Add some missing test cases for ACLE intrinsics of AArch64 NEON.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@197994 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/test/CodeGen/aarch64-neon-2velem.c b/test/CodeGen/aarch64-neon-2velem.c
index 03f7df715c536da126f4df3765deaf6b33342e3b..885aad794c035d43f70b13f7655d62983b6cd8d0 100644 (file)
--- a/test/CodeGen/aarch64-neon-2velem.c
+++ b/test/CodeGen/aarch64-neon-2velem.c
@@ -272,6 +272,30 @@ float64x2_t test_vfmsq_laneq_f64(float64x2_t a, float64x2_t b, float64x2_t v) {
   // CHECK: fmls {{v[0-9]+}}.2d, {{v[0-9]+}}.2d, {{v[0-9]+}}.d[1]
 }
 
+float32_t test_vfmas_laneq_f32(float32_t a, float32_t b, float32x4_t v) {
+  // CHECK-LABEL: test_vfmas_laneq_f32
+  return vfmas_laneq_f32(a, b, v, 3);
+  // CHECK: fmla {{s[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}.s[3]
+}
+
+float64_t test_vfmsd_lane_f64(float64_t a, float64_t b, float64x1_t v) {
+  // CHECK-LABEL: test_vfmsd_lane_f64
+  return vfmsd_lane_f64(a, b, v, 0);
+  // CHECK: fmls {{d[0-9]+}}, {{d[0-9]+}}, {{v[0-9]+}}.d[0]
+}
+
+float32_t test_vfmss_laneq_f32(float32_t a, float32_t b, float32x4_t v) {
+  // CHECK-LABEL: test_vfmss_laneq_f32
+  return vfmss_laneq_f32(a, b, v, 3);
+  // CHECK: fmls {{s[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}.s[3]
+}
+
+float64_t test_vfmsd_laneq_f64(float64_t a, float64_t b, float64x2_t v) {
+  // CHECK-LABEL: test_vfmsd_laneq_f64
+  return vfmsd_laneq_f64(a, b, v, 1);
+  // CHECK: fmls {{d[0-9]+}}, {{d[0-9]+}}, {{v[0-9]+}}.d[1]
+}
+
 int32x4_t test_vmlal_lane_s16(int32x4_t a, int16x4_t b, int16x4_t v) {
   // CHECK: test_vmlal_lane_s16
   return vmlal_lane_s16(a, b, v, 3);
@@ -1696,3 +1720,736 @@ float32x4_t test_vfmsq_n_f32(float32x4_t a, float32x4_t b, float32_t n) {
   return vfmsq_n_f32(a, b, n);
   // CHECK: fmls {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
 }
+
+int16x4_t test_vmul_n_s16(int16x4_t a, int16_t b) {
+  // CHECK-LABEL: test_vmul_n_s16
+  return vmul_n_s16(a, b);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: mul {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+int16x8_t test_vmulq_n_s16(int16x8_t a, int16_t b) {
+  // CHECK-LABEL: test_vmulq_n_s16
+  return vmulq_n_s16(a, b);
+  // CHECK: dup {{v[0-9]+}}.8h, w0
+  // CHECK: mul {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.8h
+}
+
+int32x2_t test_vmul_n_s32(int32x2_t a, int32_t b) {
+  // CHECK-LABEL: test_vmul_n_s32
+  return vmul_n_s32(a, b);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: mul {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+int32x4_t test_vmulq_n_s32(int32x4_t a, int32_t b) {
+  // CHECK-LABEL: test_vmulq_n_s32
+  return vmulq_n_s32(a, b);
+  // CHECK: dup {{v[0-9]+}}.4s, w0
+  // CHECK: mul {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+}
+
+uint16x4_t test_vmul_n_u16(uint16x4_t a, uint16_t b) {
+  // CHECK-LABEL: test_vmul_n_u16
+  return vmul_n_u16(a, b);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: mul {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+uint16x8_t test_vmulq_n_u16(uint16x8_t a, uint16_t b) {
+  // CHECK-LABEL: test_vmulq_n_u16
+  return vmulq_n_u16(a, b);
+  // CHECK: dup {{v[0-9]+}}.8h, w0
+  // CHECK: mul {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.8h
+}
+
+uint32x2_t test_vmul_n_u32(uint32x2_t a, uint32_t b) {
+  // CHECK-LABEL: test_vmul_n_u32
+  return vmul_n_u32(a, b);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: mul {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+uint32x4_t test_vmulq_n_u32(uint32x4_t a, uint32_t b) {
+  // CHECK-LABEL: test_vmulq_n_u32
+  return vmulq_n_u32(a, b);
+  // CHECK: dup {{v[0-9]+}}.4s, w0
+  // CHECK: mul {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+}
+
+int32x4_t test_vmull_n_s16(int16x4_t a, int16_t b) {
+  // CHECK-LABEL: test_vmull_n_s16
+  return vmull_n_s16(a, b);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: smull {{v[0-9]+}}.4s, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+int64x2_t test_vmull_n_s32(int32x2_t a, int32_t b) {
+  // CHECK-LABEL: test_vmull_n_s32
+  return vmull_n_s32(a, b);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: smull {{v[0-9]+}}.2d, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+uint32x4_t test_vmull_n_u16(uint16x4_t a, uint16_t b) {
+  // CHECK-LABEL: test_vmull_n_u16
+  return vmull_n_u16(a, b);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: umull {{v[0-9]+}}.4s, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+uint64x2_t test_vmull_n_u32(uint32x2_t a, uint32_t b) {
+  // CHECK-LABEL: test_vmull_n_u32
+  return vmull_n_u32(a, b);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: umull {{v[0-9]+}}.2d, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+int32x4_t test_vqdmull_n_s16(int16x4_t a, int16_t b) {
+  // CHECK-LABEL: test_vqdmull_n_s16
+  return vqdmull_n_s16(a, b);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: sqdmull {{v[0-9]+}}.4s, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+int64x2_t test_vqdmull_n_s32(int32x2_t a, int32_t b) {
+  // CHECK-LABEL: test_vqdmull_n_s32
+  return vqdmull_n_s32(a, b);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: sqdmull {{v[0-9]+}}.2d, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+int16x4_t test_vqdmulh_n_s16(int16x4_t a, int16_t b) {
+  // CHECK-LABEL: test_vqdmulh_n_s16
+  return vqdmulh_n_s16(a, b);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: sqdmulh {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+int16x8_t test_vqdmulhq_n_s16(int16x8_t a, int16_t b) {
+  // CHECK-LABEL: test_vqdmulhq_n_s16
+  return vqdmulhq_n_s16(a, b);
+  // CHECK: dup {{v[0-9]+}}.8h, w0
+  // CHECK: sqdmulh {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.8h
+}
+
+int32x2_t test_vqdmulh_n_s32(int32x2_t a, int32_t b) {
+  // CHECK-LABEL: test_vqdmulh_n_s32
+  return vqdmulh_n_s32(a, b);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: sqdmulh {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+int32x4_t test_vqdmulhq_n_s32(int32x4_t a, int32_t b) {
+  // CHECK-LABEL: test_vqdmulhq_n_s32
+  return vqdmulhq_n_s32(a, b);
+  // CHECK: dup {{v[0-9]+}}.4s, w0
+  // CHECK: sqdmulh {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+}
+
+int16x4_t test_vqrdmulh_n_s16(int16x4_t a, int16_t b) {
+  // CHECK-LABEL: test_vqrdmulh_n_s16
+  return vqrdmulh_n_s16(a, b);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: sqrdmulh {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+int16x8_t test_vqrdmulhq_n_s16(int16x8_t a, int16_t b) {
+  // CHECK-LABEL: test_vqrdmulhq_n_s16
+  return vqrdmulhq_n_s16(a, b);
+  // CHECK: dup {{v[0-9]+}}.8h, w0
+  // CHECK: sqrdmulh {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.8h
+}
+
+int32x2_t test_vqrdmulh_n_s32(int32x2_t a, int32_t b) {
+  // CHECK-LABEL: test_vqrdmulh_n_s32
+  return vqrdmulh_n_s32(a, b);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: sqrdmulh {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+int32x4_t test_vqrdmulhq_n_s32(int32x4_t a, int32_t b) {
+  // CHECK-LABEL: test_vqrdmulhq_n_s32
+  return vqrdmulhq_n_s32(a, b);
+  // CHECK: dup {{v[0-9]+}}.4s, w0
+  // CHECK: sqrdmulh {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+}
+
+int16x4_t test_vmla_n_s16(int16x4_t a, int16x4_t b, int16_t c) {
+  // CHECK-LABEL: test_vmla_n_s16
+  return vmla_n_s16(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: mla {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+int16x8_t test_vmlaq_n_s16(int16x8_t a, int16x8_t b, int16_t c) {
+  // CHECK-LABEL: test_vmlaq_n_s16
+  return vmlaq_n_s16(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.8h, w0
+  // CHECK: mla {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.8h
+}
+
+int32x2_t test_vmla_n_s32(int32x2_t a, int32x2_t b, int32_t c) {
+  // CHECK-LABEL: test_vmla_n_s32
+  return vmla_n_s32(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: mla {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+int32x4_t test_vmlaq_n_s32(int32x4_t a, int32x4_t b, int32_t c) {
+  // CHECK-LABEL: test_vmlaq_n_s32
+  return vmlaq_n_s32(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.4s, w0
+  // CHECK: mla {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+}
+
+uint16x4_t test_vmla_n_u16(uint16x4_t a, uint16x4_t b, uint16_t c) {
+  // CHECK-LABEL: test_vmla_n_u16
+  return vmla_n_u16(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: mla {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+uint16x8_t test_vmlaq_n_u16(uint16x8_t a, uint16x8_t b, uint16_t c) {
+  // CHECK-LABEL: test_vmlaq_n_u16
+  return vmlaq_n_u16(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.8h, w0
+  // CHECK: mla {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.8h
+}
+
+uint32x2_t test_vmla_n_u32(uint32x2_t a, uint32x2_t b, uint32_t c) {
+  // CHECK-LABEL: test_vmla_n_u32
+  return vmla_n_u32(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: mla {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+uint32x4_t test_vmlaq_n_u32(uint32x4_t a, uint32x4_t b, uint32_t c) {
+  // CHECK-LABEL: test_vmlaq_n_u32
+  return vmlaq_n_u32(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.4s, w0
+  // CHECK: mla {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+}
+
+int32x4_t test_vmlal_n_s16(int32x4_t a, int16x4_t b, int16_t c) {
+  // CHECK-LABEL: test_vmlal_n_s16
+  return vmlal_n_s16(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: smlal {{v[0-9]+}}.4s, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+int64x2_t test_vmlal_n_s32(int64x2_t a, int32x2_t b, int32_t c) {
+  // CHECK-LABEL: test_vmlal_n_s32
+  return vmlal_n_s32(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: smlal {{v[0-9]+}}.2d, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+uint32x4_t test_vmlal_n_u16(uint32x4_t a, uint16x4_t b, uint16_t c) {
+  // CHECK-LABEL: test_vmlal_n_u16
+  return vmlal_n_u16(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: umlal {{v[0-9]+}}.4s, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+uint64x2_t test_vmlal_n_u32(uint64x2_t a, uint32x2_t b, uint32_t c) {
+  // CHECK-LABEL: test_vmlal_n_u32
+  return vmlal_n_u32(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: umlal {{v[0-9]+}}.2d, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+int32x4_t test_vqdmlal_n_s16(int32x4_t a, int16x4_t b, int16_t c) {
+  // CHECK-LABEL: test_vqdmlal_n_s16
+  return vqdmlal_n_s16(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: sqdmlal {{v[0-9]+}}.4s, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+int64x2_t test_vqdmlal_n_s32(int64x2_t a, int32x2_t b, int32_t c) {
+  // CHECK-LABEL: test_vqdmlal_n_s32
+  return vqdmlal_n_s32(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: sqdmlal {{v[0-9]+}}.2d, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+int16x4_t test_vmls_n_s16(int16x4_t a, int16x4_t b, int16_t c) {
+  // CHECK-LABEL: test_vmls_n_s16
+  return vmls_n_s16(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: mls {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+int16x8_t test_vmlsq_n_s16(int16x8_t a, int16x8_t b, int16_t c) {
+  // CHECK-LABEL: test_vmlsq_n_s16
+  return vmlsq_n_s16(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.8h, w0
+  // CHECK: mls {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.8h
+}
+
+int32x2_t test_vmls_n_s32(int32x2_t a, int32x2_t b, int32_t c) {
+  // CHECK-LABEL: test_vmls_n_s32
+  return vmls_n_s32(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: mls {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+int32x4_t test_vmlsq_n_s32(int32x4_t a, int32x4_t b, int32_t c) {
+  // CHECK-LABEL: test_vmlsq_n_s32
+  return vmlsq_n_s32(a, b, c);
+  // CHECK: mls {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+}
+
+uint16x4_t test_vmls_n_u16(uint16x4_t a, uint16x4_t b, uint16_t c) {
+  // CHECK-LABEL: test_vmls_n_u16
+  return vmls_n_u16(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: mls {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+uint16x8_t test_vmlsq_n_u16(uint16x8_t a, uint16x8_t b, uint16_t c) {
+  // CHECK-LABEL: test_vmlsq_n_u16
+  return vmlsq_n_u16(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.8h, w0
+  // CHECK: mls {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.8h
+}
+
+uint32x2_t test_vmls_n_u32(uint32x2_t a, uint32x2_t b, uint32_t c) {
+  // CHECK-LABEL: test_vmls_n_u32
+  return vmls_n_u32(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: mls {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+uint32x4_t test_vmlsq_n_u32(uint32x4_t a, uint32x4_t b, uint32_t c) {
+  // CHECK-LABEL: test_vmlsq_n_u32
+  return vmlsq_n_u32(a, b, c);
+  // CHECK: mls {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+}
+
+int32x4_t test_vmlsl_n_s16(int32x4_t a, int16x4_t b, int16_t c) {
+  // CHECK-LABEL: test_vmlsl_n_s16
+  return vmlsl_n_s16(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: smlsl {{v[0-9]+}}.4s, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+int64x2_t test_vmlsl_n_s32(int64x2_t a, int32x2_t b, int32_t c) {
+  // CHECK-LABEL: test_vmlsl_n_s32
+  return vmlsl_n_s32(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: smlsl {{v[0-9]+}}.2d, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+uint32x4_t test_vmlsl_n_u16(uint32x4_t a, uint16x4_t b, uint16_t c) {
+  // CHECK-LABEL: test_vmlsl_n_u16
+  return vmlsl_n_u16(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: umlsl {{v[0-9]+}}.4s, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+uint64x2_t test_vmlsl_n_u32(uint64x2_t a, uint32x2_t b, uint32_t c) {
+  // CHECK-LABEL: test_vmlsl_n_u32
+  return vmlsl_n_u32(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: umlsl {{v[0-9]+}}.2d, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+int32x4_t test_vqdmlsl_n_s16(int32x4_t a, int16x4_t b, int16_t c) {
+  // CHECK-LABEL: test_vqdmlsl_n_s16
+  return vqdmlsl_n_s16(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.4h, w0
+  // CHECK: sqdmlsl {{v[0-9]+}}.4s, {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+int64x2_t test_vqdmlsl_n_s32(int64x2_t a, int32x2_t b, int32_t c) {
+  // CHECK-LABEL: test_vqdmlsl_n_s32
+  return vqdmlsl_n_s32(a, b, c);
+  // CHECK: dup {{v[0-9]+}}.2s, w0
+  // CHECK: sqdmlsl {{v[0-9]+}}.2d, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+uint16x4_t test_vmla_lane_u16_0(uint16x4_t a, uint16x4_t b, uint16x4_t v) {
+  // CHECK-LABEL: test_vmla_lane_u16_0
+  return vmla_lane_u16(a, b, v, 0);
+  // CHECK: mla {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[0]
+}
+
+uint16x8_t test_vmlaq_lane_u16_0(uint16x8_t a, uint16x8_t b, uint16x4_t v) {
+  // CHECK-LABEL: test_vmlaq_lane_u16_0
+  return vmlaq_lane_u16(a, b, v, 0);
+  // CHECK: mla {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[0]
+}
+
+uint32x2_t test_vmla_lane_u32_0(uint32x2_t a, uint32x2_t b, uint32x2_t v) {
+  // CHECK-LABEL: test_vmla_lane_u32_0
+  return vmla_lane_u32(a, b, v, 0);
+  // CHECK: mla {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+}
+
+uint32x4_t test_vmlaq_lane_u32_0(uint32x4_t a, uint32x4_t b, uint32x2_t v) {
+  // CHECK-LABEL: test_vmlaq_lane_u32_0
+  return vmlaq_lane_u32(a, b, v, 0);
+  // CHECK: mla {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+}
+
+uint16x4_t test_vmla_laneq_u16_0(uint16x4_t a, uint16x4_t b, uint16x8_t v) {
+  // CHECK-LABEL: test_vmla_laneq_u16_0
+  return vmla_laneq_u16(a, b, v, 0);
+  // CHECK: mla {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[0]
+}
+
+uint16x8_t test_vmlaq_laneq_u16_0(uint16x8_t a, uint16x8_t b, uint16x8_t v) {
+  // CHECK-LABEL: test_vmlaq_laneq_u16_0
+  return vmlaq_laneq_u16(a, b, v, 0);
+  // CHECK: mla {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[0]
+}
+
+uint32x2_t test_vmla_laneq_u32_0(uint32x2_t a, uint32x2_t b, uint32x4_t v) {
+  // CHECK-LABEL: test_vmla_laneq_u32_0
+  return vmla_laneq_u32(a, b, v, 0);
+  // CHECK: mla {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+}
+
+uint32x4_t test_vmlaq_laneq_u32_0(uint32x4_t a, uint32x4_t b, uint32x4_t v) {
+  // CHECK-LABEL: test_vmlaq_laneq_u32_0
+  return vmlaq_laneq_u32(a, b, v, 0);
+  // CHECK: mla {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+}
+
+int32x4_t test_vqdmlal_laneq_s16_0(int32x4_t a, int16x4_t b, int16x8_t v) {
+  // CHECK-LABEL: test_vqdmlal_laneq_s16_0
+  return vqdmlal_laneq_s16(a, b, v, 0);
+  // CHECK: sqdmlal {{v[0-9]+}}.4s, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[0]
+}
+
+int64x2_t test_vqdmlal_laneq_s32_0(int64x2_t a, int32x2_t b, int32x4_t v) {
+  // CHECK-LABEL: test_vqdmlal_laneq_s32_0
+  return vqdmlal_laneq_s32(a, b, v, 0);
+  // CHECK: sqdmlal {{v[0-9]+}}.2d, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+}
+
+int32x4_t test_vqdmlal_high_laneq_s16_0(int32x4_t a, int16x8_t b, int16x8_t v) {
+  // CHECK-LABEL: test_vqdmlal_high_laneq_s16_0
+  return vqdmlal_high_laneq_s16(a, b, v, 0);
+  // CHECK: sqdmlal2 {{v[0-9]+}}.4s, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[0]
+}
+
+int64x2_t test_vqdmlal_high_laneq_s32_0(int64x2_t a, int32x4_t b, int32x4_t v) {
+  // CHECK-LABEL: test_vqdmlal_high_laneq_s32_0
+  return vqdmlal_high_laneq_s32(a, b, v, 0);
+  // CHECK: sqdmlal2 {{v[0-9]+}}.2d, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+}
+
+uint16x4_t test_vmls_lane_u16_0(uint16x4_t a, uint16x4_t b, uint16x4_t v) {
+  // CHECK-LABEL: test_vmls_lane_u16_0
+  return vmls_lane_u16(a, b, v, 0);
+  // CHECK: mls {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[0]
+}
+
+uint16x8_t test_vmlsq_lane_u16_0(uint16x8_t a, uint16x8_t b, uint16x4_t v) {
+  // CHECK-LABEL: test_vmlsq_lane_u16_0
+  return vmlsq_lane_u16(a, b, v, 0);
+  // CHECK: mls {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[0]
+}
+
+uint32x2_t test_vmls_lane_u32_0(uint32x2_t a, uint32x2_t b, uint32x2_t v) {
+  // CHECK-LABEL: test_vmls_lane_u32_0
+  return vmls_lane_u32(a, b, v, 0);
+  // CHECK: mls {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+}
+
+uint32x4_t test_vmlsq_lane_u32_0(uint32x4_t a, uint32x4_t b, uint32x2_t v) {
+  // CHECK-LABEL: test_vmlsq_lane_u32_0
+  return vmlsq_lane_u32(a, b, v, 0);
+  // CHECK: mls {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+}
+
+uint16x4_t test_vmls_laneq_u16_0(uint16x4_t a, uint16x4_t b, uint16x8_t v) {
+  // CHECK-LABEL: test_vmls_laneq_u16_0
+  return vmls_laneq_u16(a, b, v, 0);
+  // CHECK: mls {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[0]
+}
+
+uint16x8_t test_vmlsq_laneq_u16_0(uint16x8_t a, uint16x8_t b, uint16x8_t v) {
+  // CHECK-LABEL: test_vmlsq_laneq_u16_0
+  return vmlsq_laneq_u16(a, b, v, 0);
+  // CHECK: mls {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[0]
+}
+
+uint32x2_t test_vmls_laneq_u32_0(uint32x2_t a, uint32x2_t b, uint32x4_t v) {
+  // CHECK-LABEL: test_vmls_laneq_u32_0
+  return vmls_laneq_u32(a, b, v, 0);
+  // CHECK: mls {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+}
+
+uint32x4_t test_vmlsq_laneq_u32_0(uint32x4_t a, uint32x4_t b, uint32x4_t v) {
+  // CHECK-LABEL: test_vmlsq_laneq_u32_0
+  return vmlsq_laneq_u32(a, b, v, 0);
+  // CHECK: mls {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+}
+
+int32x4_t test_vqdmlsl_laneq_s16_0(int32x4_t a, int16x4_t b, int16x8_t v) {
+  // CHECK-LABEL: test_vqdmlsl_laneq_s16_0
+  return vqdmlsl_laneq_s16(a, b, v, 0);
+  // CHECK: sqdmlsl {{v[0-9]+}}.4s, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[0]
+}
+
+int64x2_t test_vqdmlsl_laneq_s32_0(int64x2_t a, int32x2_t b, int32x4_t v) {
+  // CHECK-LABEL: test_vqdmlsl_laneq_s32_0
+  return vqdmlsl_laneq_s32(a, b, v, 0);
+  // CHECK: sqdmlsl {{v[0-9]+}}.2d, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+}
+
+int32x4_t test_vqdmlsl_high_laneq_s16_0(int32x4_t a, int16x8_t b, int16x8_t v) {
+  // CHECK-LABEL: test_vqdmlsl_high_laneq_s16_0
+  return vqdmlsl_high_laneq_s16(a, b, v, 0);
+  // CHECK: sqdmlsl2 {{v[0-9]+}}.4s, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[0]
+}
+
+int64x2_t test_vqdmlsl_high_laneq_s32_0(int64x2_t a, int32x4_t b, int32x4_t v) {
+  // CHECK-LABEL: test_vqdmlsl_high_laneq_s32_0
+  return vqdmlsl_high_laneq_s32(a, b, v, 0);
+  // CHECK: sqdmlsl2 {{v[0-9]+}}.2d, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+}
+
+int16x4_t test_vqdmulh_laneq_s16_0(int16x4_t a, int16x8_t v) {
+  // CHECK-LABEL: test_vqdmulh_laneq_s16_0
+  return vqdmulh_laneq_s16(a, v, 0);
+  // CHECK: sqdmulh {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[0]
+}
+
+int16x8_t test_vqdmulhq_laneq_s16_0(int16x8_t a, int16x8_t v) {
+  // CHECK-LABEL: test_vqdmulhq_laneq_s16_0
+  return vqdmulhq_laneq_s16(a, v, 0);
+  // CHECK: sqdmulh {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[0]
+}
+
+int32x2_t test_vqdmulh_laneq_s32_0(int32x2_t a, int32x4_t v) {
+  // CHECK-LABEL: test_vqdmulh_laneq_s32_0
+  return vqdmulh_laneq_s32(a, v, 0);
+  // CHECK: sqdmulh {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+}
+
+int32x4_t test_vqdmulhq_laneq_s32_0(int32x4_t a, int32x4_t v) {
+  // CHECK-LABEL: test_vqdmulhq_laneq_s32_0
+  return vqdmulhq_laneq_s32(a, v, 0);
+  // CHECK: sqdmulh {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+}
+
+int16x4_t test_vqrdmulh_laneq_s16_0(int16x4_t a, int16x8_t v) {
+  // CHECK-LABEL: test_vqrdmulh_laneq_s16_0
+  return vqrdmulh_laneq_s16(a, v, 0);
+  // CHECK: sqrdmulh {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[0]
+}
+
+int16x8_t test_vqrdmulhq_laneq_s16_0(int16x8_t a, int16x8_t v) {
+  // CHECK-LABEL: test_vqrdmulhq_laneq_s16_0
+  return vqrdmulhq_laneq_s16(a, v, 0);
+  // CHECK: sqrdmulh {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[0]
+}
+
+int32x2_t test_vqrdmulh_laneq_s32_0(int32x2_t a, int32x4_t v) {
+  // CHECK-LABEL: test_vqrdmulh_laneq_s32_0
+  return vqrdmulh_laneq_s32(a, v, 0);
+  // CHECK: sqrdmulh {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+}
+
+int32x4_t test_vqrdmulhq_laneq_s32_0(int32x4_t a, int32x4_t v) {
+  // CHECK-LABEL: test_vqrdmulhq_laneq_s32_0
+  return vqrdmulhq_laneq_s32(a, v, 0);
+  // CHECK: sqrdmulh {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+}
+
+uint16x4_t test_vmla_lane_u16(uint16x4_t a, uint16x4_t b, uint16x4_t v) {
+  // CHECK-LABEL: test_vmla_lane_u16
+  return vmla_lane_u16(a, b, v, 3);
+  // CHECK: mla {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[3]
+}
+
+uint16x8_t test_vmlaq_lane_u16(uint16x8_t a, uint16x8_t b, uint16x4_t v) {
+  // CHECK-LABEL: test_vmlaq_lane_u16
+  return vmlaq_lane_u16(a, b, v, 3);
+  // CHECK: mla {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[3]
+}
+
+uint32x2_t test_vmla_lane_u32(uint32x2_t a, uint32x2_t b, uint32x2_t v) {
+  // CHECK-LABEL: test_vmla_lane_u32
+  return vmla_lane_u32(a, b, v, 1);
+  // CHECK: mla {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[1]
+}
+
+uint32x4_t test_vmlaq_lane_u32(uint32x4_t a, uint32x4_t b, uint32x2_t v) {
+  // CHECK-LABEL: test_vmlaq_lane_u32
+  return vmlaq_lane_u32(a, b, v, 1);
+  // CHECK: mla {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[1]
+}
+
+uint16x4_t test_vmla_laneq_u16(uint16x4_t a, uint16x4_t b, uint16x8_t v) {
+  // CHECK-LABEL: test_vmla_laneq_u16
+  return vmla_laneq_u16(a, b, v, 7);
+  // CHECK: mla {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[7]
+}
+
+uint16x8_t test_vmlaq_laneq_u16(uint16x8_t a, uint16x8_t b, uint16x8_t v) {
+  // CHECK-LABEL: test_vmlaq_laneq_u16
+  return vmlaq_laneq_u16(a, b, v, 7);
+  // CHECK: mla {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[7]
+}
+
+uint32x2_t test_vmla_laneq_u32(uint32x2_t a, uint32x2_t b, uint32x4_t v) {
+  // CHECK-LABEL: test_vmla_laneq_u32
+  return vmla_laneq_u32(a, b, v, 3);
+  // CHECK: mla {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[3]
+}
+
+uint32x4_t test_vmlaq_laneq_u32(uint32x4_t a, uint32x4_t b, uint32x4_t v) {
+  // CHECK-LABEL: test_vmlaq_laneq_u32
+  return vmlaq_laneq_u32(a, b, v, 3);
+  // CHECK: mla {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[3]
+}
+
+int32x4_t test_vqdmlal_laneq_s16(int32x4_t a, int16x4_t b, int16x8_t v) {
+  // CHECK-LABEL: test_vqdmlal_laneq_s16
+  return vqdmlal_laneq_s16(a, b, v, 7);
+  // CHECK: sqdmlal {{v[0-9]+}}.4s, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[7]
+}
+
+int64x2_t test_vqdmlal_laneq_s32(int64x2_t a, int32x2_t b, int32x4_t v) {
+  // CHECK-LABEL: test_vqdmlal_laneq_s32
+  return vqdmlal_laneq_s32(a, b, v, 3);
+  // CHECK: sqdmlal {{v[0-9]+}}.2d, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[3]
+}
+
+int32x4_t test_vqdmlal_high_laneq_s16(int32x4_t a, int16x8_t b, int16x8_t v) {
+  // CHECK-LABEL: test_vqdmlal_high_laneq_s16
+  return vqdmlal_high_laneq_s16(a, b, v, 7);
+  // CHECK: sqdmlal2 {{v[0-9]+}}.4s, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[7]
+}
+
+int64x2_t test_vqdmlal_high_laneq_s32(int64x2_t a, int32x4_t b, int32x4_t v) {
+  // CHECK-LABEL: test_vqdmlal_high_laneq_s32
+  return vqdmlal_high_laneq_s32(a, b, v, 3);
+  // CHECK: sqdmlal2 {{v[0-9]+}}.2d, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[3]
+}
+
+uint16x4_t test_vmls_lane_u16(uint16x4_t a, uint16x4_t b, uint16x4_t v) {
+  // CHECK-LABEL: test_vmls_lane_u16
+  return vmls_lane_u16(a, b, v, 3);
+  // CHECK: mls {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[3]
+}
+
+uint16x8_t test_vmlsq_lane_u16(uint16x8_t a, uint16x8_t b, uint16x4_t v) {
+  // CHECK-LABEL: test_vmlsq_lane_u16
+  return vmlsq_lane_u16(a, b, v, 3);
+  // CHECK: mls {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[3]
+}
+
+uint32x2_t test_vmls_lane_u32(uint32x2_t a, uint32x2_t b, uint32x2_t v) {
+  // CHECK-LABEL: test_vmls_lane_u32
+  return vmls_lane_u32(a, b, v, 1);
+  // CHECK: mls {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[1]
+}
+
+uint32x4_t test_vmlsq_lane_u32(uint32x4_t a, uint32x4_t b, uint32x2_t v) {
+  // CHECK-LABEL: test_vmlsq_lane_u32
+  return vmlsq_lane_u32(a, b, v, 1);
+  // CHECK: mls {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[1]
+}
+
+uint16x4_t test_vmls_laneq_u16(uint16x4_t a, uint16x4_t b, uint16x8_t v) {
+  // CHECK-LABEL: test_vmls_laneq_u16
+  return vmls_laneq_u16(a, b, v, 7);
+  // CHECK: mls {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[7]
+}
+
+uint16x8_t test_vmlsq_laneq_u16(uint16x8_t a, uint16x8_t b, uint16x8_t v) {
+  // CHECK-LABEL: test_vmlsq_laneq_u16
+  return vmlsq_laneq_u16(a, b, v, 7);
+  // CHECK: mls {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[7]
+}
+
+uint32x2_t test_vmls_laneq_u32(uint32x2_t a, uint32x2_t b, uint32x4_t v) {
+  // CHECK-LABEL: test_vmls_laneq_u32
+  return vmls_laneq_u32(a, b, v, 3);
+  // CHECK: mls {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[3]
+}
+
+uint32x4_t test_vmlsq_laneq_u32(uint32x4_t a, uint32x4_t b, uint32x4_t v) {
+  // CHECK-LABEL: test_vmlsq_laneq_u32
+  return vmlsq_laneq_u32(a, b, v, 3);
+  // CHECK: mls {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[3]
+}
+
+int32x4_t test_vqdmlsl_laneq_s16(int32x4_t a, int16x4_t b, int16x8_t v) {
+  // CHECK-LABEL: test_vqdmlsl_laneq_s16
+  return vqdmlsl_laneq_s16(a, b, v, 7);
+  // CHECK: sqdmlsl {{v[0-9]+}}.4s, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[7]
+}
+
+int64x2_t test_vqdmlsl_laneq_s32(int64x2_t a, int32x2_t b, int32x4_t v) {
+  // CHECK-LABEL: test_vqdmlsl_laneq_s32
+  return vqdmlsl_laneq_s32(a, b, v, 3);
+  // CHECK: sqdmlsl {{v[0-9]+}}.2d, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[3]
+}
+
+int32x4_t test_vqdmlsl_high_laneq_s16(int32x4_t a, int16x8_t b, int16x8_t v) {
+  // CHECK-LABEL: test_vqdmlsl_high_laneq_s16
+  return vqdmlsl_high_laneq_s16(a, b, v, 7);
+  // CHECK: sqdmlsl2 {{v[0-9]+}}.4s, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[7]
+}
+
+int64x2_t test_vqdmlsl_high_laneq_s32(int64x2_t a, int32x4_t b, int32x4_t v) {
+  // CHECK-LABEL: test_vqdmlsl_high_laneq_s32
+  return vqdmlsl_high_laneq_s32(a, b, v, 3);
+  // CHECK: sqdmlsl2 {{v[0-9]+}}.2d, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[3]
+}
+
+int16x4_t test_vqdmulh_laneq_s16(int16x4_t a, int16x8_t v) {
+  // CHECK-LABEL: test_vqdmulh_laneq_s16
+  return vqdmulh_laneq_s16(a, v, 7);
+  // CHECK: sqdmulh {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[7]
+}
+
+int16x8_t test_vqdmulhq_laneq_s16(int16x8_t a, int16x8_t v) {
+  // CHECK-LABEL: test_vqdmulhq_laneq_s16
+  return vqdmulhq_laneq_s16(a, v, 7);
+  // CHECK: sqdmulh {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[7]
+}
+
+int32x2_t test_vqdmulh_laneq_s32(int32x2_t a, int32x4_t v) {
+  // CHECK-LABEL: test_vqdmulh_laneq_s32
+  return vqdmulh_laneq_s32(a, v, 3);
+  // CHECK: sqdmulh {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[3]
+}
+
+int32x4_t test_vqdmulhq_laneq_s32(int32x4_t a, int32x4_t v) {
+  // CHECK-LABEL: test_vqdmulhq_laneq_s32
+  return vqdmulhq_laneq_s32(a, v, 3);
+  // CHECK: sqdmulh {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[3]
+}
+
+int16x4_t test_vqrdmulh_laneq_s16(int16x4_t a, int16x8_t v) {
+  // CHECK-LABEL: test_vqrdmulh_laneq_s16
+  return vqrdmulh_laneq_s16(a, v, 7);
+  // CHECK: sqrdmulh {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, {{v[0-9]+}}.h[7]
+}
+
+int16x8_t test_vqrdmulhq_laneq_s16(int16x8_t a, int16x8_t v) {
+  // CHECK-LABEL: test_vqrdmulhq_laneq_s16
+  return vqrdmulhq_laneq_s16(a, v, 7);
+  // CHECK: sqrdmulh {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.h[7]
+}
+
+int32x2_t test_vqrdmulh_laneq_s32(int32x2_t a, int32x4_t v) {
+  // CHECK-LABEL: test_vqrdmulh_laneq_s32
+  return vqrdmulh_laneq_s32(a, v, 3);
+  // CHECK: sqrdmulh {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[3]
+}
+
+int32x4_t test_vqrdmulhq_laneq_s32(int32x4_t a, int32x4_t v) {
+  // CHECK-LABEL: test_vqrdmulhq_laneq_s32
+  return vqrdmulhq_laneq_s32(a, v, 3);
+  // CHECK: sqrdmulh {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[3]
+}
+

diff --git a/test/CodeGen/aarch64-neon-3v.c b/test/CodeGen/aarch64-neon-3v.c
new file mode 100644 (file)
index 0000000..fc18d06
--- /dev/null
+++ b/test/CodeGen/aarch64-neon-3v.c
@@ -0,0 +1,487 @@
+// REQUIRES: aarch64-registered-target
+// RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +neon \
+// RUN:   -S -O3 -o - %s | FileCheck %s
+
+// Test new aarch64 intrinsics and types
+
+#include <arm_neon.h>
+
+int8x8_t test_vand_s8(int8x8_t a, int8x8_t b) {
+  // CHECK-LABEL: test_vand_s8
+  return vand_s8(a, b);
+  // CHECK: and {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int8x16_t test_vandq_s8(int8x16_t a, int8x16_t b) {
+  // CHECK-LABEL: test_vandq_s8
+  return vandq_s8(a, b);
+  // CHECK: and {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int16x4_t test_vand_s16(int16x4_t a, int16x4_t b) {
+  // CHECK-LABEL: test_vand_s16
+  return vand_s16(a, b);
+  // CHECK: and {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int16x8_t test_vandq_s16(int16x8_t a, int16x8_t b) {
+  // CHECK-LABEL: test_vandq_s16
+  return vandq_s16(a, b);
+  // CHECK: and {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int32x2_t test_vand_s32(int32x2_t a, int32x2_t b) {
+  // CHECK-LABEL: test_vand_s32
+  return vand_s32(a, b);
+  // CHECK: and {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int32x4_t test_vandq_s32(int32x4_t a, int32x4_t b) {
+  // CHECK-LABEL: test_vandq_s32
+  return vandq_s32(a, b);
+  // CHECK: and {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int64x1_t test_vand_s64(int64x1_t a, int64x1_t b) {
+  // CHECK-LABEL: test_vand_s64
+  return vand_s64(a, b);
+  // CHECK: and {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int64x2_t test_vandq_s64(int64x2_t a, int64x2_t b) {
+  // CHECK-LABEL: test_vandq_s64
+  return vandq_s64(a, b);
+  // CHECK: and {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint8x8_t test_vand_u8(uint8x8_t a, uint8x8_t b) {
+  // CHECK-LABEL: test_vand_u8
+  return vand_u8(a, b);
+  // CHECK: and {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint8x16_t test_vandq_u8(uint8x16_t a, uint8x16_t b) {
+  // CHECK-LABEL: test_vandq_u8
+  return vandq_u8(a, b);
+  // CHECK: and {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint16x4_t test_vand_u16(uint16x4_t a, uint16x4_t b) {
+  // CHECK-LABEL: test_vand_u16
+  return vand_u16(a, b);
+  // CHECK: and {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint16x8_t test_vandq_u16(uint16x8_t a, uint16x8_t b) {
+  // CHECK-LABEL: test_vandq_u16
+  return vandq_u16(a, b);
+  // CHECK: and {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint32x2_t test_vand_u32(uint32x2_t a, uint32x2_t b) {
+  // CHECK-LABEL: test_vand_u32
+  return vand_u32(a, b);
+  // CHECK: and {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint32x4_t test_vandq_u32(uint32x4_t a, uint32x4_t b) {
+  // CHECK-LABEL: test_vandq_u32
+  return vandq_u32(a, b);
+  // CHECK: and {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint64x1_t test_vand_u64(uint64x1_t a, uint64x1_t b) {
+  // CHECK-LABEL: test_vand_u64
+  return vand_u64(a, b);
+  // CHECK: and {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint64x2_t test_vandq_u64(uint64x2_t a, uint64x2_t b) {
+  // CHECK-LABEL: test_vandq_u64
+  return vandq_u64(a, b);
+  // CHECK: and {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int8x8_t test_vorr_s8(int8x8_t a, int8x8_t b) {
+  // CHECK-LABEL: test_vorr_s8
+  return vorr_s8(a, b);
+  // CHECK: orr {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int8x16_t test_vorrq_s8(int8x16_t a, int8x16_t b) {
+  // CHECK-LABEL: test_vorrq_s8
+  return vorrq_s8(a, b);
+  // CHECK: orr {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int16x4_t test_vorr_s16(int16x4_t a, int16x4_t b) {
+  // CHECK-LABEL: test_vorr_s16
+  return vorr_s16(a, b);
+  // CHECK: orr {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int16x8_t test_vorrq_s16(int16x8_t a, int16x8_t b) {
+  // CHECK-LABEL: test_vorrq_s16
+  return vorrq_s16(a, b);
+  // CHECK: orr {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int32x2_t test_vorr_s32(int32x2_t a, int32x2_t b) {
+  // CHECK-LABEL: test_vorr_s32
+  return vorr_s32(a, b);
+  // CHECK: orr {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int32x4_t test_vorrq_s32(int32x4_t a, int32x4_t b) {
+  // CHECK-LABEL: test_vorrq_s32
+  return vorrq_s32(a, b);
+  // CHECK: orr {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int64x1_t test_vorr_s64(int64x1_t a, int64x1_t b) {
+  // CHECK-LABEL: test_vorr_s64
+  return vorr_s64(a, b);
+  // CHECK: orr {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int64x2_t test_vorrq_s64(int64x2_t a, int64x2_t b) {
+  // CHECK-LABEL: test_vorrq_s64
+  return vorrq_s64(a, b);
+  // CHECK: orr {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint8x8_t test_vorr_u8(uint8x8_t a, uint8x8_t b) {
+  // CHECK-LABEL: test_vorr_u8
+  return vorr_u8(a, b);
+  // CHECK: orr {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint8x16_t test_vorrq_u8(uint8x16_t a, uint8x16_t b) {
+  // CHECK-LABEL: test_vorrq_u8
+  return vorrq_u8(a, b);
+  // CHECK: orr {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint16x4_t test_vorr_u16(uint16x4_t a, uint16x4_t b) {
+  // CHECK-LABEL: test_vorr_u16
+  return vorr_u16(a, b);
+  // CHECK: orr {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint16x8_t test_vorrq_u16(uint16x8_t a, uint16x8_t b) {
+  // CHECK-LABEL: test_vorrq_u16
+  return vorrq_u16(a, b);
+  // CHECK: orr {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint32x2_t test_vorr_u32(uint32x2_t a, uint32x2_t b) {
+  // CHECK-LABEL: test_vorr_u32
+  return vorr_u32(a, b);
+  // CHECK: orr {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint32x4_t test_vorrq_u32(uint32x4_t a, uint32x4_t b) {
+  // CHECK-LABEL: test_vorrq_u32
+  return vorrq_u32(a, b);
+  // CHECK: orr {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint64x1_t test_vorr_u64(uint64x1_t a, uint64x1_t b) {
+  // CHECK-LABEL: test_vorr_u64
+  return vorr_u64(a, b);
+  // CHECK: orr {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint64x2_t test_vorrq_u64(uint64x2_t a, uint64x2_t b) {
+  // CHECK-LABEL: test_vorrq_u64
+  return vorrq_u64(a, b);
+  // CHECK: orr {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int8x8_t test_veor_s8(int8x8_t a, int8x8_t b) {
+  // CHECK-LABEL: test_veor_s8
+  return veor_s8(a, b);
+  // CHECK: eor {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int8x16_t test_veorq_s8(int8x16_t a, int8x16_t b) {
+  // CHECK-LABEL: test_veorq_s8
+  return veorq_s8(a, b);
+  // CHECK: eor {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int16x4_t test_veor_s16(int16x4_t a, int16x4_t b) {
+  // CHECK-LABEL: test_veor_s16
+  return veor_s16(a, b);
+  // CHECK: eor {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int16x8_t test_veorq_s16(int16x8_t a, int16x8_t b) {
+  // CHECK-LABEL: test_veorq_s16
+  return veorq_s16(a, b);
+  // CHECK: eor {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int32x2_t test_veor_s32(int32x2_t a, int32x2_t b) {
+  // CHECK-LABEL: test_veor_s32
+  return veor_s32(a, b);
+  // CHECK: eor {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int32x4_t test_veorq_s32(int32x4_t a, int32x4_t b) {
+  // CHECK-LABEL: test_veorq_s32
+  return veorq_s32(a, b);
+  // CHECK: eor {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int64x1_t test_veor_s64(int64x1_t a, int64x1_t b) {
+  // CHECK-LABEL: test_veor_s64
+  return veor_s64(a, b);
+  // CHECK: eor {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int64x2_t test_veorq_s64(int64x2_t a, int64x2_t b) {
+  // CHECK-LABEL: test_veorq_s64
+  return veorq_s64(a, b);
+  // CHECK: eor {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint8x8_t test_veor_u8(uint8x8_t a, uint8x8_t b) {
+  // CHECK-LABEL: test_veor_u8
+  return veor_u8(a, b);
+  // CHECK: eor {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint8x16_t test_veorq_u8(uint8x16_t a, uint8x16_t b) {
+  // CHECK-LABEL: test_veorq_u8
+  return veorq_u8(a, b);
+  // CHECK: eor {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint16x4_t test_veor_u16(uint16x4_t a, uint16x4_t b) {
+  // CHECK-LABEL: test_veor_u16
+  return veor_u16(a, b);
+  // CHECK: eor {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint16x8_t test_veorq_u16(uint16x8_t a, uint16x8_t b) {
+  // CHECK-LABEL: test_veorq_u16
+  return veorq_u16(a, b);
+  // CHECK: eor {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint32x2_t test_veor_u32(uint32x2_t a, uint32x2_t b) {
+  // CHECK-LABEL: test_veor_u32
+  return veor_u32(a, b);
+  // CHECK: eor {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint32x4_t test_veorq_u32(uint32x4_t a, uint32x4_t b) {
+  // CHECK-LABEL: test_veorq_u32
+  return veorq_u32(a, b);
+  // CHECK: eor {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint64x1_t test_veor_u64(uint64x1_t a, uint64x1_t b) {
+  // CHECK-LABEL: test_veor_u64
+  return veor_u64(a, b);
+  // CHECK: eor {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint64x2_t test_veorq_u64(uint64x2_t a, uint64x2_t b) {
+  // CHECK-LABEL: test_veorq_u64
+  return veorq_u64(a, b);
+  // CHECK: eor {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int8x8_t test_vbic_s8(int8x8_t a, int8x8_t b) {
+  // CHECK-LABEL: test_vbic_s8
+  return vbic_s8(a, b);
+  // CHECK: bic {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int8x16_t test_vbicq_s8(int8x16_t a, int8x16_t b) {
+  // CHECK-LABEL: test_vbicq_s8
+  return vbicq_s8(a, b);
+  // CHECK: bic {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int16x4_t test_vbic_s16(int16x4_t a, int16x4_t b) {
+  // CHECK-LABEL: test_vbic_s16
+  return vbic_s16(a, b);
+  // CHECK: bic {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int16x8_t test_vbicq_s16(int16x8_t a, int16x8_t b) {
+  // CHECK-LABEL: test_vbicq_s16
+  return vbicq_s16(a, b);
+  // CHECK: bic {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int32x2_t test_vbic_s32(int32x2_t a, int32x2_t b) {
+  // CHECK-LABEL: test_vbic_s32
+  return vbic_s32(a, b);
+  // CHECK: bic {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int32x4_t test_vbicq_s32(int32x4_t a, int32x4_t b) {
+  // CHECK-LABEL: test_vbicq_s32
+  return vbicq_s32(a, b);
+  // CHECK: bic {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int64x1_t test_vbic_s64(int64x1_t a, int64x1_t b) {
+  // CHECK-LABEL: test_vbic_s64
+  return vbic_s64(a, b);
+  // CHECK: bic {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int64x2_t test_vbicq_s64(int64x2_t a, int64x2_t b) {
+  // CHECK-LABEL: test_vbicq_s64
+  return vbicq_s64(a, b);
+  // CHECK: bic {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint8x8_t test_vbic_u8(uint8x8_t a, uint8x8_t b) {
+  // CHECK-LABEL: test_vbic_u8
+  return vbic_u8(a, b);
+  // CHECK: bic {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint8x16_t test_vbicq_u8(uint8x16_t a, uint8x16_t b) {
+  // CHECK-LABEL: test_vbicq_u8
+  return vbicq_u8(a, b);
+  // CHECK: bic {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint16x4_t test_vbic_u16(uint16x4_t a, uint16x4_t b) {
+  // CHECK-LABEL: test_vbic_u16
+  return vbic_u16(a, b);
+  // CHECK: bic {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint16x8_t test_vbicq_u16(uint16x8_t a, uint16x8_t b) {
+  // CHECK-LABEL: test_vbicq_u16
+  return vbicq_u16(a, b);
+  // CHECK: bic {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint32x2_t test_vbic_u32(uint32x2_t a, uint32x2_t b) {
+  // CHECK-LABEL: test_vbic_u32
+  return vbic_u32(a, b);
+  // CHECK: bic {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint32x4_t test_vbicq_u32(uint32x4_t a, uint32x4_t b) {
+  // CHECK-LABEL: test_vbicq_u32
+  return vbicq_u32(a, b);
+  // CHECK: bic {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint64x1_t test_vbic_u64(uint64x1_t a, uint64x1_t b) {
+  // CHECK-LABEL: test_vbic_u64
+  return vbic_u64(a, b);
+  // CHECK: bic {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint64x2_t test_vbicq_u64(uint64x2_t a, uint64x2_t b) {
+  // CHECK-LABEL: test_vbicq_u64
+  return vbicq_u64(a, b);
+  // CHECK: bic {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int8x8_t test_vorn_s8(int8x8_t a, int8x8_t b) {
+  // CHECK-LABEL: test_vorn_s8
+  return vorn_s8(a, b);
+  // CHECK: orn {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int8x16_t test_vornq_s8(int8x16_t a, int8x16_t b) {
+  // CHECK-LABEL: test_vornq_s8
+  return vornq_s8(a, b);
+  // CHECK: orn {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int16x4_t test_vorn_s16(int16x4_t a, int16x4_t b) {
+  // CHECK-LABEL: test_vorn_s16
+  return vorn_s16(a, b);
+  // CHECK: orn {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int16x8_t test_vornq_s16(int16x8_t a, int16x8_t b) {
+  // CHECK-LABEL: test_vornq_s16
+  return vornq_s16(a, b);
+  // CHECK: orn {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int32x2_t test_vorn_s32(int32x2_t a, int32x2_t b) {
+  // CHECK-LABEL: test_vorn_s32
+  return vorn_s32(a, b);
+  // CHECK: orn {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int32x4_t test_vornq_s32(int32x4_t a, int32x4_t b) {
+  // CHECK-LABEL: test_vornq_s32
+  return vornq_s32(a, b);
+  // CHECK: orn {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+int64x1_t test_vorn_s64(int64x1_t a, int64x1_t b) {
+  // CHECK-LABEL: test_vorn_s64
+  return vorn_s64(a, b);
+  // CHECK: orn {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+int64x2_t test_vornq_s64(int64x2_t a, int64x2_t b) {
+  // CHECK-LABEL: test_vornq_s64
+  return vornq_s64(a, b);
+  // CHECK: orn {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint8x8_t test_vorn_u8(uint8x8_t a, uint8x8_t b) {
+  // CHECK-LABEL: test_vorn_u8
+  return vorn_u8(a, b);
+  // CHECK: orn {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint8x16_t test_vornq_u8(uint8x16_t a, uint8x16_t b) {
+  // CHECK-LABEL: test_vornq_u8
+  return vornq_u8(a, b);
+  // CHECK: orn {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint16x4_t test_vorn_u16(uint16x4_t a, uint16x4_t b) {
+  // CHECK-LABEL: test_vorn_u16
+  return vorn_u16(a, b);
+  // CHECK: orn {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint16x8_t test_vornq_u16(uint16x8_t a, uint16x8_t b) {
+  // CHECK-LABEL: test_vornq_u16
+  return vornq_u16(a, b);
+  // CHECK: orn {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint32x2_t test_vorn_u32(uint32x2_t a, uint32x2_t b) {
+  // CHECK-LABEL: test_vorn_u32
+  return vorn_u32(a, b);
+  // CHECK: orn {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint32x4_t test_vornq_u32(uint32x4_t a, uint32x4_t b) {
+  // CHECK-LABEL: test_vornq_u32
+  return vornq_u32(a, b);
+  // CHECK: orn {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint64x1_t test_vorn_u64(uint64x1_t a, uint64x1_t b) {
+  // CHECK-LABEL: test_vorn_u64
+  return vorn_u64(a, b);
+  // CHECK: orn {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint64x2_t test_vornq_u64(uint64x2_t a, uint64x2_t b) {
+  // CHECK-LABEL: test_vornq_u64
+  return vornq_u64(a, b);
+  // CHECK: orn {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}

diff --git a/test/CodeGen/aarch64-neon-fma.c b/test/CodeGen/aarch64-neon-fma.c
new file mode 100644 (file)
index 0000000..b3a54be
--- /dev/null
+++ b/test/CodeGen/aarch64-neon-fma.c
@@ -0,0 +1,173 @@
+// REQUIRES: aarch64-registered-target
+// RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +neon \
+// RUN:   -ffp-contract=fast -S -O3 -o - %s | FileCheck -check-prefix=CHECK-FMA %s
+// RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +neon \
+// RUN:   -S -O3 -o - %s | FileCheck %s
+
+// Test new aarch64 intrinsics and types
+
+#include <arm_neon.h>
+
+float32x2_t test_vmla_n_f32(float32x2_t a, float32x2_t b, float32_t c) {
+  // CHECK-LABEL: test_vmla_n_f32
+  return vmla_n_f32(a, b, c);
+  // CHECK: fmul {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+  // CHECK: fadd {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+  // CHECK-FMA: dup {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+  // CHECK-FMA: fmla {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+float32x4_t test_vmlaq_n_f32(float32x4_t a, float32x4_t b, float32_t c) {
+  // CHECK-LABEL: test_vmlaq_n_f32
+  return vmlaq_n_f32(a, b, c);
+  // CHECK: fmul {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+  // CHECK: fadd {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+  // CHECK-FMA: dup {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+  // CHECK-FMA: fmla {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+}
+
+float32x4_t test_vmlsq_n_f32(float32x4_t a, float32x4_t b, float32_t c) {
+  // CHECK-LABEL: test_vmlsq_n_f32
+  return vmlsq_n_f32(a, b, c);
+  // CHECK: fmul {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+  // CHECK: fsub {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+  // CHECK-FMA: dup {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+  // CHECK-FMA: fmls {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+}
+
+float32x2_t test_vmls_n_f32(float32x2_t a, float32x2_t b, float32_t c) {
+  // CHECK-LABEL: test_vmls_n_f32
+  return vmls_n_f32(a, b, c);
+  // CHECK: fmul {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+  // CHECK: fsub {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+  // CHECK-FMA: dup {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+  // CHECK-FMA: fmls {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+float32x2_t test_vmla_lane_f32_0(float32x2_t a, float32x2_t b, float32x2_t v) {
+  // CHECK-LABEL: test_vmla_lane_f32_0
+  return vmla_lane_f32(a, b, v, 0);
+  // CHECK: fmul {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+  // CHECK: fadd {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+  // CHECK-FMA: fmla {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+}
+
+float32x4_t test_vmlaq_lane_f32_0(float32x4_t a, float32x4_t b, float32x2_t v) {
+  // CHECK-LABEL: test_vmlaq_lane_f32_0
+  return vmlaq_lane_f32(a, b, v, 0);
+  // CHECK: fmul {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+  // CHECK: fadd {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+  // CHECK-FMA: fmla {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+}
+
+float32x2_t test_vmla_laneq_f32_0(float32x2_t a, float32x2_t b, float32x4_t v) {
+  // CHECK-LABEL: test_vmla_laneq_f32_0
+  return vmla_laneq_f32(a, b, v, 0);
+  // CHECK: fmul {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+  // CHECK: fadd {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+  // CHECK-FMA: fmla {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+}
+
+float32x4_t test_vmlaq_laneq_f32_0(float32x4_t a, float32x4_t b, float32x4_t v) {
+  // CHECK-LABEL: test_vmlaq_laneq_f32_0
+  return vmlaq_laneq_f32(a, b, v, 0);
+  // CHECK: fmul {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+  // CHECK: fadd {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+  // CHECK-FMA: fmla {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+}
+
+float32x2_t test_vmls_lane_f32_0(float32x2_t a, float32x2_t b, float32x2_t v) {
+  // CHECK-LABEL: test_vmls_lane_f32_0
+  return vmls_lane_f32(a, b, v, 0);
+  // CHECK: fmul {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+  // CHECK: fsub {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+  // CHECK-FMA: fmls {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+}
+
+float32x4_t test_vmlsq_lane_f32_0(float32x4_t a, float32x4_t b, float32x2_t v) {
+  // CHECK-LABEL: test_vmlsq_lane_f32_0
+  return vmlsq_lane_f32(a, b, v, 0);
+  // CHECK: fmul {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+  // CHECK: fsub {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+  // CHECK-FMA: fmls {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+}
+
+float32x2_t test_vmls_laneq_f32_0(float32x2_t a, float32x2_t b, float32x4_t v) {
+  // CHECK-LABEL: test_vmls_laneq_f32_0
+  return vmls_laneq_f32(a, b, v, 0);
+  // CHECK: fmul {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+  // CHECK: fsub {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+  // CHECK-FMA: fmls {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[0]
+}
+
+float32x4_t test_vmlsq_laneq_f32_0(float32x4_t a, float32x4_t b, float32x4_t v) {
+  // CHECK-LABEL: test_vmlsq_laneq_f32_0
+  return vmlsq_laneq_f32(a, b, v, 0);
+  // CHECK: fmul {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+  // CHECK: fsub {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+  // CHECK-FMA: fmls {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[0]
+}
+
+float32x2_t test_vmla_lane_f32(float32x2_t a, float32x2_t b, float32x2_t v) {
+  // CHECK-LABEL: test_vmla_lane_f32
+  return vmla_lane_f32(a, b, v, 1);
+  // CHECK: fmul {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[1]
+  // CHECK: fadd {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+  // CHECK-FMA: fmla {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[1]
+}
+
+float32x4_t test_vmlaq_lane_f32(float32x4_t a, float32x4_t b, float32x2_t v) {
+  // CHECK-LABEL: test_vmlaq_lane_f32
+  return vmlaq_lane_f32(a, b, v, 1);
+  // CHECK: fmul {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[1]
+  // CHECK: fadd {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+  // CHECK-FMA: fmla {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[1]
+}
+
+float32x2_t test_vmla_laneq_f32(float32x2_t a, float32x2_t b, float32x4_t v) {
+  // CHECK-LABEL: test_vmla_laneq_f32
+  return vmla_laneq_f32(a, b, v, 3);
+  // CHECK: fmul {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[3]
+  // CHECK: fadd {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+  // CHECK-FMA: fmla {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[3]
+}
+
+float32x4_t test_vmlaq_laneq_f32(float32x4_t a, float32x4_t b, float32x4_t v) {
+  // CHECK-LABEL: test_vmlaq_laneq_f32
+  return vmlaq_laneq_f32(a, b, v, 3);
+  // CHECK: fmul {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[3]
+  // CHECK: fadd {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+  // CHECK-FMA: fmla {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[3]
+}
+
+float32x2_t test_vmls_lane_f32(float32x2_t a, float32x2_t b, float32x2_t v) {
+  // CHECK-LABEL: test_vmls_lane_f32
+  return vmls_lane_f32(a, b, v, 1);
+  // CHECK: fmul {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[1]
+  // CHECK: fsub {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+  // CHECK-FMA: fmls {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[1]
+}
+
+float32x4_t test_vmlsq_lane_f32(float32x4_t a, float32x4_t b, float32x2_t v) {
+  // CHECK-LABEL: test_vmlsq_lane_f32
+  return vmlsq_lane_f32(a, b, v, 1);
+  // CHECK: fmul {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[1]
+  // CHECK: fsub {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+  // CHECK-FMA: fmls {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[1]
+}
+float32x2_t test_vmls_laneq_f32(float32x2_t a, float32x2_t b, float32x4_t v) {
+  // CHECK-LABEL: test_vmls_laneq_f32
+  return vmls_laneq_f32(a, b, v, 3);
+  // CHECK: fmul {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[3]
+  // CHECK: fsub {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+  // CHECK-FMA: fmls {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, {{v[0-9]+}}.s[3]
+}
+
+float32x4_t test_vmlsq_laneq_f32(float32x4_t a, float32x4_t b, float32x4_t v) {
+  // CHECK-LABEL: test_vmlsq_laneq_f32
+  return vmlsq_laneq_f32(a, b, v, 3);
+  // CHECK: fmul {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[3]
+  // CHECK: fsub {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+  // CHECK-FMA: fmls {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.s[3]
+}
+

diff --git a/test/CodeGen/aarch64-neon-intrinsics.c b/test/CodeGen/aarch64-neon-intrinsics.c
index c0bca67aaa4da552c34a4000c18ac24fa6019561..35dd06ecc76f93bff2098554e26bd5867ab3edd8 100644 (file)
--- a/test/CodeGen/aarch64-neon-intrinsics.c
+++ b/test/CodeGen/aarch64-neon-intrinsics.c
@@ -1119,6 +1119,17 @@ uint16x8_t test_vtstq_p16(poly16x8_t v1, poly16x8_t v2) {
   // CHECK: cmtst {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, {{v[0-9]+}}.8h
 }
 
+uint64x1_t test_vtst_s64(int64x1_t a, int64x1_t b) {
+  // CHECK-LABEL: test_vtst_s64
+  return vtst_s64(a, b);
+  // CHECK: cmtst {{d[0-9]+}}, {{d[0-9]+}}, {{d[0-9]+}}
+}
+
+uint64x1_t test_vtst_u64(uint64x1_t a, uint64x1_t b) {
+  // CHECK-LABEL: test_vtst_u64
+  return vtst_u64(a, b);
+  // CHECK: cmtst {{d[0-9]+}}, {{d[0-9]+}}, {{d[0-9]+}}
+}
 
 uint8x8_t test_vceq_s8(int8x8_t v1, int8x8_t v2) {
   // CHECK: test_vceq_s8
@@ -2539,6 +2550,18 @@ uint64x2_t test_vqrshlq_u64(uint64x2_t a, int64x2_t b) {
 // CHECK: uqrshl {{v[0-9]+}}.2d, {{v[0-9]+}}.2d, {{v[0-9]+}}.2d
 }
 
+poly64x1_t test_vsli_n_p64(poly64x1_t a, poly64x1_t b) {
+// CHECK-LABEL: test_vsli_n_p64
+  return vsli_n_p64(a, b, 0); 
+// CHECK: sli {{d[0-9]+}}, {{d[0-9]+}}, #0
+}
+
+poly64x2_t test_vsliq_n_p64(poly64x2_t a, poly64x2_t b) {
+// CHECK-LABEL: test_vsliq_n_p64
+  return vsliq_n_p64(a, b, 0); 
+// CHECK: sli {{v[0-9]+}}.2d, {{v[0-9]+}}.2d, #0
+}
+
 int8x8_t test_vmax_s8(int8x8_t a, int8x8_t b) {
 // CHECK: test_vmax_s8
   return vmax_s8(a, b);
@@ -5771,6 +5794,18 @@ float64_t test_vrecpxd_f64(float64_t a) {
   return vrecpxd_f64(a);
 }
 
+uint32x2_t test_vrsqrte_u32(uint32x2_t a) {
+// CHECK-LABEL: test_vrsqrte_u32
+// CHECK: ursqrte {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+  return vrsqrte_u32(a);
+}
+
+uint32x4_t test_vrsqrteq_u32(uint32x4_t a) {
+// CHECK-LABEL: test_vrsqrteq_u32
+// CHECK: ursqrte {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+  return vrsqrteq_u32(a);
+}
+
 float32_t test_vrsqrtes_f32(float32_t a) {
 // CHECK: vrsqrtes_f32
 // CHECK: frsqrte {{s[0-9]+}}, {{s[0-9]+}}
@@ -8872,6 +8907,90 @@ int64_t test_vqshld_n_s64(int64_t a) {
   return (int64_t)vqshld_n_s64(a, 63);
 }
 
+int8x8_t test_vqshl_n_s8(int8x8_t a) {
+  // CHECK-LABEL: test_vqshl_n_s8
+  return vqshl_n_s8(a, 0);
+  // CHECK: sqshl {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, #0
+}
+
+int8x16_t test_vqshlq_n_s8(int8x16_t a) {
+  // CHECK-LABEL: test_vqshlq_n_s8
+  return vqshlq_n_s8(a, 0);
+  // CHECK: sqshl {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, #0
+}
+
+int16x4_t test_vqshl_n_s16(int16x4_t a) {
+  // CHECK-LABEL: test_vqshl_n_s16
+  return vqshl_n_s16(a, 0);
+  // CHECK: sqshl {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, #0
+}
+
+int16x8_t test_vqshlq_n_s16(int16x8_t a) {
+  // CHECK-LABEL: test_vqshlq_n_s16
+  return vqshlq_n_s16(a, 0);
+  // CHECK: sqshl {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, #0
+}
+
+int32x2_t test_vqshl_n_s32(int32x2_t a) {
+  // CHECK-LABEL: test_vqshl_n_s32
+  return vqshl_n_s32(a, 0);
+  // CHECK: sqshl {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, #0
+}
+
+int32x4_t test_vqshlq_n_s32(int32x4_t a) {
+  // CHECK-LABEL: test_vqshlq_n_s32
+  return vqshlq_n_s32(a, 0);
+  // CHECK: sqshl {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, #0
+}
+
+int64x2_t test_vqshlq_n_s64(int64x2_t a) {
+  // CHECK-LABEL: test_vqshlq_n_s64
+  return vqshlq_n_s64(a, 0);
+  // CHECK: sqshl {{v[0-9]+}}.2d, {{v[0-9]+}}.2d, #0
+}
+
+uint8x8_t test_vqshl_n_u8(uint8x8_t a) {
+  // CHECK-LABEL: test_vqshl_n_u8
+  return vqshl_n_u8(a, 0);
+  // CHECK: uqshl {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, #0
+}
+
+uint8x16_t test_vqshlq_n_u8(uint8x16_t a) {
+  // CHECK-LABEL: test_vqshlq_n_u8
+  return vqshlq_n_u8(a, 0);
+  // CHECK: uqshl {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, #0
+}
+
+uint16x4_t test_vqshl_n_u16(uint16x4_t a) {
+  // CHECK-LABEL: test_vqshl_n_u16
+  return vqshl_n_u16(a, 0);
+  // CHECK: uqshl {{v[0-9]+}}.4h, {{v[0-9]+}}.4h, #0
+}
+
+uint16x8_t test_vqshlq_n_u16(uint16x8_t a) {
+  // CHECK-LABEL: test_vqshlq_n_u16
+  return vqshlq_n_u16(a, 0);
+  // CHECK: uqshl {{v[0-9]+}}.8h, {{v[0-9]+}}.8h, #0
+}
+
+uint32x2_t test_vqshl_n_u32(uint32x2_t a) {
+  // CHECK-LABEL: test_vqshl_n_u32
+  return vqshl_n_u32(a, 0);
+  // CHECK: uqshl {{v[0-9]+}}.2s, {{v[0-9]+}}.2s, #0
+}
+
+uint32x4_t test_vqshlq_n_u32(uint32x4_t a) {
+  // CHECK-LABEL: test_vqshlq_n_u32
+  return vqshlq_n_u32(a, 0);
+  // CHECK: uqshl {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, #0
+}
+
+uint64x2_t test_vqshlq_n_u64(uint64x2_t a) {
+  // CHECK-LABEL: test_vqshlq_n_u64
+  return vqshlq_n_u64(a, 0);
+  // CHECK: uqshl {{v[0-9]+}}.2d, {{v[0-9]+}}.2d, #0
+}
+
 int64x1_t test_vqshl_n_s64(int64x1_t a) {
 // CHECK: test_vqshl_n_s64
 // CHECK: sqshl d{{[0-9]+}}, d{{[0-9]+}}, #1
@@ -11350,6 +11469,48 @@ uint64x1_t test_vsqadd_u64(uint64x1_t a, int64x1_t b) {
   // CHECK: usqadd d{{[0-9]+}}, d{{[0-9]+}}
 }
 
+uint8x8_t test_vsqadd_u8(uint8x8_t a, int8x8_t b) {
+  // CHECK-LABEL: test_vsqadd_u8
+  return vsqadd_u8(a, b);
+  // CHECK: usqadd {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
+}
+
+uint8x16_t test_vsqaddq_u8(uint8x16_t a, int8x16_t b) {
+  // CHECK-LABEL: test_vsqaddq_u8
+  return vsqaddq_u8(a, b);
+  // CHECK: usqadd {{v[0-9]+}}.16b, {{v[0-9]+}}.16b
+}
+
+uint16x4_t test_vsqadd_u16(uint16x4_t a, int16x4_t b) {
+  // CHECK-LABEL: test_vsqadd_u16
+  return vsqadd_u16(a, b);
+  // CHECK: usqadd {{v[0-9]+}}.4h, {{v[0-9]+}}.4h
+}
+
+uint16x8_t test_vsqaddq_u16(uint16x8_t a, int16x8_t b) {
+  // CHECK-LABEL: test_vsqaddq_u16
+  return vsqaddq_u16(a, b);
+  // CHECK: usqadd {{v[0-9]+}}.8h, {{v[0-9]+}}.8h
+}
+
+uint32x2_t test_vsqadd_u32(uint32x2_t a, int32x2_t b) {
+  // CHECK-LABEL: test_vsqadd_u32
+  return vsqadd_u32(a, b);
+  // CHECK: usqadd {{v[0-9]+}}.2s, {{v[0-9]+}}.2s
+}
+
+uint32x4_t test_vsqaddq_u32(uint32x4_t a, int32x4_t b) {
+  // CHECK-LABEL: test_vsqaddq_u32
+  return vsqaddq_u32(a, b);
+  // CHECK: usqadd {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+}
+
+uint64x2_t test_vsqaddq_u64(uint64x2_t a, int64x2_t b) {
+  // CHECK-LABEL: test_vsqaddq_u64
+  return vsqaddq_u64(a, b);
+  // CHECK: usqadd {{v[0-9]+}}.2d, {{v[0-9]+}}.2d
+}
+
 int64x1_t test_vabs_s64(int64x1_t a) {
   // CHECK-LABEL: test_vabs_s64
   return vabs_s64(a);

diff --git a/test/CodeGen/aarch64-neon-ldst-one.c b/test/CodeGen/aarch64-neon-ldst-one.c
index ec634e4f50e91ac211f5efa1567bfc3b5713781c..737810ee18d83d3312c710c79512e50cc9d786b6 100644 (file)
--- a/test/CodeGen/aarch64-neon-ldst-one.c
+++ b/test/CodeGen/aarch64-neon-ldst-one.c
@@ -900,6 +900,36 @@ poly64x1_t test_vld1_lane_p64(poly64_t  *a, poly64x1_t b) {
   // CHECK: ld1r {v{{[0-9]+}}.1d}, [{{x[0-9]+|sp}}]
 }
 
+int8x16x2_t test_vld2q_lane_s8(int8_t const * ptr, int8x16x2_t src) {
+  // CHECK-LABEL: test_vld2q_lane_s8
+  return vld2q_lane_s8(ptr, src, 15); 
+  // CHECK: ld2 {{{v[0-9]+}}.b, {{v[0-9]+}}.b}[15], [x0]
+}
+
+uint8x16x2_t test_vld2q_lane_u8(uint8_t const * ptr, uint8x16x2_t src) {
+  // CHECK-LABEL: test_vld2q_lane_u8
+  return vld2q_lane_u8(ptr, src, 15); 
+  // CHECK: ld2 {{{v[0-9]+}}.b, {{v[0-9]+}}.b}[15], [x0]
+}
+
+poly8x16x2_t test_vld2q_lane_p8(poly8_t const * ptr, poly8x16x2_t src) {
+  // CHECK-LABEL: test_vld2q_lane_p8
+  return vld2q_lane_p8(ptr, src, 15); 
+  // CHECK: ld2 {{{v[0-9]+}}.b, {{v[0-9]+}}.b}[15], [x0]
+}
+
+int8x16x3_t test_vld3q_lane_s8(int8_t const * ptr, int8x16x3_t src) {
+  // CHECK-LABEL: test_vld3q_lane_s8
+  return vld3q_lane_s8(ptr, src, 15);
+  // CHECK: ld3 {{{v[0-9]+}}.b, {{v[0-9]+}}.b, {{v[0-9]+}}.b}[15], [x0]
+}
+
+uint8x16x3_t test_vld3q_lane_u8(uint8_t const * ptr, uint8x16x3_t src) {
+  // CHECK-LABEL: test_vld3q_lane_u8
+  return vld3q_lane_u8(ptr, src, 15);
+  // CHECK: ld3 {{{v[0-9]+}}.b, {{v[0-9]+}}.b, {{v[0-9]+}}.b}[15], [x0]
+}
+
 uint16x8x2_t test_vld2q_lane_u16(uint16_t  *a, uint16x8x2_t b) {
   // CHECK-LABEL: test_vld2q_lane_u16
   return vld2q_lane_u16(a, b, 7);