1 //===-- PPCJITInfo.cpp - Implement the JIT interfaces for the PowerPC -----===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the JIT interfaces for the 32-bit PowerPC target. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #define DEBUG_TYPE "jit" 15 #include "PPCJITInfo.h" 16 #include "PPCRelocations.h" 17 #include "PPCTargetMachine.h" 18 #include "llvm/IR/Function.h" 19 #include "llvm/Support/Debug.h" 20 #include "llvm/Support/ErrorHandling.h" 21 #include "llvm/Support/Memory.h" 22 #include "llvm/Support/raw_ostream.h" 23 using namespace llvm; 24 25 static TargetJITInfo::JITCompilerFn JITCompilerFunction; 26 27 #define BUILD_ADDIS(RD,RS,IMM16) \ 28 ((15 << 26) | ((RD) << 21) | ((RS) << 16) | ((IMM16) & 65535)) 29 #define BUILD_ORI(RD,RS,UIMM16) \ 30 ((24 << 26) | ((RS) << 21) | ((RD) << 16) | ((UIMM16) & 65535)) 31 #define BUILD_ORIS(RD,RS,UIMM16) \ 32 ((25 << 26) | ((RS) << 21) | ((RD) << 16) | ((UIMM16) & 65535)) 33 #define BUILD_RLDICR(RD,RS,SH,ME) \ 34 ((30 << 26) | ((RS) << 21) | ((RD) << 16) | (((SH) & 31) << 11) | \ 35 (((ME) & 63) << 6) | (1 << 2) | ((((SH) >> 5) & 1) << 1)) 36 #define BUILD_MTSPR(RS,SPR) \ 37 ((31 << 26) | ((RS) << 21) | ((SPR) << 16) | (467 << 1)) 38 #define BUILD_BCCTRx(BO,BI,LINK) \ 39 ((19 << 26) | ((BO) << 21) | ((BI) << 16) | (528 << 1) | ((LINK) & 1)) 40 #define BUILD_B(TARGET, LINK) \ 41 ((18 << 26) | (((TARGET) & 0x00FFFFFF) << 2) | ((LINK) & 1)) 42 43 // Pseudo-ops 44 #define BUILD_LIS(RD,IMM16) BUILD_ADDIS(RD,0,IMM16) 45 #define BUILD_SLDI(RD,RS,IMM6) BUILD_RLDICR(RD,RS,IMM6,63-IMM6) 46 #define BUILD_MTCTR(RS) BUILD_MTSPR(RS,9) 47 #define BUILD_BCTR(LINK) BUILD_BCCTRx(20,0,LINK) 48 49 static void EmitBranchToAt(uint64_t At, uint64_t To, bool isCall, bool is64Bit){ 50 intptr_t Offset = ((intptr_t)To - (intptr_t)At) >> 2; 51 unsigned *AtI = (unsigned*)(intptr_t)At; 52 53 if (Offset >= -(1 << 23) && Offset < (1 << 23)) { // In range? 54 AtI[0] = BUILD_B(Offset, isCall); // b/bl target 55 } else if (!is64Bit) { 56 AtI[0] = BUILD_LIS(12, To >> 16); // lis r12, hi16(address) 57 AtI[1] = BUILD_ORI(12, 12, To); // ori r12, r12, lo16(address) 58 AtI[2] = BUILD_MTCTR(12); // mtctr r12 59 AtI[3] = BUILD_BCTR(isCall); // bctr/bctrl 60 } else { 61 AtI[0] = BUILD_LIS(12, To >> 48); // lis r12, hi16(address) 62 AtI[1] = BUILD_ORI(12, 12, To >> 32); // ori r12, r12, lo16(address) 63 AtI[2] = BUILD_SLDI(12, 12, 32); // sldi r12, r12, 32 64 AtI[3] = BUILD_ORIS(12, 12, To >> 16); // oris r12, r12, hi16(address) 65 AtI[4] = BUILD_ORI(12, 12, To); // ori r12, r12, lo16(address) 66 AtI[5] = BUILD_MTCTR(12); // mtctr r12 67 AtI[6] = BUILD_BCTR(isCall); // bctr/bctrl 68 } 69 } 70 71 extern "C" void PPC32CompilationCallback(); 72 extern "C" void PPC64CompilationCallback(); 73 74 // The first clause of the preprocessor directive looks wrong, but it is 75 // necessary when compiling this code on non-PowerPC hosts. 76 #if (!defined(__ppc__) && !defined(__powerpc__)) || defined(__powerpc64__) || defined(__ppc64__) 77 void PPC32CompilationCallback() { 78 llvm_unreachable("This is not a 32bit PowerPC, you can't execute this!"); 79 } 80 #elif !defined(__ELF__) 81 // CompilationCallback stub - We can't use a C function with inline assembly in 82 // it, because we the prolog/epilog inserted by GCC won't work for us. Instead, 83 // write our own wrapper, which does things our way, so we have complete control 84 // over register saving and restoring. 85 asm( 86 ".text\n" 87 ".align 2\n" 88 ".globl _PPC32CompilationCallback\n" 89 "_PPC32CompilationCallback:\n" 90 // Make space for 8 ints r[3-10] and 13 doubles f[1-13] and the 91 // FIXME: need to save v[0-19] for altivec? 92 // FIXME: could shrink frame 93 // Set up a proper stack frame 94 // FIXME Layout 95 // PowerPC32 ABI linkage - 24 bytes 96 // parameters - 32 bytes 97 // 13 double registers - 104 bytes 98 // 8 int registers - 32 bytes 99 "mflr r0\n" 100 "stw r0, 8(r1)\n" 101 "stwu r1, -208(r1)\n" 102 // Save all int arg registers 103 "stw r10, 204(r1)\n" "stw r9, 200(r1)\n" 104 "stw r8, 196(r1)\n" "stw r7, 192(r1)\n" 105 "stw r6, 188(r1)\n" "stw r5, 184(r1)\n" 106 "stw r4, 180(r1)\n" "stw r3, 176(r1)\n" 107 // Save all call-clobbered FP regs. 108 "stfd f13, 168(r1)\n" "stfd f12, 160(r1)\n" 109 "stfd f11, 152(r1)\n" "stfd f10, 144(r1)\n" 110 "stfd f9, 136(r1)\n" "stfd f8, 128(r1)\n" 111 "stfd f7, 120(r1)\n" "stfd f6, 112(r1)\n" 112 "stfd f5, 104(r1)\n" "stfd f4, 96(r1)\n" 113 "stfd f3, 88(r1)\n" "stfd f2, 80(r1)\n" 114 "stfd f1, 72(r1)\n" 115 // Arguments to Compilation Callback: 116 // r3 - our lr (address of the call instruction in stub plus 4) 117 // r4 - stub's lr (address of instruction that called the stub plus 4) 118 // r5 - is64Bit - always 0. 119 "mr r3, r0\n" 120 "lwz r2, 208(r1)\n" // stub's frame 121 "lwz r4, 8(r2)\n" // stub's lr 122 "li r5, 0\n" // 0 == 32 bit 123 "bl _LLVMPPCCompilationCallback\n" 124 "mtctr r3\n" 125 // Restore all int arg registers 126 "lwz r10, 204(r1)\n" "lwz r9, 200(r1)\n" 127 "lwz r8, 196(r1)\n" "lwz r7, 192(r1)\n" 128 "lwz r6, 188(r1)\n" "lwz r5, 184(r1)\n" 129 "lwz r4, 180(r1)\n" "lwz r3, 176(r1)\n" 130 // Restore all FP arg registers 131 "lfd f13, 168(r1)\n" "lfd f12, 160(r1)\n" 132 "lfd f11, 152(r1)\n" "lfd f10, 144(r1)\n" 133 "lfd f9, 136(r1)\n" "lfd f8, 128(r1)\n" 134 "lfd f7, 120(r1)\n" "lfd f6, 112(r1)\n" 135 "lfd f5, 104(r1)\n" "lfd f4, 96(r1)\n" 136 "lfd f3, 88(r1)\n" "lfd f2, 80(r1)\n" 137 "lfd f1, 72(r1)\n" 138 // Pop 3 frames off the stack and branch to target 139 "lwz r1, 208(r1)\n" 140 "lwz r2, 8(r1)\n" 141 "mtlr r2\n" 142 "bctr\n" 143 ); 144 145 #else 146 // ELF PPC 32 support 147 148 // CompilationCallback stub - We can't use a C function with inline assembly in 149 // it, because we the prolog/epilog inserted by GCC won't work for us. Instead, 150 // write our own wrapper, which does things our way, so we have complete control 151 // over register saving and restoring. 152 asm( 153 ".text\n" 154 ".align 2\n" 155 ".globl PPC32CompilationCallback\n" 156 "PPC32CompilationCallback:\n" 157 // Make space for 8 ints r[3-10] and 8 doubles f[1-8] and the 158 // FIXME: need to save v[0-19] for altivec? 159 // FIXME: could shrink frame 160 // Set up a proper stack frame 161 // FIXME Layout 162 // 8 double registers - 64 bytes 163 // 8 int registers - 32 bytes 164 "mflr 0\n" 165 "stw 0, 4(1)\n" 166 "stwu 1, -104(1)\n" 167 // Save all int arg registers 168 "stw 10, 100(1)\n" "stw 9, 96(1)\n" 169 "stw 8, 92(1)\n" "stw 7, 88(1)\n" 170 "stw 6, 84(1)\n" "stw 5, 80(1)\n" 171 "stw 4, 76(1)\n" "stw 3, 72(1)\n" 172 // Save all call-clobbered FP regs. 173 "stfd 8, 64(1)\n" 174 "stfd 7, 56(1)\n" "stfd 6, 48(1)\n" 175 "stfd 5, 40(1)\n" "stfd 4, 32(1)\n" 176 "stfd 3, 24(1)\n" "stfd 2, 16(1)\n" 177 "stfd 1, 8(1)\n" 178 // Arguments to Compilation Callback: 179 // r3 - our lr (address of the call instruction in stub plus 4) 180 // r4 - stub's lr (address of instruction that called the stub plus 4) 181 // r5 - is64Bit - always 0. 182 "mr 3, 0\n" 183 "lwz 5, 104(1)\n" // stub's frame 184 "lwz 4, 4(5)\n" // stub's lr 185 "li 5, 0\n" // 0 == 32 bit 186 "bl LLVMPPCCompilationCallback\n" 187 "mtctr 3\n" 188 // Restore all int arg registers 189 "lwz 10, 100(1)\n" "lwz 9, 96(1)\n" 190 "lwz 8, 92(1)\n" "lwz 7, 88(1)\n" 191 "lwz 6, 84(1)\n" "lwz 5, 80(1)\n" 192 "lwz 4, 76(1)\n" "lwz 3, 72(1)\n" 193 // Restore all FP arg registers 194 "lfd 8, 64(1)\n" 195 "lfd 7, 56(1)\n" "lfd 6, 48(1)\n" 196 "lfd 5, 40(1)\n" "lfd 4, 32(1)\n" 197 "lfd 3, 24(1)\n" "lfd 2, 16(1)\n" 198 "lfd 1, 8(1)\n" 199 // Pop 3 frames off the stack and branch to target 200 "lwz 1, 104(1)\n" 201 "lwz 0, 4(1)\n" 202 "mtlr 0\n" 203 "bctr\n" 204 ); 205 #endif 206 207 #if !defined(__powerpc64__) && !defined(__ppc64__) 208 void PPC64CompilationCallback() { 209 llvm_unreachable("This is not a 64bit PowerPC, you can't execute this!"); 210 } 211 #else 212 # ifdef __ELF__ 213 asm( 214 ".text\n" 215 ".align 2\n" 216 ".globl PPC64CompilationCallback\n" 217 ".section \".opd\",\"aw\",@progbits\n" 218 ".align 3\n" 219 "PPC64CompilationCallback:\n" 220 ".quad .L.PPC64CompilationCallback,.TOC.@tocbase,0\n" 221 ".size PPC64CompilationCallback,24\n" 222 ".previous\n" 223 ".align 4\n" 224 ".type PPC64CompilationCallback,@function\n" 225 ".L.PPC64CompilationCallback:\n" 226 # else 227 asm( 228 ".text\n" 229 ".align 2\n" 230 ".globl _PPC64CompilationCallback\n" 231 "_PPC64CompilationCallback:\n" 232 # endif 233 // Make space for 8 ints r[3-10] and 13 doubles f[1-13] and the 234 // FIXME: need to save v[0-19] for altivec? 235 // Set up a proper stack frame 236 // Layout 237 // PowerPC64 ABI linkage - 48 bytes 238 // parameters - 64 bytes 239 // 13 double registers - 104 bytes 240 // 8 int registers - 64 bytes 241 "mflr 0\n" 242 "std 0, 16(1)\n" 243 "stdu 1, -280(1)\n" 244 // Save all int arg registers 245 "std 10, 272(1)\n" "std 9, 264(1)\n" 246 "std 8, 256(1)\n" "std 7, 248(1)\n" 247 "std 6, 240(1)\n" "std 5, 232(1)\n" 248 "std 4, 224(1)\n" "std 3, 216(1)\n" 249 // Save all call-clobbered FP regs. 250 "stfd 13, 208(1)\n" "stfd 12, 200(1)\n" 251 "stfd 11, 192(1)\n" "stfd 10, 184(1)\n" 252 "stfd 9, 176(1)\n" "stfd 8, 168(1)\n" 253 "stfd 7, 160(1)\n" "stfd 6, 152(1)\n" 254 "stfd 5, 144(1)\n" "stfd 4, 136(1)\n" 255 "stfd 3, 128(1)\n" "stfd 2, 120(1)\n" 256 "stfd 1, 112(1)\n" 257 // Arguments to Compilation Callback: 258 // r3 - our lr (address of the call instruction in stub plus 4) 259 // r4 - stub's lr (address of instruction that called the stub plus 4) 260 // r5 - is64Bit - always 1. 261 "mr 3, 0\n" // return address (still in r0) 262 "ld 5, 280(1)\n" // stub's frame 263 "ld 4, 16(5)\n" // stub's lr 264 "li 5, 1\n" // 1 == 64 bit 265 # ifdef __ELF__ 266 "bl LLVMPPCCompilationCallback\n" 267 "nop\n" 268 # else 269 "bl _LLVMPPCCompilationCallback\n" 270 # endif 271 "mtctr 3\n" 272 // Restore all int arg registers 273 "ld 10, 272(1)\n" "ld 9, 264(1)\n" 274 "ld 8, 256(1)\n" "ld 7, 248(1)\n" 275 "ld 6, 240(1)\n" "ld 5, 232(1)\n" 276 "ld 4, 224(1)\n" "ld 3, 216(1)\n" 277 // Restore all FP arg registers 278 "lfd 13, 208(1)\n" "lfd 12, 200(1)\n" 279 "lfd 11, 192(1)\n" "lfd 10, 184(1)\n" 280 "lfd 9, 176(1)\n" "lfd 8, 168(1)\n" 281 "lfd 7, 160(1)\n" "lfd 6, 152(1)\n" 282 "lfd 5, 144(1)\n" "lfd 4, 136(1)\n" 283 "lfd 3, 128(1)\n" "lfd 2, 120(1)\n" 284 "lfd 1, 112(1)\n" 285 // Pop 3 frames off the stack and branch to target 286 "ld 1, 280(1)\n" 287 "ld 0, 16(1)\n" 288 "mtlr 0\n" 289 // XXX: any special TOC handling in the ELF case for JIT? 290 "bctr\n" 291 ); 292 #endif 293 294 extern "C" { 295 LLVM_LIBRARY_VISIBILITY void * 296 LLVMPPCCompilationCallback(unsigned *StubCallAddrPlus4, 297 unsigned *OrigCallAddrPlus4, 298 bool is64Bit) { 299 // Adjust the pointer to the address of the call instruction in the stub 300 // emitted by emitFunctionStub, rather than the instruction after it. 301 unsigned *StubCallAddr = StubCallAddrPlus4 - 1; 302 unsigned *OrigCallAddr = OrigCallAddrPlus4 - 1; 303 304 void *Target = JITCompilerFunction(StubCallAddr); 305 306 // Check to see if *OrigCallAddr is a 'bl' instruction, and if we can rewrite 307 // it to branch directly to the destination. If so, rewrite it so it does not 308 // need to go through the stub anymore. 309 unsigned OrigCallInst = *OrigCallAddr; 310 if ((OrigCallInst >> 26) == 18) { // Direct call. 311 intptr_t Offset = ((intptr_t)Target - (intptr_t)OrigCallAddr) >> 2; 312 313 if (Offset >= -(1 << 23) && Offset < (1 << 23)) { // In range? 314 // Clear the original target out. 315 OrigCallInst &= (63 << 26) | 3; 316 // Fill in the new target. 317 OrigCallInst |= (Offset & ((1 << 24)-1)) << 2; 318 // Replace the call. 319 *OrigCallAddr = OrigCallInst; 320 } 321 } 322 323 // Assert that we are coming from a stub that was created with our 324 // emitFunctionStub. 325 if ((*StubCallAddr >> 26) == 18) 326 StubCallAddr -= 3; 327 else { 328 assert((*StubCallAddr >> 26) == 19 && "Call in stub is not indirect!"); 329 StubCallAddr -= is64Bit ? 9 : 6; 330 } 331 332 // Rewrite the stub with an unconditional branch to the target, for any users 333 // who took the address of the stub. 334 EmitBranchToAt((intptr_t)StubCallAddr, (intptr_t)Target, false, is64Bit); 335 sys::Memory::InvalidateInstructionCache(StubCallAddr, 7*4); 336 337 // Put the address of the target function to call and the address to return to 338 // after calling the target function in a place that is easy to get on the 339 // stack after we restore all regs. 340 return Target; 341 } 342 } 343 344 345 346 TargetJITInfo::LazyResolverFn 347 PPCJITInfo::getLazyResolverFunction(JITCompilerFn Fn) { 348 JITCompilerFunction = Fn; 349 return is64Bit ? PPC64CompilationCallback : PPC32CompilationCallback; 350 } 351 352 TargetJITInfo::StubLayout PPCJITInfo::getStubLayout() { 353 // The stub contains up to 10 4-byte instructions, aligned at 4 bytes: 3 354 // instructions to save the caller's address if this is a lazy-compilation 355 // stub, plus a 1-, 4-, or 7-instruction sequence to load an arbitrary address 356 // into a register and jump through it. 357 StubLayout Result = {10*4, 4}; 358 return Result; 359 } 360 361 #if (defined(__POWERPC__) || defined (__ppc__) || defined(_POWER)) && \ 362 defined(__APPLE__) 363 extern "C" void sys_icache_invalidate(const void *Addr, size_t len); 364 #endif 365 366 void *PPCJITInfo::emitFunctionStub(const Function* F, void *Fn, 367 JITCodeEmitter &JCE) { 368 // If this is just a call to an external function, emit a branch instead of a 369 // call. The code is the same except for one bit of the last instruction. 370 if (Fn != (void*)(intptr_t)PPC32CompilationCallback && 371 Fn != (void*)(intptr_t)PPC64CompilationCallback) { 372 void *Addr = (void*)JCE.getCurrentPCValue(); 373 JCE.emitWordBE(0); 374 JCE.emitWordBE(0); 375 JCE.emitWordBE(0); 376 JCE.emitWordBE(0); 377 JCE.emitWordBE(0); 378 JCE.emitWordBE(0); 379 JCE.emitWordBE(0); 380 EmitBranchToAt((intptr_t)Addr, (intptr_t)Fn, false, is64Bit); 381 sys::Memory::InvalidateInstructionCache(Addr, 7*4); 382 return Addr; 383 } 384 385 void *Addr = (void*)JCE.getCurrentPCValue(); 386 if (is64Bit) { 387 JCE.emitWordBE(0xf821ffb1); // stdu r1,-80(r1) 388 JCE.emitWordBE(0x7d6802a6); // mflr r11 389 JCE.emitWordBE(0xf9610060); // std r11, 96(r1) 390 } else if (TM.getSubtargetImpl()->isDarwinABI()){ 391 JCE.emitWordBE(0x9421ffe0); // stwu r1,-32(r1) 392 JCE.emitWordBE(0x7d6802a6); // mflr r11 393 JCE.emitWordBE(0x91610028); // stw r11, 40(r1) 394 } else { 395 JCE.emitWordBE(0x9421ffe0); // stwu r1,-32(r1) 396 JCE.emitWordBE(0x7d6802a6); // mflr r11 397 JCE.emitWordBE(0x91610024); // stw r11, 36(r1) 398 } 399 intptr_t BranchAddr = (intptr_t)JCE.getCurrentPCValue(); 400 JCE.emitWordBE(0); 401 JCE.emitWordBE(0); 402 JCE.emitWordBE(0); 403 JCE.emitWordBE(0); 404 JCE.emitWordBE(0); 405 JCE.emitWordBE(0); 406 JCE.emitWordBE(0); 407 EmitBranchToAt(BranchAddr, (intptr_t)Fn, true, is64Bit); 408 sys::Memory::InvalidateInstructionCache(Addr, 10*4); 409 return Addr; 410 } 411 412 413 void PPCJITInfo::relocate(void *Function, MachineRelocation *MR, 414 unsigned NumRelocs, unsigned char* GOTBase) { 415 for (unsigned i = 0; i != NumRelocs; ++i, ++MR) { 416 unsigned *RelocPos = (unsigned*)Function + MR->getMachineCodeOffset()/4; 417 intptr_t ResultPtr = (intptr_t)MR->getResultPointer(); 418 switch ((PPC::RelocationType)MR->getRelocationType()) { 419 default: llvm_unreachable("Unknown relocation type!"); 420 case PPC::reloc_pcrel_bx: 421 // PC-relative relocation for b and bl instructions. 422 ResultPtr = (ResultPtr-(intptr_t)RelocPos) >> 2; 423 assert(ResultPtr >= -(1 << 23) && ResultPtr < (1 << 23) && 424 "Relocation out of range!"); 425 *RelocPos |= (ResultPtr & ((1 << 24)-1)) << 2; 426 break; 427 case PPC::reloc_pcrel_bcx: 428 // PC-relative relocation for BLT,BLE,BEQ,BGE,BGT,BNE, or other 429 // bcx instructions. 430 ResultPtr = (ResultPtr-(intptr_t)RelocPos) >> 2; 431 assert(ResultPtr >= -(1 << 13) && ResultPtr < (1 << 13) && 432 "Relocation out of range!"); 433 *RelocPos |= (ResultPtr & ((1 << 14)-1)) << 2; 434 break; 435 case PPC::reloc_absolute_high: // high bits of ref -> low 16 of instr 436 case PPC::reloc_absolute_low: { // low bits of ref -> low 16 of instr 437 ResultPtr += MR->getConstantVal(); 438 439 // If this is a high-part access, get the high-part. 440 if (MR->getRelocationType() == PPC::reloc_absolute_high) { 441 // If the low part will have a carry (really a borrow) from the low 442 // 16-bits into the high 16, add a bit to borrow from. 443 if (((int)ResultPtr << 16) < 0) 444 ResultPtr += 1 << 16; 445 ResultPtr >>= 16; 446 } 447 448 // Do the addition then mask, so the addition does not overflow the 16-bit 449 // immediate section of the instruction. 450 unsigned LowBits = (*RelocPos + ResultPtr) & 65535; 451 unsigned HighBits = *RelocPos & ~65535; 452 *RelocPos = LowBits | HighBits; // Slam into low 16-bits 453 break; 454 } 455 case PPC::reloc_absolute_low_ix: { // low bits of ref -> low 14 of instr 456 ResultPtr += MR->getConstantVal(); 457 // Do the addition then mask, so the addition does not overflow the 16-bit 458 // immediate section of the instruction. 459 unsigned LowBits = (*RelocPos + ResultPtr) & 0xFFFC; 460 unsigned HighBits = *RelocPos & 0xFFFF0003; 461 *RelocPos = LowBits | HighBits; // Slam into low 14-bits. 462 break; 463 } 464 } 465 } 466 } 467 468 void PPCJITInfo::replaceMachineCodeForFunction(void *Old, void *New) { 469 EmitBranchToAt((intptr_t)Old, (intptr_t)New, false, is64Bit); 470 sys::Memory::InvalidateInstructionCache(Old, 7*4); 471 } 472