clang/test/CodeGen/arm-neon-misc.c

   1 // REQUIRES: arm-registered-target
   2 // RUN: %clang_cc1 -triple thumbv7-apple-darwin \
   3 // RUN:   -target-abi apcs-gnu \
   4 // RUN:   -target-cpu cortex-a8 \
   5 // RUN:   -mfloat-abi soft \
   6 // RUN:   -target-feature +soft-float-abi \
   7 // RUN:   -ffreestanding \
   8 // RUN:   -emit-llvm -w -o - %s | FileCheck %s
   9
  10 #include <arm_neon.h>
  11
  12 // Radar 11998303: Avoid using i64 types for vld1q_lane and vst1q_lane Neon
  13 // intrinsics with <2 x i64> vectors to avoid poor code for i64 in the backend.
  14 void t1(uint64_t *src, uint8_t *dst) {
  15 // CHECK: @t1
  16   uint64x2_t q = vld1q_u64(src);
  17 // CHECK: call <2 x i64> @llvm.arm.neon.vld1.v2i64.p0
  18   vst1q_lane_u64(dst, q, 1);
  19 // CHECK: shufflevector <2 x i64>
  20 // CHECK: call void @llvm.arm.neon.vst1.p0.v1i64
  21 }
  22
  23 void t2(uint64_t *src1, uint8_t *src2, uint64x2_t *dst) {
  24 // CHECK: @t2
  25     uint64x2_t q = vld1q_u64(src1);
  26 // CHECK: call <2 x i64> @llvm.arm.neon.vld1.v2i64.p0
  27     q = vld1q_lane_u64(src2, q, 0);
  28 // CHECK: shufflevector <2 x i64>
  29 // CHECK: call <1 x i64> @llvm.arm.neon.vld1.v1i64.p0
  30 // CHECK: shufflevector <1 x i64>
  31     *dst = q;
  32 // CHECK: store <2 x i64>
  33 }