1; RUN: llc < %s -O2 -mtriple=x86_64-linux-android -mattr=+mmx | FileCheck %s 2; RUN: llc < %s -O2 -mtriple=x86_64-linux-gnu -mattr=+mmx | FileCheck %s 3 4; These tests were generated from simplified libm C code. 5; When compiled for the x86_64-linux-android target, 6; long double is mapped to f128 type that should be passed 7; in SSE registers. When the f128 type calling convention 8; problem was fixed, old llvm code failed to handle f128 values 9; in several f128/i128 type operations. These unit tests hopefully 10; will catch regression in any future change in this area. 11; To modified or enhance these test cases, please consult libm 12; code pattern and compile with -target x86_64-linux-android 13; to generate IL. The __float128 keyword if not accepted by 14; clang, just define it to "long double". 15; 16 17; typedef long double __float128; 18; union IEEEl2bits { 19; __float128 e; 20; struct { 21; unsigned long manl :64; 22; unsigned long manh :48; 23; unsigned int exp :15; 24; unsigned int sign :1; 25; } bits; 26; struct { 27; unsigned long manl :64; 28; unsigned long manh :48; 29; unsigned int expsign :16; 30; } xbits; 31; }; 32 33; C code: 34; void foo(__float128 x); 35; void TestUnionLD1(__float128 s, unsigned long n) { 36; union IEEEl2bits u; 37; __float128 w; 38; u.e = s; 39; u.bits.manh = n; 40; w = u.e; 41; foo(w); 42; } 43define void @TestUnionLD1(fp128 %s, i64 %n) #0 { 44entry: 45 %0 = bitcast fp128 %s to i128 46 %1 = zext i64 %n to i128 47 %bf.value = shl nuw i128 %1, 64 48 %bf.shl = and i128 %bf.value, 5192296858534809181786422619668480 49 %bf.clear = and i128 %0, -5192296858534809181786422619668481 50 %bf.set = or i128 %bf.shl, %bf.clear 51 %2 = bitcast i128 %bf.set to fp128 52 tail call void @foo(fp128 %2) #2 53 ret void 54; CHECK-LABEL: TestUnionLD1: 55; CHECK: movaps %xmm0, -24(%rsp) 56; CHECK-NEXT: movq -24(%rsp), %rax 57; CHECK-NEXT: movabsq $281474976710655, %rcx 58; CHECK-NEXT: andq %rdi, %rcx 59; CHECK-NEXT: movabsq $-281474976710656, %rdx 60; CHECK-NEXT: andq -16(%rsp), %rdx 61; CHECK-NEXT: movq %rax, -40(%rsp) 62; CHECK-NEXT: orq %rcx, %rdx 63; CHECK-NEXT: movq %rdx, -32(%rsp) 64; CHECK-NEXT: movaps -40(%rsp), %xmm0 65; CHECK-NEXT: jmp foo 66} 67 68; C code: 69; __float128 TestUnionLD2(__float128 s) { 70; union IEEEl2bits u; 71; __float128 w; 72; u.e = s; 73; u.bits.manl = 0; 74; w = u.e; 75; return w; 76; } 77define fp128 @TestUnionLD2(fp128 %s) #0 { 78entry: 79 %0 = bitcast fp128 %s to i128 80 %bf.clear = and i128 %0, -18446744073709551616 81 %1 = bitcast i128 %bf.clear to fp128 82 ret fp128 %1 83; CHECK-LABEL: TestUnionLD2: 84; CHECK: movaps %xmm0, -24(%rsp) 85; CHECK-NEXT: movq -16(%rsp), %rax 86; CHECK-NEXT: movq %rax, -32(%rsp) 87; CHECK-NEXT: movq $0, -40(%rsp) 88; CHECK-NEXT: movaps -40(%rsp), %xmm0 89; CHECK-NEXT: retq 90} 91 92; C code: 93; __float128 TestI128_1(__float128 x) 94; { 95; union IEEEl2bits z; 96; z.e = x; 97; z.bits.sign = 0; 98; return (z.e < 0.1L) ? 1.0L : 2.0L; 99; } 100define fp128 @TestI128_1(fp128 %x) #0 { 101entry: 102 %0 = bitcast fp128 %x to i128 103 %bf.clear = and i128 %0, 170141183460469231731687303715884105727 104 %1 = bitcast i128 %bf.clear to fp128 105 %cmp = fcmp olt fp128 %1, 0xL999999999999999A3FFB999999999999 106 %cond = select i1 %cmp, fp128 0xL00000000000000003FFF000000000000, fp128 0xL00000000000000004000000000000000 107 ret fp128 %cond 108; CHECK-LABEL: TestI128_1: 109; CHECK: movaps %xmm0, 110; CHECK: movabsq $9223372036854775807, 111; CHECK: callq __lttf2 112; CHECK: testl %eax, %eax 113; CHECK: movaps {{.*}}, %xmm0 114; CHECK: retq 115} 116 117; C code: 118; __float128 TestI128_2(__float128 x, __float128 y) 119; { 120; unsigned short hx; 121; union IEEEl2bits ge_u; 122; ge_u.e = x; 123; hx = ge_u.xbits.expsign; 124; return (hx & 0x8000) == 0 ? x : y; 125; } 126define fp128 @TestI128_2(fp128 %x, fp128 %y) #0 { 127entry: 128 %0 = bitcast fp128 %x to i128 129 %cmp = icmp sgt i128 %0, -1 130 %cond = select i1 %cmp, fp128 %x, fp128 %y 131 ret fp128 %cond 132; CHECK-LABEL: TestI128_2: 133; CHECK: movaps %xmm0, -24(%rsp) 134; CHECK-NEXT: cmpq $0, -16(%rsp) 135; CHECK-NEXT: jns 136; CHECK: movaps %xmm1, %xmm0 137; CHECK: retq 138} 139 140; C code: 141; __float128 TestI128_3(__float128 x, int *ex) 142; { 143; union IEEEl2bits u; 144; u.e = x; 145; if (u.bits.exp == 0) { 146; u.e *= 0x1.0p514; 147; u.bits.exp = 0x3ffe; 148; } 149; return (u.e); 150; } 151define fp128 @TestI128_3(fp128 %x, i32* nocapture readnone %ex) #0 { 152entry: 153 %0 = bitcast fp128 %x to i128 154 %bf.cast = and i128 %0, 170135991163610696904058773219554885632 155 %cmp = icmp eq i128 %bf.cast, 0 156 br i1 %cmp, label %if.then, label %if.end 157 158if.then: ; preds = %entry 159 %mul = fmul fp128 %x, 0xL00000000000000004201000000000000 160 %1 = bitcast fp128 %mul to i128 161 %bf.clear4 = and i128 %1, -170135991163610696904058773219554885633 162 %bf.set = or i128 %bf.clear4, 85060207136517546210586590865283612672 163 br label %if.end 164 165if.end: ; preds = %if.then, %entry 166 %u.sroa.0.0 = phi i128 [ %bf.set, %if.then ], [ %0, %entry ] 167 %2 = bitcast i128 %u.sroa.0.0 to fp128 168 ret fp128 %2 169; CHECK-LABEL: TestI128_3: 170; CHECK: movaps %xmm0, 171; CHECK: movabsq $9223090561878065152, 172; CHECK: testq 173; CHECK: callq __multf3 174; CHECK-NEXT: movaps %xmm0 175; CHECK: movabsq $-9223090561878065153, 176; CHECK: movabsq $4611123068473966592, 177; CHECK: retq 178} 179 180; C code: 181; __float128 TestI128_4(__float128 x) 182; { 183; union IEEEl2bits u; 184; __float128 df; 185; u.e = x; 186; u.xbits.manl = 0; 187; df = u.e; 188; return x + df; 189; } 190define fp128 @TestI128_4(fp128 %x) #0 { 191entry: 192 %0 = bitcast fp128 %x to i128 193 %bf.clear = and i128 %0, -18446744073709551616 194 %1 = bitcast i128 %bf.clear to fp128 195 %add = fadd fp128 %1, %x 196 ret fp128 %add 197; CHECK-LABEL: TestI128_4: 198; CHECK: movaps %xmm0, %xmm1 199; CHECK-NEXT: movaps %xmm1, 16(%rsp) 200; CHECK-NEXT: movq 24(%rsp), %rax 201; CHECK-NEXT: movq %rax, 8(%rsp) 202; CHECK-NEXT: movq $0, (%rsp) 203; CHECK-NEXT: movaps (%rsp), %xmm0 204; CHECK-NEXT: callq __addtf3 205; CHECK: retq 206} 207 208@v128 = common global i128 0, align 16 209@v128_2 = common global i128 0, align 16 210 211; C code: 212; unsigned __int128 v128, v128_2; 213; void TestShift128_2() { 214; v128 = ((v128 << 96) | v128_2); 215; } 216define void @TestShift128_2() #2 { 217entry: 218 %0 = load i128, i128* @v128, align 16 219 %shl = shl i128 %0, 96 220 %1 = load i128, i128* @v128_2, align 16 221 %or = or i128 %shl, %1 222 store i128 %or, i128* @v128, align 16 223 ret void 224; CHECK-LABEL: TestShift128_2: 225; CHECK: movq v128(%rip), %rax 226; CHECK-NEXT: shlq $32, %rax 227; CHECK-NEXT: movq v128_2(%rip), %rcx 228; CHECK-NEXT: orq v128_2+8(%rip), %rax 229; CHECK-NEXT: movq %rcx, v128(%rip) 230; CHECK-NEXT: movq %rax, v128+8(%rip) 231; CHECK-NEXT: retq 232} 233 234define fp128 @acosl(fp128 %x) #0 { 235entry: 236 %0 = bitcast fp128 %x to i128 237 %bf.clear = and i128 %0, -18446744073709551616 238 %1 = bitcast i128 %bf.clear to fp128 239 %add = fadd fp128 %1, %x 240 ret fp128 %add 241; CHECK-LABEL: acosl: 242; CHECK: movaps %xmm0, %xmm1 243; CHECK-NEXT: movaps %xmm1, 16(%rsp) 244; CHECK-NEXT: movq 24(%rsp), %rax 245; CHECK-NEXT: movq %rax, 8(%rsp) 246; CHECK-NEXT: movq $0, (%rsp) 247; CHECK-NEXT: movaps (%rsp), %xmm0 248; CHECK-NEXT: callq __addtf3 249; CHECK: retq 250} 251 252; Compare i128 values and check i128 constants. 253define fp128 @TestComp(fp128 %x, fp128 %y) #0 { 254entry: 255 %0 = bitcast fp128 %x to i128 256 %cmp = icmp sgt i128 %0, -1 257 %cond = select i1 %cmp, fp128 %x, fp128 %y 258 ret fp128 %cond 259; CHECK-LABEL: TestComp: 260; CHECK: movaps %xmm0, -24(%rsp) 261; CHECK-NEXT: cmpq $0, -16(%rsp) 262; CHECK-NEXT: jns 263; CHECK: movaps %xmm1, %xmm0 264; CHECK: retq 265} 266 267declare void @foo(fp128) #1 268 269; Test logical operations on fp128 values. 270define fp128 @TestFABS_LD(fp128 %x) #0 { 271entry: 272 %call = tail call fp128 @fabsl(fp128 %x) #2 273 ret fp128 %call 274; CHECK-LABEL: TestFABS_LD 275; CHECK: andps {{.*}}, %xmm0 276; CHECK-NEXT: retq 277} 278 279declare fp128 @fabsl(fp128) #1 280 281declare fp128 @copysignl(fp128, fp128) #1 282 283; Test more complicated logical operations generated from copysignl. 284define void @TestCopySign({ fp128, fp128 }* noalias nocapture sret %agg.result, { fp128, fp128 }* byval nocapture readonly align 16 %z) #0 { 285entry: 286 %z.realp = getelementptr inbounds { fp128, fp128 }, { fp128, fp128 }* %z, i64 0, i32 0 287 %z.real = load fp128, fp128* %z.realp, align 16 288 %z.imagp = getelementptr inbounds { fp128, fp128 }, { fp128, fp128 }* %z, i64 0, i32 1 289 %z.imag4 = load fp128, fp128* %z.imagp, align 16 290 %cmp = fcmp ogt fp128 %z.real, %z.imag4 291 %sub = fsub fp128 %z.imag4, %z.imag4 292 br i1 %cmp, label %if.then, label %cleanup 293 294if.then: ; preds = %entry 295 %call = tail call fp128 @fabsl(fp128 %sub) #2 296 br label %cleanup 297 298cleanup: ; preds = %entry, %if.then 299 %z.real.sink = phi fp128 [ %z.real, %if.then ], [ %sub, %entry ] 300 %call.sink = phi fp128 [ %call, %if.then ], [ %z.real, %entry ] 301 %call5 = tail call fp128 @copysignl(fp128 %z.real.sink, fp128 %z.imag4) #2 302 %0 = getelementptr inbounds { fp128, fp128 }, { fp128, fp128 }* %agg.result, i64 0, i32 0 303 %1 = getelementptr inbounds { fp128, fp128 }, { fp128, fp128 }* %agg.result, i64 0, i32 1 304 store fp128 %call.sink, fp128* %0, align 16 305 store fp128 %call5, fp128* %1, align 16 306 ret void 307; CHECK-LABEL: TestCopySign 308; CHECK-NOT: call 309; CHECK: callq __subtf3 310; CHECK-NOT: call 311; CHECK: callq __gttf2 312; CHECK-NOT: call 313; CHECK: andps {{.*}}, %xmm0 314; CHECK: retq 315} 316 317 318attributes #0 = { nounwind uwtable "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+ssse3,+sse3,+popcnt,+sse,+sse2,+sse4.1,+sse4.2" "unsafe-fp-math"="false" "use-soft-float"="false" } 319attributes #1 = { "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+ssse3,+sse3,+popcnt,+sse,+sse2,+sse4.1,+sse4.2" "unsafe-fp-math"="false" "use-soft-float"="false" } 320attributes #2 = { nounwind readnone } 321