1; Tests various aspects of x86 immediate encoding. Some encodings are shorter. 2; For example, the encoding is shorter for 8-bit immediates or when using EAX. 3; This assumes that EAX is chosen as the first free register in O2 mode. 4 5; RUN: %p2i --filetype=obj --disassemble -i %s --args -O2 | FileCheck %s 6 7define internal i32 @testXor8Imm8(i32 %arg) { 8entry: 9 %arg_i8 = trunc i32 %arg to i8 10 %result_i8 = xor i8 %arg_i8, 127 11 %result = zext i8 %result_i8 to i32 12 ret i32 %result 13} 14; CHECK-LABEL: testXor8Imm8 15; CHECK: 34 7f xor al 16 17define internal i32 @testXor8Imm8Neg(i32 %arg) { 18entry: 19 %arg_i8 = trunc i32 %arg to i8 20 %result_i8 = xor i8 %arg_i8, -128 21 %result = zext i8 %result_i8 to i32 22 ret i32 %result 23} 24; CHECK-LABEL: testXor8Imm8Neg 25; CHECK: 34 80 xor al 26 27define internal i32 @testXor8Imm8NotEAX(i32 %arg, i32 %arg2, i32 %arg3) { 28entry: 29 %arg_i8 = trunc i32 %arg to i8 30 %arg2_i8 = trunc i32 %arg2 to i8 31 %arg3_i8 = trunc i32 %arg3 to i8 32 %x1 = xor i8 %arg_i8, 127 33 %x2 = xor i8 %arg2_i8, 127 34 %x3 = xor i8 %arg3_i8, 127 35 %x4 = add i8 %x1, %x2 36 %x5 = add i8 %x4, %x3 37 %result = zext i8 %x5 to i32 38 ret i32 %result 39} 40; CHECK-LABEL: testXor8Imm8NotEAX 41; CHECK: 80 f{{[1-3]}} 7f xor {{[^a]}}l 42 43define internal i32 @testXor16Imm8(i32 %arg) { 44entry: 45 %arg_i16 = trunc i32 %arg to i16 46 %result_i16 = xor i16 %arg_i16, 127 47 %result = zext i16 %result_i16 to i32 48 ret i32 %result 49} 50; CHECK-LABEL: testXor16Imm8 51; CHECK: 66 83 f0 7f xor ax 52 53define internal i32 @testXor16Imm8Neg(i32 %arg) { 54entry: 55 %arg_i16 = trunc i32 %arg to i16 56 %result_i16 = xor i16 %arg_i16, -128 57 %result = zext i16 %result_i16 to i32 58 ret i32 %result 59} 60; CHECK-LABEL: testXor16Imm8Neg 61; CHECK: 66 83 f0 80 xor ax 62 63define internal i32 @testXor16Imm16Eax(i32 %arg) { 64entry: 65 %arg_i16 = trunc i32 %arg to i16 66 %tmp = xor i16 %arg_i16, 1024 67 %result_i16 = add i16 %tmp, 1 68 %result = zext i16 %result_i16 to i32 69 ret i32 %result 70} 71; CHECK-LABEL: testXor16Imm16Eax 72; CHECK: 66 35 00 04 xor ax 73; CHECK-NEXT: add ax,0x1 74 75define internal i32 @testXor16Imm16NegEax(i32 %arg) { 76entry: 77 %arg_i16 = trunc i32 %arg to i16 78 %tmp = xor i16 %arg_i16, -256 79 %result_i16 = add i16 %tmp, 1 80 %result = zext i16 %result_i16 to i32 81 ret i32 %result 82} 83; CHECK-LABEL: testXor16Imm16NegEax 84; CHECK: 66 35 00 ff xor ax 85; CHECK-NEXT: add ax,0x1 86 87define internal i32 @testXor16Imm16NotEAX(i32 %arg_i32, i32 %arg2_i32, i32 %arg3_i32) { 88entry: 89 %arg = trunc i32 %arg_i32 to i16 90 %arg2 = trunc i32 %arg2_i32 to i16 91 %arg3 = trunc i32 %arg3_i32 to i16 92 %x = xor i16 %arg, 32767 93 %x2 = xor i16 %arg2, 32767 94 %x3 = xor i16 %arg3, 32767 95 %add1 = add i16 %x, %x2 96 %add2 = add i16 %add1, %x3 97 %result = zext i16 %add2 to i32 98 ret i32 %result 99} 100; CHECK-LABEL: testXor16Imm16NotEAX 101; CHECK: 66 81 f{{[1-3]}} ff 7f xor {{[^a]}}x 102; CHECK-NEXT: 66 81 f{{[1-3]}} ff 7f xor {{[^a]}}x 103 104define internal i32 @testXor32Imm8(i32 %arg) { 105entry: 106 %result = xor i32 %arg, 127 107 ret i32 %result 108} 109; CHECK-LABEL: testXor32Imm8 110; CHECK: 83 f0 7f xor eax 111 112define internal i32 @testXor32Imm8Neg(i32 %arg) { 113entry: 114 %result = xor i32 %arg, -128 115 ret i32 %result 116} 117; CHECK-LABEL: testXor32Imm8Neg 118; CHECK: 83 f0 80 xor eax 119 120define internal i32 @testXor32Imm32Eax(i32 %arg) { 121entry: 122 %result = xor i32 %arg, 16777216 123 ret i32 %result 124} 125; CHECK-LABEL: testXor32Imm32Eax 126; CHECK: 35 00 00 00 01 xor eax 127 128define internal i32 @testXor32Imm32NegEax(i32 %arg) { 129entry: 130 %result = xor i32 %arg, -256 131 ret i32 %result 132} 133; CHECK-LABEL: testXor32Imm32NegEax 134; CHECK: 35 00 ff ff ff xor eax 135 136define internal i32 @testXor32Imm32NotEAX(i32 %arg, i32 %arg2, i32 %arg3) { 137entry: 138 %x = xor i32 %arg, 32767 139 %x2 = xor i32 %arg2, 32767 140 %x3 = xor i32 %arg3, 32767 141 %add1 = add i32 %x, %x2 142 %add2 = add i32 %add1, %x3 143 ret i32 %add2 144} 145; CHECK-LABEL: testXor32Imm32NotEAX 146; CHECK: 81 f{{[1-3]}} ff 7f 00 00 xor e{{[^a]}}x, 147 148; Should be similar for add, sub, etc., so sample a few. 149 150define internal i32 @testAdd8Imm8(i32 %arg) { 151entry: 152 %arg_i8 = trunc i32 %arg to i8 153 %result_i8 = add i8 %arg_i8, 126 154 %result = zext i8 %result_i8 to i32 155 ret i32 %result 156} 157; CHECK-LABEL: testAdd8Imm8 158; CHECK: 04 7e add al 159 160define internal i32 @testSub8Imm8(i32 %arg) { 161entry: 162 %arg_i8 = trunc i32 %arg to i8 163 %result_i8 = sub i8 %arg_i8, 125 164 %result = zext i8 %result_i8 to i32 165 ret i32 %result 166} 167; CHECK-LABEL: testSub8Imm8 168; CHECK: 2c 7d sub al 169 170; imul has some shorter 8-bit immediate encodings. 171; It also has a shorter encoding for eax, but we don't do that yet. 172 173define internal i32 @testMul16Imm8(i32 %arg) { 174entry: 175 %arg_i16 = trunc i32 %arg to i16 176 %tmp = mul i16 %arg_i16, 99 177 %result_i16 = add i16 %tmp, 1 178 %result = zext i16 %result_i16 to i32 179 ret i32 %result 180} 181; CHECK-LABEL: testMul16Imm8 182; CHECK: 66 6b c0 63 imul ax,ax 183; CHECK-NEXT: add ax,0x1 184 185define internal i32 @testMul16Imm8Neg(i32 %arg) { 186entry: 187 %arg_i16 = trunc i32 %arg to i16 188 %tmp = mul i16 %arg_i16, -111 189 %result_i16 = add i16 %tmp, 1 190 %result = zext i16 %result_i16 to i32 191 ret i32 %result 192} 193; CHECK-LABEL: testMul16Imm8Neg 194; CHECK: 66 6b c0 91 imul ax,ax 195; CHECK-NEXT: add ax,0x1 196 197define internal i32 @testMul16Imm16(i32 %arg) { 198entry: 199 %arg_i16 = trunc i32 %arg to i16 200 %tmp = mul i16 %arg_i16, 1025 201 %result_i16 = add i16 %tmp, 1 202 %result = zext i16 %result_i16 to i32 203 ret i32 %result 204} 205; CHECK-LABEL: testMul16Imm16 206; CHECK: 66 69 c0 01 04 imul ax,ax 207; CHECK-NEXT: add ax,0x1 208 209define internal i32 @testMul16Imm16Neg(i32 %arg) { 210entry: 211 %arg_i16 = trunc i32 %arg to i16 212 %tmp = mul i16 %arg_i16, -255 213 %result_i16 = add i16 %tmp, 1 214 %result = zext i16 %result_i16 to i32 215 ret i32 %result 216} 217; CHECK-LABEL: testMul16Imm16Neg 218; CHECK: 66 69 c0 01 ff imul ax,ax,0xff01 219; CHECK-NEXT: add ax,0x1 220 221define internal i32 @testMul32Imm8(i32 %arg) { 222entry: 223 %result = mul i32 %arg, 99 224 ret i32 %result 225} 226; CHECK-LABEL: testMul32Imm8 227; CHECK: 6b c0 63 imul eax,eax 228 229define internal i32 @testMul32Imm8Neg(i32 %arg) { 230entry: 231 %result = mul i32 %arg, -111 232 ret i32 %result 233} 234; CHECK-LABEL: testMul32Imm8Neg 235; CHECK: 6b c0 91 imul eax,eax 236 237define internal i32 @testMul32Imm16(i32 %arg) { 238entry: 239 %result = mul i32 %arg, 1025 240 ret i32 %result 241} 242; CHECK-LABEL: testMul32Imm16 243; CHECK: 69 c0 01 04 00 00 imul eax,eax 244 245define internal i32 @testMul32Imm16Neg(i32 %arg) { 246entry: 247 %result = mul i32 %arg, -255 248 ret i32 %result 249} 250; CHECK-LABEL: testMul32Imm16Neg 251; CHECK: 69 c0 01 ff ff ff imul eax,eax,0xffffff01 252 253define internal i32 @testMul32Imm32ThreeAddress(i32 %a) { 254entry: 255 %mul = mul i32 232, %a 256 %add = add i32 %mul, %a 257 ret i32 %add 258} 259; CHECK-LABEL: testMul32Imm32ThreeAddress 260; CHECK: 69 c8 e8 00 00 00 imul ecx,eax,0xe8 261 262define internal i32 @testMul32Mem32Imm32ThreeAddress(i32 %addr_arg) { 263entry: 264 %__1 = inttoptr i32 %addr_arg to i32* 265 %a = load i32, i32* %__1, align 1 266 %mul = mul i32 232, %a 267 ret i32 %mul 268} 269; CHECK-LABEL: testMul32Mem32Imm32ThreeAddress 270; CHECK: 69 00 e8 00 00 00 imul eax,DWORD PTR [eax],0xe8 271 272define internal i32 @testMul32Imm8ThreeAddress(i32 %a) { 273entry: 274 %mul = mul i32 127, %a 275 %add = add i32 %mul, %a 276 ret i32 %add 277} 278; CHECK-LABEL: testMul32Imm8ThreeAddress 279; CHECK: 6b c8 7f imul ecx,eax,0x7f 280 281define internal i32 @testMul32Mem32Imm8ThreeAddress(i32 %addr_arg) { 282entry: 283 %__1 = inttoptr i32 %addr_arg to i32* 284 %a = load i32, i32* %__1, align 1 285 %mul = mul i32 127, %a 286 ret i32 %mul 287} 288; CHECK-LABEL: testMul32Mem32Imm8ThreeAddress 289; CHECK: 6b 00 7f imul eax,DWORD PTR [eax],0x7f 290 291define internal i32 @testMul16Imm16ThreeAddress(i32 %a) { 292entry: 293 %arg_i16 = trunc i32 %a to i16 294 %mul = mul i16 232, %arg_i16 295 %add = add i16 %mul, %arg_i16 296 %result = zext i16 %add to i32 297 ret i32 %result 298} 299; CHECK-LABEL: testMul16Imm16ThreeAddress 300; CHECK: 66 69 c8 e8 00 imul cx,ax,0xe8 301 302define internal i32 @testMul16Mem16Imm16ThreeAddress(i32 %addr_arg) { 303entry: 304 %__1 = inttoptr i32 %addr_arg to i16* 305 %a = load i16, i16* %__1, align 1 306 %mul = mul i16 232, %a 307 %result = zext i16 %mul to i32 308 ret i32 %result 309} 310; CHECK-LABEL: testMul16Mem16Imm16ThreeAddress 311; CHECK: 66 69 00 e8 00 imul ax,WORD PTR [eax],0xe8 312 313define internal i32 @testMul16Imm8ThreeAddress(i32 %a) { 314entry: 315 %arg_i16 = trunc i32 %a to i16 316 %mul = mul i16 127, %arg_i16 317 %add = add i16 %mul, %arg_i16 318 %result = zext i16 %add to i32 319 ret i32 %result 320} 321; CHECK-LABEL: testMul16Imm8ThreeAddress 322; CHECK: 66 6b c8 7f imul cx,ax,0x7f 323 324define internal i32 @testMul16Mem16Imm8ThreeAddress(i32 %addr_arg) { 325entry: 326 %__1 = inttoptr i32 %addr_arg to i16* 327 %a = load i16, i16* %__1, align 1 328 %mul = mul i16 127, %a 329 %result = zext i16 %mul to i32 330 ret i32 %result 331} 332; CHECK-LABEL: testMul16Mem16Imm8ThreeAddress 333; CHECK: 66 6b 00 7f imul ax,WORD PTR [eax],0x7f 334 335; The GPR shift instructions either allow an 8-bit immediate or 336; have a special encoding for "1". 337define internal i32 @testShl16Imm8(i32 %arg) { 338entry: 339 %arg_i16 = trunc i32 %arg to i16 340 %tmp = shl i16 %arg_i16, 13 341 %result = zext i16 %tmp to i32 342 ret i32 %result 343} 344; CHECK-LABEL: testShl16Imm8 345; CHECK: 66 c1 e0 0d shl ax,0xd 346 347define internal i32 @testShl16Imm1(i32 %arg) { 348entry: 349 %arg_i16 = trunc i32 %arg to i16 350 %tmp = shl i16 %arg_i16, 1 351 %result = zext i16 %tmp to i32 352 ret i32 %result 353} 354; CHECK-LABEL: testShl16Imm1 355; CHECK: 66 d1 e0 shl ax 356 357; Currently the "test" instruction is used for 64-bit shifts, and 358; for ctlz 64-bit, so we use those to test the "test" instruction. 359; One optimization for "test": the "test" instruction is essentially a 360; bitwise AND that doesn't modify the two source operands, so for immediates 361; under 8-bits and registers with 8-bit variants we can use the shorter form. 362 363define internal i64 @test_via_shl64Bit(i64 %a, i64 %b) { 364entry: 365 %shl = shl i64 %a, %b 366 ret i64 %shl 367} 368; CHECK-LABEL: test_via_shl64Bit 369; CHECK: 0f a5 c2 shld edx,eax,cl 370; CHECK: d3 e0 shl eax,cl 371; CHECK: f6 c1 20 test cl,0x20 372 373; Test a few register encodings of "test". 374declare i64 @llvm.ctlz.i64(i64, i1) 375 376define internal i64 @test_via_ctlz_64(i64 %x, i64 %y, i64 %z, i64 %w) { 377entry: 378 %r = call i64 @llvm.ctlz.i64(i64 %x, i1 false) 379 %r2 = call i64 @llvm.ctlz.i64(i64 %y, i1 false) 380 %r3 = call i64 @llvm.ctlz.i64(i64 %z, i1 false) 381 %r4 = call i64 @llvm.ctlz.i64(i64 %w, i1 false) 382 %res1 = add i64 %r, %r2 383 %res2 = add i64 %r3, %r4 384 %res = add i64 %res1, %res2 385 ret i64 %res 386} 387; CHECK-LABEL: test_via_ctlz_64 388; CHECK-DAG: 85 c0 test eax,eax 389; CHECK-DAG: 85 db test ebx,ebx 390; CHECK-DAG: 85 f6 test esi,esi 391