ghc/rts/StgCRun.c

   1 /* -----------------------------------------------------------------------------
   2  *
   3  * (c) The GHC Team, 1998-2003
   4  *
   5  * STG-to-C glue.
   6  *
   7  * To run an STG function from C land, call
   8  *
   9  *              rv = StgRun(f,BaseReg);
  10  *
  11  * where "f" is the STG function to call, and BaseReg is the address of the
  12  * RegTable for this run (we might have separate RegTables if we're running
  13  * multiple threads on an SMP machine).
  14  *
  15  * In the end, "f" must JMP to StgReturn (defined below),
  16  * passing the return-value "rv" in R1,
  17  * to return to the caller of StgRun returning "rv" in
  18  * the whatever way C returns a value.
  19  *
  20  * NOTE: StgRun/StgReturn do *NOT* load or store Hp or any
  21  * other registers (other than saving the C callee-saves
  22  * registers).  Instead, the called function "f" must do that
  23  * in STG land.
  24  *
  25  * GCC will have assumed that pushing/popping of C-stack frames is
  26  * going on when it generated its code, and used stack space
  27  * accordingly.  However, we actually {\em post-process away} all
  28  * such stack-framery (see \tr{ghc/driver/ghc-asm.lprl}). Things will
  29  * be OK however, if we initially make sure there are
  30  * @RESERVED_C_STACK_BYTES@ on the C-stack to begin with, for local
  31  * variables.
  32  *
  33  * -------------------------------------------------------------------------- */
  34
  35 #include "PosixSource.h"
  36
  37
  38 /*
  39  * We define the following (unused) global register variables, because for
  40  * some reason gcc generates sub-optimal code for StgRun() on the Alpha
  41  * (unnecessarily saving extra registers on the stack) if we don't.
  42  *
  43  * Why do it at the top of this file, rather than near StgRun() below?  Because
  44  * gcc doesn't let us define global register variables after any function
  45  * definition has been read.  Any point after #include "Stg.h" would be too
  46  * late.
  47  *
  48  * We define alpha_EXTRA_CAREFUL here to save $s6, $f8 and $f9 -- registers
  49  * that we don't use but which are callee-save registers.  The __divq() routine
  50  * in libc.a clobbers $s6.
  51  */
  52 #include "ghcconfig.h"
  53 #ifdef alpha_HOST_ARCH
  54 #define alpha_EXTRA_CAREFUL
  55 register long   fake_ra __asm__("$26");
  56 register long   fake_gp __asm__("$29");
  57 #ifdef alpha_EXTRA_CAREFUL
  58 register long   fake_s6 __asm__("$15");
  59 register double fake_f8 __asm__("$f8");
  60 register double fake_f9 __asm__("$f9");
  61 #endif
  62 #endif
  63
  64 /* include Stg.h first because we want real machine regs in here: we
  65  * have to get the value of R1 back from Stg land to C land intact.
  66  */
  67 #include "Stg.h"
  68 #include "Rts.h"
  69 #include "StgRun.h"
  70 #include "RtsFlags.h"
  71 #include "OSThreads.h"
  72 #include "Capability.h"
  73
  74 #ifdef DEBUG
  75 #include "RtsUtils.h"
  76 #include "Printer.h"
  77 #endif
  78
  79 #ifdef USE_MINIINTERPRETER
  80
  81 /* -----------------------------------------------------------------------------
  82    any architecture (using miniinterpreter)
  83    -------------------------------------------------------------------------- */
  84
  85 StgThreadReturnCode StgRun(StgFunPtr f, StgRegTable *basereg STG_UNUSED)
  86 {
  87     while (f) {
  88         if (RtsFlags[0].DebugFlags.interpreter) {
  89             debugBelch("Jumping to ");
  90             printPtr((P_)f); fflush(stdout);
  91             debugBelch("\n");
  92         }
  93         f = (StgFunPtr) (f)();
  94     }
  95     return (StgThreadReturnCode)R1.i;
  96 }
  97
  98 StgFunPtr StgReturn(void)
  99 {
 100     return 0;
 101 }
 102
 103 #else /* !USE_MINIINTERPRETER */
 104
 105 #ifdef LEADING_UNDERSCORE
 106 #define STG_RETURN "_StgReturn"
 107 #else
 108 #define STG_RETURN "StgReturn"
 109 #endif
 110
 111 /* -----------------------------------------------------------------------------
 112    x86 architecture
 113    -------------------------------------------------------------------------- */
 114
 115 #ifdef i386_HOST_ARCH
 116
 117 StgThreadReturnCode
 118 StgRun(StgFunPtr f, StgRegTable *basereg) {
 119
 120     unsigned char space[ RESERVED_C_STACK_BYTES + 4*sizeof(void *) ];
 121     StgThreadReturnCode r;
 122
 123     __asm__ volatile (
 124         /*
 125          * save callee-saves registers on behalf of the STG code.
 126          */
 127         "movl %%esp, %%eax\n\t"
 128         "addl %4, %%eax\n\t"
 129         "movl %%ebx,0(%%eax)\n\t"
 130         "movl %%esi,4(%%eax)\n\t"
 131         "movl %%edi,8(%%eax)\n\t"
 132         "movl %%ebp,12(%%eax)\n\t"
 133         /*
 134          * Set BaseReg
 135          */
 136         "movl %3,%%ebx\n\t"
 137         /*
 138          * grab the function argument from the stack, and jump to it.
 139          */
 140         "movl %2,%%eax\n\t"
 141         "jmp *%%eax\n\t"
 142
 143         ".global " STG_RETURN "\n"
 144         STG_RETURN ":\n\t"
 145
 146         "movl %%esi, %%eax\n\t"   /* Return value in R1  */
 147
 148         /*
 149          * restore callee-saves registers.  (Don't stomp on %%eax!)
 150          */
 151         "movl %%esp, %%edx\n\t"
 152         "addl %4, %%edx\n\t"
 153         "movl 0(%%edx),%%ebx\n\t"       /* restore the registers saved above */
 154         "movl 4(%%edx),%%esi\n\t"
 155         "movl 8(%%edx),%%edi\n\t"
 156         "movl 12(%%edx),%%ebp\n\t"
 157
 158       : "=&a" (r), "=m" (space)
 159       : "m" (f), "m" (basereg), "i" (RESERVED_C_STACK_BYTES)
 160       : "edx" /* stomps on %edx */
 161     );
 162
 163     return r;
 164 }
 165
 166 #endif
 167
 168 /* ----------------------------------------------------------------------------
 169    x86-64 is almost the same as plain x86.
 170
 171    I've done it using entirely inline assembler, because I couldn't
 172    get gcc to generate the correct subtraction from %rsp by using
 173    the local array variable trick.  It didn't seem to reserve
 174    enough space.  Oh well, it's not much harder this way.
 175
 176    ------------------------------------------------------------------------- */
 177
 178 #ifdef x86_64_HOST_ARCH
 179
 180 extern StgThreadReturnCode StgRun(StgFunPtr f, StgRegTable *basereg);
 181
 182 static void StgRunIsImplementedInAssembler(void)
 183 {
 184     __asm__ volatile (
 185         /*
 186          * save callee-saves registers on behalf of the STG code.
 187          */
 188         ".globl StgRun\n"
 189         "StgRun:\n\t"
 190         "subq %0, %%rsp\n\t"
 191         "movq %%rsp, %%rax\n\t"
 192         "addq %0-48, %%rax\n\t"
 193         "movq %%rbx,0(%%rax)\n\t"
 194         "movq %%rbp,8(%%rax)\n\t"
 195         "movq %%r12,16(%%rax)\n\t"
 196         "movq %%r13,24(%%rax)\n\t"
 197         "movq %%r14,32(%%rax)\n\t"
 198         "movq %%r15,40(%%rax)\n\t"
 199         /*
 200          * Set BaseReg
 201          */
 202         "movq %%rsi,%%rbx\n\t"
 203         /*
 204          * grab the function argument from the stack, and jump to it.
 205          */
 206         "movq %%rdi,%%rax\n\t"
 207         "jmp *%%rax\n\t"
 208
 209         ".global " STG_RETURN "\n"
 210         STG_RETURN ":\n\t"
 211
 212         "movq %%r13, %%rax\n\t"   /* Return value in R1  */
 213
 214         /*
 215          * restore callee-saves registers.  (Don't stomp on %%rax!)
 216          */
 217         "movq %%rsp, %%rdx\n\t"
 218         "addq %0-48, %%rdx\n\t"
 219         "movq 0(%%rdx),%%rbx\n\t"       /* restore the registers saved above */
 220         "movq 8(%%rdx),%%rbp\n\t"
 221         "movq 16(%%rdx),%%r12\n\t"
 222         "movq 24(%%rdx),%%r13\n\t"
 223         "movq 32(%%rdx),%%r14\n\t"
 224         "movq 40(%%rdx),%%r15\n\t"
 225         "addq %0, %%rsp\n\t"
 226         "retq"
 227
 228         : : "i"(RESERVED_C_STACK_BYTES+48+8 /*stack frame size*/));
 229     /*
 230        HACK alert!
 231
 232        The x86_64 ABI specifies that on a procedure call, %rsp is
 233        aligned on a 16-byte boundary + 8.  That is, the first
 234        argument on the stack after the return address will be
 235        16-byte aligned.
 236
 237        Which should be fine: RESERVED_C_STACK_BYTES+48 is a multiple
 238        of 16 bytes.
 239
 240        BUT... when we do a C-call from STG land, gcc likes to put the
 241        stack alignment adjustment in the prolog.  eg. if we're calling
 242        a function with arguments in regs, gcc will insert 'subq $8,%rsp'
 243        in the prolog, to keep %rsp aligned (the return address is 8
 244        bytes, remember).  The mangler throws away the prolog, so we
 245        lose the stack alignment.
 246
 247        The hack is to add this extra 8 bytes to our %rsp adjustment
 248        here, so that throughout STG code, %rsp is 16-byte aligned,
 249        ready for a C-call.
 250
 251        A quick way to see if this is wrong is to compile this code:
 252
 253           main = System.Exit.exitWith ExitSuccess
 254
 255        And run it with +RTS -sstderr.  The stats code in the RTS, in
 256        particular statsPrintf(), relies on the stack alignment because
 257        it saves the %xmm regs on the stack, so it'll fall over if the
 258        stack isn't aligned, and calling exitWith from Haskell invokes
 259        shutdownHaskellAndExit using a C call.
 260
 261        Future gcc releases will almost certainly break this hack...
 262     */
 263 }
 264
 265 #endif /* x86-64 */
 266
 267 /* -----------------------------------------------------------------------------
 268    Sparc architecture
 269
 270    --
 271    OLD COMMENT from GHC-3.02:
 272
 273    We want tailjumps to be calls, because `call xxx' is the only Sparc
 274    branch that allows an arbitrary label as a target.  (Gcc's ``goto
 275    *target'' construct ends up loading the label into a register and
 276    then jumping, at the cost of two extra instructions for the 32-bit
 277    load.)
 278
 279    When entering the threaded world, we stash our return address in a
 280    known location so that \tr{%i7} is available as an extra
 281    callee-saves register.  Of course, we have to restore this when
 282    coming out of the threaded world.
 283
 284    I hate this god-forsaken architecture.  Since the top of the
 285    reserved stack space is used for globals and the bottom is reserved
 286    for outgoing arguments, we have to stick our return address
 287    somewhere in the middle.  Currently, I'm allowing 100 extra
 288    outgoing arguments beyond the first 6.  --JSM
 289
 290    Updated info (GHC 4.06): we don't appear to use %i7 any more, so
 291    I'm not sure whether we still need to save it.  Incedentally, what
 292    does the last paragraph above mean when it says "the top of the
 293    stack is used for globals"?  What globals?  --SDM
 294
 295    Updated info (GHC 4.08.2): not saving %i7 any more (see below).
 296    -------------------------------------------------------------------------- */
 297
 298 #ifdef sparc_HOST_ARCH
 299
 300 StgThreadReturnCode
 301 StgRun(StgFunPtr f, StgRegTable *basereg) {
 302
 303     unsigned char space[RESERVED_C_STACK_BYTES];
 304 #if 0
 305     register void *i7 __asm__("%i7");
 306     ((void **)(space))[100] = i7;
 307 #endif
 308     f();
 309     __asm__ volatile (
 310             ".align 4\n"
 311             ".global " STG_RETURN "\n"
 312             STG_RETURN ":"
 313             : : : "l0","l1","l2","l3","l4","l5","l6","l7");
 314     /* we tell the C compiler that l0-l7 are clobbered on return to
 315      * StgReturn, otherwise it tries to use these to save eg. the
 316      * address of space[100] across the call.  The correct thing
 317      * to do would be to save all the callee-saves regs, but we
 318      * can't be bothered to do that.
 319      *
 320      * The code that gcc generates for this little fragment is now
 321      * terrible.  We could do much better by coding it directly in
 322      * assembler.
 323      */
 324 #if 0
 325     /* updated 4.08.2: we don't save %i7 in the middle of the reserved
 326      * space any more, since gcc tries to save its address across the
 327      * call to f(), this gets clobbered in STG land and we end up
 328      * dereferencing a bogus pointer in StgReturn.
 329      */
 330     __asm__ volatile ("ld %1,%0"
 331                       : "=r" (i7) : "m" (((void **)(space))[100]));
 332 #endif
 333     return (StgThreadReturnCode)R1.i;
 334 }
 335
 336 #endif
 337
 338 /* -----------------------------------------------------------------------------
 339    alpha architecture
 340
 341    "The stack pointer (SP) must at all times denote an address that has octaword
 342     alignment. (This restriction has the side effect that the in-memory portion
 343     of the argument list, if any, will start on an octaword boundary.) Note that
 344     the stack grows toward lower addresses. During a procedure invocation, SP
 345     can never be set to a value that is higher than the value of SP at entry to
 346     that procedure invocation.
 347
 348    "The contents of the stack, located above the portion of the argument list
 349     (if any) that is passed in memory, belong to the calling procedure. Because
 350     they are part of the calling procedure, they should not be read or written
 351     by the called procedure, except as specified by indirect arguments or
 352     language-controlled up-level references.
 353
 354    "The SP value might be used by the hardware when raising exceptions and
 355     asynchronous interrupts. It must be assumed that the contents of the stack
 356     below the current SP value and within the stack for the current thread are
 357     continually and unpredictably modified, as specified in the _Alpha
 358     Architecture Reference Manual_, and as a result of asynchronous software
 359     actions."
 360
 361    -- Compaq Computer Corporation, Houston. Tru64 UNIX Calling Standard for
 362       Alpha Systems, 5.1 edition, August 2000, section 3.2.1.  http://www.
 363       tru64unix.compaq.com/docs/base_doc/DOCUMENTATION/V51_PDF/ARH9MBTE.PDF
 364    -------------------------------------------------------------------------- */
 365
 366 #ifdef alpha_HOST_ARCH
 367
 368 StgThreadReturnCode
 369 StgRun(StgFunPtr f, StgRegTable *basereg)
 370 {
 371     register long   real_ra __asm__("$26"); volatile long   save_ra;
 372     register long   real_gp __asm__("$29"); volatile long   save_gp;
 373
 374     register long   real_s0 __asm__("$9" ); volatile long   save_s0;
 375     register long   real_s1 __asm__("$10"); volatile long   save_s1;
 376     register long   real_s2 __asm__("$11"); volatile long   save_s2;
 377     register long   real_s3 __asm__("$12"); volatile long   save_s3;
 378     register long   real_s4 __asm__("$13"); volatile long   save_s4;
 379     register long   real_s5 __asm__("$14"); volatile long   save_s5;
 380 #ifdef alpha_EXTRA_CAREFUL
 381     register long   real_s6 __asm__("$15"); volatile long   save_s6;
 382 #endif
 383
 384     register double real_f2 __asm__("$f2"); volatile double save_f2;
 385     register double real_f3 __asm__("$f3"); volatile double save_f3;
 386     register double real_f4 __asm__("$f4"); volatile double save_f4;
 387     register double real_f5 __asm__("$f5"); volatile double save_f5;
 388     register double real_f6 __asm__("$f6"); volatile double save_f6;
 389     register double real_f7 __asm__("$f7"); volatile double save_f7;
 390 #ifdef alpha_EXTRA_CAREFUL
 391     register double real_f8 __asm__("$f8"); volatile double save_f8;
 392     register double real_f9 __asm__("$f9"); volatile double save_f9;
 393 #endif
 394
 395     register StgFunPtr real_pv __asm__("$27");
 396
 397     StgThreadReturnCode ret;
 398
 399     save_ra = real_ra;
 400     save_gp = real_gp;
 401
 402     save_s0 = real_s0;
 403     save_s1 = real_s1;
 404     save_s2 = real_s2;
 405     save_s3 = real_s3;
 406     save_s4 = real_s4;
 407     save_s5 = real_s5;
 408 #ifdef alpha_EXTRA_CAREFUL
 409     save_s6 = real_s6;
 410 #endif
 411
 412     save_f2 = real_f2;
 413     save_f3 = real_f3;
 414     save_f4 = real_f4;
 415     save_f5 = real_f5;
 416     save_f6 = real_f6;
 417     save_f7 = real_f7;
 418 #ifdef alpha_EXTRA_CAREFUL
 419     save_f8 = real_f8;
 420     save_f9 = real_f9;
 421 #endif
 422
 423     real_pv = f;
 424
 425     __asm__ volatile(   "lda $30,-%0($30)"      "\n"
 426                 "\t"    "jmp ($27)"             "\n"
 427                 "\t"    ".align 3"              "\n"
 428                 ".globl " STG_RETURN            "\n"
 429                 STG_RETURN ":"                  "\n"
 430                 "\t"    "lda $30,%0($30)"       "\n"
 431                 : : "K" (RESERVED_C_STACK_BYTES));
 432
 433     ret = real_s5;
 434
 435     real_s0 = save_s0;
 436     real_s1 = save_s1;
 437     real_s2 = save_s2;
 438     real_s3 = save_s3;
 439     real_s4 = save_s4;
 440     real_s5 = save_s5;
 441 #ifdef alpha_EXTRA_CAREFUL
 442     real_s6 = save_s6;
 443 #endif
 444
 445     real_f2 = save_f2;
 446     real_f3 = save_f3;
 447     real_f4 = save_f4;
 448     real_f5 = save_f5;
 449     real_f6 = save_f6;
 450     real_f7 = save_f7;
 451 #ifdef alpha_EXTRA_CAREFUL
 452     real_f8 = save_f8;
 453     real_f9 = save_f9;
 454 #endif
 455
 456     real_ra = save_ra;
 457     real_gp = save_gp;
 458
 459     return ret;
 460 }
 461
 462 #endif /* alpha_HOST_ARCH */
 463
 464 /* -----------------------------------------------------------------------------
 465    HP-PA architecture
 466    -------------------------------------------------------------------------- */
 467
 468 #ifdef hppa1_1_HOST_ARCH
 469
 470 StgThreadReturnCode
 471 StgRun(StgFunPtr f, StgRegTable *basereg)
 472 {
 473     StgChar space[RESERVED_C_STACK_BYTES+16*sizeof(long)+10*sizeof(double)];
 474     StgThreadReturnCode ret;
 475
 476     __asm__ volatile ("ldo %0(%%r30),%%r19\n"
 477                       "\tstw %%r3, 0(0,%%r19)\n"
 478                       "\tstw %%r4, 4(0,%%r19)\n"
 479                       "\tstw %%r5, 8(0,%%r19)\n"
 480                       "\tstw %%r6,12(0,%%r19)\n"
 481                       "\tstw %%r7,16(0,%%r19)\n"
 482                       "\tstw %%r8,20(0,%%r19)\n"
 483                       "\tstw %%r9,24(0,%%r19)\n"
 484                       "\tstw %%r10,28(0,%%r19)\n"
 485                       "\tstw %%r11,32(0,%%r19)\n"
 486                       "\tstw %%r12,36(0,%%r19)\n"
 487                       "\tstw %%r13,40(0,%%r19)\n"
 488                       "\tstw %%r14,44(0,%%r19)\n"
 489                       "\tstw %%r15,48(0,%%r19)\n"
 490                       "\tstw %%r16,52(0,%%r19)\n"
 491                       "\tstw %%r17,56(0,%%r19)\n"
 492                       "\tstw %%r18,60(0,%%r19)\n"
 493                       "\tldo 80(%%r19),%%r19\n"
 494                       "\tfstds %%fr12,-16(0,%%r19)\n"
 495                       "\tfstds %%fr13, -8(0,%%r19)\n"
 496                       "\tfstds %%fr14,  0(0,%%r19)\n"
 497                       "\tfstds %%fr15,  8(0,%%r19)\n"
 498                       "\tldo 32(%%r19),%%r19\n"
 499                       "\tfstds %%fr16,-16(0,%%r19)\n"
 500                       "\tfstds %%fr17, -8(0,%%r19)\n"
 501                       "\tfstds %%fr18,  0(0,%%r19)\n"
 502                       "\tfstds %%fr19,  8(0,%%r19)\n"
 503                       "\tldo 32(%%r19),%%r19\n"
 504                       "\tfstds %%fr20,-16(0,%%r19)\n"
 505                       "\tfstds %%fr21, -8(0,%%r19)\n" : :
 506                       "n" (-(116 * sizeof(long) + 10 * sizeof(double))) : "%r19"
 507                       );
 508
 509     f();
 510
 511     __asm__ volatile (".align 4\n"
 512                       "\t.EXPORT " STG_RETURN ",CODE\n"
 513                       "\t.EXPORT " STG_RETURN ",ENTRY,PRIV_LEV=3\n"
 514                       STG_RETURN "\n"
 515                       /* "\tldo %0(%%r3),%%r19\n" */
 516                       "\tldo %1(%%r30),%%r19\n"
 517                       "\tcopy %%r11, %0\n"  /* save R1 */
 518                       "\tldw  0(0,%%r19),%%r3\n"
 519                       "\tldw  4(0,%%r19),%%r4\n"
 520                       "\tldw  8(0,%%r19),%%r5\n"
 521                       "\tldw 12(0,%%r19),%%r6\n"
 522                       "\tldw 16(0,%%r19),%%r7\n"
 523                       "\tldw 20(0,%%r19),%%r8\n"
 524                       "\tldw 24(0,%%r19),%%r9\n"
 525                       "\tldw 28(0,%%r19),%%r10\n"
 526                       "\tldw 32(0,%%r19),%%r11\n"
 527                       "\tldw 36(0,%%r19),%%r12\n"
 528                       "\tldw 40(0,%%r19),%%r13\n"
 529                       "\tldw 44(0,%%r19),%%r14\n"
 530                       "\tldw 48(0,%%r19),%%r15\n"
 531                       "\tldw 52(0,%%r19),%%r16\n"
 532                       "\tldw 56(0,%%r19),%%r17\n"
 533                       "\tldw 60(0,%%r19),%%r18\n"
 534                       "\tldo 80(%%r19),%%r19\n"
 535                       "\tfldds -16(0,%%r19),%%fr12\n"
 536                       "\tfldds  -8(0,%%r19),%%fr13\n"
 537                       "\tfldds   0(0,%%r19),%%fr14\n"
 538                       "\tfldds   8(0,%%r19),%%fr15\n"
 539                       "\tldo 32(%%r19),%%r19\n"
 540                       "\tfldds -16(0,%%r19),%%fr16\n"
 541                       "\tfldds  -8(0,%%r19),%%fr17\n"
 542                       "\tfldds   0(0,%%r19),%%fr18\n"
 543                       "\tfldds   8(0,%%r19),%%fr19\n"
 544                       "\tldo 32(%%r19),%%r19\n"
 545                       "\tfldds -16(0,%%r19),%%fr20\n"
 546                       "\tfldds  -8(0,%%r19),%%fr21\n"
 547                          : "=r" (ret)
 548                          : "n" (-(116 * sizeof(long) + 10 * sizeof(double)))
 549                          : "%r19"
 550                       );
 551
 552     return ret;
 553 }
 554
 555 #endif /* hppa1_1_HOST_ARCH */
 556
 557 /* -----------------------------------------------------------------------------
 558    PowerPC architecture
 559
 560    Everything is in assembler, so we don't have to deal with GCC...
 561
 562    -------------------------------------------------------------------------- */
 563
 564 #ifdef powerpc_HOST_ARCH
 565
 566 extern StgThreadReturnCode StgRun(StgFunPtr f, StgRegTable *basereg);
 567
 568 #ifdef darwin_HOST_OS
 569 static void StgRunIsImplementedInAssembler(void)
 570 {
 571 #if HAVE_SUBSECTIONS_VIA_SYMBOLS
 572             // if the toolchain supports deadstripping, we have to
 573             // prevent it here (it tends to get confused here).
 574         __asm__ volatile (".no_dead_strip _StgRunIsImplementedInAssembler");
 575 #endif
 576         __asm__ volatile (
 577                 "\n.globl _StgRun\n"
 578                 "_StgRun:\n"
 579                 "\tmflr r0\n"
 580                 "\tbl saveFP # f14\n"
 581                 "\tstmw r13,-220(r1)\n"
 582                 "\tstwu r1,-%0(r1)\n"
 583                 "\tmr r27,r4\n" // BaseReg == r27
 584                 "\tmtctr r3\n"
 585                 "\tmr r12,r3\n"
 586                 "\tbctr\n"
 587                 ".globl _StgReturn\n"
 588                 "_StgReturn:\n"
 589                 "\tmr r3,r14\n"
 590                 "\tla r1,%0(r1)\n"
 591                 "\tlmw r13,-220(r1)\n"
 592                 "\tb restFP # f14\n"
 593         : : "i"(RESERVED_C_STACK_BYTES+224 /*stack frame size*/));
 594 }
 595 #else
 596
 597 // This version is for PowerPC Linux.
 598
 599 // Differences from the Darwin/Mac OS X version:
 600 // *) Different Assembler Syntax
 601 // *) Doesn't use Register Saving Helper Functions (although they exist somewhere)
 602 // *) We may not access positive stack offsets
 603 //    (no "Red Zone" as in the Darwin ABI)
 604 // *) The Link Register is saved to a different offset in the caller's stack frame
 605 //    (Linux: 4(r1), Darwin 8(r1))
 606
 607 static void StgRunIsImplementedInAssembler(void)
 608 {
 609         __asm__ volatile (
 610                 "\t.globl StgRun\n"
 611                 "\t.type StgRun,@function\n"
 612                 "StgRun:\n"
 613                 "\tmflr 0\n"
 614                 "\tstw 0,4(1)\n"
 615                 "\tmr 5,1\n"
 616                 "\tstwu 1,-%0(1)\n"
 617                 "\tstmw 13,-220(5)\n"
 618                 "\tstfd 14,-144(5)\n"
 619                 "\tstfd 15,-136(5)\n"
 620                 "\tstfd 16,-128(5)\n"
 621                 "\tstfd 17,-120(5)\n"
 622                 "\tstfd 18,-112(5)\n"
 623                 "\tstfd 19,-104(5)\n"
 624                 "\tstfd 20,-96(5)\n"
 625                 "\tstfd 21,-88(5)\n"
 626                 "\tstfd 22,-80(5)\n"
 627                 "\tstfd 23,-72(5)\n"
 628                 "\tstfd 24,-64(5)\n"
 629                 "\tstfd 25,-56(5)\n"
 630                 "\tstfd 26,-48(5)\n"
 631                 "\tstfd 27,-40(5)\n"
 632                 "\tstfd 28,-32(5)\n"
 633                 "\tstfd 29,-24(5)\n"
 634                 "\tstfd 30,-16(5)\n"
 635                 "\tstfd 31,-8(5)\n"
 636                 "\tmr 27,4\n"  // BaseReg == r27
 637                 "\tmtctr 3\n"
 638                 "\tmr 12,3\n"
 639                 "\tbctr\n"
 640                 ".globl StgReturn\n"
 641                 "\t.type StgReturn,@function\n"
 642                 "StgReturn:\n"
 643                 "\tmr 3,14\n"
 644                 "\tla 5,%0(1)\n"
 645                 "\tlmw 13,-220(5)\n"
 646                 "\tlfd 14,-144(5)\n"
 647                 "\tlfd 15,-136(5)\n"
 648                 "\tlfd 16,-128(5)\n"
 649                 "\tlfd 17,-120(5)\n"
 650                 "\tlfd 18,-112(5)\n"
 651                 "\tlfd 19,-104(5)\n"
 652                 "\tlfd 20,-96(5)\n"
 653                 "\tlfd 21,-88(5)\n"
 654                 "\tlfd 22,-80(5)\n"
 655                 "\tlfd 23,-72(5)\n"
 656                 "\tlfd 24,-64(5)\n"
 657                 "\tlfd 25,-56(5)\n"
 658                 "\tlfd 26,-48(5)\n"
 659                 "\tlfd 27,-40(5)\n"
 660                 "\tlfd 28,-32(5)\n"
 661                 "\tlfd 29,-24(5)\n"
 662                 "\tlfd 30,-16(5)\n"
 663                 "\tlfd 31,-8(5)\n"
 664                 "\tmr 1,5\n"
 665                 "\tlwz 0,4(1)\n"
 666                 "\tmtlr 0\n"
 667                 "\tblr\n"
 668         : : "i"(RESERVED_C_STACK_BYTES+224 /*stack frame size*/));
 669 }
 670 #endif
 671
 672 #endif
 673
 674 /* -----------------------------------------------------------------------------
 675    PowerPC 64 architecture
 676
 677    Everything is in assembler, so we don't have to deal with GCC...
 678
 679    -------------------------------------------------------------------------- */
 680
 681 #ifdef powerpc64_HOST_ARCH
 682
 683 #ifdef linux_HOST_OS
 684 extern StgThreadReturnCode StgRun(StgFunPtr f, StgRegTable *basereg);
 685
 686 static void StgRunIsImplementedInAssembler(void)
 687 {
 688         // r0 volatile
 689         // r1 stack pointer
 690         // r2 toc - needs to be saved
 691         // r3-r10 argument passing, volatile
 692         // r11, r12 very volatile (not saved across cross-module calls)
 693         // r13 thread local state (never modified, don't need to save)
 694         // r14-r31 callee-save
 695         __asm__ volatile (
 696                 ".section \".opd\",\"aw\"\n"
 697                 ".align 3\n"
 698                 ".globl StgRun\n"
 699                 "StgRun:\n"
 700                         "\t.quad\t.StgRun,.TOC.@tocbase,0\n"
 701                         "\t.size StgRun,24\n"
 702                 ".globl StgReturn\n"
 703                 "StgReturn:\n"
 704                         "\t.quad\t.StgReturn,.TOC.@tocbase,0\n"
 705                         "\t.size StgReturn,24\n"
 706                 ".previous\n"
 707                 ".globl .StgRun\n"
 708                 ".type .StgRun,@function\n"
 709                 ".StgRun:\n"
 710                         "\tmflr 0\n"
 711                         "\tmr 5, 1\n"
 712                         "\tstd 0, 16(1)\n"
 713                         "\tstdu 1, -%0(1)\n"
 714                         "\tstd 2, -296(5)\n"
 715                         "\tstd 14, -288(5)\n"
 716                         "\tstd 15, -280(5)\n"
 717                         "\tstd 16, -272(5)\n"
 718                         "\tstd 17, -264(5)\n"
 719                         "\tstd 18, -256(5)\n"
 720                         "\tstd 19, -248(5)\n"
 721                         "\tstd 20, -240(5)\n"
 722                         "\tstd 21, -232(5)\n"
 723                         "\tstd 22, -224(5)\n"
 724                         "\tstd 23, -216(5)\n"
 725                         "\tstd 24, -208(5)\n"
 726                         "\tstd 25, -200(5)\n"
 727                         "\tstd 26, -192(5)\n"
 728                         "\tstd 27, -184(5)\n"
 729                         "\tstd 28, -176(5)\n"
 730                         "\tstd 29, -168(5)\n"
 731                         "\tstd 30, -160(5)\n"
 732                         "\tstd 31, -152(5)\n"
 733                         "\tstfd 14, -144(5)\n"
 734                         "\tstfd 15, -136(5)\n"
 735                         "\tstfd 16, -128(5)\n"
 736                         "\tstfd 17, -120(5)\n"
 737                         "\tstfd 18, -112(5)\n"
 738                         "\tstfd 19, -104(5)\n"
 739                         "\tstfd 20, -96(5)\n"
 740                         "\tstfd 21, -88(5)\n"
 741                         "\tstfd 22, -80(5)\n"
 742                         "\tstfd 23, -72(5)\n"
 743                         "\tstfd 24, -64(5)\n"
 744                         "\tstfd 25, -56(5)\n"
 745                         "\tstfd 26, -48(5)\n"
 746                         "\tstfd 27, -40(5)\n"
 747                         "\tstfd 28, -32(5)\n"
 748                         "\tstfd 29, -24(5)\n"
 749                         "\tstfd 30, -16(5)\n"
 750                         "\tstfd 31, -8(5)\n"
 751                         "\tmr 27, 4\n"  // BaseReg == r27
 752                         "\tld 2, 8(3)\n"
 753                         "\tld 3, 0(3)\n"
 754                         "\tmtctr 3\n"
 755                         "\tbctr\n"
 756                 ".globl .StgReturn\n"
 757                 ".type .StgReturn,@function\n"
 758                 ".StgReturn:\n"
 759                         "\tmr 3,14\n"
 760                         "\tla 5, %0(1)\n" // load address == addi r5, r1, %0
 761                         "\tld 2, -296(5)\n"
 762                         "\tld 14, -288(5)\n"
 763                         "\tld 15, -280(5)\n"
 764                         "\tld 16, -272(5)\n"
 765                         "\tld 17, -264(5)\n"
 766                         "\tld 18, -256(5)\n"
 767                         "\tld 19, -248(5)\n"
 768                         "\tld 20, -240(5)\n"
 769                         "\tld 21, -232(5)\n"
 770                         "\tld 22, -224(5)\n"
 771                         "\tld 23, -216(5)\n"
 772                         "\tld 24, -208(5)\n"
 773                         "\tld 25, -200(5)\n"
 774                         "\tld 26, -192(5)\n"
 775                         "\tld 27, -184(5)\n"
 776                         "\tld 28, -176(5)\n"
 777                         "\tld 29, -168(5)\n"
 778                         "\tld 30, -160(5)\n"
 779                         "\tld 31, -152(5)\n"
 780                         "\tlfd 14, -144(5)\n"
 781                         "\tlfd 15, -136(5)\n"
 782                         "\tlfd 16, -128(5)\n"
 783                         "\tlfd 17, -120(5)\n"
 784                         "\tlfd 18, -112(5)\n"
 785                         "\tlfd 19, -104(5)\n"
 786                         "\tlfd 20, -96(5)\n"
 787                         "\tlfd 21, -88(5)\n"
 788                         "\tlfd 22, -80(5)\n"
 789                         "\tlfd 23, -72(5)\n"
 790                         "\tlfd 24, -64(5)\n"
 791                         "\tlfd 25, -56(5)\n"
 792                         "\tlfd 26, -48(5)\n"
 793                         "\tlfd 27, -40(5)\n"
 794                         "\tlfd 28, -32(5)\n"
 795                         "\tlfd 29, -24(5)\n"
 796                         "\tlfd 30, -16(5)\n"
 797                         "\tlfd 31, -8(5)\n"
 798                         "\tmr 1, 5\n"
 799                         "\tld 0, 16(1)\n"
 800                         "\tmtlr 0\n"
 801                         "\tblr\n"
 802         : : "i"(RESERVED_C_STACK_BYTES+304 /*stack frame size*/));
 803 }
 804 #else // linux_HOST_OS
 805 #error Only linux support for power64 right now.
 806 #endif
 807
 808 #endif
 809
 810 /* -----------------------------------------------------------------------------
 811    IA64 architecture
 812
 813    Again, in assembler - so we can fiddle with the register stack, and because
 814    gcc doesn't handle asm-clobbered callee-saves correctly.
 815
 816    loc0  - loc15: preserved locals
 817    loc16 - loc28: STG registers
 818            loc29: saved ar.pfs
 819            loc30: saved b0
 820            loc31: saved gp (gcc 3.3 uses this slot)
 821    -------------------------------------------------------------------------- */
 822
 823 #ifdef ia64_HOST_ARCH
 824
 825 /* the memory stack is rarely used, so 16K is excessive */
 826 #undef RESERVED_C_STACK_BYTES
 827 #define RESERVED_C_STACK_BYTES 1024
 828
 829 #if ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 3)) || (__GNUC__ > 3)
 830 /* gcc 3.3+: leave an extra slot for gp saves */
 831 #define LOCALS 32
 832 #else
 833 #define LOCALS 31
 834 #endif
 835
 836 static void StgRunIsImplementedInAssembler(void)
 837 {
 838     __asm__ volatile(
 839                 ".global StgRun\n"
 840                 "StgRun:\n"
 841                 "\talloc loc29 = ar.pfs, 0, %1, 8, 0\n" /* setup register frame */
 842                 "\tld8 r18 = [r32],8\n"                 /* get procedure address */
 843                 "\tadds sp = -%0, sp ;;\n"              /* setup stack */
 844                 "\tld8 gp = [r32]\n"                    /* get procedure GP */
 845                 "\tadds r16 = %0-(6*16), sp\n"
 846                 "\tadds r17 = %0-(5*16), sp ;;\n"
 847                 "\tstf.spill [r16] = f16,32\n"          /* spill callee-saved fp regs */
 848                 "\tstf.spill [r17] = f17,32\n"
 849                 "\tmov b6 = r18 ;;\n"                   /* set target address */
 850                 "\tstf.spill [r16] = f18,32\n"
 851                 "\tstf.spill [r17] = f19,32\n"
 852                 "\tmov loc30 = b0 ;;\n"                 /* save return address */
 853                 "\tstf.spill [r16] = f20,32\n"
 854                 "\tstf.spill [r17] = f21,32\n"
 855                 "\tbr.few b6 ;;\n"                      /* branch to function */
 856                 ".global StgReturn\n"
 857                 "StgReturn:\n"
 858                 "\tmov r8 = loc16\n"            /* return value in r8 */
 859                 "\tadds r16 = %0-(6*16), sp\n"
 860                 "\tadds r17 = %0-(5*16), sp ;;\n"
 861                 "\tldf.fill f16 = [r16],32\n"   /* start restoring fp regs */
 862                 "\tldf.fill f17 = [r17],32\n"
 863                 "\tmov ar.pfs = loc29 ;;\n"     /* restore register frame */
 864                 "\tldf.fill f18 = [r16],32\n"
 865                 "\tldf.fill f19 = [r17],32\n"
 866                 "\tmov b0 = loc30 ;;\n"         /* restore return address */
 867                 "\tldf.fill f20 = [r16],32\n"
 868                 "\tldf.fill f21 = [r17],32\n"
 869                 "\tadds sp = %0, sp\n"          /* restore stack */
 870                 "\tbr.ret.sptk.many b0 ;;\n"    /* return */
 871         : : "i"(RESERVED_C_STACK_BYTES + 6*16), "i"(LOCALS));
 872 }
 873
 874 #endif
 875
 876 #endif /* !USE_MINIINTERPRETER */