Home | History | Annotate | Download | only in CodeGen 
      1 // RUN: %clang_cc1 -triple aarch64-linux-gnu -target-feature +neon -target-abi aapcs -ffreestanding -fallow-half-arguments-and-returns -emit-llvm -w -o - %s | FileCheck %s
      2 
      3 // AAPCS clause C.8 says: If the argument has an alignment of 16 then the NGRN
      4 // is rounded up to the next even number.
      5 
      6 // CHECK: void @test1(i32 %x0, i128 %x2_x3, i128 %x4_x5, i128 %x6_x7, i128 %sp.coerce)
      7 typedef union { __int128 a; } Small;
      8 void test1(int x0, __int128 x2_x3, __int128 x4_x5, __int128 x6_x7, Small sp) {
      9 }
     10 
     11 
     12 // CHECK: void @test2(i32 %x0, i128 %x2_x3.coerce, i32 %x4, i128 %x6_x7.coerce, i32 %sp, i128 %sp16.coerce)
     13 void test2(int x0, Small x2_x3, int x4, Small x6_x7, int sp, Small sp16) {
     14 }
     15 
     16 // We coerce HFAs into a contiguous [N x double] type if they're going on the
     17 // stack in order to avoid holes. Make sure we get all of them, and not just the
     18 // first:
     19 
     20 // CHECK: void @test3([4 x float] %s0_s3.coerce, float %s4, [4 x float] %sp.coerce, [4 x float] %sp16.coerce)
     21 typedef struct { float arr[4]; } HFA;
     22 void test3(HFA s0_s3, float s4, HFA sp, HFA sp16) {
     23 }
     24 
     25 
     26 // However, we shouldn't perform the [N x double] coercion on types which have
     27 // sufficient alignment to avoid holes on their own. We could coerce to [N x
     28 // fp128] or something, but leaving them as-is retains more information for
     29 // users to debug.
     30 
     31 //  CHECK: void @test4([3 x <16 x i8>] %v0_v2.coerce, [3 x <16 x i8>] %v3_v5.coerce, [3 x <16 x i8>] %sp.coerce, double %sp48, [3 x <16 x i8>] %sp64.coerce)
     32 typedef __attribute__((neon_vector_type(16))) signed char int8x16_t;
     33 typedef struct { int8x16_t arr[3]; } BigHFA;
     34 void test4(BigHFA v0_v2, BigHFA v3_v5, BigHFA sp, double sp48, BigHFA sp64) {
     35 }
     36 
     37 // It's the job of the argument *consumer* to perform the required sign & zero
     38 // extensions under AAPCS. There shouldn't be
     39 
     40 // CHECK: define i8 @test5(i8 %a, i16 %b)
     41 unsigned char test5(unsigned char a, signed short b) {
     42 }
     43 
     44 // __fp16 can be used as a function argument or return type (ACLE 2.0)
     45 // CHECK: define half @test_half(half %{{.*}})
     46 __fp16 test_half(__fp16 A) { }
     47 
     48 // __fp16 is a base type for homogeneous floating-point aggregates for AArch64 (but not 32-bit ARM).
     49 // CHECK: define %struct.HFA_half @test_half_hfa([4 x half] %{{.*}})
     50 struct HFA_half { __fp16 a[4]; };
     51 struct HFA_half test_half_hfa(struct HFA_half A) { }
     52