[AArch64] Change representation of SABD in RTL

Richard raised a concern about the RTL we use to represent the AdvSIMD SABD
(vector signed absolute difference) instruction.
We currently represent it as ABS (MINUS op1 op2).

This isn't exactly what SABD does. ABS treats its input as a signed value
and returns the absolute of that.

For example:
(sabd:QI 64 -128) == 192 (unsigned) aka -64 (signed)
whereas
(minus:QI 64 -128) == 192 (unsigned) aka -64 (signed), (abs ...) of that is 64.

A better way to describe the instruction is with MINUS (SMAX (op1 op2) SMIN (op1 op2)).
This patch implements that, and also implements similar semantics for the UABD instruction
that uses UMAX and UMIN.

That way for the example above we'll have:
(minus:QI (smax:QI (64 -128)) (smin:QI (64 -128))) == (minus:QI 64 -128) == 192 (or -64 signed) which matches
what SABD does.

	* config/aarch64/iterators.md (max_opp): New code_attr.
	(USMAX): New code iterator.
	* config/aarch64/predicates.md (aarch64_smin): New predicate.
	(aarch64_smax): Likewise.
	* config/aarch64/aarch64-simd.md (abd<mode>_3): Rename to...
	(*aarch64_<su>abd<mode>_3): ... Change RTL representation to
	MINUS (MAX MIN).

	* gcc.target/aarch64/abd_1.c: New test.
	* gcc.dg/sabd_1.c: Likewise.

From-SVN: r268658

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 1969ff7..0984790 100644 (file)
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,13 @@
+2019-02-07  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
+
+       * config/aarch64/iterators.md (max_opp): New code_attr.
+       (USMAX): New code iterator.
+       * config/aarch64/predicates.md (aarch64_smin): New predicate.
+       (aarch64_smax): Likewise.
+       * config/aarch64/aarch64-simd.md (abd<mode>_3): Rename to...
+       (*aarch64_<su>abd<mode>_3): ... Change RTL representation to
+       MINUS (MAX MIN).
+
 2019-02-07  H.J. Lu  <hongjiu.lu@intel.com>
 
        PR target/89229

diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md
index cae9a8f..e3852c5 100644 (file)
--- a/gcc/config/aarch64/aarch64-simd.md
+++ b/gcc/config/aarch64/aarch64-simd.md
@@ -705,13 +705,22 @@
   [(set_attr "type" "neon_abs<q>")]
 )
 
-(define_insn "abd<mode>_3"
+;; It's tempting to represent SABD as ABS (MINUS op1 op2).
+;; This isn't accurate as ABS treats always its input as a signed value.
+;; So (ABS:QI (minus:QI 64 -128)) == (ABS:QI (192 or -64 signed)) == 64.
+;; Whereas SABD would return 192 (-64 signed) on the above example.
+;; Use MINUS ([us]max (op1, op2), [us]min (op1, op2)) instead.
+(define_insn "*aarch64_<su>abd<mode>_3"
   [(set (match_operand:VDQ_BHSI 0 "register_operand" "=w")
-       (abs:VDQ_BHSI (minus:VDQ_BHSI
-                      (match_operand:VDQ_BHSI 1 "register_operand" "w")
-                      (match_operand:VDQ_BHSI 2 "register_operand" "w"))))]
-  "TARGET_SIMD"
-  "sabd\t%0.<Vtype>, %1.<Vtype>, %2.<Vtype>"
+       (minus:VDQ_BHSI
+         (USMAX:VDQ_BHSI
+           (match_operand:VDQ_BHSI 1 "register_operand" "w")
+           (match_operand:VDQ_BHSI 2 "register_operand" "w"))
+         (match_operator 3 "aarch64_<max_opp>"
+           [(match_dup 1)
+            (match_dup 2)])))]
+  "TARGET_SIMD"
+  "<su>abd\t%0.<Vtype>, %1.<Vtype>, %2.<Vtype>"
   [(set_attr "type" "neon_abd<q>")]
 )
 

diff --git a/gcc/config/aarch64/iterators.md b/gcc/config/aarch64/iterators.md
index 85fa161..6caeeac 100644 (file)
--- a/gcc/config/aarch64/iterators.md
+++ b/gcc/config/aarch64/iterators.md
@@ -1052,6 +1052,9 @@
 
 (define_code_attr f16mac [(plus "a") (minus "s")])
 
+;; Map smax to smin and umax to umin.
+(define_code_attr max_opp [(smax "smin") (umax "umin")])
+
 ;; The number of subvectors in an SVE_STRUCT.
 (define_mode_attr vector_count [(VNx32QI "2") (VNx16HI "2")
                                (VNx8SI  "2") (VNx4DI  "2")
@@ -1200,6 +1203,9 @@
 
 (define_code_iterator FMAXMIN [smax smin])
 
+;; Signed and unsigned max operations.
+(define_code_iterator USMAX [smax umax])
+
 ;; Code iterator for variants of vector max and min.
 (define_code_iterator ADDSUB [plus minus])
 

diff --git a/gcc/config/aarch64/predicates.md b/gcc/config/aarch64/predicates.md
index 855cf7b..b8e6d23 100644 (file)
--- a/gcc/config/aarch64/predicates.md
+++ b/gcc/config/aarch64/predicates.md
@@ -319,6 +319,12 @@
   (ior (match_operand 0 "register_operand")
        (match_operand 0 "const_scalar_int_operand")))
 
+(define_predicate "aarch64_smin"
+  (match_code "smin"))
+
+(define_predicate "aarch64_umin"
+  (match_code "umin"))
+
 ;; True for integer comparisons and for FP comparisons other than LTGT or UNEQ.
 (define_special_predicate "aarch64_comparison_operator"
   (match_code "eq,ne,le,lt,ge,gt,geu,gtu,leu,ltu,unordered,

diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 5004a5e..9951b17 100644 (file)
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,8 @@
+2019-02-07  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
+
+       * gcc.target/aarch64/abd_1.c: New test.
+       * gcc.dg/sabd_1.c: Likewise.
+
 2019-02-07  Dominique d'Humieres  <dominiq@gcc.gnu.org>
 
        PR fortran/52789

diff --git a/gcc/testsuite/gcc.dg/sabd_1.c b/gcc/testsuite/gcc.dg/sabd_1.c
new file mode 100644 (file)
index 0000000..587e305
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/sabd_1.c
@@ -0,0 +1,47 @@
+/* { dg-do run } */
+/* { dg-options "-O3 -fwrapv" } */
+/* Make sure vectorized absolute difference behaves same as scalar version.  */
+
+#define N 16
+signed char a[] = {-100, -100, -100, -100,-100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100 };
+signed char b[] = { 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100 };
+
+signed char out[N];
+
+__attribute__ ((noinline,noipa))
+void
+foo (void)
+{
+  for (int i = 0; i < N; i++)
+    {
+      signed char diff = b[i] - a[i];
+      out[i] = diff > 0 ? diff : -diff;
+    }
+}
+
+signed char out2[N];
+
+__attribute__ ((noinline,noipa))
+void
+foo_scalar (void)
+{
+  for (int i = 0; i < N; i++)
+    {
+      asm volatile ("");
+      signed char diff = b[i] - a[i];
+      out2[i] = diff > 0 ? diff : -diff;
+    }
+}
+
+int
+main (void)
+{
+  foo ();
+  foo_scalar ();
+  for (int i = 0; i < N; i++)
+    if (out[i] != out2[i])
+      __builtin_abort ();
+
+  return 0;
+}
+

diff --git a/gcc/testsuite/gcc.target/aarch64/abd_1.c b/gcc/testsuite/gcc.target/aarch64/abd_1.c
new file mode 100644 (file)
index 0000000..a27cb5e
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/abd_1.c
@@ -0,0 +1,36 @@
+/* { dg-do compile } */
+/* { dg-options "-O3" } */
+
+#pragma GCC target "+nosve"
+
+#define MAX(x, y) ((x) > (y) ? (x) : (y))
+#define MIN(x, y) ((x) < (y) ? (x) : (y))
+#define N 1024
+
+#define FUNC(T)                                                                \
+void                                                                   \
+sabd_##T (signed T * restrict a, signed T * restrict b,                \
+               signed T * restrict out)                                \
+{                                                                      \
+  for (int i = 0; i < N; i++)                                          \
+    out[i] = MAX (a[i], b[i]) - MIN (a[i], b[i]);                      \
+}                                                                      \
+                                                                       \
+void                                                                   \
+uabd_##T (unsigned T * restrict a, unsigned T * restrict b,    \
+                 unsigned T * restrict out)                            \
+{                                                                      \
+  for (int i = 0; i < N; i++)                                          \
+    out[i] = MAX (a[i], b[i]) - MIN (a[i], b[i]);                      \
+}
+
+FUNC(char)
+FUNC(short)
+FUNC(int)
+
+/* { dg-final { scan-assembler-times "sabd\\tv\[0-9\]+\.16b, v\[0-9\]+\.16b, v\[0-9\]+\.16b" 1 } } */
+/* { dg-final { scan-assembler-times "uabd\\tv\[0-9\]+\.16b, v\[0-9\]+\.16b, v\[0-9\]+\.16b" 1 } } */
+/* { dg-final { scan-assembler-times "sabd\\tv\[0-9\]+\.8h, v\[0-9\]+\.8h, v\[0-9\]+\.8h" 1 } } */
+/* { dg-final { scan-assembler-times "uabd\\tv\[0-9\]+\.8h, v\[0-9\]+\.8h, v\[0-9\]+\.8h" 1 } } */
+/* { dg-final { scan-assembler-times "sabd\\tv\[0-9\]+\.4s, v\[0-9\]+\.4s, v\[0-9\]+\.4s" 1 } } */
+/* { dg-final { scan-assembler-times "uabd\\tv\[0-9\]+\.4s, v\[0-9\]+\.4s, v\[0-9\]+\.4s" 1 } } */