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 #ifdef darwin_TARGET_OS
 118 #define STG_GLOBAL ".globl "
 119 #else
 120 #define STG_GLOBAL ".global "
 121 #endif
 122
 123 StgThreadReturnCode
 124 StgRun(StgFunPtr f, StgRegTable *basereg) {
 125
 126     unsigned char space[ RESERVED_C_STACK_BYTES + 4*sizeof(void *) ];
 127     StgThreadReturnCode r;
 128
 129     __asm__ volatile (
 130         /*
 131          * save callee-saves registers on behalf of the STG code.
 132          */
 133         "movl %%esp, %%eax\n\t"
 134         "addl %4, %%eax\n\t"
 135         "movl %%ebx,0(%%eax)\n\t"
 136         "movl %%esi,4(%%eax)\n\t"
 137         "movl %%edi,8(%%eax)\n\t"
 138         "movl %%ebp,12(%%eax)\n\t"
 139         /*
 140          * Set BaseReg
 141          */
 142         "movl %3,%%ebx\n\t"
 143         /*
 144          * grab the function argument from the stack, and jump to it.
 145          */
 146         "movl %2,%%eax\n\t"
 147         "jmp *%%eax\n\t"
 148
 149         STG_GLOBAL STG_RETURN "\n"
 150         STG_RETURN ":\n\t"
 151
 152         "movl %%esi, %%eax\n\t"   /* Return value in R1  */
 153
 154         /*
 155          * restore callee-saves registers.  (Don't stomp on %%eax!)
 156          */
 157         "movl %%esp, %%edx\n\t"
 158         "addl %4, %%edx\n\t"
 159         "movl 0(%%edx),%%ebx\n\t"       /* restore the registers saved above */
 160         "movl 4(%%edx),%%esi\n\t"
 161         "movl 8(%%edx),%%edi\n\t"
 162         "movl 12(%%edx),%%ebp\n\t"
 163
 164       : "=&a" (r), "=m" (space)
 165       : "m" (f), "m" (basereg), "i" (RESERVED_C_STACK_BYTES)
 166       : "edx" /* stomps on %edx */
 167     );
 168
 169     return r;
 170 }
 171
 172 #endif
 173
 174 /* ----------------------------------------------------------------------------
 175    x86-64 is almost the same as plain x86.
 176
 177    I've done it using entirely inline assembler, because I couldn't
 178    get gcc to generate the correct subtraction from %rsp by using
 179    the local array variable trick.  It didn't seem to reserve
 180    enough space.  Oh well, it's not much harder this way.
 181
 182    ------------------------------------------------------------------------- */
 183
 184 #ifdef x86_64_HOST_ARCH
 185
 186 extern StgThreadReturnCode StgRun(StgFunPtr f, StgRegTable *basereg);
 187
 188 static void StgRunIsImplementedInAssembler(void)
 189 {
 190     __asm__ volatile (
 191         /*
 192          * save callee-saves registers on behalf of the STG code.
 193          */
 194         ".globl StgRun\n"
 195         "StgRun:\n\t"
 196         "subq %0, %%rsp\n\t"
 197         "movq %%rsp, %%rax\n\t"
 198         "addq %0-48, %%rax\n\t"
 199         "movq %%rbx,0(%%rax)\n\t"
 200         "movq %%rbp,8(%%rax)\n\t"
 201         "movq %%r12,16(%%rax)\n\t"
 202         "movq %%r13,24(%%rax)\n\t"
 203         "movq %%r14,32(%%rax)\n\t"
 204         "movq %%r15,40(%%rax)\n\t"
 205         /*
 206          * Set BaseReg
 207          */
 208         "movq %%rsi,%%rbx\n\t"
 209         /*
 210          * grab the function argument from the stack, and jump to it.
 211          */
 212         "movq %%rdi,%%rax\n\t"
 213         "jmp *%%rax\n\t"
 214
 215         ".global " STG_RETURN "\n"
 216         STG_RETURN ":\n\t"
 217
 218         "movq %%r13, %%rax\n\t"   /* Return value in R1  */
 219
 220         /*
 221          * restore callee-saves registers.  (Don't stomp on %%rax!)
 222          */
 223         "movq %%rsp, %%rdx\n\t"
 224         "addq %0-48, %%rdx\n\t"
 225         "movq 0(%%rdx),%%rbx\n\t"       /* restore the registers saved above */
 226         "movq 8(%%rdx),%%rbp\n\t"
 227         "movq 16(%%rdx),%%r12\n\t"
 228         "movq 24(%%rdx),%%r13\n\t"
 229         "movq 32(%%rdx),%%r14\n\t"
 230         "movq 40(%%rdx),%%r15\n\t"
 231         "addq %0, %%rsp\n\t"
 232         "retq"
 233
 234         : : "i"(RESERVED_C_STACK_BYTES+48+8 /*stack frame size*/));
 235     /*
 236        HACK alert!
 237
 238        The x86_64 ABI specifies that on a procedure call, %rsp is
 239        aligned on a 16-byte boundary + 8.  That is, the first
 240        argument on the stack after the return address will be
 241        16-byte aligned.
 242
 243        Which should be fine: RESERVED_C_STACK_BYTES+48 is a multiple
 244        of 16 bytes.
 245
 246        BUT... when we do a C-call from STG land, gcc likes to put the
 247        stack alignment adjustment in the prolog.  eg. if we're calling
 248        a function with arguments in regs, gcc will insert 'subq $8,%rsp'
 249        in the prolog, to keep %rsp aligned (the return address is 8
 250        bytes, remember).  The mangler throws away the prolog, so we
 251        lose the stack alignment.
 252
 253        The hack is to add this extra 8 bytes to our %rsp adjustment
 254        here, so that throughout STG code, %rsp is 16-byte aligned,
 255        ready for a C-call.
 256
 257        A quick way to see if this is wrong is to compile this code:
 258
 259           main = System.Exit.exitWith ExitSuccess
 260
 261        And run it with +RTS -sstderr.  The stats code in the RTS, in
 262        particular statsPrintf(), relies on the stack alignment because
 263        it saves the %xmm regs on the stack, so it'll fall over if the
 264        stack isn't aligned, and calling exitWith from Haskell invokes
 265        shutdownHaskellAndExit using a C call.
 266
 267        Future gcc releases will almost certainly break this hack...
 268     */
 269 }
 270
 271 #endif /* x86-64 */
 272
 273 /* -----------------------------------------------------------------------------
 274    Sparc architecture
 275
 276    --
 277    OLD COMMENT from GHC-3.02:
 278
 279    We want tailjumps to be calls, because `call xxx' is the only Sparc
 280    branch that allows an arbitrary label as a target.  (Gcc's ``goto
 281    *target'' construct ends up loading the label into a register and
 282    then jumping, at the cost of two extra instructions for the 32-bit
 283    load.)
 284
 285    When entering the threaded world, we stash our return address in a
 286    known location so that \tr{%i7} is available as an extra
 287    callee-saves register.  Of course, we have to restore this when
 288    coming out of the threaded world.
 289
 290    I hate this god-forsaken architecture.  Since the top of the
 291    reserved stack space is used for globals and the bottom is reserved
 292    for outgoing arguments, we have to stick our return address
 293    somewhere in the middle.  Currently, I'm allowing 100 extra
 294    outgoing arguments beyond the first 6.  --JSM
 295
 296    Updated info (GHC 4.06): we don't appear to use %i7 any more, so
 297    I'm not sure whether we still need to save it.  Incedentally, what
 298    does the last paragraph above mean when it says "the top of the
 299    stack is used for globals"?  What globals?  --SDM
 300
 301    Updated info (GHC 4.08.2): not saving %i7 any more (see below).
 302    -------------------------------------------------------------------------- */
 303
 304 #ifdef sparc_HOST_ARCH
 305
 306 StgThreadReturnCode
 307 StgRun(StgFunPtr f, StgRegTable *basereg) {
 308
 309     unsigned char space[RESERVED_C_STACK_BYTES];
 310 #if 0
 311     register void *i7 __asm__("%i7");
 312     ((void **)(space))[100] = i7;
 313 #endif
 314     f();
 315     __asm__ volatile (
 316             ".align 4\n"
 317             ".global " STG_RETURN "\n"
 318             STG_RETURN ":"
 319             : : : "l0","l1","l2","l3","l4","l5","l6","l7");
 320     /* we tell the C compiler that l0-l7 are clobbered on return to
 321      * StgReturn, otherwise it tries to use these to save eg. the
 322      * address of space[100] across the call.  The correct thing
 323      * to do would be to save all the callee-saves regs, but we
 324      * can't be bothered to do that.
 325      *
 326      * The code that gcc generates for this little fragment is now
 327      * terrible.  We could do much better by coding it directly in
 328      * assembler.
 329      */
 330 #if 0
 331     /* updated 4.08.2: we don't save %i7 in the middle of the reserved
 332      * space any more, since gcc tries to save its address across the
 333      * call to f(), this gets clobbered in STG land and we end up
 334      * dereferencing a bogus pointer in StgReturn.
 335      */
 336     __asm__ volatile ("ld %1,%0"
 337                       : "=r" (i7) : "m" (((void **)(space))[100]));
 338 #endif
 339     return (StgThreadReturnCode)R1.i;
 340 }
 341
 342 #endif
 343
 344 /* -----------------------------------------------------------------------------
 345    alpha architecture
 346
 347    "The stack pointer (SP) must at all times denote an address that has octaword
 348     alignment. (This restriction has the side effect that the in-memory portion
 349     of the argument list, if any, will start on an octaword boundary.) Note that
 350     the stack grows toward lower addresses. During a procedure invocation, SP
 351     can never be set to a value that is higher than the value of SP at entry to
 352     that procedure invocation.
 353
 354    "The contents of the stack, located above the portion of the argument list
 355     (if any) that is passed in memory, belong to the calling procedure. Because
 356     they are part of the calling procedure, they should not be read or written
 357     by the called procedure, except as specified by indirect arguments or
 358     language-controlled up-level references.
 359
 360    "The SP value might be used by the hardware when raising exceptions and
 361     asynchronous interrupts. It must be assumed that the contents of the stack
 362     below the current SP value and within the stack for the current thread are
 363     continually and unpredictably modified, as specified in the _Alpha
 364     Architecture Reference Manual_, and as a result of asynchronous software
 365     actions."
 366
 367    -- Compaq Computer Corporation, Houston. Tru64 UNIX Calling Standard for
 368       Alpha Systems, 5.1 edition, August 2000, section 3.2.1.  http://www.
 369       tru64unix.compaq.com/docs/base_doc/DOCUMENTATION/V51_PDF/ARH9MBTE.PDF
 370    -------------------------------------------------------------------------- */
 371
 372 #ifdef alpha_HOST_ARCH
 373
 374 StgThreadReturnCode
 375 StgRun(StgFunPtr f, StgRegTable *basereg)
 376 {
 377     register long   real_ra __asm__("$26"); volatile long   save_ra;
 378     register long   real_gp __asm__("$29"); volatile long   save_gp;
 379
 380     register long   real_s0 __asm__("$9" ); volatile long   save_s0;
 381     register long   real_s1 __asm__("$10"); volatile long   save_s1;
 382     register long   real_s2 __asm__("$11"); volatile long   save_s2;
 383     register long   real_s3 __asm__("$12"); volatile long   save_s3;
 384     register long   real_s4 __asm__("$13"); volatile long   save_s4;
 385     register long   real_s5 __asm__("$14"); volatile long   save_s5;
 386 #ifdef alpha_EXTRA_CAREFUL
 387     register long   real_s6 __asm__("$15"); volatile long   save_s6;
 388 #endif
 389
 390     register double real_f2 __asm__("$f2"); volatile double save_f2;
 391     register double real_f3 __asm__("$f3"); volatile double save_f3;
 392     register double real_f4 __asm__("$f4"); volatile double save_f4;
 393     register double real_f5 __asm__("$f5"); volatile double save_f5;
 394     register double real_f6 __asm__("$f6"); volatile double save_f6;
 395     register double real_f7 __asm__("$f7"); volatile double save_f7;
 396 #ifdef alpha_EXTRA_CAREFUL
 397     register double real_f8 __asm__("$f8"); volatile double save_f8;
 398     register double real_f9 __asm__("$f9"); volatile double save_f9;
 399 #endif
 400
 401     register StgFunPtr real_pv __asm__("$27");
 402
 403     StgThreadReturnCode ret;
 404
 405     save_ra = real_ra;
 406     save_gp = real_gp;
 407
 408     save_s0 = real_s0;
 409     save_s1 = real_s1;
 410     save_s2 = real_s2;
 411     save_s3 = real_s3;
 412     save_s4 = real_s4;
 413     save_s5 = real_s5;
 414 #ifdef alpha_EXTRA_CAREFUL
 415     save_s6 = real_s6;
 416 #endif
 417
 418     save_f2 = real_f2;
 419     save_f3 = real_f3;
 420     save_f4 = real_f4;
 421     save_f5 = real_f5;
 422     save_f6 = real_f6;
 423     save_f7 = real_f7;
 424 #ifdef alpha_EXTRA_CAREFUL
 425     save_f8 = real_f8;
 426     save_f9 = real_f9;
 427 #endif
 428
 429     real_pv = f;
 430
 431     __asm__ volatile(   "lda $30,-%0($30)"      "\n"
 432                 "\t"    "jmp ($27)"             "\n"
 433                 "\t"    ".align 3"              "\n"
 434                 ".globl " STG_RETURN            "\n"
 435                 STG_RETURN ":"                  "\n"
 436                 "\t"    "lda $30,%0($30)"       "\n"
 437                 : : "K" (RESERVED_C_STACK_BYTES));
 438
 439     ret = real_s5;
 440
 441     real_s0 = save_s0;
 442     real_s1 = save_s1;
 443     real_s2 = save_s2;
 444     real_s3 = save_s3;
 445     real_s4 = save_s4;
 446     real_s5 = save_s5;
 447 #ifdef alpha_EXTRA_CAREFUL
 448     real_s6 = save_s6;
 449 #endif
 450
 451     real_f2 = save_f2;
 452     real_f3 = save_f3;
 453     real_f4 = save_f4;
 454     real_f5 = save_f5;
 455     real_f6 = save_f6;
 456     real_f7 = save_f7;
 457 #ifdef alpha_EXTRA_CAREFUL
 458     real_f8 = save_f8;
 459     real_f9 = save_f9;
 460 #endif
 461
 462     real_ra = save_ra;
 463     real_gp = save_gp;
 464
 465     return ret;
 466 }
 467
 468 #endif /* alpha_HOST_ARCH */
 469
 470 /* -----------------------------------------------------------------------------
 471    HP-PA architecture
 472    -------------------------------------------------------------------------- */
 473
 474 #ifdef hppa1_1_HOST_ARCH
 475
 476 StgThreadReturnCode
 477 StgRun(StgFunPtr f, StgRegTable *basereg)
 478 {
 479     StgChar space[RESERVED_C_STACK_BYTES+16*sizeof(long)+10*sizeof(double)];
 480     StgThreadReturnCode ret;
 481
 482     __asm__ volatile ("ldo %0(%%r30),%%r19\n"
 483                       "\tstw %%r3, 0(0,%%r19)\n"
 484                       "\tstw %%r4, 4(0,%%r19)\n"
 485                       "\tstw %%r5, 8(0,%%r19)\n"
 486                       "\tstw %%r6,12(0,%%r19)\n"
 487                       "\tstw %%r7,16(0,%%r19)\n"
 488                       "\tstw %%r8,20(0,%%r19)\n"
 489                       "\tstw %%r9,24(0,%%r19)\n"
 490                       "\tstw %%r10,28(0,%%r19)\n"
 491                       "\tstw %%r11,32(0,%%r19)\n"
 492                       "\tstw %%r12,36(0,%%r19)\n"
 493                       "\tstw %%r13,40(0,%%r19)\n"
 494                       "\tstw %%r14,44(0,%%r19)\n"
 495                       "\tstw %%r15,48(0,%%r19)\n"
 496                       "\tstw %%r16,52(0,%%r19)\n"
 497                       "\tstw %%r17,56(0,%%r19)\n"
 498                       "\tstw %%r18,60(0,%%r19)\n"
 499                       "\tldo 80(%%r19),%%r19\n"
 500                       "\tfstds %%fr12,-16(0,%%r19)\n"
 501                       "\tfstds %%fr13, -8(0,%%r19)\n"
 502                       "\tfstds %%fr14,  0(0,%%r19)\n"
 503                       "\tfstds %%fr15,  8(0,%%r19)\n"
 504                       "\tldo 32(%%r19),%%r19\n"
 505                       "\tfstds %%fr16,-16(0,%%r19)\n"
 506                       "\tfstds %%fr17, -8(0,%%r19)\n"
 507                       "\tfstds %%fr18,  0(0,%%r19)\n"
 508                       "\tfstds %%fr19,  8(0,%%r19)\n"
 509                       "\tldo 32(%%r19),%%r19\n"
 510                       "\tfstds %%fr20,-16(0,%%r19)\n"
 511                       "\tfstds %%fr21, -8(0,%%r19)\n" : :
 512                       "n" (-(116 * sizeof(long) + 10 * sizeof(double))) : "%r19"
 513                       );
 514
 515     f();
 516
 517     __asm__ volatile (".align 4\n"
 518                       "\t.EXPORT " STG_RETURN ",CODE\n"
 519                       "\t.EXPORT " STG_RETURN ",ENTRY,PRIV_LEV=3\n"
 520                       STG_RETURN "\n"
 521                       /* "\tldo %0(%%r3),%%r19\n" */
 522                       "\tldo %1(%%r30),%%r19\n"
 523                       "\tcopy %%r11, %0\n"  /* save R1 */
 524                       "\tldw  0(0,%%r19),%%r3\n"
 525                       "\tldw  4(0,%%r19),%%r4\n"
 526                       "\tldw  8(0,%%r19),%%r5\n"
 527                       "\tldw 12(0,%%r19),%%r6\n"
 528                       "\tldw 16(0,%%r19),%%r7\n"
 529                       "\tldw 20(0,%%r19),%%r8\n"
 530                       "\tldw 24(0,%%r19),%%r9\n"
 531                       "\tldw 28(0,%%r19),%%r10\n"
 532                       "\tldw 32(0,%%r19),%%r11\n"
 533                       "\tldw 36(0,%%r19),%%r12\n"
 534                       "\tldw 40(0,%%r19),%%r13\n"
 535                       "\tldw 44(0,%%r19),%%r14\n"
 536                       "\tldw 48(0,%%r19),%%r15\n"
 537                       "\tldw 52(0,%%r19),%%r16\n"
 538                       "\tldw 56(0,%%r19),%%r17\n"
 539                       "\tldw 60(0,%%r19),%%r18\n"
 540                       "\tldo 80(%%r19),%%r19\n"
 541                       "\tfldds -16(0,%%r19),%%fr12\n"
 542                       "\tfldds  -8(0,%%r19),%%fr13\n"
 543                       "\tfldds   0(0,%%r19),%%fr14\n"
 544                       "\tfldds   8(0,%%r19),%%fr15\n"
 545                       "\tldo 32(%%r19),%%r19\n"
 546                       "\tfldds -16(0,%%r19),%%fr16\n"
 547                       "\tfldds  -8(0,%%r19),%%fr17\n"
 548                       "\tfldds   0(0,%%r19),%%fr18\n"
 549                       "\tfldds   8(0,%%r19),%%fr19\n"
 550                       "\tldo 32(%%r19),%%r19\n"
 551                       "\tfldds -16(0,%%r19),%%fr20\n"
 552                       "\tfldds  -8(0,%%r19),%%fr21\n"
 553                          : "=r" (ret)
 554                          : "n" (-(116 * sizeof(long) + 10 * sizeof(double)))
 555                          : "%r19"
 556                       );
 557
 558     return ret;
 559 }
 560
 561 #endif /* hppa1_1_HOST_ARCH */
 562
 563 /* -----------------------------------------------------------------------------
 564    PowerPC architecture
 565
 566    Everything is in assembler, so we don't have to deal with GCC...
 567
 568    -------------------------------------------------------------------------- */
 569
 570 #ifdef powerpc_HOST_ARCH
 571
 572 extern StgThreadReturnCode StgRun(StgFunPtr f, StgRegTable *basereg);
 573
 574 #ifdef darwin_HOST_OS
 575 static void StgRunIsImplementedInAssembler(void)
 576 {
 577 #if HAVE_SUBSECTIONS_VIA_SYMBOLS
 578             // if the toolchain supports deadstripping, we have to
 579             // prevent it here (it tends to get confused here).
 580         __asm__ volatile (".no_dead_strip _StgRunIsImplementedInAssembler");
 581 #endif
 582         __asm__ volatile (
 583                 "\n.globl _StgRun\n"
 584                 "_StgRun:\n"
 585                 "\tmflr r0\n"
 586                 "\tbl saveFP # f14\n"
 587                 "\tstmw r13,-220(r1)\n"
 588                 "\tstwu r1,-%0(r1)\n"
 589                 "\tmr r27,r4\n" // BaseReg == r27
 590                 "\tmtctr r3\n"
 591                 "\tmr r12,r3\n"
 592                 "\tbctr\n"
 593                 ".globl _StgReturn\n"
 594                 "_StgReturn:\n"
 595                 "\tmr r3,r14\n"
 596                 "\tla r1,%0(r1)\n"
 597                 "\tlmw r13,-220(r1)\n"
 598                 "\tb restFP # f14\n"
 599         : : "i"(RESERVED_C_STACK_BYTES+224 /*stack frame size*/));
 600 }
 601 #else
 602
 603 // This version is for PowerPC Linux.
 604
 605 // Differences from the Darwin/Mac OS X version:
 606 // *) Different Assembler Syntax
 607 // *) Doesn't use Register Saving Helper Functions (although they exist somewhere)
 608 // *) We may not access positive stack offsets
 609 //    (no "Red Zone" as in the Darwin ABI)
 610 // *) The Link Register is saved to a different offset in the caller's stack frame
 611 //    (Linux: 4(r1), Darwin 8(r1))
 612
 613 static void StgRunIsImplementedInAssembler(void)
 614 {
 615         __asm__ volatile (
 616                 "\t.globl StgRun\n"
 617                 "\t.type StgRun,@function\n"
 618                 "StgRun:\n"
 619                 "\tmflr 0\n"
 620                 "\tstw 0,4(1)\n"
 621                 "\tmr 5,1\n"
 622                 "\tstwu 1,-%0(1)\n"
 623                 "\tstmw 13,-220(5)\n"
 624                 "\tstfd 14,-144(5)\n"
 625                 "\tstfd 15,-136(5)\n"
 626                 "\tstfd 16,-128(5)\n"
 627                 "\tstfd 17,-120(5)\n"
 628                 "\tstfd 18,-112(5)\n"
 629                 "\tstfd 19,-104(5)\n"
 630                 "\tstfd 20,-96(5)\n"
 631                 "\tstfd 21,-88(5)\n"
 632                 "\tstfd 22,-80(5)\n"
 633                 "\tstfd 23,-72(5)\n"
 634                 "\tstfd 24,-64(5)\n"
 635                 "\tstfd 25,-56(5)\n"
 636                 "\tstfd 26,-48(5)\n"
 637                 "\tstfd 27,-40(5)\n"
 638                 "\tstfd 28,-32(5)\n"
 639                 "\tstfd 29,-24(5)\n"
 640                 "\tstfd 30,-16(5)\n"
 641                 "\tstfd 31,-8(5)\n"
 642                 "\tmr 27,4\n"  // BaseReg == r27
 643                 "\tmtctr 3\n"
 644                 "\tmr 12,3\n"
 645                 "\tbctr\n"
 646                 ".globl StgReturn\n"
 647                 "\t.type StgReturn,@function\n"
 648                 "StgReturn:\n"
 649                 "\tmr 3,14\n"
 650                 "\tla 5,%0(1)\n"
 651                 "\tlmw 13,-220(5)\n"
 652                 "\tlfd 14,-144(5)\n"
 653                 "\tlfd 15,-136(5)\n"
 654                 "\tlfd 16,-128(5)\n"
 655                 "\tlfd 17,-120(5)\n"
 656                 "\tlfd 18,-112(5)\n"
 657                 "\tlfd 19,-104(5)\n"
 658                 "\tlfd 20,-96(5)\n"
 659                 "\tlfd 21,-88(5)\n"
 660                 "\tlfd 22,-80(5)\n"
 661                 "\tlfd 23,-72(5)\n"
 662                 "\tlfd 24,-64(5)\n"
 663                 "\tlfd 25,-56(5)\n"
 664                 "\tlfd 26,-48(5)\n"
 665                 "\tlfd 27,-40(5)\n"
 666                 "\tlfd 28,-32(5)\n"
 667                 "\tlfd 29,-24(5)\n"
 668                 "\tlfd 30,-16(5)\n"
 669                 "\tlfd 31,-8(5)\n"
 670                 "\tmr 1,5\n"
 671                 "\tlwz 0,4(1)\n"
 672                 "\tmtlr 0\n"
 673                 "\tblr\n"
 674         : : "i"(RESERVED_C_STACK_BYTES+224 /*stack frame size*/));
 675 }
 676 #endif
 677
 678 #endif
 679
 680 /* -----------------------------------------------------------------------------
 681    PowerPC 64 architecture
 682
 683    Everything is in assembler, so we don't have to deal with GCC...
 684
 685    -------------------------------------------------------------------------- */
 686
 687 #ifdef powerpc64_HOST_ARCH
 688
 689 #ifdef linux_HOST_OS
 690 extern StgThreadReturnCode StgRun(StgFunPtr f, StgRegTable *basereg);
 691
 692 static void StgRunIsImplementedInAssembler(void)
 693 {
 694         // r0 volatile
 695         // r1 stack pointer
 696         // r2 toc - needs to be saved
 697         // r3-r10 argument passing, volatile
 698         // r11, r12 very volatile (not saved across cross-module calls)
 699         // r13 thread local state (never modified, don't need to save)
 700         // r14-r31 callee-save
 701         __asm__ volatile (
 702                 ".section \".opd\",\"aw\"\n"
 703                 ".align 3\n"
 704                 ".globl StgRun\n"
 705                 "StgRun:\n"
 706                         "\t.quad\t.StgRun,.TOC.@tocbase,0\n"
 707                         "\t.size StgRun,24\n"
 708                 ".globl StgReturn\n"
 709                 "StgReturn:\n"
 710                         "\t.quad\t.StgReturn,.TOC.@tocbase,0\n"
 711                         "\t.size StgReturn,24\n"
 712                 ".previous\n"
 713                 ".globl .StgRun\n"
 714                 ".type .StgRun,@function\n"
 715                 ".StgRun:\n"
 716                         "\tmflr 0\n"
 717                         "\tmr 5, 1\n"
 718                         "\tstd 0, 16(1)\n"
 719                         "\tstdu 1, -%0(1)\n"
 720                         "\tstd 2, -296(5)\n"
 721                         "\tstd 14, -288(5)\n"
 722                         "\tstd 15, -280(5)\n"
 723                         "\tstd 16, -272(5)\n"
 724                         "\tstd 17, -264(5)\n"
 725                         "\tstd 18, -256(5)\n"
 726                         "\tstd 19, -248(5)\n"
 727                         "\tstd 20, -240(5)\n"
 728                         "\tstd 21, -232(5)\n"
 729                         "\tstd 22, -224(5)\n"
 730                         "\tstd 23, -216(5)\n"
 731                         "\tstd 24, -208(5)\n"
 732                         "\tstd 25, -200(5)\n"
 733                         "\tstd 26, -192(5)\n"
 734                         "\tstd 27, -184(5)\n"
 735                         "\tstd 28, -176(5)\n"
 736                         "\tstd 29, -168(5)\n"
 737                         "\tstd 30, -160(5)\n"
 738                         "\tstd 31, -152(5)\n"
 739                         "\tstfd 14, -144(5)\n"
 740                         "\tstfd 15, -136(5)\n"
 741                         "\tstfd 16, -128(5)\n"
 742                         "\tstfd 17, -120(5)\n"
 743                         "\tstfd 18, -112(5)\n"
 744                         "\tstfd 19, -104(5)\n"
 745                         "\tstfd 20, -96(5)\n"
 746                         "\tstfd 21, -88(5)\n"
 747                         "\tstfd 22, -80(5)\n"
 748                         "\tstfd 23, -72(5)\n"
 749                         "\tstfd 24, -64(5)\n"
 750                         "\tstfd 25, -56(5)\n"
 751                         "\tstfd 26, -48(5)\n"
 752                         "\tstfd 27, -40(5)\n"
 753                         "\tstfd 28, -32(5)\n"
 754                         "\tstfd 29, -24(5)\n"
 755                         "\tstfd 30, -16(5)\n"
 756                         "\tstfd 31, -8(5)\n"
 757                         "\tmr 27, 4\n"  // BaseReg == r27
 758                         "\tld 2, 8(3)\n"
 759                         "\tld 3, 0(3)\n"
 760                         "\tmtctr 3\n"
 761                         "\tbctr\n"
 762                 ".globl .StgReturn\n"
 763                 ".type .StgReturn,@function\n"
 764                 ".StgReturn:\n"
 765                         "\tmr 3,14\n"
 766                         "\tla 5, %0(1)\n" // load address == addi r5, r1, %0
 767                         "\tld 2, -296(5)\n"
 768                         "\tld 14, -288(5)\n"
 769                         "\tld 15, -280(5)\n"
 770                         "\tld 16, -272(5)\n"
 771                         "\tld 17, -264(5)\n"
 772                         "\tld 18, -256(5)\n"
 773                         "\tld 19, -248(5)\n"
 774                         "\tld 20, -240(5)\n"
 775                         "\tld 21, -232(5)\n"
 776                         "\tld 22, -224(5)\n"
 777                         "\tld 23, -216(5)\n"
 778                         "\tld 24, -208(5)\n"
 779                         "\tld 25, -200(5)\n"
 780                         "\tld 26, -192(5)\n"
 781                         "\tld 27, -184(5)\n"
 782                         "\tld 28, -176(5)\n"
 783                         "\tld 29, -168(5)\n"
 784                         "\tld 30, -160(5)\n"
 785                         "\tld 31, -152(5)\n"
 786                         "\tlfd 14, -144(5)\n"
 787                         "\tlfd 15, -136(5)\n"
 788                         "\tlfd 16, -128(5)\n"
 789                         "\tlfd 17, -120(5)\n"
 790                         "\tlfd 18, -112(5)\n"
 791                         "\tlfd 19, -104(5)\n"
 792                         "\tlfd 20, -96(5)\n"
 793                         "\tlfd 21, -88(5)\n"
 794                         "\tlfd 22, -80(5)\n"
 795                         "\tlfd 23, -72(5)\n"
 796                         "\tlfd 24, -64(5)\n"
 797                         "\tlfd 25, -56(5)\n"
 798                         "\tlfd 26, -48(5)\n"
 799                         "\tlfd 27, -40(5)\n"
 800                         "\tlfd 28, -32(5)\n"
 801                         "\tlfd 29, -24(5)\n"
 802                         "\tlfd 30, -16(5)\n"
 803                         "\tlfd 31, -8(5)\n"
 804                         "\tmr 1, 5\n"
 805                         "\tld 0, 16(1)\n"
 806                         "\tmtlr 0\n"
 807                         "\tblr\n"
 808         : : "i"(RESERVED_C_STACK_BYTES+304 /*stack frame size*/));
 809 }
 810 #else // linux_HOST_OS
 811 #error Only linux support for power64 right now.
 812 #endif
 813
 814 #endif
 815
 816 /* -----------------------------------------------------------------------------
 817    IA64 architecture
 818
 819    Again, in assembler - so we can fiddle with the register stack, and because
 820    gcc doesn't handle asm-clobbered callee-saves correctly.
 821
 822    loc0  - loc15: preserved locals
 823    loc16 - loc28: STG registers
 824            loc29: saved ar.pfs
 825            loc30: saved b0
 826            loc31: saved gp (gcc 3.3 uses this slot)
 827    -------------------------------------------------------------------------- */
 828
 829 #ifdef ia64_HOST_ARCH
 830
 831 /* the memory stack is rarely used, so 16K is excessive */
 832 #undef RESERVED_C_STACK_BYTES
 833 #define RESERVED_C_STACK_BYTES 1024
 834
 835 #if ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 3)) || (__GNUC__ > 3)
 836 /* gcc 3.3+: leave an extra slot for gp saves */
 837 #define LOCALS 32
 838 #else
 839 #define LOCALS 31
 840 #endif
 841
 842 static void StgRunIsImplementedInAssembler(void)
 843 {
 844     __asm__ volatile(
 845                 ".global StgRun\n"
 846                 "StgRun:\n"
 847                 "\talloc loc29 = ar.pfs, 0, %1, 8, 0\n" /* setup register frame */
 848                 "\tld8 r18 = [r32],8\n"                 /* get procedure address */
 849                 "\tadds sp = -%0, sp ;;\n"              /* setup stack */
 850                 "\tld8 gp = [r32]\n"                    /* get procedure GP */
 851                 "\tadds r16 = %0-(6*16), sp\n"
 852                 "\tadds r17 = %0-(5*16), sp ;;\n"
 853                 "\tstf.spill [r16] = f16,32\n"          /* spill callee-saved fp regs */
 854                 "\tstf.spill [r17] = f17,32\n"
 855                 "\tmov b6 = r18 ;;\n"                   /* set target address */
 856                 "\tstf.spill [r16] = f18,32\n"
 857                 "\tstf.spill [r17] = f19,32\n"
 858                 "\tmov loc30 = b0 ;;\n"                 /* save return address */
 859                 "\tstf.spill [r16] = f20,32\n"
 860                 "\tstf.spill [r17] = f21,32\n"
 861                 "\tbr.few b6 ;;\n"                      /* branch to function */
 862                 ".global StgReturn\n"
 863                 "StgReturn:\n"
 864                 "\tmov r8 = loc16\n"            /* return value in r8 */
 865                 "\tadds r16 = %0-(6*16), sp\n"
 866                 "\tadds r17 = %0-(5*16), sp ;;\n"
 867                 "\tldf.fill f16 = [r16],32\n"   /* start restoring fp regs */
 868                 "\tldf.fill f17 = [r17],32\n"
 869                 "\tmov ar.pfs = loc29 ;;\n"     /* restore register frame */
 870                 "\tldf.fill f18 = [r16],32\n"
 871                 "\tldf.fill f19 = [r17],32\n"
 872                 "\tmov b0 = loc30 ;;\n"         /* restore return address */
 873                 "\tldf.fill f20 = [r16],32\n"
 874                 "\tldf.fill f21 = [r17],32\n"
 875                 "\tadds sp = %0, sp\n"          /* restore stack */
 876                 "\tbr.ret.sptk.many b0 ;;\n"    /* return */
 877         : : "i"(RESERVED_C_STACK_BYTES + 6*16), "i"(LOCALS));
 878 }
 879
 880 #endif
 881
 882 #endif /* !USE_MINIINTERPRETER */