit destination.  All of the nir_op_*2*64
code generation paths respected this, but nir_op_b2[fi] did not.

Previous to commit a68dd47b911 it was not possible to generate such an
instruction from the GLSL path.  It may have been possible from SPIR-V,
but it's not clear.  The aforementioned patch converts a 64-bit
nir_op_fsign into a sequence of operations including a nir_op_b2f with a
64-bit result.  This "just works" everywhere except these Atom parts.

This problem was not detected during normal CI testing because the Atom
parts are not included in developer builds.

v2 (idr): Make the patch compile, and make some cosmetic changes.  Add a
commit message.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=108319
Fixes: a68dd47b911 "nir/algebraic: Simplify fsat of fsign"
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

diff --git a/src/intel/compiler/brw_fs_nir.cpp b/src/intel/compiler/brw_fs_nir.cpp
index 12b087a..7930205 100644 (file)
--- a/src/intel/compiler/brw_fs_nir.cpp
+++ b/src/intel/compiler/brw_fs_nir.cpp
@@ -793,6 +793,11 @@ fs_visitor::nir_emit_alu(const fs_builder &bld, nir_alu_instr *instr)
       inst->saturate = instr->dest.saturate;
       break;
 
+   case nir_op_b2i:
+   case nir_op_b2f:
+      op[0].type = BRW_REGISTER_TYPE_D;
+      op[0].negate = !op[0].negate;
+      /* fallthrough */
    case nir_op_f2f64:
    case nir_op_f2i64:
    case nir_op_f2u64:
@@ -1213,11 +1218,6 @@ fs_visitor::nir_emit_alu(const fs_builder &bld, nir_alu_instr *instr)
       inst->saturate = instr->dest.saturate;
       break;
 
-   case nir_op_b2i:
-   case nir_op_b2f:
-      bld.MOV(result, negate(op[0]));
-      break;
-
    case nir_op_i2b:
    case nir_op_f2b: {
       uint32_t bit_size = nir_src_bit_size(instr->src[0].src);