[xstormy16 PATCH] Efficient HImode rotate left by a single bit.

[xstormy16 PATCH] Efficient HImode rotate left by a single bit.

[Roger Sayle]

[-- Attachment #1: Type: text/plain, Size: 1270 bytes --]

This patch contains some minor tweak to xstormy16's machine description
most significantly providing a pattern for HImode rotate left by a single
bit that requires only two instructions.

unsigned short foo(unsigned short x)
{
  return (x << 1) | (x >> 15);
}

currently with -O2 generates:
foo:    mov r7,r2
        shr r7,#15
        shl r2,#1
        or r2,r7
        ret

with this patch, GCC now generates:
foo:    shl r2,#1 | adc r2,#0
        ret

Additionally neghi2 is converted to a define_insn (so that the RTL
optimizers see the negation semantics), and HImode rotations by
8-bits can now be recognized and implemented using swpb.

This patch has been tested by building a cross-compiler to xstormy16-elf
from x86_64-pc-linux-gnu and confirming the new test cases pass.
Ok for mainline?


2023-04-29  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
        * config/stormy16/stormy16.md (neghi2): Convert from a define_expand
        to a define_insn.
        (*rotatehi_1): New define_insn for efficient 2 insn sequence.
        (*rotatehi_8, *rotaterthi_8): New define_insn to emit a swpb.

gcc/testsuite/ChangeLog
        * gcc.target/xstormy16/neghi2.c: New test case.
        * gcc.target/rotatehi-1.c: Likewise.


Thanks in advance,
Roger
--


[-- Attachment #2: patchxs2.txt --]
[-- Type: text/plain, Size: 2283 bytes --]

diff --git a/gcc/config/stormy16/stormy16.md b/gcc/config/stormy16/stormy16.md
index b2e86ee..be1ee04 100644
--- a/gcc/config/stormy16/stormy16.md
+++ b/gcc/config/stormy16/stormy16.md
@@ -514,13 +518,13 @@
 
 ;; Negation
 
-(define_expand "neghi2"
-  [(set (match_operand:HI 0 "register_operand" "")
-	(not:HI (match_operand:HI 1 "register_operand" "")))
-   (parallel [(set (match_dup 0) (plus:HI (match_dup 0) (const_int 1)))
+(define_insn "neghi2"
+  [(parallel [(set (match_operand:HI 0 "register_operand" "=r")
+		   (neg:HI (match_operand:HI 1 "register_operand" "0")))
 	      (clobber (reg:BI CARRY_REG))])]
   ""
-  "")
+  "not %0 | add %0,#1"
+  [(set_attr "length" "4")])
 \f
 ;; ::::::::::::::::::::
 ;; ::
@@ -554,6 +558,24 @@
    (clobber (reg:BI CARRY_REG))]
   ""
   "shr %0,%2")
+
+;; HImode rotate left by 1 bit
+(define_insn "*rotatehi_1"
+  [(set (match_operand:HI 0 "register_operand" "=r")
+	(rotate:HI (match_operand:HI 1 "register_operand" "0")
+		   (const_int 1)))
+   (clobber (reg:BI CARRY_REG))]
+  ""
+  "shl %0,#1 | adc %0,#0"
+  [(set_attr "length" "4")])
+
+;; HImode rotate left by 8 bits
+(define_insn "*<code>hi_8"
+  [(set (match_operand:HI 0 "register_operand" "=r")
+	(any_rotate:HI (match_operand:HI 1 "register_operand" "0")
+		       (const_int 8)))]
+  ""
+  "swpb %0")
 \f
 ;; ::::::::::::::::::::
 ;; ::
diff --git a/gcc/testsuite/gcc.target/xstormy16/neghi2.c b/gcc/testsuite/gcc.target/xstormy16/neghi2.c
new file mode 100644
index 0000000..dd3dd1e
--- /dev/null
+++ b/gcc/testsuite/gcc.target/xstormy16/neghi2.c
@@ -0,0 +1,8 @@
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+
+short neg(short x)
+{
+  return -x;
+}
+/* { dg-final { scan-assembler "not r2 | add r2,#1" } } */
diff --git a/gcc/testsuite/gcc.target/xstormy16/rotatehi-1.c b/gcc/testsuite/gcc.target/xstormy16/rotatehi-1.c
new file mode 100644
index 0000000..586e7dc
--- /dev/null
+++ b/gcc/testsuite/gcc.target/xstormy16/rotatehi-1.c
@@ -0,0 +1,10 @@
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+
+unsigned short foo(unsigned short x)
+{
+  return (x << 1) | (x >> 15);
+}
+
+/* { dg-final { scan-assembler "shl r2,#1" } } */
+/* { dg-final { scan-assembler "adc r2,#0" } } */

Re: [xstormy16 PATCH] Efficient HImode rotate left by a single bit.

[Jeff Law]

On 4/29/23 10:25, Roger Sayle wrote:
> This patch contains some minor tweak to xstormy16's machine description
> most significantly providing a pattern for HImode rotate left by a single
> bit that requires only two instructions.
> 
> unsigned short foo(unsigned short x)
> {
>    return (x << 1) | (x >> 15);
> }
> 
> currently with -O2 generates:
> foo:    mov r7,r2
>          shr r7,#15
>          shl r2,#1
>          or r2,r7
>          ret
> 
> with this patch, GCC now generates:
> foo:    shl r2,#1 | adc r2,#0
>          ret
> 
> Additionally neghi2 is converted to a define_insn (so that the RTL
> optimizers see the negation semantics), and HImode rotations by
> 8-bits can now be recognized and implemented using swpb.
> 
> This patch has been tested by building a cross-compiler to xstormy16-elf
> from x86_64-pc-linux-gnu and confirming the new test cases pass.
> Ok for mainline?
> 
> 
> 2023-04-29  Roger Sayle  <roger@nextmovesoftware.com>
> 
> gcc/ChangeLog
>          * config/stormy16/stormy16.md (neghi2): Convert from a define_expand
>          to a define_insn.
>          (*rotatehi_1): New define_insn for efficient 2 insn sequence.
>          (*rotatehi_8, *rotaterthi_8): New define_insn to emit a swpb.
> 
> gcc/testsuite/ChangeLog
>          * gcc.target/xstormy16/neghi2.c: New test case.
>          * gcc.target/rotatehi-1.c: Likewise.
It may be the case that exposing negation as a not + add sequence was 
thought to potentially produce better code by exposing the component 
instructions.  Or it may have simply been the case that nobody 
considered the tradeoffs.


Either way, I think the patch is fine.  As is always the case, figure 
~24hrs after committing we'll have test results.



jeff