diff mbox series

[committed] Fix testsuite fallout on stormy16 after IOR->PLUS change

Message ID a2d91174-5e85-4852-8804-046fc8c67a52@gmail.com
State New
Headers show
Series [committed] Fix testsuite fallout on stormy16 after IOR->PLUS change | expand

Commit Message

Jeff Law June 21, 2024, 10 p.m. UTC
More minor fallout from the IOR->PLUS change a little while ago.  This 
time on xstormy16.

The pattern to swap nibbles actually tries to handle all the cases of 
IOR, XOR and PLUS.  But when we generate PLUS earlier in the pipeline, 
the simplifications/canonicalizations are slightly different resulting 
in the pattern not matching.

This patch adds an alternate pattern which matches what we get now. 
Basically it looks like QImode rotate by 4, zero extended to HI.

Run in my tester to verify the regression was fixed.  Pushing to the trunk.

Jeff
* config/stormy16/stormy16.md (swpn_zext): New pattern.

Comments

Georg-Johann Lay June 22, 2024, 12:01 p.m. UTC | #1
Am 22.06.24 um 00:00 schrieb Jeff Law:
> More minor fallout from the IOR->PLUS change a little while ago.  This 
> time on xstormy16.
> 
> The pattern to swap nibbles actually tries to handle all the cases of 
> IOR, XOR and PLUS.  But when we generate PLUS earlier in the pipeline, 
> the simplifications/canonicalizations are slightly different resulting 
> in the pattern not matching.

The avr backend has quite some (combine) patterns that combine IOR
and other operations like AND to bit move instructions etc.

Does this mean all such patterns are void now?

Johann

> This patch adds an alternate pattern which matches what we get now. 
> Basically it looks like QImode rotate by 4, zero extended to HI.
> 
> Run in my tester to verify the regression was fixed.  Pushing to the trunk.
> 
> Jeff
Jeff Law June 22, 2024, 2:11 p.m. UTC | #2
On 6/22/24 6:01 AM, Georg-Johann Lay wrote:
> 
> 
> Am 22.06.24 um 00:00 schrieb Jeff Law:
>> More minor fallout from the IOR->PLUS change a little while ago.  This 
>> time on xstormy16.
>>
>> The pattern to swap nibbles actually tries to handle all the cases of 
>> IOR, XOR and PLUS.  But when we generate PLUS earlier in the pipeline, 
>> the simplifications/canonicalizations are slightly different resulting 
>> in the pattern not matching.
> 
> The avr backend has quite some (combine) patterns that combine IOR
> and other operations like AND to bit move instructions etc.
> 
> Does this mean all such patterns are void now?
Certainly not void, though you may be missing optimizations.  The key 
properties to look for.

Rotation like patterns, but the target doesn't have a named pattern for 
rotation.  So an IOR with ASHIFT/LSHIFTRT operands and appropriate shift 
counts.

Those rotation patterns have other embedded operations such as a sign 
extension, logical ands to mask off bits and the like.


Jeff
diff mbox series

Patch

diff --git a/gcc/config/stormy16/stormy16.md b/gcc/config/stormy16/stormy16.md
index 7f12679847b..62318345cdc 100644
--- a/gcc/config/stormy16/stormy16.md
+++ b/gcc/config/stormy16/stormy16.md
@@ -1363,6 +1363,20 @@  (define_insn "*swpn_zext"
   "swpn %0 | and %0,#255"
   [(set_attr "length" "6")])
 
+;; Alternate form when we use PLUS instead of IOR early in the
+;; expanders.
+(define_insn "*swpn_zext"
+  [(set (match_operand:HI 0 "register_operand" "=r")
+	(zero_extend:HI
+	  (subreg:QI
+	    (any_or_plus:HI
+	      (ashift:HI (match_operand:HI 1 "register_operand" "0")
+			 (const_int 4))
+	      (lshiftrt:HI (match_dup 1) (const_int 4))) 0)))]
+  ""
+  "swpn %0 | and %0,#255"
+  [(set_attr "length" "6")])
+
 (define_insn "*swpn_sext"
   [(set (match_operand:HI 0 "register_operand" "=r")
 	(sign_extend:HI