An x86 backend improvement that I'm working results in combine attempting
to recognize:

(set (reg:DI 87 [ xD.2846 ])
     (ior:DI (subreg:DI (ashift:TI (zero_extend:TI (reg:DI 92))
                                   (const_int 64 [0x40])) 0)
             (reg:DI 91)))

where the lowpart SUBREG has difficulty seeing through the (hi<<64)
that the lowpart must be zero.  Rather than workaround this in the
backend, the better fix is to teach simplify-rtx that
lowpart((hi<<64)|lo) -> lo and highpart((hi<<64)|lo) -> hi, so that
all backends benefit.  Reducing the number of places where the
middle-end generates a SUBREG of something other than REG is a
good thing.

This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
and make -k check, both with and without --target_board=unix{-m32}
with no new failures, except for pr78904-1b.c, for which a backend
solution has just been proposed.  Ok for mainline?


2023-06-18  Roger Sayle  <ro...@nextmovesoftware.com>

gcc/ChangeLog
        * simplify-rtx.cc (simplify_subreg):  Optimize lowpart SUBREGs
        of ASHIFT to const0_rtx with sufficiently large shift count.
        Optimize highpart SUBREGs of ASHIFT as the shift operand when
        the shift count is the correct offset.  Optimize SUBREGs of
        multi-word logic operations if the SUBREGs of both operands
        can be simplified.


Thanks in advance,
Roger
--

diff --git a/gcc/simplify-rtx.cc b/gcc/simplify-rtx.cc
index 21b7eb4..6715247 100644
--- a/gcc/simplify-rtx.cc
+++ b/gcc/simplify-rtx.cc
@@ -7746,6 +7746,38 @@ simplify_context::simplify_subreg (machine_mode 
outermode, rtx op,
        return CONST0_RTX (outermode);
     }
 
+  /* Optimize SUBREGS of scalar integral ASHIFT by a valid constant.  */
+  if (GET_CODE (op) == ASHIFT
+      && SCALAR_INT_MODE_P (innermode)
+      && CONST_INT_P (XEXP (op, 1))
+      && INTVAL (XEXP (op, 1)) > 0
+      && known_gt (GET_MODE_BITSIZE (innermode), INTVAL (XEXP (op, 1))))
+    {
+      HOST_WIDE_INT val = INTVAL (XEXP (op, 1));
+      /* A lowpart SUBREG of a ASHIFT by a constant may fold to zero.  */
+      if (known_eq (subreg_lowpart_offset (outermode, innermode), byte)
+         && known_le (GET_MODE_BITSIZE (outermode), val))
+        return CONST0_RTX (outermode);
+      /* Optimize the highpart SUBREG of a suitable ASHIFT (ZERO_EXTEND).  */
+      if (GET_CODE (XEXP (op, 0)) == ZERO_EXTEND
+         && GET_MODE (XEXP (XEXP (op, 0), 0)) == outermode
+         && known_eq (GET_MODE_BITSIZE (outermode), val)
+         && known_eq (GET_MODE_BITSIZE (innermode), 2 * val)
+         && known_eq (subreg_highpart_offset (outermode, innermode), byte))
+       return XEXP (XEXP (op, 0), 0);
+    }
+
+  /* Attempt to simplify WORD_MODE SUBREGs of bitwise expressions.  */
+  if (outermode == word_mode
+      && (GET_CODE (op) == IOR || GET_CODE (op) == XOR || GET_CODE (op) == AND)
+      && SCALAR_INT_MODE_P (innermode))
+    {
+      rtx op0 = simplify_subreg (outermode, XEXP (op, 0), innermode, byte);
+      rtx op1 = simplify_subreg (outermode, XEXP (op, 1), innermode, byte);
+      if (op0 && op1)
+       return simplify_gen_binary (GET_CODE (op), outermode, op0, op1);
+    }
+
   scalar_int_mode int_outermode, int_innermode;
   if (is_a <scalar_int_mode> (outermode, &int_outermode)
       && is_a <scalar_int_mode> (innermode, &int_innermode)

Reply via email to