[committed,AArch64] Add SVE conditional conversion patterns
diff mbox series

Message ID mpt7e7gm2tv.fsf@arm.com
State New
Headers show
Series
  • [committed,AArch64] Add SVE conditional conversion patterns
Related show

Commit Message

Richard Sandiford Aug. 14, 2019, 10:57 a.m. UTC
This patch adds patterns to match conditional conversions between
integers and like-sized floats.  The patterns are actually more
general than that, but the other combinations can only be tested
via the ACLE.

Tested on aarch64-linux-gnu (with and without SVE) and aarch64_be-elf.
Applied as r274478.

Richard


2019-08-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
	* config/aarch64/aarch64-sve.md
	(*cond_<SVE_COND_FCVTI:optab>_nontrunc<SVE_F:mode><SVE_HSDI:mode>)
	(*cond_<SVE_COND_ICVTF:optab>_nonextend<SVE_HSDI:mode><SVE_F:mode>):
	New patterns.

gcc/testsuite/
	* gcc.target/aarch64/sve/cond_convert_1.c: New test.
	* gcc.target/aarch64/sve/cond_convert_1_run.c: Likewise.
	* gcc.target/aarch64/sve/cond_convert_2.c: Likewise.
	* gcc.target/aarch64/sve/cond_convert_2_run.c: Likewise.
	* gcc.target/aarch64/sve/cond_convert_3.c: Likewise.
	* gcc.target/aarch64/sve/cond_convert_3_run.c: Likewise.
	* gcc.target/aarch64/sve/cond_convert_4.c: Likewise.
	* gcc.target/aarch64/sve/cond_convert_4_run.c: Likewise.
	* gcc.target/aarch64/sve/cond_convert_5.c: Likewise.
	* gcc.target/aarch64/sve/cond_convert_5_run.c: Likewise.
	* gcc.target/aarch64/sve/cond_convert_6.c: Likewise.
	* gcc.target/aarch64/sve/cond_convert_6_run.c: Likewise.

Patch
diff mbox series

Index: gcc/config/aarch64/aarch64-sve.md
===================================================================
--- gcc/config/aarch64/aarch64-sve.md	2019-08-14 11:53:04.636898923 +0100
+++ gcc/config/aarch64/aarch64-sve.md	2019-08-14 11:55:33.251813494 +0100
@@ -4071,6 +4071,39 @@  (define_insn "*aarch64_sve_<optab>_trunc
   "fcvtz<su>\t%0.<VNx4SI_ONLY:Vetype>, %1/m, %2.<VNx2DF_ONLY:Vetype>"
 )
 
+;; Predicated float-to-integer conversion with merging, either to the same
+;; width or wider.
+;;
+;; The first alternative doesn't need the earlyclobber, but the only case
+;; it would help is the uninteresting one in which operands 2 and 3 are
+;; the same register (despite having different modes).  Making all the
+;; alternatives earlyclobber makes things more consistent for the
+;; register allocator.
+(define_insn_and_rewrite "*cond_<optab>_nontrunc<SVE_F:mode><SVE_HSDI:mode>"
+  [(set (match_operand:SVE_HSDI 0 "register_operand" "=&w, &w, ?&w")
+	(unspec:SVE_HSDI
+	  [(match_operand:<SVE_HSDI:VPRED> 1 "register_operand" "Upl, Upl, Upl")
+	   (unspec:SVE_HSDI
+	     [(match_operand 4)
+	      (match_operand:SI 5 "aarch64_sve_gp_strictness")
+	      (match_operand:SVE_F 2 "register_operand" "w, w, w")]
+	     SVE_COND_FCVTI)
+	   (match_operand:SVE_HSDI 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
+	  UNSPEC_SEL))]
+  "TARGET_SVE
+   && <SVE_HSDI:elem_bits> >= <SVE_F:elem_bits>
+   && aarch64_sve_pred_dominates_p (&operands[4], operands[1])"
+  "@
+   fcvtz<su>\t%0.<SVE_HSDI:Vetype>, %1/m, %2.<SVE_F:Vetype>
+   movprfx\t%0.<SVE_HSDI:Vetype>, %1/z, %2.<SVE_HSDI:Vetype>\;fcvtz<su>\t%0.<SVE_HSDI:Vetype>, %1/m, %2.<SVE_F:Vetype>
+   movprfx\t%0, %3\;fcvtz<su>\t%0.<SVE_HSDI:Vetype>, %1/m, %2.<SVE_F:Vetype>"
+  "&& !rtx_equal_p (operands[1], operands[4])"
+  {
+    operands[4] = copy_rtx (operands[1]);
+  }
+  [(set_attr "movprfx" "*,yes,yes")]
+)
+
 ;; -------------------------------------------------------------------------
 ;; ---- [INT<-FP] Packs
 ;; -------------------------------------------------------------------------
@@ -4155,6 +4188,39 @@  (define_insn "aarch64_sve_<optab>_extend
   "<su>cvtf\t%0.<VNx2DF_ONLY:Vetype>, %1/m, %2.<VNx4SI_ONLY:Vetype>"
 )
 
+;; Predicated integer-to-float conversion with merging, either to the same
+;; width or narrower.
+;;
+;; The first alternative doesn't need the earlyclobber, but the only case
+;; it would help is the uninteresting one in which operands 2 and 3 are
+;; the same register (despite having different modes).  Making all the
+;; alternatives earlyclobber makes things more consistent for the
+;; register allocator.
+(define_insn_and_rewrite "*cond_<optab>_nonextend<SVE_HSDI:mode><SVE_F:mode>"
+  [(set (match_operand:SVE_F 0 "register_operand" "=&w, &w, ?&w")
+	(unspec:SVE_F
+	  [(match_operand:<SVE_HSDI:VPRED> 1 "register_operand" "Upl, Upl, Upl")
+	   (unspec:SVE_F
+	     [(match_operand 4)
+	      (match_operand:SI 5 "aarch64_sve_gp_strictness")
+	      (match_operand:SVE_HSDI 2 "register_operand" "w, w, w")]
+	     SVE_COND_ICVTF)
+	   (match_operand:SVE_F 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
+	  UNSPEC_SEL))]
+  "TARGET_SVE
+   && <SVE_HSDI:elem_bits> >= <SVE_F:elem_bits>
+   && aarch64_sve_pred_dominates_p (&operands[4], operands[1])"
+  "@
+   <su>cvtf\t%0.<SVE_F:Vetype>, %1/m, %2.<SVE_HSDI:Vetype>
+   movprfx\t%0.<SVE_HSDI:Vetype>, %1/z, %2.<SVE_HSDI:Vetype>\;<su>cvtf\t%0.<SVE_F:Vetype>, %1/m, %2.<SVE_HSDI:Vetype>
+   movprfx\t%0, %3\;<su>cvtf\t%0.<SVE_F:Vetype>, %1/m, %2.<SVE_HSDI:Vetype>"
+  "&& !rtx_equal_p (operands[1], operands[4])"
+  {
+    operands[4] = copy_rtx (operands[1]);
+  }
+  [(set_attr "movprfx" "*,yes,yes")]
+)
+
 ;; -------------------------------------------------------------------------
 ;; ---- [FP<-INT] Packs
 ;; -------------------------------------------------------------------------
Index: gcc/testsuite/gcc.target/aarch64/sve/cond_convert_1.c
===================================================================
--- /dev/null	2019-07-30 08:53:31.317691683 +0100
+++ gcc/testsuite/gcc.target/aarch64/sve/cond_convert_1.c	2019-08-14 11:55:33.251813494 +0100
@@ -0,0 +1,37 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-vectorize -fno-trapping-math" } */
+
+#include <stdint.h>
+
+#define DEF_LOOP(FLOAT_TYPE, INT_TYPE)				\
+  void __attribute__ ((noipa))					\
+  test_##INT_TYPE (FLOAT_TYPE *__restrict r,			\
+		   INT_TYPE *__restrict a,			\
+		   FLOAT_TYPE *__restrict b,			\
+		   INT_TYPE *__restrict pred, int n)		\
+  {								\
+    for (int i = 0; i < n; ++i)					\
+      r[i] = pred[i] ? (FLOAT_TYPE) a[i] : b[i];		\
+  }
+
+#define TEST_ALL(T) \
+  T (_Float16, int16_t) \
+  T (_Float16, uint16_t) \
+  T (float, int32_t) \
+  T (float, uint32_t) \
+  T (double, int64_t) \
+  T (double, uint64_t)
+
+TEST_ALL (DEF_LOOP)
+
+/* { dg-final { scan-assembler-times {\tscvtf\tz[0-9]+\.h, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tucvtf\tz[0-9]+\.h, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tscvtf\tz[0-9]+\.s, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tucvtf\tz[0-9]+\.s, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tscvtf\tz[0-9]+\.d, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tucvtf\tz[0-9]+\.d, p[0-7]/m,} 1 } } */
+
+/* { dg-final { scan-assembler-not {\tmov\tz} } } */
+/* At the moment we don't manage to avoid using MOVPRFX.  */
+/* { dg-final { scan-assembler-not {\tmovprfx\t} { xfail *-*-* } } } */
+/* { dg-final { scan-assembler-not {\tsel\t} } } */
Index: gcc/testsuite/gcc.target/aarch64/sve/cond_convert_1_run.c
===================================================================
--- /dev/null	2019-07-30 08:53:31.317691683 +0100
+++ gcc/testsuite/gcc.target/aarch64/sve/cond_convert_1_run.c	2019-08-14 11:55:33.251813494 +0100
@@ -0,0 +1,29 @@ 
+/* { dg-do run { target { aarch64_sve_hw } } } */
+/* { dg-options "-O2 -ftree-vectorize -ftrapping-math" } */
+
+#include "cond_convert_1.c"
+
+#define N 99
+
+#define TEST_LOOP(FLOAT_TYPE, INT_TYPE)				\
+  {								\
+    FLOAT_TYPE r[N], b[N];					\
+    INT_TYPE a[N], pred[N];					\
+    for (int i = 0; i < N; ++i)					\
+      {								\
+	a[i] = (i & 1 ? i : 3 * i) * (i % 3 == 0 ? 1 : -1);	\
+	b[i] = (i % 9) * (i % 7 + 1);				\
+	pred[i] = (i % 7 < 4);					\
+	asm volatile ("" ::: "memory");				\
+      }								\
+    test_##INT_TYPE (r, a, b, pred, N);				\
+    for (int i = 0; i < N; ++i)					\
+      if (r[i] != (pred[i] ? (FLOAT_TYPE) a[i] : b[i]))		\
+	__builtin_abort ();					\
+  }
+
+int main ()
+{
+  TEST_ALL (TEST_LOOP)
+  return 0;
+}
Index: gcc/testsuite/gcc.target/aarch64/sve/cond_convert_2.c
===================================================================
--- /dev/null	2019-07-30 08:53:31.317691683 +0100
+++ gcc/testsuite/gcc.target/aarch64/sve/cond_convert_2.c	2019-08-14 11:55:33.251813494 +0100
@@ -0,0 +1,36 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-vectorize -fno-trapping-math" } */
+
+#include <stdint.h>
+
+#define DEF_LOOP(FLOAT_TYPE, INT_TYPE)				\
+  void __attribute__ ((noipa))					\
+  test_##INT_TYPE (FLOAT_TYPE *__restrict r,			\
+		   INT_TYPE *__restrict a,			\
+		   INT_TYPE *__restrict pred, int n)		\
+  {								\
+    for (int i = 0; i < n; ++i)					\
+      r[i] = pred[i] ? (FLOAT_TYPE) a[i] : 1.0;			\
+  }
+
+#define TEST_ALL(T) \
+  T (_Float16, int16_t) \
+  T (_Float16, uint16_t) \
+  T (float, int32_t) \
+  T (float, uint32_t) \
+  T (double, int64_t) \
+  T (double, uint64_t)
+
+TEST_ALL (DEF_LOOP)
+
+/* { dg-final { scan-assembler-times {\tscvtf\tz[0-9]+\.h, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tucvtf\tz[0-9]+\.h, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tscvtf\tz[0-9]+\.s, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tucvtf\tz[0-9]+\.s, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tscvtf\tz[0-9]+\.d, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tucvtf\tz[0-9]+\.d, p[0-7]/m,} 1 } } */
+
+/* { dg-final { scan-assembler-times {\tmovprfx\tz[0-9]+, z[0-9]+\n} 6 } } */
+
+/* { dg-final { scan-assembler-not {\tmov\tz[^\n]*z} } } */
+/* { dg-final { scan-assembler-not {\tsel\t} } } */
Index: gcc/testsuite/gcc.target/aarch64/sve/cond_convert_2_run.c
===================================================================
--- /dev/null	2019-07-30 08:53:31.317691683 +0100
+++ gcc/testsuite/gcc.target/aarch64/sve/cond_convert_2_run.c	2019-08-14 11:55:33.251813494 +0100
@@ -0,0 +1,28 @@ 
+/* { dg-do run { target { aarch64_sve_hw } } } */
+/* { dg-options "-O2 -ftree-vectorize -ftrapping-math" } */
+
+#include "cond_convert_2.c"
+
+#define N 99
+
+#define TEST_LOOP(FLOAT_TYPE, INT_TYPE)				\
+  {								\
+    FLOAT_TYPE r[N];						\
+    INT_TYPE a[N], pred[N];					\
+    for (int i = 0; i < N; ++i)					\
+      {								\
+	a[i] = (i & 1 ? i : 3 * i) * (i % 3 == 0 ? 1 : -1);	\
+	pred[i] = (i % 7 < 4);					\
+	asm volatile ("" ::: "memory");				\
+      }								\
+    test_##INT_TYPE (r, a, pred, N);				\
+    for (int i = 0; i < N; ++i)					\
+      if (r[i] != (pred[i] ? (FLOAT_TYPE) a[i] : 1.0))		\
+	__builtin_abort ();					\
+  }
+
+int main ()
+{
+  TEST_ALL (TEST_LOOP)
+  return 0;
+}
Index: gcc/testsuite/gcc.target/aarch64/sve/cond_convert_3.c
===================================================================
--- /dev/null	2019-07-30 08:53:31.317691683 +0100
+++ gcc/testsuite/gcc.target/aarch64/sve/cond_convert_3.c	2019-08-14 11:55:33.251813494 +0100
@@ -0,0 +1,40 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-vectorize -fno-trapping-math" } */
+
+#include <stdint.h>
+
+#define DEF_LOOP(FLOAT_TYPE, INT_TYPE)				\
+  void __attribute__ ((noipa))					\
+  test_##INT_TYPE (FLOAT_TYPE *__restrict r,			\
+		   INT_TYPE *__restrict a,			\
+		   INT_TYPE *__restrict pred, int n)		\
+  {								\
+    for (int i = 0; i < n; ++i)					\
+      r[i] = pred[i] ? (FLOAT_TYPE) a[i] : 0.0;			\
+  }
+
+#define TEST_ALL(T) \
+  T (_Float16, int16_t) \
+  T (_Float16, uint16_t) \
+  T (float, int32_t) \
+  T (float, uint32_t) \
+  T (double, int64_t) \
+  T (double, uint64_t)
+
+TEST_ALL (DEF_LOOP)
+
+/* { dg-final { scan-assembler-times {\tscvtf\tz[0-9]+\.h, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tucvtf\tz[0-9]+\.h, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tscvtf\tz[0-9]+\.s, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tucvtf\tz[0-9]+\.s, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tscvtf\tz[0-9]+\.d, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tucvtf\tz[0-9]+\.d, p[0-7]/m,} 1 } } */
+
+/* Really we should be able to use MOVPRFX /z here, but at the moment
+   we're relying on combine to merge a SEL and an arithmetic operation,
+   and the SEL doesn't allow the "false" value to be zero when the "true"
+   value is a register.  */
+/* { dg-final { scan-assembler-times {\tmovprfx\tz[0-9]+, z[0-9]+\n} 6 } } */
+
+/* { dg-final { scan-assembler-not {\tmov\tz[^\n]*z} } } */
+/* { dg-final { scan-assembler-not {\tsel\t} } } */
Index: gcc/testsuite/gcc.target/aarch64/sve/cond_convert_3_run.c
===================================================================
--- /dev/null	2019-07-30 08:53:31.317691683 +0100
+++ gcc/testsuite/gcc.target/aarch64/sve/cond_convert_3_run.c	2019-08-14 11:55:33.251813494 +0100
@@ -0,0 +1,28 @@ 
+/* { dg-do run { target { aarch64_sve_hw } } } */
+/* { dg-options "-O2 -ftree-vectorize -ftrapping-math" } */
+
+#include "cond_convert_3.c"
+
+#define N 99
+
+#define TEST_LOOP(FLOAT_TYPE, INT_TYPE)				\
+  {								\
+    FLOAT_TYPE r[N];						\
+    INT_TYPE a[N], pred[N];					\
+    for (int i = 0; i < N; ++i)					\
+      {								\
+	a[i] = (i & 1 ? i : 3 * i) * (i % 3 == 0 ? 1 : -1);	\
+	pred[i] = (i % 7 < 4);					\
+	asm volatile ("" ::: "memory");				\
+      }								\
+    test_##INT_TYPE (r, a, pred, N);				\
+    for (int i = 0; i < N; ++i)					\
+      if (r[i] != (pred[i] ? (FLOAT_TYPE) a[i] : 0.0))		\
+	__builtin_abort ();					\
+  }
+
+int main ()
+{
+  TEST_ALL (TEST_LOOP)
+  return 0;
+}
Index: gcc/testsuite/gcc.target/aarch64/sve/cond_convert_4.c
===================================================================
--- /dev/null	2019-07-30 08:53:31.317691683 +0100
+++ gcc/testsuite/gcc.target/aarch64/sve/cond_convert_4.c	2019-08-14 11:55:33.251813494 +0100
@@ -0,0 +1,37 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-vectorize -fno-trapping-math" } */
+
+#include <stdint.h>
+
+#define DEF_LOOP(FLOAT_TYPE, INT_TYPE)				\
+  void __attribute__ ((noipa))					\
+  test_##INT_TYPE (INT_TYPE *__restrict r,			\
+		   FLOAT_TYPE *__restrict a,			\
+		   INT_TYPE *__restrict b,			\
+		   INT_TYPE *__restrict pred, int n)		\
+  {								\
+    for (int i = 0; i < n; ++i)					\
+      r[i] = pred[i] ? (INT_TYPE) a[i] : b[i];			\
+  }
+
+#define TEST_ALL(T) \
+  T (_Float16, int16_t) \
+  T (_Float16, uint16_t) \
+  T (float, int32_t) \
+  T (float, uint32_t) \
+  T (double, int64_t) \
+  T (double, uint64_t)
+
+TEST_ALL (DEF_LOOP)
+
+/* { dg-final { scan-assembler-times {\tfcvtzs\tz[0-9]+\.h, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzu\tz[0-9]+\.h, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzs\tz[0-9]+\.s, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzu\tz[0-9]+\.s, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzs\tz[0-9]+\.d, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzu\tz[0-9]+\.d, p[0-7]/m,} 1 } } */
+
+/* { dg-final { scan-assembler-not {\tmov\tz} } } */
+/* At the moment we don't manage to avoid using MOVPRFX.  */
+/* { dg-final { scan-assembler-not {\tmovprfx\t} { xfail *-*-* } } } */
+/* { dg-final { scan-assembler-not {\tsel\t} } } */
Index: gcc/testsuite/gcc.target/aarch64/sve/cond_convert_4_run.c
===================================================================
--- /dev/null	2019-07-30 08:53:31.317691683 +0100
+++ gcc/testsuite/gcc.target/aarch64/sve/cond_convert_4_run.c	2019-08-14 11:55:33.251813494 +0100
@@ -0,0 +1,29 @@ 
+/* { dg-do run { target { aarch64_sve_hw } } } */
+/* { dg-options "-O2 -ftree-vectorize -ftrapping-math" } */
+
+#include "cond_convert_4.c"
+
+#define N 99
+
+#define TEST_LOOP(FLOAT_TYPE, INT_TYPE)				\
+  {								\
+    INT_TYPE r[N], b[N], pred[N];				\
+    FLOAT_TYPE a[N];						\
+    for (int i = 0; i < N; ++i)					\
+      {								\
+	a[i] = (i & 1 ? i : 3 * i) * (i % 3 == 0 ? 1 : -1);	\
+	b[i] = (i % 9) * (i % 7 + 1);				\
+	pred[i] = (i % 7 < 4);					\
+	asm volatile ("" ::: "memory");				\
+      }								\
+    test_##INT_TYPE (r, a, b, pred, N);				\
+    for (int i = 0; i < N; ++i)					\
+      if (r[i] != (pred[i] ? (INT_TYPE) a[i] : b[i]))		\
+	__builtin_abort ();					\
+  }
+
+int main ()
+{
+  TEST_ALL (TEST_LOOP)
+  return 0;
+}
Index: gcc/testsuite/gcc.target/aarch64/sve/cond_convert_5.c
===================================================================
--- /dev/null	2019-07-30 08:53:31.317691683 +0100
+++ gcc/testsuite/gcc.target/aarch64/sve/cond_convert_5.c	2019-08-14 11:55:33.251813494 +0100
@@ -0,0 +1,36 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-vectorize -fno-trapping-math" } */
+
+#include <stdint.h>
+
+#define DEF_LOOP(FLOAT_TYPE, INT_TYPE)				\
+  void __attribute__ ((noipa))					\
+  test_##INT_TYPE (INT_TYPE *__restrict r,			\
+		   FLOAT_TYPE *__restrict a,			\
+		   INT_TYPE *__restrict pred, int n)		\
+  {								\
+    for (int i = 0; i < n; ++i)					\
+      r[i] = pred[i] ? (INT_TYPE) a[i] : 72;			\
+  }
+
+#define TEST_ALL(T) \
+  T (_Float16, int16_t) \
+  T (_Float16, uint16_t) \
+  T (float, int32_t) \
+  T (float, uint32_t) \
+  T (double, int64_t) \
+  T (double, uint64_t)
+
+TEST_ALL (DEF_LOOP)
+
+/* { dg-final { scan-assembler-times {\tfcvtzs\tz[0-9]+\.h, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzu\tz[0-9]+\.h, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzs\tz[0-9]+\.s, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzu\tz[0-9]+\.s, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzs\tz[0-9]+\.d, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzu\tz[0-9]+\.d, p[0-7]/m,} 1 } } */
+
+/* { dg-final { scan-assembler-times {\tmovprfx\tz[0-9]+, z[0-9]+\n} 6 } } */
+
+/* { dg-final { scan-assembler-not {\tmov\tz[^\n]*z} } } */
+/* { dg-final { scan-assembler-not {\tsel\t} } } */
Index: gcc/testsuite/gcc.target/aarch64/sve/cond_convert_5_run.c
===================================================================
--- /dev/null	2019-07-30 08:53:31.317691683 +0100
+++ gcc/testsuite/gcc.target/aarch64/sve/cond_convert_5_run.c	2019-08-14 11:55:33.251813494 +0100
@@ -0,0 +1,28 @@ 
+/* { dg-do run { target { aarch64_sve_hw } } } */
+/* { dg-options "-O2 -ftree-vectorize -ftrapping-math" } */
+
+#include "cond_convert_5.c"
+
+#define N 99
+
+#define TEST_LOOP(FLOAT_TYPE, INT_TYPE)				\
+  {								\
+    INT_TYPE r[N], pred[N];					\
+    FLOAT_TYPE a[N];						\
+    for (int i = 0; i < N; ++i)					\
+      {								\
+	a[i] = (i & 1 ? i : 3 * i) * (i % 3 == 0 ? 1 : -1);	\
+	pred[i] = (i % 7 < 4);					\
+	asm volatile ("" ::: "memory");				\
+      }								\
+    test_##INT_TYPE (r, a, pred, N);				\
+    for (int i = 0; i < N; ++i)					\
+      if (r[i] != (pred[i] ? (INT_TYPE) a[i] : 72))		\
+	__builtin_abort ();					\
+  }
+
+int main ()
+{
+  TEST_ALL (TEST_LOOP)
+  return 0;
+}
Index: gcc/testsuite/gcc.target/aarch64/sve/cond_convert_6.c
===================================================================
--- /dev/null	2019-07-30 08:53:31.317691683 +0100
+++ gcc/testsuite/gcc.target/aarch64/sve/cond_convert_6.c	2019-08-14 11:55:33.251813494 +0100
@@ -0,0 +1,40 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-vectorize -fno-trapping-math" } */
+
+#include <stdint.h>
+
+#define DEF_LOOP(FLOAT_TYPE, INT_TYPE)				\
+  void __attribute__ ((noipa))					\
+  test_##INT_TYPE (INT_TYPE *__restrict r,			\
+		   FLOAT_TYPE *__restrict a,			\
+		   INT_TYPE *__restrict pred, int n)		\
+  {								\
+    for (int i = 0; i < n; ++i)					\
+      r[i] = pred[i] ? (INT_TYPE) a[i] : 0;			\
+  }
+
+#define TEST_ALL(T) \
+  T (_Float16, int16_t) \
+  T (_Float16, uint16_t) \
+  T (float, int32_t) \
+  T (float, uint32_t) \
+  T (double, int64_t) \
+  T (double, uint64_t)
+
+TEST_ALL (DEF_LOOP)
+
+/* { dg-final { scan-assembler-times {\tfcvtzs\tz[0-9]+\.h, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzu\tz[0-9]+\.h, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzs\tz[0-9]+\.s, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzu\tz[0-9]+\.s, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzs\tz[0-9]+\.d, p[0-7]/m,} 1 } } */
+/* { dg-final { scan-assembler-times {\tfcvtzu\tz[0-9]+\.d, p[0-7]/m,} 1 } } */
+
+/* Really we should be able to use MOVPRFX /z here, but at the moment
+   we're relying on combine to merge a SEL and an arithmetic operation,
+   and the SEL doesn't allow the "false" value to be zero when the "true"
+   value is a register.  */
+/* { dg-final { scan-assembler-times {\tmovprfx\tz[0-9]+, z[0-9]+\n} 6 } } */
+
+/* { dg-final { scan-assembler-not {\tmov\tz[^\n]*z} } } */
+/* { dg-final { scan-assembler-not {\tsel\t} } } */
Index: gcc/testsuite/gcc.target/aarch64/sve/cond_convert_6_run.c
===================================================================
--- /dev/null	2019-07-30 08:53:31.317691683 +0100
+++ gcc/testsuite/gcc.target/aarch64/sve/cond_convert_6_run.c	2019-08-14 11:55:33.251813494 +0100
@@ -0,0 +1,28 @@ 
+/* { dg-do run { target { aarch64_sve_hw } } } */
+/* { dg-options "-O2 -ftree-vectorize -ftrapping-math" } */
+
+#include "cond_convert_6.c"
+
+#define N 99
+
+#define TEST_LOOP(FLOAT_TYPE, INT_TYPE)				\
+  {								\
+    INT_TYPE r[N], pred[N];					\
+    FLOAT_TYPE a[N];						\
+    for (int i = 0; i < N; ++i)					\
+      {								\
+	a[i] = (i & 1 ? i : 3 * i) * (i % 3 == 0 ? 1 : -1);	\
+	pred[i] = (i % 7 < 4);					\
+	asm volatile ("" ::: "memory");				\
+      }								\
+    test_##INT_TYPE (r, a, pred, N);				\
+    for (int i = 0; i < N; ++i)					\
+      if (r[i] != (pred[i] ? (INT_TYPE) a[i] : 0))		\
+	__builtin_abort ();					\
+  }
+
+int main ()
+{
+  TEST_ALL (TEST_LOOP)
+  return 0;
+}