[PATCH] Use gen_lowpart_or_truncate in simplify_shift_const_1

[Adam Nemet <anemet at caviumnetworks dot com>]

This was uncovered by a MIPS64 patch that I am working on to let
backends turn explicit truncates into implicit ones.  With the patch I
got an UNPREDICTABLE exception in gcc.c-torture/execute/simd-2.c on
the simulator.  This exception occurs if a 32-bit mode operation is
performed on a register that has not been properly truncated.

The problem is in simplify_shift_const_1.  It does this
transformation:

  (ashift:SI (subreg:SI (lshiftrt:DI (reg:DI 436) 48)) 16)
  -> (subreg:SI (ashift:DI (lshiftrt:DI (reg:DI 436) 48) 16)))

If truncation from DI to SI requires explicit truncation we cannot
just lift the subreg unconditionally here.

I am changing the call to gen_lowpart to gen_lowpart_or_truncate which
will only generate a SUBREG if the value is known have already been
truncated.

There are lot of other places where we use gen_lowpart which might not
be correct.  Changing them to use gen_lowpart_or_truncate would hurt
performance at this point.  I hope that after the above mentioned
patch this will not be as intrusive.

This patch was bootstrapped and tested on i686-pc-linux-gnu and tested
on mipsisa64-elf with and without the patch triggering the bug.

OK to install?

Adam

	* combine.c (simplify_shift_const_1): Use gen_lowpart_or_truncate
	instead of gen_lowpart.

Index: combine.c
===================================================================
--- combine.c	(revision 111703)
+++ combine.c	(working copy)
@@ -9199,7 +9199,7 @@ simplify_shift_const_1 (enum rtx_code co
 				GET_MODE_MASK (result_mode) >> orig_count);
 
   /* Do the remainder of the processing in RESULT_MODE.  */
-  x = gen_lowpart (result_mode, x);
+  x = gen_lowpart_or_truncate (result_mode, x);
 
   /* If COMPLEMENT_P is set, we have to complement X before doing the outer
      operation.  */