r216474 - [test/CodeGen/ARM] Update arm_neon_intrinsics test case to actually test the

Quentin Colombet qcolombet at apple.com
Tue Aug 26 11:43:32 PDT 2014


Author: qcolombet
Date: Tue Aug 26 13:43:31 2014
New Revision: 216474

URL: http://llvm.org/viewvc/llvm-project?rev=216474&view=rev
Log:
[test/CodeGen/ARM] Update arm_neon_intrinsics test case to actually test the
lowering of the intrinsics.
Prior to this commit, most of the copy-related intrinsics could be optimized
away. The situation is still not ideal as there are several possibilities to
lower a given intrinsic. Currently, we match LLVM behavior.

Modified:
    cfe/trunk/test/CodeGen/arm_neon_intrinsics.c

Modified: cfe/trunk/test/CodeGen/arm_neon_intrinsics.c
URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/test/CodeGen/arm_neon_intrinsics.c?rev=216474&r1=216473&r2=216474&view=diff
==============================================================================
--- cfe/trunk/test/CodeGen/arm_neon_intrinsics.c (original)
+++ cfe/trunk/test/CodeGen/arm_neon_intrinsics.c Tue Aug 26 13:43:31 2014
@@ -1474,79 +1474,109 @@ poly8x16_t test_vcntq_p8(poly8x16_t a) {
 
 
 // CHECK-LABEL: test_vcombine_s8
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
 int8x16_t test_vcombine_s8(int8x8_t a, int8x8_t b) {
   return vcombine_s8(a, b);
 }
 
 // CHECK-LABEL: test_vcombine_s16
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
 int16x8_t test_vcombine_s16(int16x4_t a, int16x4_t b) {
   return vcombine_s16(a, b);
 }
 
 // CHECK-LABEL: test_vcombine_s32
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
 int32x4_t test_vcombine_s32(int32x2_t a, int32x2_t b) {
   return vcombine_s32(a, b);
 }
 
 // CHECK-LABEL: test_vcombine_s64
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
 int64x2_t test_vcombine_s64(int64x1_t a, int64x1_t b) {
   return vcombine_s64(a, b);
 }
 
 // CHECK-LABEL: test_vcombine_f16
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
 float16x8_t test_vcombine_f16(float16x4_t a, float16x4_t b) {
   return vcombine_f16(a, b);
 }
 
 // CHECK-LABEL: test_vcombine_f32
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
 float32x4_t test_vcombine_f32(float32x2_t a, float32x2_t b) {
   return vcombine_f32(a, b);
 }
 
 // CHECK-LABEL: test_vcombine_u8
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
 uint8x16_t test_vcombine_u8(uint8x8_t a, uint8x8_t b) {
   return vcombine_u8(a, b);
 }
 
 // CHECK-LABEL: test_vcombine_u16
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
 uint16x8_t test_vcombine_u16(uint16x4_t a, uint16x4_t b) {
   return vcombine_u16(a, b);
 }
 
 // CHECK-LABEL: test_vcombine_u32
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
 uint32x4_t test_vcombine_u32(uint32x2_t a, uint32x2_t b) {
   return vcombine_u32(a, b);
 }
 
 // CHECK-LABEL: test_vcombine_u64
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
 uint64x2_t test_vcombine_u64(uint64x1_t a, uint64x1_t b) {
   return vcombine_u64(a, b);
 }
 
 // CHECK-LABEL: test_vcombine_p8
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
 poly8x16_t test_vcombine_p8(poly8x8_t a, poly8x8_t b) {
   return vcombine_p8(a, b);
 }
 
 // CHECK-LABEL: test_vcombine_p16
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
+// CHECK: vmov d{{[0-9]+}}, r{{[0-9]+}}, r{{[0-9]+}}
 poly16x8_t test_vcombine_p16(poly16x4_t a, poly16x4_t b) {
   return vcombine_p16(a, b);
 }
 
 
 // CHECK-LABEL: test_vcreate_s8
+// CHECK: vmov [[REG:d[0-9]+]], r0, r1
+// CHECK: vclz.i8 d{{[0-9]+}}, [[REG]]
 int8x8_t test_vcreate_s8(uint64_t a) {
-  return vcreate_s8(a);
+  return vclz_s8(vcreate_s8(a));
 }
 
 // CHECK-LABEL: test_vcreate_s16
+// CHECK: vmov [[REG:d[0-9]+]], r0, r1
+// CHECK: vclz.i16 d{{[0-9]+}}, [[REG]]
 int16x4_t test_vcreate_s16(uint64_t a) {
-  return vcreate_s16(a);
+  return vclz_s16(vcreate_s16(a));
 }
 
 // CHECK-LABEL: test_vcreate_s32
+// CHECK: vmov [[REG:d[0-9]+]], r0, r1
+// CHECK: vclz.i32 d{{[0-9]+}}, [[REG]]
 int32x2_t test_vcreate_s32(uint64_t a) {
-  return vcreate_s32(a);
+  return vclz_s32(vcreate_s32(a));
 }
 
 // CHECK-LABEL: test_vcreate_f16
@@ -1560,38 +1590,59 @@ float32x2_t test_vcreate_f32(uint64_t a)
 }
 
 // CHECK-LABEL: test_vcreate_u8
+// CHECK: vmov [[REG:d[0-9]+]], r0, r1
+// CHECK: vclz.i8 d{{[0-9]+}}, [[REG]]
 uint8x8_t test_vcreate_u8(uint64_t a) {
-  return vcreate_u8(a);
+  return vclz_s8(vcreate_u8(a));
 }
 
 // CHECK-LABEL: test_vcreate_u16
+// CHECK: vmov [[REG:d[0-9]+]], r0, r1
+// CHECK: vclz.i16 d{{[0-9]+}}, [[REG]]
 uint16x4_t test_vcreate_u16(uint64_t a) {
-  return vcreate_u16(a);
+  return vclz_s16(vcreate_u16(a));
 }
 
 // CHECK-LABEL: test_vcreate_u32
+// CHECK: vmov [[REG:d[0-9]+]], r0, r1
+// CHECK: vclz.i32 d{{[0-9]+}}, [[REG]]
 uint32x2_t test_vcreate_u32(uint64_t a) {
-  return vcreate_u32(a);
+  return vclz_s32(vcreate_u32(a));
 }
 
+
+// We have two ways of lowering that.  Either with one 'vmov d, r, r' or
+// with two 'vmov d[],r'.  LLVM does the latter. We may want to be less
+// strict about the matching pattern if it starts causing problem.
 // CHECK-LABEL: test_vcreate_u64
+// CHECK: vmov.32 [[REG:d[0-9]+]][0], r0
+// CHECK: vmov.32 [[REG]][1], r1
 uint64x1_t test_vcreate_u64(uint64_t a) {
-  return vcreate_u64(a);
+  uint64x1_t tmp = vcreate_u64(a);
+  return vadd_u64(tmp, tmp);
+
 }
 
 // CHECK-LABEL: test_vcreate_p8
+// CHECK: vmov [[REG:d[0-9]+]], r0, r1
+// CHECK: vcnt.8 d{{[0-9]+}}, [[REG]]
 poly8x8_t test_vcreate_p8(uint64_t a) {
-  return vcreate_p8(a);
+  return vcnt_p8(vcreate_p8(a));
 }
 
 // CHECK-LABEL: test_vcreate_p16
+// CHECK: vmov [[REG:d[0-9]+]], r0, r1
 poly16x4_t test_vcreate_p16(uint64_t a) {
-  return vcreate_p16(a);
+  poly16x4_t tmp = vcreate_p16(a);
+  return vbsl_p16(tmp, tmp, tmp);
 }
 
 // CHECK-LABEL: test_vcreate_s64
+// CHECK: vmov.32 [[REG:d[0-9]+]][0], r0
+// CHECK: vmov.32 [[REG]][1], r1
 int64x1_t test_vcreate_s64(uint64_t a) {
-  return vcreate_s64(a);
+  int64x1_t tmp = vcreate_s64(a);
+  return vadd_s64(tmp, tmp);
 }
 
 
@@ -1963,27 +2014,32 @@ float32x4_t test_vdupq_n_f32(float32_t a
 }
 
 // CHECK-LABEL: test_vdup_n_s64
-// The optimizer is able to get rid of all moves now.
+// CHECK: vmov
 int64x1_t test_vdup_n_s64(int64_t a) {
-  return vdup_n_s64(a);
+  int64x1_t tmp = vdup_n_s64(a);
+  return vadd_s64(tmp, tmp);
 }
 
 // CHECK-LABEL: test_vdup_n_u64
-// The optimizer is able to get rid of all moves now.
+// CHECK: vmov
 uint64x1_t test_vdup_n_u64(uint64_t a) {
-  return vdup_n_u64(a);
+  int64x1_t tmp = vdup_n_u64(a);
+  return vadd_s64(tmp, tmp);
+
 }
 
 // CHECK-LABEL: test_vdupq_n_s64
-// The optimizer is able to get rid of all moves now.
+// CHECK: vmov
 int64x2_t test_vdupq_n_s64(int64_t a) {
-  return vdupq_n_s64(a);
+  int64x2_t tmp = vdupq_n_s64(a);
+  return vaddq_s64(tmp, tmp);
 }
 
 // CHECK-LABEL: test_vdupq_n_u64
-// The optimizer is able to get rid of all moves now.
+// CHECK: vmov
 uint64x2_t test_vdupq_n_u64(uint64_t a) {
-  return vdupq_n_u64(a);
+  int64x2_t tmp = vdupq_n_u64(a);
+  return vaddq_u64(tmp, tmp);
 }
 
 
@@ -4969,15 +5025,19 @@ float32x4_t test_vmovq_n_f32(float32_t a
 }
 
 // CHECK-LABEL: test_vmov_n_s64
-// The optimizer is able to remove all moves now.
+// CHECK: vmov.32 [[REG:d[0-9]+]][0], r0
+// CHECK: vmov.32 [[REG]][1], r1
 int64x1_t test_vmov_n_s64(int64_t a) {
-  return vmov_n_s64(a);
+  int64x1_t tmp = vmov_n_s64(a);
+  return vadd_s64(tmp, tmp);
 }
 
 // CHECK-LABEL: test_vmov_n_u64
-// The optimizer is able to remove all moves now.
+// CHECK: vmov.32 [[REG:d[0-9]+]][0], r0
+// CHECK: vmov.32 [[REG]][1], r1
 uint64x1_t test_vmov_n_u64(uint64_t a) {
-  return vmov_n_u64(a);
+  uint64x1_t tmp = vmov_n_u64(a);
+  return vadd_u64(tmp, tmp);
 }
 
 // CHECK-LABEL: test_vmovq_n_s64





More information about the cfe-commits mailing list