[RFC] Do not consider volatile asms as optimization barriers #1

Richard Sandiford rdsandiford@googlemail.com
Tue Mar 4 07:27:00 GMT 2014


Richard Sandiford <rdsandiford@googlemail.com> writes:
> I'll run a full test overnight, but does this look like it might be
> a way out, at least for 4.9?

FWIW, it passed testing on x86_64-linux-gnu ({,-m32}, all,ada).
Here it is again with an updated cselib.c comment.  OK to install?

Thanks,
Richard


gcc/
	* builtins.c (expand_builtin_setjmp_receiver): Use and clobber
	hard_frame_pointer_rtx.
	* cse.c (cse_insn): Remove volatile check.
	* cselib.c (cselib_process_insn): Likewise.
	* dse.c (scan_insn): Likewise.

Index: gcc/builtins.c
===================================================================
--- gcc/builtins.c	2014-03-03 21:47:59.749026019 +0000
+++ gcc/builtins.c	2014-03-03 21:48:00.550030853 +0000
@@ -910,18 +910,27 @@ expand_builtin_setjmp_receiver (rtx rece
 #ifdef HAVE_nonlocal_goto
   if (! HAVE_nonlocal_goto)
 #endif
-    /* First adjust our frame pointer to its actual value.  It was
-       previously set to the start of the virtual area corresponding to
-       the stacked variables when we branched here and now needs to be
-       adjusted to the actual hardware fp value.
-
-       Assignments to virtual registers are converted by
-       instantiate_virtual_regs into the corresponding assignment
-       to the underlying register (fp in this case) that makes
-       the original assignment true.
-       So the following insn will actually be decrementing fp by
-       STARTING_FRAME_OFFSET.  */
-    emit_move_insn (virtual_stack_vars_rtx, hard_frame_pointer_rtx);
+    {
+      /* First adjust our frame pointer to its actual value.  It was
+	 previously set to the start of the virtual area corresponding to
+	 the stacked variables when we branched here and now needs to be
+	 adjusted to the actual hardware fp value.
+
+	 Assignments to virtual registers are converted by
+	 instantiate_virtual_regs into the corresponding assignment
+	 to the underlying register (fp in this case) that makes
+	 the original assignment true.
+	 So the following insn will actually be decrementing fp by
+	 STARTING_FRAME_OFFSET.  */
+      emit_move_insn (virtual_stack_vars_rtx, hard_frame_pointer_rtx);
+
+      /* Restoring the frame pointer also modifies the hard frame pointer.
+	 Mark it used (so that the previous assignment remains live once
+	 the frame pointer is eliminated) and clobbered (to represent the
+	 implicit update from the assignment).  */
+      emit_use (hard_frame_pointer_rtx);
+      emit_clobber (hard_frame_pointer_rtx);
+    }
 
 #if !HARD_FRAME_POINTER_IS_ARG_POINTER
   if (fixed_regs[ARG_POINTER_REGNUM])
@@ -965,8 +974,7 @@ expand_builtin_setjmp_receiver (rtx rece
 
   /* We must not allow the code we just generated to be reordered by
      scheduling.  Specifically, the update of the frame pointer must
-     happen immediately, not later.  Similarly, we must block
-     (frame-related) register values to be used across this code.  */
+     happen immediately, not later.  */
   emit_insn (gen_blockage ());
 }
 
Index: gcc/cse.c
===================================================================
--- gcc/cse.c	2014-03-03 21:47:59.869026741 +0000
+++ gcc/cse.c	2014-03-03 21:48:00.625031305 +0000
@@ -5664,11 +5664,6 @@ cse_insn (rtx insn)
 	  invalidate (XEXP (dest, 0), GET_MODE (dest));
       }
 
-  /* A volatile ASM or an UNSPEC_VOLATILE invalidates everything.  */
-  if (NONJUMP_INSN_P (insn)
-      && volatile_insn_p (PATTERN (insn)))
-    flush_hash_table ();
-
   /* Don't cse over a call to setjmp; on some machines (eg VAX)
      the regs restored by the longjmp come from a later time
      than the setjmp.  */
Index: gcc/cselib.c
===================================================================
--- gcc/cselib.c	2014-03-03 21:47:59.870026748 +0000
+++ gcc/cselib.c	2014-03-03 22:09:24.211994918 +0000
@@ -2626,12 +2626,10 @@ cselib_process_insn (rtx insn)
 
   cselib_current_insn = insn;
 
-  /* Forget everything at a CODE_LABEL, a volatile insn, or a setjmp.  */
+  /* Forget everything at a CODE_LABEL or a setjmp.  */
   if ((LABEL_P (insn)
        || (CALL_P (insn)
-	   && find_reg_note (insn, REG_SETJMP, NULL))
-       || (NONJUMP_INSN_P (insn)
-	   && volatile_insn_p (PATTERN (insn))))
+	   && find_reg_note (insn, REG_SETJMP, NULL)))
       && !cselib_preserve_constants)
     {
       cselib_reset_table (next_uid);
Index: gcc/dse.c
===================================================================
--- gcc/dse.c	2014-03-03 21:47:59.871026754 +0000
+++ gcc/dse.c	2014-03-03 21:48:00.627031317 +0000
@@ -2470,16 +2470,6 @@ scan_insn (bb_info_t bb_info, rtx insn)
       return;
     }
 
-  /* Cselib clears the table for this case, so we have to essentially
-     do the same.  */
-  if (NONJUMP_INSN_P (insn)
-      && volatile_insn_p (PATTERN (insn)))
-    {
-      add_wild_read (bb_info);
-      insn_info->cannot_delete = true;
-      return;
-    }
-
   /* Look at all of the uses in the insn.  */
   note_uses (&PATTERN (insn), check_mem_read_use, bb_info);
 



More information about the Gcc-patches mailing list