intel/fs: work around gen12 lower-precision source modifier limitation

GEN:BUG:1604601757 prevents source modifiers for multiplication of
lower precision integers.

lower_mul_dword_inst() splits 32x32 multiplication into 32x16, and
needs to eliminate source modifiers in this case.

Closes: #3329
Reviewed-by: Francisco Jerez <currojerez@riseup.net>

diff --git a/src/intel/compiler/brw_fs.cpp b/src/intel/compiler/brw_fs.cpp
index a89c53ea5a51b19cbae5df7305ec75975ef792f2..dbf7391baef516bae1820e8103cbc6e2d32343db 100644 (file)
--- a/src/intel/compiler/brw_fs.cpp
+++ b/src/intel/compiler/brw_fs.cpp
@@ -3930,7 +3930,20 @@ fs_visitor::lower_mul_dword_inst(fs_inst *inst, bblock_t *block)
       high.offset = inst->dst.offset % REG_SIZE;
 
       if (devinfo->gen >= 7) {
-         if (inst->src[1].abs)
+         /* From GEN:BUG:1604601757:
+          *
+          * "When multiplying a DW and any lower precision integer, source modifier
+          *  is not supported."
+          *
+          * An unsupported negate modifier on src[1] would ordinarily be
+          * lowered by the subsequent lower_regioning pass.  In this case that
+          * pass would spawn another dword multiply.  Instead, lower the
+          * modifier first.
+          */
+         const bool source_mods_unsupported = (devinfo->gen >= 12);
+
+         if (inst->src[1].abs || (inst->src[1].negate &&
+                                  source_mods_unsupported))
             lower_src_modifiers(this, block, inst, 1);
 
          if (inst->src[1].file == IMM) {