1 ; We run nvvm-reflect (and then optimize) this module twice, once with metadata 2 ; that enables FTZ, and again with metadata that disables it. 3 4 ; RUN: cat %s > %t.noftz 5 ; RUN: echo '!0 = !{i32 4, !"nvvm-reflect-ftz", i32 0}' >> %t.noftz 6 ; RUN: opt %t.noftz -S -nvvm-reflect -O2 \ 7 ; RUN: | FileCheck %s --check-prefix=USE_FTZ_0 --check-prefix=CHECK 8 9 ; RUN: cat %s > %t.ftz 10 ; RUN: echo '!0 = !{i32 4, !"nvvm-reflect-ftz", i32 1}' >> %t.ftz 11 ; RUN: opt %t.ftz -S -nvvm-reflect -O2 \ 12 ; RUN: | FileCheck %s --check-prefix=USE_FTZ_1 --check-prefix=CHECK 13 14 @str = private unnamed_addr addrspace(4) constant [11 x i8] c"__CUDA_FTZ\00" 15 16 declare i32 @__nvvm_reflect(i8*) 17 declare i8* @llvm.nvvm.ptr.constant.to.gen.p0i8.p4i8(i8 addrspace(4)*) 18 19 ; CHECK-LABEL: @foo 20 define float @foo(float %a, float %b) { 21 ; CHECK-NOT: call i32 @__nvvm_reflect 22 %ptr = tail call i8* @llvm.nvvm.ptr.constant.to.gen.p0i8.p4i8(i8 addrspace(4)* getelementptr inbounds ([11 x i8], [11 x i8] addrspace(4)* @str, i32 0, i32 0)) 23 %reflect = tail call i32 @__nvvm_reflect(i8* %ptr) 24 %cmp = icmp ugt i32 %reflect, 0 25 br i1 %cmp, label %use_mul, label %use_add 26 27 use_mul: 28 ; USE_FTZ_1: fmul float %a, %b 29 ; USE_FTZ_0-NOT: fadd float %a, %b 30 %ret1 = fmul float %a, %b 31 br label %exit 32 33 use_add: 34 ; USE_FTZ_0: fadd float %a, %b 35 ; USE_FTZ_1-NOT: fmul float %a, %b 36 %ret2 = fadd float %a, %b 37 br label %exit 38 39 exit: 40 %ret = phi float [%ret1, %use_mul], [%ret2, %use_add] 41 ret float %ret 42 } 43 44 declare i32 @llvm.nvvm.reflect.p0i8(i8*) 45 46 ; CHECK-LABEL: define i32 @intrinsic 47 define i32 @intrinsic() { 48 ; CHECK-NOT: call i32 @llvm.nvvm.reflect 49 ; USE_FTZ_0: ret i32 0 50 ; USE_FTZ_1: ret i32 1 51 %ptr = tail call i8* @llvm.nvvm.ptr.constant.to.gen.p0i8.p4i8(i8 addrspace(4)* getelementptr inbounds ([11 x i8], [11 x i8] addrspace(4)* @str, i32 0, i32 0)) 52 %reflect = tail call i32 @llvm.nvvm.reflect.p0i8(i8* %ptr) 53 ret i32 %reflect 54 } 55 56 ; CUDA-7.0 passes __nvvm_reflect argument slightly differently. 57 ; Verify that it works, too 58 59 @"$str" = private addrspace(1) constant [11 x i8] c"__CUDA_FTZ\00" 60 61 ; CHECK-LABEL: @bar 62 define float @bar(float %a, float %b) { 63 ; CHECK-NOT: call i32 @__nvvm_reflect 64 %reflect = call i32 @__nvvm_reflect(i8* addrspacecast (i8 addrspace(1)* getelementptr inbounds ([11 x i8], [11 x i8] addrspace(1)* @"$str", i32 0, i32 0) to i8*)) 65 %cmp = icmp ne i32 %reflect, 0 66 br i1 %cmp, label %use_mul, label %use_add 67 68 use_mul: 69 ; USE_FTZ_1: fmul float %a, %b 70 ; USE_FTZ_0-NOT: fadd float %a, %b 71 %ret1 = fmul float %a, %b 72 br label %exit 73 74 use_add: 75 ; USE_FTZ_0: fadd float %a, %b 76 ; USE_FTZ_1-NOT: fmul float %a, %b 77 %ret2 = fadd float %a, %b 78 br label %exit 79 80 exit: 81 %ret = phi float [%ret1, %use_mul], [%ret2, %use_add] 82 ret float %ret 83 } 84 85 !llvm.module.flags = !{!0} 86 ; A module flag is added to the end of this file by the RUN lines at the top. 87
