1 // RUN: %clang_cc1 -triple mips-unknown-linux -o - -O1 -emit-llvm %s | FileCheck %s -check-prefix=ALL -check-prefix=O32 2 // RUN: %clang_cc1 -triple mipsel-unknown-linux -o - -O1 -emit-llvm %s | FileCheck %s -check-prefix=ALL -check-prefix=O32 3 // RUN: %clang_cc1 -triple mips64-unknown-linux -o - -O1 -emit-llvm -target-abi n32 %s | FileCheck %s -check-prefix=ALL -check-prefix=N32 -check-prefix=NEW 4 // RUN: %clang_cc1 -triple mips64-unknown-linux -o - -O1 -emit-llvm -target-abi n32 %s | FileCheck %s -check-prefix=ALL -check-prefix=N32 -check-prefix=NEW 5 // RUN: %clang_cc1 -triple mips64-unknown-linux -o - -O1 -emit-llvm %s | FileCheck %s -check-prefix=ALL -check-prefix=N64 -check-prefix=NEW 6 // RUN: %clang_cc1 -triple mips64el-unknown-linux -o - -O1 -emit-llvm %s | FileCheck %s -check-prefix=ALL -check-prefix=N64 -check-prefix=NEW 7 8 #include <stdarg.h> 9 10 typedef int v4i32 __attribute__ ((__vector_size__ (16))); 11 12 int test_i32(char *fmt, ...) { 13 va_list va; 14 15 va_start(va, fmt); 16 int v = va_arg(va, int); 17 va_end(va); 18 19 return v; 20 } 21 22 // ALL-LABEL: define i32 @test_i32(i8*{{.*}} %fmt, ...) 23 // 24 // O32: %va = alloca i8*, align [[PTRALIGN:4]] 25 // N32: %va = alloca i8*, align [[PTRALIGN:4]] 26 // N64: %va = alloca i8*, align [[PTRALIGN:8]] 27 // 28 // ALL: [[VA1:%.+]] = bitcast i8** %va to i8* 29 // ALL: call void @llvm.va_start(i8* [[VA1]]) 30 // 31 // O32: [[TMP0:%.+]] = bitcast i8** %va to i32** 32 // O32: [[AP_CUR:%.+]] = load i32** [[TMP0]], align [[PTRALIGN]] 33 // NEW: [[TMP0:%.+]] = bitcast i8** %va to i64** 34 // NEW: [[AP_CUR:%.+]] = load i64** [[TMP0]], align [[PTRALIGN]] 35 // 36 // O32: [[TMP1:%.+]] = getelementptr i32* [[AP_CUR]], i32 1 37 // O32: [[AP_NEXT:%.+]] = bitcast i32* [[TMP1]] to i8* 38 // N32: [[TMP1:%.+]] = getelementptr i64* [[AP_CUR]], {{i32|i64}} 1 39 // N64: [[TMP1:%.+]] = getelementptr i64* [[AP_CUR]], {{i32|i64}} 1 40 // NEW: [[AP_NEXT:%.+]] = bitcast i64* [[TMP1]] to i8* 41 // 42 // ALL: store i8* [[AP_NEXT]], i8** %va, align [[PTRALIGN]] 43 // 44 // O32: [[ARG1:%.+]] = load i32* [[AP_CUR]], align 4 45 // NEW: [[TMP2:%.+]] = load i64* [[AP_CUR]], align 8 46 // NEW: [[ARG1:%.+]] = trunc i64 [[TMP2]] to i32 47 // 48 // ALL: call void @llvm.va_end(i8* [[VA1]]) 49 // ALL: ret i32 [[ARG1]] 50 // ALL: } 51 52 int test_i32_2args(char *fmt, ...) { 53 va_list va; 54 55 va_start(va, fmt); 56 int v1 = va_arg(va, int); 57 int v2 = va_arg(va, int); 58 va_end(va); 59 60 return v1 + v2; 61 } 62 63 // ALL-LABEL: define i32 @test_i32_2args(i8*{{.*}} %fmt, ...) 64 // 65 // ALL: %va = alloca i8*, align [[PTRALIGN]] 66 // ALL: [[VA1:%.+]] = bitcast i8** %va to i8* 67 // ALL: call void @llvm.va_start(i8* [[VA1]]) 68 // 69 // O32: [[TMP0:%.+]] = bitcast i8** %va to i32** 70 // O32: [[AP_CUR:%.+]] = load i32** [[TMP0]], align [[PTRALIGN]] 71 // NEW: [[TMP0:%.+]] = bitcast i8** %va to i64** 72 // NEW: [[AP_CUR:%.+]] = load i64** [[TMP0]], align [[PTRALIGN]] 73 // 74 // O32: [[TMP1:%.+]] = getelementptr i32* [[AP_CUR]], i32 1 75 // O32: [[AP_NEXT:%.+]] = bitcast i32* [[TMP1]] to i8* 76 // N32: [[TMP1:%.+]] = getelementptr i64* [[AP_CUR]], [[INTPTR_T:i32]] 1 77 // FIXME: N32 optimised this bitcast out. Why only for this ABI? 78 // N64: [[TMP1:%.+]] = getelementptr i64* [[AP_CUR]], [[INTPTR_T:i64]] 1 79 // N64: [[AP_NEXT:%.+]] = bitcast i64* [[TMP1]] to i8* 80 // 81 // O32: store i8* [[AP_NEXT]], i8** %va, align [[PTRALIGN]] 82 // FIXME: N32 optimised this store out. Why only for this ABI? 83 // N64: store i8* [[AP_NEXT]], i8** %va, align [[PTRALIGN]] 84 // 85 // O32: [[ARG1:%.+]] = load i32* [[AP_CUR]], align 4 86 // NEW: [[TMP3:%.+]] = load i64* [[AP_CUR]], align 8 87 // NEW: [[ARG1:%.+]] = trunc i64 [[TMP3]] to i32 88 // 89 // O32: [[TMP2:%.+]] = getelementptr i32* [[AP_CUR]], i32 2 90 // O32: [[AP_NEXT:%.+]] = bitcast i32* [[TMP2]] to i8* 91 // NEW: [[TMP2:%.+]] = getelementptr i64* [[AP_CUR]], [[INTPTR_T]] 2 92 // NEW: [[AP_NEXT:%.+]] = bitcast i64* [[TMP2]] to i8* 93 // 94 // ALL: store i8* [[AP_NEXT]], i8** %va, align [[PTRALIGN]] 95 // 96 // O32: [[ARG2:%.+]] = load i32* [[TMP1]], align 4 97 // NEW: [[TMP4:%.+]] = load i64* [[TMP1]], align 8 98 // NEW: [[ARG2:%.+]] = trunc i64 [[TMP4]] to i32 99 // 100 // ALL: call void @llvm.va_end(i8* [[VA1]]) 101 // ALL: [[ADD:%.+]] = add nsw i32 [[ARG2]], [[ARG1]] 102 // ALL: ret i32 [[ADD]] 103 // ALL: } 104 105 long long test_i64(char *fmt, ...) { 106 va_list va; 107 108 va_start(va, fmt); 109 long long v = va_arg(va, long long); 110 va_end(va); 111 112 return v; 113 } 114 115 // ALL-LABEL: define i64 @test_i64(i8*{{.*}} %fmt, ...) 116 // 117 // ALL: %va = alloca i8*, align [[PTRALIGN]] 118 // ALL: [[VA1:%.+]] = bitcast i8** %va to i8* 119 // ALL: call void @llvm.va_start(i8* [[VA1]]) 120 // 121 // O32: [[AP_CUR:%.+]] = load i8** %va, align [[PTRALIGN]] 122 // NEW: [[TMP0:%.+]] = bitcast i8** %va to i64** 123 // NEW: [[AP_CUR:%.+]] = load i64** [[TMP0]], align [[PTRALIGN]] 124 // 125 // i64 is 8-byte aligned, while this is within O32's stack alignment there's no 126 // guarantee that the offset is still 8-byte aligned after earlier reads. 127 // O32: [[PTR0:%.+]] = ptrtoint i8* [[AP_CUR]] to [[INTPTR_T:i32]] 128 // O32: [[PTR1:%.+]] = add i32 [[PTR0]], 7 129 // O32: [[PTR2:%.+]] = and i32 [[PTR1]], -8 130 // O32: [[PTR3:%.+]] = inttoptr [[INTPTR_T]] [[PTR2]] to i64* 131 // O32: [[PTR4:%.+]] = inttoptr [[INTPTR_T]] [[PTR2]] to i8* 132 // 133 // O32: [[AP_NEXT:%.+]] = getelementptr i8* [[PTR4]], [[INTPTR_T]] 8 134 // N32: [[TMP1:%.+]] = getelementptr i64* [[AP_CUR]], i32 1 135 // N64: [[TMP1:%.+]] = getelementptr i64* [[AP_CUR]], i64 1 136 // NEW: [[AP_NEXT:%.+]] = bitcast i64* [[TMP1]] to i8* 137 // 138 // ALL: store i8* [[AP_NEXT]], i8** %va, align [[PTRALIGN]] 139 // 140 // O32: [[ARG1:%.+]] = load i64* [[PTR3]], align 8 141 // NEW: [[ARG1:%.+]] = load i64* [[AP_CUR]], align 8 142 // 143 // ALL: call void @llvm.va_end(i8* [[VA1]]) 144 // ALL: ret i64 [[ARG1]] 145 // ALL: } 146 147 int test_v4i32(char *fmt, ...) { 148 va_list va; 149 150 va_start(va, fmt); 151 v4i32 v = va_arg(va, v4i32); 152 va_end(va); 153 154 return v[0]; 155 } 156 157 // ALL-LABEL: define i32 @test_v4i32(i8*{{.*}} %fmt, ...) 158 // 159 // ALL: %va = alloca i8*, align [[PTRALIGN]] 160 // ALL: [[VA1:%.+]] = bitcast i8** %va to i8* 161 // ALL: call void @llvm.va_start(i8* [[VA1]]) 162 // ALL: [[AP_CUR:%.+]] = load i8** %va, align [[PTRALIGN]] 163 // 164 // O32: [[PTR0:%.+]] = ptrtoint i8* [[AP_CUR]] to [[INTPTR_T:i32]] 165 // N32: [[PTR0:%.+]] = ptrtoint i8* [[AP_CUR]] to [[INTPTR_T:i32]] 166 // N64: [[PTR0:%.+]] = ptrtoint i8* [[AP_CUR]] to [[INTPTR_T:i64]] 167 // 168 // Vectors are 16-byte aligned, however the O32 ABI has a maximum alignment of 169 // 8-bytes since the base of the stack is 8-byte aligned. 170 // O32: [[PTR1:%.+]] = add i32 [[PTR0]], 7 171 // O32: [[PTR2:%.+]] = and i32 [[PTR1]], -8 172 // 173 // NEW: [[PTR1:%.+]] = add [[INTPTR_T]] [[PTR0]], 15 174 // NEW: [[PTR2:%.+]] = and [[INTPTR_T]] [[PTR1]], -16 175 // 176 // ALL: [[PTR3:%.+]] = inttoptr [[INTPTR_T]] [[PTR2]] to <4 x i32>* 177 // ALL: [[PTR4:%.+]] = inttoptr [[INTPTR_T]] [[PTR2]] to i8* 178 // ALL: [[AP_NEXT:%.+]] = getelementptr i8* [[PTR4]], [[INTPTR_T]] 16 179 // ALL: store i8* [[AP_NEXT]], i8** %va, align [[PTRALIGN]] 180 // ALL: [[PTR5:%.+]] = load <4 x i32>* [[PTR3]], align 16 181 // ALL: call void @llvm.va_end(i8* [[VA1]]) 182 // ALL: [[VECEXT:%.+]] = extractelement <4 x i32> [[PTR5]], i32 0 183 // ALL: ret i32 [[VECEXT]] 184 // ALL: } 185