xtensa: Fix PR target/78118

It started failing after the following commit: 32e90dc6a0cda45 ("PR
rtl-optimization/61047").

The change that made xtensa backend go ICE looks completely unrelated,
and indeed, the issue is caused by the side effect of
compute_frame_size() function call hidden in the
INITIAL_ELIMINATION_OFFSET macro. This call updates the value of the
xtensa_current_frame_size static variable, used in "return" instruction
predicate. Prior to the change the value of xtensa_current_frame_size was
set to 0 after the end of epilogue generation, which enabled the "return"
instruction for the CALL0 ABI, but after the change the additional
INITIAL_ELIMINATION_OFFSET calls make xtensa_current_frame_size non-zero
and "return" pattern unavailable.

Get rid of the global xtensa_current_frame_size and
xtensa_callee_save_size variables by moving them into the
machine_function structure. Implement predicate for the "return" pattern
as a function. Don't communicate completion of epilogue generation
through zeroing of xtensa_current_frame_size, add explicit epilogue_done
variable to the machine_function structure. Don't update stack frame
layout after the completion of reload.

2016-11-01  Max Filippov  <jcmvbkbc@gmail.com>
gcc/
	* config/xtensa/xtensa-protos.h
	(xtensa_use_return_instruction_p): New prototype.
	* config/xtensa/xtensa.c (xtensa_current_frame_size,
	xtensa_callee_save_size): Remove.
	(struct machine_function): Add new fields: current_frame_size,
	callee_save_size, frame_laid_out and epilogue_done.
	(compute_frame_size, xtensa_expand_prologue,
	xtensa_expand_epilogue): Replace xtensa_callee_save_size with
	cfun->machine->callee_save_size and xtensa_current_frame_size
	with cfun->machine->current_frame_size.
	(compute_frame_size): Update cfun->machine->frame_laid_out and
	don't update frame layout after reload completion.
	(xtensa_expand_epilogue): Set cfun->machine->epilogue_done
	instead of zeroing xtensa_current_frame_size.
	(xtensa_use_return_instruction_p): New function.
	* config/xtensa/xtensa.h (xtensa_current_frame_size): Remove
	declaration.
	(INITIAL_ELIMINATION_OFFSET): Use return value of
	compute_frame_size instead of xtensa_current_frame_size value.
	* config/xtensa/xtensa.md ("return" pattern): Use new predicate
	function xtensa_use_return_instruction_p instead of inline code.

From-SVN: r241748

gcc/ChangeLog

+2016-11-01  Max Filippov  <jcmvbkbc@gmail.com>
+
+	* config/xtensa/xtensa-protos.h
+	(xtensa_use_return_instruction_p): New prototype.
+	* config/xtensa/xtensa.c (xtensa_current_frame_size,
+	xtensa_callee_save_size): Remove.
+	(struct machine_function): Add new fields: current_frame_size,
+	callee_save_size, frame_laid_out and epilogue_done.
+	(compute_frame_size, xtensa_expand_prologue,
+	xtensa_expand_epilogue): Replace xtensa_callee_save_size with
+	cfun->machine->callee_save_size and xtensa_current_frame_size
+	with cfun->machine->current_frame_size.
+	(compute_frame_size): Update cfun->machine->frame_laid_out and
+	don't update frame layout after reload completion.
+	(xtensa_expand_epilogue): Set cfun->machine->epilogue_done
+	instead of zeroing xtensa_current_frame_size.
+	(xtensa_use_return_instruction_p): New function.
+	* config/xtensa/xtensa.h (xtensa_current_frame_size): Remove
+	declaration.
+	(INITIAL_ELIMINATION_OFFSET): Use return value of
+	compute_frame_size instead of xtensa_current_frame_size value.
+	* config/xtensa/xtensa.md ("return" pattern): Use new predicate
+	function xtensa_use_return_instruction_p instead of inline code.
+
 2016-11-01  Jakub Jelinek  <jakub@redhat.com>
 
 	* tree.h (BLOCK_IN_COLD_SECTION_P): Define.

gcc/config/xtensa/xtensa-protos.h

@@ -68,6 +68,7 @@ extern rtx xtensa_return_addr (int, rtx);
 extern void xtensa_setup_frame_addresses (void);
 extern int xtensa_dbx_register_number (int);
 extern long compute_frame_size (int);
+extern bool xtensa_use_return_instruction_p (void);
 extern void xtensa_expand_prologue (void);
 extern void xtensa_expand_epilogue (void);
 extern void order_regs_for_local_alloc (void);

gcc/config/xtensa/xtensa.c

@@ -79,11 +79,6 @@ enum internal_test
    can support a given mode.  */
 char xtensa_hard_regno_mode_ok[(int) MAX_MACHINE_MODE][FIRST_PSEUDO_REGISTER];
 
-/* Current frame size calculated by compute_frame_size.  */
-unsigned xtensa_current_frame_size;
-/* Callee-save area size in the current frame calculated by compute_frame_size. */
-int xtensa_callee_save_size;
-
 /* Largest block move to handle in-line.  */
 #define LARGEST_MOVE_RATIO 15
 
@@ -95,6 +90,13 @@ struct GTY(()) machine_function
   bool vararg_a7;
   rtx vararg_a7_copy;
   rtx_insn *set_frame_ptr_insn;
+  /* Current frame size calculated by compute_frame_size.  */
+  unsigned current_frame_size;
+  /* Callee-save area size in the current frame calculated by
+     compute_frame_size.  */
+  int callee_save_size;
+  bool frame_laid_out;
+  bool epilogue_done;
 };
 
 /* Vector, indexed by hard register number, which contains 1 for a
@@ -2632,24 +2634,29 @@ compute_frame_size (int size)
 {
   int regno;
 
+  if (reload_completed && cfun->machine->frame_laid_out)
+    return cfun->machine->current_frame_size;
+
   /* Add space for the incoming static chain value.  */
   if (cfun->static_chain_decl != NULL)
     size += (1 * UNITS_PER_WORD);
 
-  xtensa_callee_save_size = 0;
+  cfun->machine->callee_save_size = 0;
   for (regno = 0; regno < FIRST_PSEUDO_REGISTER; ++regno)
     {
       if (xtensa_call_save_reg(regno))
-	xtensa_callee_save_size += UNITS_PER_WORD;
+	cfun->machine->callee_save_size += UNITS_PER_WORD;
     }
 
-  xtensa_current_frame_size =
+  cfun->machine->current_frame_size =
     XTENSA_STACK_ALIGN (size
-			+ xtensa_callee_save_size
+			+ cfun->machine->callee_save_size
 			+ crtl->outgoing_args_size
 			+ (WINDOW_SIZE * UNITS_PER_WORD));
-  xtensa_callee_save_size = XTENSA_STACK_ALIGN (xtensa_callee_save_size);
-  return xtensa_current_frame_size;
+  cfun->machine->callee_save_size =
+    XTENSA_STACK_ALIGN (cfun->machine->callee_save_size);
+  cfun->machine->frame_laid_out = true;
+  return cfun->machine->current_frame_size;
 }
 
 
@@ -2703,6 +2710,7 @@ xtensa_expand_prologue (void)
     {
       int regno;
       HOST_WIDE_INT offset = 0;
+      int callee_save_size = cfun->machine->callee_save_size;
 
       /* -128 is a limit of single addi instruction. */
       if (total_size > 0 && total_size <= 128)
@@ -2716,7 +2724,7 @@ xtensa_expand_prologue (void)
 	  add_reg_note (insn, REG_FRAME_RELATED_EXPR, note_rtx);
 	  offset = total_size - UNITS_PER_WORD;
 	}
-      else if (xtensa_callee_save_size)
+      else if (callee_save_size)
 	{
 	  /* 1020 is maximal s32i offset, if the frame is bigger than that
 	   * we move sp to the end of callee-saved save area, save and then
@@ -2724,13 +2732,13 @@ xtensa_expand_prologue (void)
 	  if (total_size > 1024)
 	    {
 	      insn = emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx,
-					    GEN_INT (-xtensa_callee_save_size)));
+					    GEN_INT (-callee_save_size)));
 	      RTX_FRAME_RELATED_P (insn) = 1;
 	      note_rtx = gen_rtx_SET (stack_pointer_rtx,
 				      plus_constant (Pmode, stack_pointer_rtx,
-						     -xtensa_callee_save_size));
+						     -callee_save_size));
 	      add_reg_note (insn, REG_FRAME_RELATED_EXPR, note_rtx);
-	      offset = xtensa_callee_save_size - UNITS_PER_WORD;
+	      offset = callee_save_size - UNITS_PER_WORD;
 	    }
 	  else
 	    {
@@ -2766,13 +2774,13 @@ xtensa_expand_prologue (void)
 	{
 	  rtx tmp_reg = gen_rtx_REG (Pmode, A9_REG);
 	  emit_move_insn (tmp_reg, GEN_INT (total_size -
-					    xtensa_callee_save_size));
+					    callee_save_size));
 	  insn = emit_insn (gen_subsi3 (stack_pointer_rtx,
 					stack_pointer_rtx, tmp_reg));
 	  RTX_FRAME_RELATED_P (insn) = 1;
 	  note_rtx = gen_rtx_SET (stack_pointer_rtx,
 				  plus_constant (Pmode, stack_pointer_rtx,
-						 xtensa_callee_save_size -
+						 callee_save_size -
 						 total_size));
 	  add_reg_note (insn, REG_FRAME_RELATED_EXPR, note_rtx);
 	}
@@ -2840,21 +2848,21 @@ xtensa_expand_epilogue (void)
       int regno;
       HOST_WIDE_INT offset;
 
-      if (xtensa_current_frame_size > (frame_pointer_needed ? 127 : 1024))
+      if (cfun->machine->current_frame_size > (frame_pointer_needed ? 127 : 1024))
 	{
 	  rtx tmp_reg = gen_rtx_REG (Pmode, A9_REG);
-	  emit_move_insn (tmp_reg, GEN_INT (xtensa_current_frame_size -
-					    xtensa_callee_save_size));
+	  emit_move_insn (tmp_reg, GEN_INT (cfun->machine->current_frame_size -
+					    cfun->machine->callee_save_size));
 	  emit_insn (gen_addsi3 (stack_pointer_rtx, frame_pointer_needed ?
 				 hard_frame_pointer_rtx : stack_pointer_rtx,
 				 tmp_reg));
-	  offset = xtensa_callee_save_size - UNITS_PER_WORD;
+	  offset = cfun->machine->callee_save_size - UNITS_PER_WORD;
 	}
       else
 	{
 	  if (frame_pointer_needed)
 	    emit_move_insn (stack_pointer_rtx, hard_frame_pointer_rtx);
-	  offset = xtensa_current_frame_size - UNITS_PER_WORD;
+	  offset = cfun->machine->current_frame_size - UNITS_PER_WORD;
 	}
 
       /* Prevent reordering of saved a0 update and loading it back from
@@ -2874,16 +2882,16 @@ xtensa_expand_epilogue (void)
 	    }
 	}
 
-      if (xtensa_current_frame_size > 0)
+      if (cfun->machine->current_frame_size > 0)
 	{
 	  if (frame_pointer_needed || /* always reachable with addi */
-	      xtensa_current_frame_size > 1024 ||
-	      xtensa_current_frame_size <= 127)
+	      cfun->machine->current_frame_size > 1024 ||
+	      cfun->machine->current_frame_size <= 127)
 	    {
-	      if (xtensa_current_frame_size <= 127)
-		offset = xtensa_current_frame_size;
+	      if (cfun->machine->current_frame_size <= 127)
+		offset = cfun->machine->current_frame_size;
 	      else
-		offset = xtensa_callee_save_size;
+		offset = cfun->machine->callee_save_size;
 
 	      emit_insn (gen_addsi3 (stack_pointer_rtx,
 				     stack_pointer_rtx,
@@ -2892,7 +2900,8 @@ xtensa_expand_epilogue (void)
 	  else
 	    {
 	      rtx tmp_reg = gen_rtx_REG (Pmode, A9_REG);
-	      emit_move_insn (tmp_reg, GEN_INT (xtensa_current_frame_size));
+	      emit_move_insn (tmp_reg,
+			      GEN_INT (cfun->machine->current_frame_size));
 	      emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx,
 				     tmp_reg));
 	    }
@@ -2903,11 +2912,22 @@ xtensa_expand_epilogue (void)
 				  stack_pointer_rtx,
 				  EH_RETURN_STACKADJ_RTX));
     }
-  xtensa_current_frame_size = 0;
-  xtensa_callee_save_size = 0;
+  cfun->machine->epilogue_done = true;
   emit_jump_insn (gen_return ());
 }
 
+bool
+xtensa_use_return_instruction_p (void)
+{
+  if (!reload_completed)
+    return false;
+  if (TARGET_WINDOWED_ABI)
+    return true;
+  if (compute_frame_size (get_frame_size ()) == 0)
+    return true;
+  return cfun->machine->epilogue_done;
+}
+
 void
 xtensa_set_return_address (rtx address, rtx scratch)
 {

gcc/config/xtensa/xtensa.h

@@ -23,8 +23,6 @@ along with GCC; see the file COPYING3.  If not see
 
 /* External variables defined in xtensa.c.  */
 
-extern unsigned xtensa_current_frame_size;
-
 /* Macros used in the machine description to select various Xtensa
    configuration options.  */
 #ifndef XCHAL_HAVE_MUL32_HIGH
@@ -477,14 +475,14 @@ enum reg_class
 /* Specify the initial difference between the specified pair of registers.  */
 #define INITIAL_ELIMINATION_OFFSET(FROM, TO, OFFSET)			\
   do {									\
-    compute_frame_size (get_frame_size ());				\
+    long frame_size = compute_frame_size (get_frame_size ());		\
     switch (FROM)							\
       {									\
       case FRAME_POINTER_REGNUM:					\
         (OFFSET) = 0;							\
 	break;								\
       case ARG_POINTER_REGNUM:						\
-        (OFFSET) = xtensa_current_frame_size;				\
+        (OFFSET) = frame_size;						\
 	break;								\
       default:								\
 	gcc_unreachable ();						\

gcc/config/xtensa/xtensa.md

@@ -1663,7 +1663,7 @@
 (define_insn "return"
   [(return)
    (use (reg:SI A0_REG))]
-  "(TARGET_WINDOWED_ABI || !xtensa_current_frame_size) && reload_completed"
+  "xtensa_use_return_instruction_p ()"
 {
   return TARGET_WINDOWED_ABI ?
       (TARGET_DENSITY ? "retw.n" : "retw") :