clang/test/CodeGenCXX/align-avx-complete-objects.cpp

   1 // RUN: %clang_cc1 %std_cxx11-14 %s -O0 -triple=x86_64-apple-darwin -target-feature +avx2 -fmax-type-align=16 -emit-llvm -o - -Werror | FileCheck %s --check-prefixes=CHECK,PRE17
   2 // RUN: %clang_cc1 %std_cxx17- %s -O0 -triple=x86_64-apple-darwin -target-feature +avx2 -fmax-type-align=16 -emit-llvm -o - -Werror | FileCheck %s --check-prefixes=CHECK,CXX17
   3
   4 typedef float AVX2Float __attribute__((__vector_size__(32)));
   5
   6
   7 #if __cplusplus < 202002L
   8 volatile
   9 #endif
  10 float TestAlign(void)
  11 {
  12        volatile AVX2Float *p = new AVX2Float;
  13         *p = *p;
  14         AVX2Float r = *p;
  15         return r[0];
  16 }
  17
  18 // CHECK: [[R:%.*]] = alloca <8 x float>, align 32
  19 // PRE17-NEXT:  [[CALL:%.*]] = call noalias noundef nonnull ptr @_Znwm(i64 noundef 32)
  20 // CXX17-NEXT:  [[CALL:%.*]] = call noalias noundef nonnull align 32 ptr @_ZnwmSt11align_val_t(i64 noundef 32, i64 noundef 32)
  21 // CHECK-NEXT:  store ptr [[CALL]], ptr [[P:%.*]], align 8
  22 // CHECK-NEXT:  [[ONE:%.*]] = load ptr, ptr [[P]], align 8
  23 // CHECK-NEXT:  [[TWO:%.*]] = load volatile <8 x float>, ptr [[ONE]], align 16
  24 // CHECK-NEXT:  [[THREE:%.*]] = load ptr, ptr [[P]], align 8
  25 // CHECK-NEXT:  store volatile <8 x float> [[TWO]], ptr [[THREE]], align 16
  26 // CHECK-NEXT:  [[FOUR:%.*]] = load ptr, ptr [[P]], align 8
  27 // CHECK-NEXT:  [[FIVE:%.*]] = load volatile <8 x float>, ptr [[FOUR]], align 16
  28 // CHECK-NEXT:  store <8 x float> [[FIVE]], ptr [[R]], align 32
  29 // CHECK-NEXT:  [[SIX:%.*]] = load <8 x float>, ptr [[R]], align 32
  30 // CHECK-NEXT:  [[VECEXT:%.*]] = extractelement <8 x float> [[SIX]], i32 0
  31 // CHECK-NEXT:  ret float [[VECEXT]]
  32
  33 typedef float AVX2Float_Explicitly_aligned __attribute__((__vector_size__(32))) __attribute__((aligned (32)));
  34
  35 typedef AVX2Float_Explicitly_aligned AVX2Float_indirect;
  36
  37 typedef AVX2Float_indirect AVX2Float_use_existing_align;
  38
  39 #if __cplusplus < 202002L
  40 volatile
  41 #endif
  42 float TestAlign2(void)
  43 {
  44        volatile AVX2Float_use_existing_align *p = new AVX2Float_use_existing_align;
  45         *p = *p;
  46         AVX2Float_use_existing_align r = *p;
  47         return r[0];
  48 }
  49
  50 // CHECK: [[R:%.*]] = alloca <8 x float>, align 32
  51 // PRE17-NEXT:  [[CALL:%.*]] = call noalias noundef nonnull ptr @_Znwm(i64 noundef 32)
  52 // CXX17-NEXT:  [[CALL:%.*]] = call noalias noundef nonnull align 32 ptr @_ZnwmSt11align_val_t(i64 noundef 32, i64 noundef 32)
  53 // CHECK-NEXT:  store ptr [[CALL]], ptr [[P:%.*]], align 8
  54 // CHECK-NEXT:  [[ONE:%.*]] = load ptr, ptr [[P]], align 8
  55 // CHECK-NEXT:  [[TWO:%.*]] = load volatile <8 x float>, ptr [[ONE]], align 32
  56 // CHECK-NEXT:  [[THREE:%.*]] = load ptr, ptr [[P]], align 8
  57 // CHECK-NEXT:  store volatile <8 x float> [[TWO]], ptr [[THREE]], align 32
  58 // CHECK-NEXT:  [[FOUR:%.*]] = load ptr, ptr [[P]], align 8
  59 // CHECK-NEXT:  [[FIVE:%.*]] = load volatile <8 x float>, ptr [[FOUR]], align 32
  60 // CHECK-NEXT:  store <8 x float> [[FIVE]], ptr [[R]], align 32
  61 // CHECK-NEXT:  [[SIX:%.*]] = load <8 x float>, ptr [[R]], align 32
  62 // CHECK-NEXT:  [[VECEXT:%.*]] = extractelement <8 x float> [[SIX]], i32 0
  63 // CHECK-NEXT:  ret float [[VECEXT]]