RFC: peephole vs RTX_FRAME_RELATED_P

Mon Dec 19 18:23:00 GMT 2005

On i386 we replace (add sp -4) with (push reg). This generates faster
and smaller code.
However, we are not copying RTX_FRAME_RELATED_P from the old
instructions to the new, and so we are not emitting unwind information
for the stack pointer adjustment. The breaks stack traces on gcj, and
I suspect it breaks a bunch of stuff elsewhere too.
This very crude patch sets RTX_FRAME_RELATED_P on every one of the new
instructions if any of the old instructions had RTX_FRAME_RELATED_P
set. It seems to do the trick, but I suspect there must be a more
subtle way to do it.
Can anyone suggest a neater way to do this?
Without this patch, for
java.lang.Throwable.Throwable(java.lang.String) we get
00a326a8 <java.lang.Throwable.Throwable(java.lang.String)>:
 a326a8: 56 push %esi
 a326a9: 53 push %ebx
 a326aa: 50 push %eax <=== ** This is the insn generated by peephole **
 a326ab: e8 00 00 00 00 call a326b0 <java.lang.Throwable.Throwable(java.lang.String)+0x8>
 a326b0: 5b pop %ebx
00058f08 0000001c 0003252c FDE cie=000269e0 pc=00a328f8..00a3292e			 
 Augmentation data: 00 00 00 00				 			 
								 			 
 DW_CFA_advance_loc: 1 to 00a328f9				 			 
 DW_CFA_def_cfa_offset: 8					 			 
 DW_CFA_advance_loc: 1 to 00a328fa				 			 
 DW_CFA_def_cfa_offset: 12					 			 
 DW_CFA_offset: r3 at cfa-12					 			 
 DW_CFA_offset: r6 at cfa-8					 			 
 DW_CFA_nop							 			 
								 
after the patch, we get
 
0005946c 00000020 000328e4 FDE cie=00026b8c pc=00a326a8..00a326de
 Augmentation data: 00 00 00 00
 DW_CFA_advance_loc: 1 to 00a326a9
 DW_CFA_def_cfa_offset: 8
 DW_CFA_advance_loc: 1 to 00a326aa
 DW_CFA_def_cfa_offset: 12
 DW_CFA_advance_loc: 1 to 00a326ab
 DW_CFA_def_cfa_offset: 16
 DW_CFA_offset: r0 at cfa-16
 DW_CFA_offset: r3 at cfa-12
 DW_CFA_offset: r6 at cfa-8
Here's a trivial test case:
public class Hello {
 public static void main(String[] args) {
 System.out.println("Hello, World!");
	new Throwable().printStackTrace();
 }
}
 $ ~/gcc/install/bin/gcj Hello.java --main=Hello -g
 $ LD_LIBRARY_PATH=~/gcc/install/lib ./a.out
Hello, World!
java.lang.Throwable
 at Hello.main (Hello.java:5)
Andrew.
2005年12月19日 Andrew Haley <aph@redhat.com>
	* recog.c (peephole2_optimize): Copy RTX_FRAME_RELATED_P from old
	instructions to new instructions.
Index: recog.c
===================================================================

--- recog.c	(revision 108424)
+++ recog.c	(working copy)
@@ -3107,6 +3107,7 @@
 	 int match_len;
 	 rtx note;
 	 bool was_call = false;
+	 bool frame_related = false;
 
 	 /* Record this insn. */
 	 if (--peep2_current < 0)
@@ -3122,6 +3123,14 @@
 	 try = peephole2_insns (PATTERN (insn), insn, &match_len);
 	 if (try != NULL)
 		{
+		 {
+		 rtx old_insn = insn;
+		 for (i = 0; i <= match_len; ++i)
+		 {
+			frame_related |= RTX_FRAME_RELATED_P (old_insn);
+			old_insn = NEXT_INSN (old_insn);
+		 }		 
+		 }
 		 /* If we are splitting a CALL_INSN, look for the CALL_INSN
 		 in SEQ and copy our CALL_INSN_FUNCTION_USAGE and other
 		 cfg-related call notes. */
@@ -3179,6 +3188,16 @@
 		 break;
 		 }
 
+		 if (frame_related)
+		 {
+		 rtx new_insn = try;
+		 while (new_insn != NULL_RTX)
+			{			 
+			 RTX_FRAME_RELATED_P (new_insn) = true;
+			 new_insn = NEXT_INSN (new_insn);
+			}			 
+		 }
+
 		 i = match_len + peep2_current;
 		 if (i >= MAX_INSNS_PER_PEEP2 + 1)
 		 i -= MAX_INSNS_PER_PEEP2 + 1;
