[AArch64][SVE] Remove unnecessary PTRUEs from FP arithmetic

When using the unpredicated all-register forms of FADD, FSUB and FMUL,
the rtl patterns would still have the predicate operand we created for
the other forms.  This patch splits the patterns after reload in order
to get rid of the predicate, like we already do for WHILE.

2018-12-07  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
	* config/aarch64/iterators.md (SVE_UNPRED_FP_BINARY): New code
	iterator.
	(sve_fp_op): Handle minus and mult.
	* config/aarch64/aarch64-sve.md (*add<mode>3, *sub<mode>3)
	(*mul<mode>3): Split the patterns after reload if we don't
	need the predicate operand.
	(*post_ra_<sve_fp_op><mode>3): New pattern.

gcc/testsuite/
	* gcc.target/aarch64/sve/pred_elim_1.c: New test.

From-SVN: r266891

5 files changed, 75 insertions, 6 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index dbe8b5f..d1a4025 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,13 @@
+2018-12-07  Richard Sandiford  <richard.sandiford@arm.com>
+
+	* config/aarch64/iterators.md (SVE_UNPRED_FP_BINARY): New code
+	iterator.
+	(sve_fp_op): Handle minus and mult.
+	* config/aarch64/aarch64-sve.md (*add<mode>3, *sub<mode>3)
+	(*mul<mode>3): Split the patterns after reload if we don't
+	need the predicate operand.
+	(*post_ra_<sve_fp_op><mode>3): New pattern.
+
 2018-12-07  Bin Cheng  <bin.cheng@linux.alibaba.com>
 
 	* profile-count.h (profile_count::oeprator>=): Fix typo by inverting
diff --git a/gcc/config/aarch64/aarch64-sve.md b/gcc/config/aarch64/aarch64-sve.md
index 5cd591b..edc6cff 100644
--- a/gcc/config/aarch64/aarch64-sve.md
+++ b/gcc/config/aarch64/aarch64-sve.md
@@ -2194,7 +2194,7 @@
 )
 
 ;; Floating-point addition predicated with a PTRUE.
-(define_insn "*add<mode>3"
+(define_insn_and_split "*add<mode>3"
   [(set (match_operand:SVE_F 0 "register_operand" "=w, w, w")
 	(unspec:SVE_F
 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
@@ -2206,7 +2206,12 @@
   "@
    fadd\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
    fsub\t%0.<Vetype>, %1/m, %0.<Vetype>, #%N3
-   fadd\t%0.<Vetype>, %2.<Vetype>, %3.<Vetype>"
+   #"
+  ; Split the unpredicated form after reload, so that we don't have
+  ; the unnecessary PTRUE.
+  "&& reload_completed
+   && register_operand (operands[3], <MODE>mode)"
+  [(set (match_dup 0) (plus:SVE_F (match_dup 2) (match_dup 3)))]
 )
 
 ;; Unpredicated floating-point subtraction.
@@ -2225,7 +2230,7 @@
 )
 
 ;; Floating-point subtraction predicated with a PTRUE.
-(define_insn "*sub<mode>3"
+(define_insn_and_split "*sub<mode>3"
   [(set (match_operand:SVE_F 0 "register_operand" "=w, w, w, w")
 	(unspec:SVE_F
 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl")
@@ -2240,7 +2245,13 @@
    fsub\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
    fadd\t%0.<Vetype>, %1/m, %0.<Vetype>, #%N3
    fsubr\t%0.<Vetype>, %1/m, %0.<Vetype>, #%2
-   fsub\t%0.<Vetype>, %2.<Vetype>, %3.<Vetype>"
+   #"
+  ; Split the unpredicated form after reload, so that we don't have
+  ; the unnecessary PTRUE.
+  "&& reload_completed
+   && register_operand (operands[2], <MODE>mode)
+   && register_operand (operands[3], <MODE>mode)"
+  [(set (match_dup 0) (minus:SVE_F (match_dup 2) (match_dup 3)))]
 )
 
 ;; Unpredicated floating-point multiplication.
@@ -2259,7 +2270,7 @@
 )
 
 ;; Floating-point multiplication predicated with a PTRUE.
-(define_insn "*mul<mode>3"
+(define_insn_and_split "*mul<mode>3"
   [(set (match_operand:SVE_F 0 "register_operand" "=w, w")
 	(unspec:SVE_F
 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
@@ -2270,9 +2281,25 @@
   "TARGET_SVE"
   "@
    fmul\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
-   fmul\t%0.<Vetype>, %2.<Vetype>, %3.<Vetype>"
+   #"
+  ; Split the unpredicated form after reload, so that we don't have
+  ; the unnecessary PTRUE.
+  "&& reload_completed
+   && register_operand (operands[3], <MODE>mode)"
+  [(set (match_dup 0) (mult:SVE_F (match_dup 2) (match_dup 3)))]
 )
 
+;; Unpredicated floating-point binary operations (post-RA only).
+;; These are generated by splitting a predicated instruction whose
+;; predicate is unused.
+(define_insn "*post_ra_<sve_fp_op><mode>3"
+  [(set (match_operand:SVE_F 0 "register_operand" "=w")
+	(SVE_UNPRED_FP_BINARY:SVE_F
+	  (match_operand:SVE_F 1 "register_operand" "w")
+	  (match_operand:SVE_F 2 "register_operand" "w")))]
+  "TARGET_SVE && reload_completed"
+  "<sve_fp_op>\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>")
+
 ;; Unpredicated fma (%0 = (%1 * %2) + %3).
 (define_expand "fma<mode>4"
   [(set (match_operand:SVE_F 0 "register_operand")
diff --git a/gcc/config/aarch64/iterators.md b/gcc/config/aarch64/iterators.md
index 524e4e6..a807557 100644
--- a/gcc/config/aarch64/iterators.md
+++ b/gcc/config/aarch64/iterators.md
@@ -1220,6 +1220,9 @@
 ;; SVE integer binary division operations.
 (define_code_iterator SVE_INT_BINARY_SD [div udiv])
 
+;; SVE floating-point operations with an unpredicated all-register form.
+(define_code_iterator SVE_UNPRED_FP_BINARY [plus minus mult])
+
 ;; SVE integer comparisons.
 (define_code_iterator SVE_INT_CMP [lt le eq ne ge gt ltu leu geu gtu])
 
@@ -1423,6 +1426,8 @@
 
 ;; The floating-point SVE instruction that implements an rtx code.
 (define_code_attr sve_fp_op [(plus "fadd")
+			     (minus "fsub")
+			     (mult "fmul")
 			     (neg "fneg")
 			     (abs "fabs")
 			     (sqrt "fsqrt")])
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 03fc100..996cacd 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,7 @@
+2018-12-07  Richard Sandiford  <richard.sandiford@arm.com>
+
+	* gcc.target/aarch64/sve/pred_elim_1.c: New test.
+
 2018-12-07  Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>
 
 	* gcc.target/i386/ipa-stack-alignment-2.c: Add
diff --git a/gcc/testsuite/gcc.target/aarch64/sve/pred_elim_1.c b/gcc/testsuite/gcc.target/aarch64/sve/pred_elim_1.c
new file mode 100644
index 0000000..6b0faf1
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/sve/pred_elim_1.c
@@ -0,0 +1,23 @@
+/* { dg-options "-O2 -ftree-vectorize" } */
+
+#define TEST_OP(NAME, TYPE, OP)				\
+  void							\
+  NAME##_##TYPE (TYPE *restrict a, TYPE *restrict b,	\
+		 TYPE *restrict c, int n)		\
+  {							\
+    for (int i = 0; i < n; ++i)				\
+      a[i] = b[i] OP c[i];				\
+  }
+
+#define TEST_TYPE(TYPE) \
+  TEST_OP (add, TYPE, +) \
+  TEST_OP (sub, TYPE, -) \
+  TEST_OP (mult, TYPE, *) \
+
+TEST_TYPE (float)
+TEST_TYPE (double)
+
+/* { dg-final { scan-assembler-times {\tfadd\t} 2 } } */
+/* { dg-final { scan-assembler-times {\tfsub\t} 2 } } */
+/* { dg-final { scan-assembler-times {\tfmul\t} 2 } } */
+/* { dg-final { scan-assembler-not {\tptrue\t} } } */