1 // Test with fast math
2 // RUN: %clang_cc1 -triple i386-pc-linux-gnu -emit-llvm -DFAST \
3 // RUN: -mreassociate \
4 // RUN: -o - %s | FileCheck --check-prefixes CHECK,CHECKFAST,CHECKNP %s
5 //
6 // RUN: %clang_cc1 -triple aarch64-unknown-linux-gnu -emit-llvm -DFAST \
7 // RUN: -mreassociate \
8 // RUN: -o - %s | FileCheck --check-prefixes CHECK,CHECKFAST,CHECKNP %s
9 //
10 // Test with fast math and fprotect-parens
11 // RUN: %clang_cc1 -triple i386-pc-linux-gnu -emit-llvm -DFAST \
12 // RUN: -mreassociate -fprotect-parens -ffp-contract=on\
13 // RUN: -o - %s | FileCheck --check-prefixes CHECK,CHECKFAST,CHECKPP %s
14 //
15 // RUN: %clang_cc1 -triple aarch64-unknown-linux-gnu -emit-llvm -DFAST \
16 // RUN: -mreassociate -fprotect-parens -ffp-contract=on\
17 // RUN: -o - %s | FileCheck --check-prefixes CHECK,CHECKFAST,CHECKPP %s
18 //
19 // Test without fast math: llvm intrinsic not created
20 // RUN: %clang_cc1 -triple i386-pc-linux-gnu -emit-llvm -fprotect-parens\
21 // RUN: -o - %s | FileCheck --implicit-check-not="llvm.arithmetic.fence" %s
22 //
23 // RUN: %clang_cc1 -triple aarch64-unknown-linux-gnu -emit-llvm -fprotect-parens\
24 // RUN: -o - %s | FileCheck --implicit-check-not="llvm.arithmetic.fence" %s
25 //
26 // Test with fast math on spir target
27 // RUN: %clang_cc1 -triple spir64 -emit-llvm -DFAST \
28 // RUN: -mreassociate -o - %s \
29 // RUN: | FileCheck --check-prefixes CHECK,CHECKFAST,CHECKNP %s
30 //
31
32 int v;
addit(float a,float b)33 int addit(float a, float b) {
34 // CHECK: define {{.*}}@addit(float noundef %a, float noundef %b) #0 {
35 _Complex double cd, cd1;
36 cd = __arithmetic_fence(cd1);
37 // CHECKFAST: call{{.*}} double @llvm.arithmetic.fence.f64({{.*}}real)
38 // CHECKFAST: call{{.*}} double @llvm.arithmetic.fence.f64({{.*}}imag)
39 // Vector should be supported.
40 typedef float __v2f32 __attribute__((__vector_size__(8)));
41 __v2f32 vec1, vec2;
42 vec1 = __arithmetic_fence(vec2);
43 // CHECKFAST: call{{.*}} <2 x float> @llvm.arithmetic.fence.v2f32
44 vec2 = (vec2 + vec1);
45 // CHECKPP: call{{.*}} <2 x float> @llvm.arithmetic.fence.v2f32
46
47 v = __arithmetic_fence(a + b);
48 // CHECKFAST: call{{.*}} float @llvm.arithmetic.fence.f32(float %add{{.*}})
49
50 v = (a + b);
51 // CHECKPP: call{{.*}} float @llvm.arithmetic.fence.f32(float %add{{.*}})
52 v = a + (b*b);
53 // CHECKPP: fmul reassoc
54 // CHECKPP-NEXT: call{{.*}} float @llvm.arithmetic.fence.f32(float %mul)
55 // CHECKNP: fmul
56 // CHECKNP: fadd
57 v = b + a*a;
58 // CHECKPP: call{{.*}} float @llvm.fmuladd.f32
59 // CHECKNP: fmul
60 // CHECKNP: fadd
61 v = b + __arithmetic_fence(a*a); // Fence blocks recognition of FMA
62 // CHECKPP: fmul
63 // CHECKNP: fmul
64
65 b = (a);
66 (a) = b;
67 // CHECK-NEXT fptosi
68 // CHECK-NEXT store i32
69 // CHECK-NEXT load float
70 // CHECK-NEXT store float
71 // CHECK-NEXT load float
72 // CHECK-NEXT store float
73 return 0;
74 // CHECK-NEXT ret i32 0
75 }
addit1(int a,int b)76 int addit1(int a, int b) {
77 // CHECK: define {{.*}}@addit1(i32 noundef %a, i32 noundef %b{{.*}}
78 v = (a + b);
79 // CHECK-NOT: call{{.*}} float @llvm.arithmetic.fence.int(float noundef %add)
80 return 0;
81 }
82 #ifdef FAST
83 #pragma float_control(precise, on)
subit(float a,float b,float * fp)84 int subit(float a, float b, float *fp) {
85 // CHECKFAST: define {{.*}}@subit(float noundef %a, float noundef %b{{.*}}
86 *fp = __arithmetic_fence(a - b);
87 *fp = (a + b);
88 // CHECK-NOT: call{{.*}} float @llvm.arithmetic.fence.f32(float noundef %add)
89 return 0;
90 }
91 #endif
92