[committed] PR 35838: Fix subreg offset in dse.c:find_shift_sequence

[committed] PR 35838: Fix subreg offset in dse.c:find_shift_sequence

[Richard Sandiford]

In the fix for PR 33927, I'd wrongly thought that the first subreg
created by find_shift_sequence would be a paradoxical one.  This isn't
so: we can do the shift in a narrower mode than the store mode if:

   shift + GET_MODE_SIZE (read_mode) <= GET_MODE_SIZE (store_mode) / 2

This caused a wrong code bug on PA.

Bootstrapped & regression-tested on x86_64-linux-gnu ({,-m32}).
Dave Anglin confirmed that it fixes the PR (thanks).  Applied as obvious.

Richard


gcc/
	PR rtl-optimization/35838
	* dse.c (find_shift_sequence): Use subreg_lowpart_offset to work
	out the byte offset of the first subreg.

Index: gcc/dse.c
===================================================================
--- gcc/dse.c	2008-04-14 12:20:03.000000000 +0100
+++ gcc/dse.c	2008-04-14 12:20:27.000000000 +0100
@@ -1446,7 +1446,7 @@ find_shift_sequence (int access_size,
        new_mode = GET_MODE_WIDER_MODE (new_mode))
     {
       rtx target, new_reg, shift_seq, insn, new_lhs;
-      int cost;
+      int cost, offset;
 
       /* Try a wider mode if truncating the store mode to NEW_MODE
 	 requires a real instruction.  */
@@ -1460,8 +1460,9 @@ find_shift_sequence (int access_size,
       if (!CONSTANT_P (store_info->rhs)
 	  && !MODES_TIEABLE_P (new_mode, store_mode))
 	continue;
+      offset = subreg_lowpart_offset (new_mode, store_mode);
       new_lhs = simplify_gen_subreg (new_mode, copy_rtx (store_info->rhs),
-				     store_mode, 0);
+				     store_mode, offset);
       if (new_lhs == NULL_RTX)
 	continue;