Add performance related explanatory comment.

	* config/ia64/ia64.md (insv): Add comment.


git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@33836 138bc75d-0d04-0410-961f-82ee72b054a4

1 files changed, 15 insertions, 0 deletions

diff --git a/gcc/config/ia64/ia64.md b/gcc/config/ia64/ia64.md
index 7ccf89dd110..4a55b030a91 100644
--- a/gcc/config/ia64/ia64.md
+++ b/gcc/config/ia64/ia64.md
@@ -684,6 +684,21 @@
 	  DONE;
 	}
 
+      /* We could handle remaining cases by emitting multiple dep
+	 instructions.
+
+	 If we need more than two dep instructions then we lose.  A 6
+	 insn sequence mov mask1,mov mask2,shl;;and,and;;or is better than
+	 mov;;dep,shr;;dep,shr;;dep.  The former can be executed in 3 cycles,
+	 the latter is 6 cycles on an Itanium (TM) processor, because there is
+	 only one function unit that can execute dep and shr immed.
+
+	 If we only need two dep instruction, then we still lose.
+	 mov;;dep,shr;;dep is still 4 cycles.  Even if we optimize away
+	 the unnecessary mov, this is still undesirable because it will be
+	 hard to optimize, and it creates unnecessary pressure on the I0
+	 function unit.  */
+
       FAIL;
 
 #if 0