intel/compiler: implement dummy mov for Wa_18035690555

Signed-off-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37804>

src/intel/compiler/brw/brw_shader.cpp

@@ -956,6 +956,8 @@ brw_allocate_registers(brw_shader &s, bool allow_spilling)
    if (s.failed)
       return;
 
+   brw_workaround_emit_dummy_mov_mulmac(s);
+
    OPT(brw_lower_scoreboard);
 }
 

src/intel/compiler/brw/brw_shader.h

@@ -381,6 +381,7 @@ bool brw_workaround_emit_dummy_mov_instruction(brw_shader &s);
 bool brw_workaround_memory_fence_before_eot(brw_shader &s);
 bool brw_workaround_nomask_control_flow(brw_shader &s);
 bool brw_workaround_source_arf_before_eot(brw_shader &s);
+bool brw_workaround_emit_dummy_mov_mulmac(brw_shader &s);
 
 /* Helpers. */
 unsigned brw_get_lowered_simd_width(const brw_shader *shader,

src/intel/compiler/brw/brw_workaround.cpp

@@ -36,6 +36,60 @@ brw_workaround_emit_dummy_mov_instruction(brw_shader &s)
    return true;
 }
 
+/* Wa_18035690555
+ *
+ * If we have mul <-> mac or macl <-> mach and src1 is the same in current
+ * and previous inst, we need to insert a dummy mov in between. We can skip
+ * issue 2 mentioned in wa as macl is not used by our compiler.
+ *
+ * Other conditions listed in the issue for mul <-> mac case:
+ *    "prev instruction src1 has regioning/scalar" (not flat)
+ *    "current instruction src1 is flat and shares the same src1 as prev"
+ */
+bool
+brw_workaround_emit_dummy_mov_mulmac(brw_shader &s)
+{
+   if (!intel_needs_workaround(s.devinfo, 18035690555))
+      return false;
+
+#define IS_MUL_CLASS(x) \
+   (x->opcode == BRW_OPCODE_MUL || x->opcode == BRW_OPCODE_MAC)
+#define IS_MACL_CLASS(x) \
+   (x->opcode == BRW_OPCODE_MACH)
+
+#define IS_FLAT(x, i) (x->dst.subnr == x->src[i].subnr && \
+                       x->src[i].is_contiguous())
+
+   brw_inst *prev_inst = NULL;
+   bool progress = false;
+   foreach_block_and_inst_safe (block, brw_inst, inst, s.cfg) {
+      if (prev_inst &&
+          inferred_exec_pipe(s.devinfo, inst) ==
+          inferred_exec_pipe(s.devinfo, prev_inst) &&
+          ((IS_MUL_CLASS(inst) && IS_MUL_CLASS(prev_inst)) ||
+           (IS_MACL_CLASS(inst) && IS_MACL_CLASS(prev_inst))) &&
+          (phys_nr(s.devinfo, inst->src[1]) ==
+           phys_nr(s.devinfo, prev_inst->src[1])) &&
+          (IS_FLAT(inst, 1) && !IS_FLAT(prev_inst, 1))) {
+         /* Insert dummy mov between prev and current inst. */
+         const brw_builder ubld = brw_builder(prev_inst).exec_all().group(8, 0);
+         ubld.MOV(ubld.null_reg_ud(), brw_imm_ud(0u));
+         progress = true;
+      }
+      prev_inst = inst;
+   }
+
+   if (progress) {
+      s.invalidate_analysis(BRW_DEPENDENCY_INSTRUCTIONS |
+                            BRW_DEPENDENCY_VARIABLES);
+   }
+#undef IS_FLAT
+#undef IS_MUL_CLASS
+#undef IS_MACL_CLASS
+
+   return progress;
+}
+
 static bool
 needs_dummy_fence(const intel_device_info *devinfo, const brw_inst *inst)
 {