re PR rtl-optimization/33927 (replace_read in dse.c could handle cases where GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode) (and the size is the same))

gcc/
	PR rtl-optimization/33927
	* Makefile.in (dse.o): Depend on $(TM_P_H).
	* expr.h (extract_low_bits): Declare.
	* expmed.c (extract_low_bits): New function.
	* rtlhooks.c (gen_lowpart_general): Generalize SUBREG handling.
	* dse.c: Include tm_p.h.
	(find_shift_sequence): Remove the read_reg argument and return the
	read value.  Emit the instructions instead of returning them.
	Iterate on new_mode rather than calculating it each time.
	Check MODES_TIEABLE_P.  Use simplify_gen_subreg to convert the
	source to NEW_MODE and extract_low_bits to convert the shifted
	value to READ_MODE.
	(replace_read): Allow the load and store to have different mode
	classes.  Use extract_low_bits when SHIFT == 0.  Create the shift
	or extraction instructions before trying the replacement.  Update
	dump-file code accordingly, avoiding use of REGNO (store_info->rhs).

gcc/testsuite/
	* gcc.target/mips/dse-1.c: Add checks for zeros.

From-SVN: r133452

1 files changed, 60 insertions, 0 deletions

diff --git a/gcc/expmed.c b/gcc/expmed.c
index 04071d3..2f2e3cc 100644
--- a/gcc/expmed.c
+++ b/gcc/expmed.c
@@ -2098,6 +2098,66 @@ extract_split_bit_field (rtx op0, unsigned HOST_WIDE_INT bitsize,
 		       NULL_RTX, 0);
 }
 
+/* Try to read the low bits of SRC as an rvalue of mode MODE, preserving
+   the bit pattern.  SRC_MODE is the mode of SRC; if this is smaller than
+   MODE, fill the upper bits with zeros.  Fail if the layout of either
+   mode is unknown (as for CC modes) or if the extraction would involve
+   unprofitable mode punning.  Return the value on success, otherwise
+   return null.
+
+   This is different from gen_lowpart* in these respects:
+
+     - the returned value must always be considered an rvalue
+
+     - when MODE is wider than SRC_MODE, the extraction involves
+       a zero extension
+
+     - when MODE is smaller than SRC_MODE, the extraction involves
+       a truncation (and is thus subject to TRULY_NOOP_TRUNCATION).
+
+   In other words, this routine performs a computation, whereas the
+   gen_lowpart* routines are conceptually lvalue or rvalue subreg
+   operations.  */
+
+rtx
+extract_low_bits (enum machine_mode mode, enum machine_mode src_mode, rtx src)
+{
+  enum machine_mode int_mode, src_int_mode;
+
+  if (mode == src_mode)
+    return src;
+
+  if (CONSTANT_P (src))
+    return simplify_gen_subreg (mode, src, src_mode,
+				subreg_lowpart_offset (mode, src_mode));
+
+  if (GET_MODE_CLASS (mode) == MODE_CC || GET_MODE_CLASS (src_mode) == MODE_CC)
+    return NULL_RTX;
+
+  if (GET_MODE_BITSIZE (mode) == GET_MODE_BITSIZE (src_mode)
+      && MODES_TIEABLE_P (mode, src_mode))
+    {
+      rtx x = gen_lowpart_common (mode, src);
+      if (x)
+        return x;
+    }
+
+  src_int_mode = int_mode_for_mode (src_mode);
+  int_mode = int_mode_for_mode (mode);
+  if (src_int_mode == BLKmode || int_mode == BLKmode)
+    return NULL_RTX;
+
+  if (!MODES_TIEABLE_P (src_int_mode, src_mode))
+    return NULL_RTX;
+  if (!MODES_TIEABLE_P (int_mode, mode))
+    return NULL_RTX;
+
+  src = gen_lowpart (src_int_mode, src);
+  src = convert_modes (int_mode, src_int_mode, src, true);
+  src = gen_lowpart (mode, src);
+  return src;
+}
+
 /* Add INC into TARGET.  */
 
 void