[rs6000] better use of unaligned vsx in memset() expansion
diff mbox series

Message ID 0e5a2fa3-47df-47d4-89cb-5c421a1e366b@linux.ibm.com
State New
Headers show
Series
  • [rs6000] better use of unaligned vsx in memset() expansion
Related show

Commit Message

Aaron Sawdey Nov. 26, 2018, 9:08 p.m. UTC
When I previously added the use of unaligned vsx stores to inline expansion
of memset, I didn't do a good job of managing boundary conditions. The intention
was to only use unaligned vsx if the block being cleared was more than 32 bytes.
What it actually did was to prevent the use of unaligned vsx for the last 32
bytes of any block being cleared. So this change puts the test up front so it
is not affected by the decrement of bytes.

OK for trunk if regstrap passes?

Thanks!
   Aaron



2018-11-26  Aaron Sawdey  <acsawdey@linux.ibm.com>

	* config/rs6000/rs6000-string.c (expand_block_clear): Change how
	we determine if unaligned vsx is ok.

Comments

Segher Boessenkool Nov. 26, 2018, 10:29 p.m. UTC | #1
On Mon, Nov 26, 2018 at 03:08:32PM -0600, Aaron Sawdey wrote:
> When I previously added the use of unaligned vsx stores to inline expansion
> of memset, I didn't do a good job of managing boundary conditions. The intention
> was to only use unaligned vsx if the block being cleared was more than 32 bytes.
> What it actually did was to prevent the use of unaligned vsx for the last 32
> bytes of any block being cleared. So this change puts the test up front so it
> is not affected by the decrement of bytes.

Oh wow.  Yes, that isn't so great.  Okay for trunk (and whatever backports).
Thanks,


Segher


> 2018-11-26  Aaron Sawdey  <acsawdey@linux.ibm.com>
> 
> 	* config/rs6000/rs6000-string.c (expand_block_clear): Change how
> 	we determine if unaligned vsx is ok.
Aaron Sawdey Nov. 28, 2018, 7:24 p.m. UTC | #2
The first version of this had a big bug and cleared past the requested bytes.
This version passes regstrap on ppc64le(power7/8/9), ppc64be(power6/7/8),
and ppc32(power8).

OK for trunk (and 8 backport after a week)?

Thanks!
   Aaron

Index: gcc/config/rs6000/rs6000-string.c
===================================================================
--- gcc/config/rs6000/rs6000-string.c	(revision 266524)
+++ gcc/config/rs6000/rs6000-string.c	(working copy)
@@ -85,6 +85,8 @@
   if (! optimize_size && bytes > 8 * clear_step)
     return 0;

+  bool unaligned_vsx_ok = (bytes >= 32 && TARGET_EFFICIENT_UNALIGNED_VSX);
+
   for (offset = 0; bytes > 0; offset += clear_bytes, bytes -= clear_bytes)
     {
       machine_mode mode = BLKmode;
@@ -91,8 +93,7 @@
       rtx dest;

       if (TARGET_ALTIVEC
-	  && ((bytes >= 16 && align >= 128)
-	      || (bytes >= 32 && TARGET_EFFICIENT_UNALIGNED_VSX)))
+	  && (bytes >= 16 && ( align >= 128 || unaligned_vsx_ok)))
 	{
 	  clear_bytes = 16;
 	  mode = V4SImode;


On 11/26/18 4:29 PM, Segher Boessenkool wrote:
> On Mon, Nov 26, 2018 at 03:08:32PM -0600, Aaron Sawdey wrote:
>> When I previously added the use of unaligned vsx stores to inline expansion
>> of memset, I didn't do a good job of managing boundary conditions. The intention
>> was to only use unaligned vsx if the block being cleared was more than 32 bytes.
>> What it actually did was to prevent the use of unaligned vsx for the last 32
>> bytes of any block being cleared. So this change puts the test up front so it
>> is not affected by the decrement of bytes.
> 
> Oh wow.  Yes, that isn't so great.  Okay for trunk (and whatever backports).
> Thanks,
> 
> 
> Segher
> 
> 
>> 2018-11-26  Aaron Sawdey  <acsawdey@linux.ibm.com>
>>
>> 	* config/rs6000/rs6000-string.c (expand_block_clear): Change how
>> 	we determine if unaligned vsx is ok.
>
Segher Boessenkool Nov. 28, 2018, 8 p.m. UTC | #3
On Wed, Nov 28, 2018 at 01:24:01PM -0600, Aaron Sawdey wrote:
> The first version of this had a big bug and cleared past the requested bytes.
> This version passes regstrap on ppc64le(power7/8/9), ppc64be(power6/7/8),
> and ppc32(power8).
> 
> OK for trunk (and 8 backport after a week)?

> @@ -91,8 +93,7 @@
>        rtx dest;
> 
>        if (TARGET_ALTIVEC
> -	  && ((bytes >= 16 && align >= 128)
> -	      || (bytes >= 32 && TARGET_EFFICIENT_UNALIGNED_VSX)))
> +	  && (bytes >= 16 && ( align >= 128 || unaligned_vsx_ok)))

Please remove the stray space?  Okay for trunk and later for 8, thanks!


Segher

Patch
diff mbox series

Index: gcc/config/rs6000/rs6000-string.c
===================================================================
--- gcc/config/rs6000/rs6000-string.c	(revision 266219)
+++ gcc/config/rs6000/rs6000-string.c	(working copy)
@@ -85,14 +85,14 @@ 
   if (! optimize_size && bytes > 8 * clear_step)
     return 0;

+  bool unaligned_vsx_ok = (bytes >= 32 && TARGET_EFFICIENT_UNALIGNED_VSX);
+
   for (offset = 0; bytes > 0; offset += clear_bytes, bytes -= clear_bytes)
     {
       machine_mode mode = BLKmode;
       rtx dest;

-      if (TARGET_ALTIVEC
-	  && ((bytes >= 16 && align >= 128)
-	      || (bytes >= 32 && TARGET_EFFICIENT_UNALIGNED_VSX)))
+      if (TARGET_ALTIVEC && ((bytes >= 16 && align >= 128) || unaligned_vsx_ok))
 	{
 	  clear_bytes = 16;
 	  mode = V4SImode;