diff mbox

Optimise CONCAT handling in emit_group_load

Message ID 871syc3e2n.fsf@e105548-lin.cambridge.arm.com
State New
Headers show

Commit Message

Richard Sandiford Nov. 15, 2016, 4:25 p.m. UTC 
The CONCAT handling in emit_group_load chooses between doing
an extraction from a single component or forcing the whole
thing to memory and extracting from there.  The condition for
the former (more efficient) option was:

	  if ((bytepos == 0 && bytelen == slen0)
	      || (bytepos != 0 && bytepos + bytelen <= slen))

On the one hand this seems dangerous, since the second line
allows bit ranges that start in the first component and leak
into the second.  On the other hand it seems strange to allow
references that start after the first byte of the second
component but not those that start after the first byte
of the first component.  This led to a pessimisation of
things like gcc.dg/builtins-54.c for hppa64-hp-hpux11.23.

This patch simply checks whether the reference is contained
within a single component.  It also makes sure that we do
an extraction on anything that doesn't span the whole
component (even if it's constant).

Tested on aarch64-linux-gnu and x86_64-linux-gnu.  OK to install?

Thanks,
Richard


[ This patch is part of the SVE series posted here:
  https://gcc.gnu.org/ml/gcc/2016-11/msg00030.html ]

gcc/
2016-11-15  Richard Sandiford  <richard.sandiford@arm.com>
	    Alan Hayward  <alan.hayward@arm.com>
	    David Sherwood  <david.sherwood@arm.com>

	* expr.c (emit_group_load_1): Tighten check for whether an
	access involves only one operand of a CONCAT.  Use extract_bit_field
	for constants if the bit range does span the whole operand.

Comments

Eric Botcazou Nov. 16, 2016, 7:21 a.m. UTC | #1 
> 2016-11-15  Richard Sandiford  <richard.sandiford@arm.com>
> 	    Alan Hayward  <alan.hayward@arm.com>
> 	    David Sherwood  <david.sherwood@arm.com>
> 
> 	* expr.c (emit_group_load_1): Tighten check for whether an
> 	access involves only one operand of a CONCAT.  Use extract_bit_field
> 	for constants if the bit range does span the whole operand.

OK, thanks.
diff mbox

Patch

diff --git a/gcc/expr.c b/gcc/expr.c
index 0b0946d..985c2b3 100644
--- a/gcc/expr.c
+++ b/gcc/expr.c
@@ -2175,19 +2175,22 @@  emit_group_load_1 (rtx *tmps, rtx dst, rtx orig_src, tree type, int ssize)
 	{
 	  unsigned int slen = GET_MODE_SIZE (GET_MODE (src));
 	  unsigned int slen0 = GET_MODE_SIZE (GET_MODE (XEXP (src, 0)));
+	  unsigned int elt = bytepos / slen0;
+	  unsigned int subpos = bytepos % slen0;
 
-	  if ((bytepos == 0 && bytelen == slen0)
-	      || (bytepos != 0 && bytepos + bytelen <= slen))
+	  if (subpos + bytelen <= slen0)
 	    {
 	      /* The following assumes that the concatenated objects all
 		 have the same size.  In this case, a simple calculation
 		 can be used to determine the object and the bit field
 		 to be extracted.  */
-	      tmps[i] = XEXP (src, bytepos / slen0);
-	      if (! CONSTANT_P (tmps[i])
-		  && (!REG_P (tmps[i]) || GET_MODE (tmps[i]) != mode))
+	      tmps[i] = XEXP (src, elt);
+	      if (subpos != 0
+		  || subpos + bytelen != slen0
+		  || (!CONSTANT_P (tmps[i])
+		      && (!REG_P (tmps[i]) || GET_MODE (tmps[i]) != mode)))
 		tmps[i] = extract_bit_field (tmps[i], bytelen * BITS_PER_UNIT,
-					     (bytepos % slen0) * BITS_PER_UNIT,
+					     subpos * BITS_PER_UNIT,
 					     1, NULL_RTX, mode, mode, false);
 	    }
 	  else