Hi,
this patch implements vector copysign using vector select on S/390.
Regtested and bootstrapped on s390x.
Regards
Robin
--
gcc/ChangeLog:
2019-02-07 Robin Dapp <[email protected]>
* config/s390/vector.md: Implement vector copysign.
gcc/testsuite/ChangeLog:
2019-02-07 Robin Dapp <[email protected]>
* gcc.target/s390/vector/vec-copysign-execute.c: New test.
* gcc.target/s390/vector/vec-copysign.c: New test.
diff --git a/gcc/config/s390/vector.md b/gcc/config/s390/vector.md
index c9ffab4c8c2..820372eca29 100644
--- a/gcc/config/s390/vector.md
+++ b/gcc/config/s390/vector.md
@@ -1362,6 +1362,31 @@
operands[4] = CONST0_RTX (V2DImode);
})
+; Vector copysign, implement using vector select
+(define_expand "copysign<mode>3"
+ [(set (match_operand:VFT 0 "register_operand" "")
+ (if_then_else:VFT
+ (eq (match_dup 3)
+ (match_dup 4))
+ (match_operand:VFT 1 "register_operand" "")
+ (match_operand:VFT 2 "register_operand" "")))]
+ "TARGET_VX"
+{
+ int sz = GET_MODE_BITSIZE (GET_MODE_INNER (<MODE>mode));
+ int prec = GET_MODE_PRECISION (GET_MODE_INNER (<tointvec>mode));
+ wide_int mask_val = wi::shwi (1l << (sz - 1), prec);
+
+ rtx mask = gen_reg_rtx (<tointvec>mode);
+
+ int nunits = GET_MODE_NUNITS (<tointvec>mode);
+ rtvec v = rtvec_alloc (nunits);
+ for (int i = 0; i < nunits; i++)
+ RTVEC_ELT (v, i) = GEN_INT (mask_val.to_shwi ());
+
+ mask = gen_rtx_CONST_VECTOR (<tointvec>mode, v);
+ operands[3] = force_reg (<tointvec>mode, mask);
+ operands[4] = CONST0_RTX (<tointvec>mode);
+})
;;
;; Integer compares
diff --git a/gcc/testsuite/gcc.target/s390/vector/vec-copysign-execute.c b/gcc/testsuite/gcc.target/s390/vector/vec-copysign-execute.c
new file mode 100644
index 00000000000..a8d675d3a72
--- /dev/null
+++ b/gcc/testsuite/gcc.target/s390/vector/vec-copysign-execute.c
@@ -0,0 +1,74 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ftree-vectorize -mzarch -march=z13" } */
+
+#include <math.h>
+#include <assert.h>
+
+#define N 20
+
+double a[N] = {-0.1, -3.2, -6.3, -9.4, -12.5, -15.6, -18.7, -21.8, 24.9,
+ 27.1, 30.2, 33.3, 36.4, 39.5, 42.6, nan("123"), __DBL_MIN__ / 2.0,
+ -nan ("1"), __DBL_MAX__ * 2.0, -__DBL_MAX__ * 1e199};
+double b[N] = {-1.2, 3.4, -5.6, 7.8, -9.0, 1.0, -2.0, 3.0, -4.0, -5.0, 6.0,
+ 7.0, -8.0, -9.0, 10.0, -11.0, -1., 0., -0., 1.3};
+double r[N];
+double r2[N];
+
+void
+foo (void)
+{
+ for (int i = 0; i < N; i++)
+ r[i] = copysign (a[i], b[i]);
+}
+
+__attribute__((optimize("no-tree-vectorize")))
+void
+check (void)
+{
+ for (int i = 0; i < N; i++)
+ {
+ r2[i] = copysign (a[i], b[i]);
+ assert (r[i] == r2[i]
+ || (isnan (r[i]) && isnan (r2[i])
+ && signbit (r[i]) == signbit (r2[i])));
+ }
+}
+
+float af[N] = {-0.1, -3.2, -6.3, -9.4, -12.5, -15.6, -18.7, -21.8, 24.9,
+ 27.1, 30.2, 33.3, 36.4, 39.5, 42.6, nan("123"), __DBL_MIN__ / 2.0,
+ -nan ("1"), __DBL_MAX__ * 2.0, -__DBL_MAX__ * 1e199};
+float bf[N] = {-1.2, 3.4, -5.6, 7.8, -9.0, 1.0, -2.0, 3.0, -4.0, -5.0, 6.0,
+ 7.0, -8.0, -9.0, 10.0, -11.0, -1., 0., -0., 1.3};
+float rf[N];
+float rf2[N];
+
+__attribute__ ((__target__ ("arch=z14")))
+void
+foof (void)
+{
+ for (int i = 0; i < N; i++)
+ rf[i] = copysignf (af[i], bf[i]);
+}
+
+__attribute__((optimize("no-tree-vectorize")))
+void
+checkf (void)
+{
+ for (int i = 0; i < N; i++)
+ {
+ rf2[i] = copysignf (af[i], bf[i]);
+ assert (rf[i] == rf2[i]
+ || (isnan (rf[i]) && isnan (rf2[i])
+ && signbit (rf[i]) == signbit (rf2[i])));
+ }
+}
+
+int main()
+{
+ foo ();
+ check ();
+
+ foof ();
+ checkf ();
+ return r[0];
+}
diff --git a/gcc/testsuite/gcc.target/s390/vector/vec-copysign.c b/gcc/testsuite/gcc.target/s390/vector/vec-copysign.c
new file mode 100644
index 00000000000..64c6970c23e
--- /dev/null
+++ b/gcc/testsuite/gcc.target/s390/vector/vec-copysign.c
@@ -0,0 +1,38 @@
+/* { dg-do compile { target { s390*-*-* } } } */
+/* { dg-options "-O2 -ftree-vectorize -mzarch" } */
+/* { dg-final { scan-assembler-times "vgmg" 1 } } */
+/* { dg-final { scan-assembler-times "vgmf" 1 } } */
+/* { dg-final { scan-assembler-times "vsel" 2 } } */
+
+#include <math.h>
+
+#define N 20
+
+double a[N] = {-0.1, -3.2, -6.3, -9.4, -12.5, -15.6, -18.7, -21.8, 24.9,
+ 27.1, 30.2, 33.3, 36.4, 39.5, 42.6, nan ("123"), __DBL_MIN__ / 2.0,
+ -nan ("1"), __DBL_MAX__ * 2.0, -__DBL_MAX__ * 1e199};
+double b[N] = {-1.2, 3.4, -5.6, 7.8, -9.0, 1.0, -2.0, 3.0, -4.0, -5.0, 6.0,
+ 7.0, -8.0, -9.0, 10.0, -11.0, -1., 0., -0., 1.3};
+double r[N];
+float af[N] = {-0.1, -3.2, -6.3, -9.4, -12.5, -15.6, -18.7, -21.8, 24.9,
+ 27.1, 30.2, 33.3, 36.4, 39.5, 42.6, nan ("123"), __DBL_MIN__ / 2.0,
+ -nan ("1"), __DBL_MAX__ * 2.0, -__DBL_MAX__ * 1e199};
+float bf[N] = {-1.2, 3.4, -5.6, 7.8, -9.0, 1.0, -2.0, 3.0, -4.0, -5.0, 6.0,
+ 7.0, -8.0, -9.0, 10.0, -11.0, -1., 0., -0., 1.3};
+float rf[N];
+
+__attribute__ ((__target__ ("arch=z13")))
+void
+foo (void)
+{
+ for (int i = 0; i < N; i++)
+ r[i] = copysign (a[i], b[i]);
+}
+
+__attribute__ ((__target__ ("arch=z14")))
+void
+foof (void)
+{
+ for (int i = 0; i < N; i++)
+ rf[i] = copysignf (af[i], bf[i]);
+}
