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
 145          */
 146         "movl %2,%%eax\n\t"
 147
 148 #if darwin_TARGET_OS
 149         /*
 150          * Darwin: keep the stack aligned
 151          */
 152         "subl $12,%%esp\n\t"
 153 #endif
 154
 155         /*
 156          * jump to it
 157          */
 158         "jmp *%%eax\n\t"
 159
 160         STG_GLOBAL STG_RETURN "\n"
 161         STG_RETURN ":\n\t"
 162
 163 #if darwin_TARGET_OS
 164         /*
 165          * Darwin: keep the stack aligned
 166          */
 167         "addl $12,%%esp\n\t"
 168 #endif
 169
 170         "movl %%esi, %%eax\n\t"   /* Return value in R1  */
 171
 172         /*
 173          * restore callee-saves registers.  (Don't stomp on %%eax!)
 174          */
 175         "movl %%esp, %%edx\n\t"
 176         "addl %4, %%edx\n\t"
 177         "movl 0(%%edx),%%ebx\n\t"       /* restore the registers saved above */
 178         "movl 4(%%edx),%%esi\n\t"
 179         "movl 8(%%edx),%%edi\n\t"
 180         "movl 12(%%edx),%%ebp\n\t"
 181
 182       : "=&a" (r), "=m" (space)
 183       : "m" (f), "m" (basereg), "i" (RESERVED_C_STACK_BYTES)
 184       : "edx" /* stomps on %edx */
 185     );
 186
 187     return r;
 188 }
 189
 190 #endif
 191
 192 /* ----------------------------------------------------------------------------
 193    x86-64 is almost the same as plain x86.
 194
 195    I've done it using entirely inline assembler, because I couldn't
 196    get gcc to generate the correct subtraction from %rsp by using
 197    the local array variable trick.  It didn't seem to reserve
 198    enough space.  Oh well, it's not much harder this way.
 199
 200    ------------------------------------------------------------------------- */
 201
 202 #ifdef x86_64_HOST_ARCH
 203
 204 extern StgThreadReturnCode StgRun(StgFunPtr f, StgRegTable *basereg);
 205
 206 static void StgRunIsImplementedInAssembler(void)
 207 {
 208     __asm__ volatile (
 209         /*
 210          * save callee-saves registers on behalf of the STG code.
 211          */
 212         ".globl StgRun\n"
 213         "StgRun:\n\t"
 214         "subq %0, %%rsp\n\t"
 215         "movq %%rsp, %%rax\n\t"
 216         "addq %0-48, %%rax\n\t"
 217         "movq %%rbx,0(%%rax)\n\t"
 218         "movq %%rbp,8(%%rax)\n\t"
 219         "movq %%r12,16(%%rax)\n\t"
 220         "movq %%r13,24(%%rax)\n\t"
 221         "movq %%r14,32(%%rax)\n\t"
 222         "movq %%r15,40(%%rax)\n\t"
 223         /*
 224          * Set BaseReg
 225          */
 226         "movq %%rsi,%%rbx\n\t"
 227         /*
 228          * grab the function argument from the stack, and jump to it.
 229          */
 230         "movq %%rdi,%%rax\n\t"
 231         "jmp *%%rax\n\t"
 232
 233         ".global " STG_RETURN "\n"
 234         STG_RETURN ":\n\t"
 235
 236         "movq %%r13, %%rax\n\t"   /* Return value in R1  */
 237
 238         /*
 239          * restore callee-saves registers.  (Don't stomp on %%rax!)
 240          */
 241         "movq %%rsp, %%rdx\n\t"
 242         "addq %0-48, %%rdx\n\t"
 243         "movq 0(%%rdx),%%rbx\n\t"       /* restore the registers saved above */
 244         "movq 8(%%rdx),%%rbp\n\t"
 245         "movq 16(%%rdx),%%r12\n\t"
 246         "movq 24(%%rdx),%%r13\n\t"
 247         "movq 32(%%rdx),%%r14\n\t"
 248         "movq 40(%%rdx),%%r15\n\t"
 249         "addq %0, %%rsp\n\t"
 250         "retq"
 251
 252         : : "i"(RESERVED_C_STACK_BYTES+48+8 /*stack frame size*/));
 253     /*
 254        HACK alert!
 255
 256        The x86_64 ABI specifies that on a procedure call, %rsp is
 257        aligned on a 16-byte boundary + 8.  That is, the first
 258        argument on the stack after the return address will be
 259        16-byte aligned.
 260
 261        Which should be fine: RESERVED_C_STACK_BYTES+48 is a multiple
 262        of 16 bytes.
 263
 264        BUT... when we do a C-call from STG land, gcc likes to put the
 265        stack alignment adjustment in the prolog.  eg. if we're calling
 266        a function with arguments in regs, gcc will insert 'subq $8,%rsp'
 267        in the prolog, to keep %rsp aligned (the return address is 8
 268        bytes, remember).  The mangler throws away the prolog, so we
 269        lose the stack alignment.
 270
 271        The hack is to add this extra 8 bytes to our %rsp adjustment
 272        here, so that throughout STG code, %rsp is 16-byte aligned,
 273        ready for a C-call.
 274
 275        A quick way to see if this is wrong is to compile this code:
 276
 277           main = System.Exit.exitWith ExitSuccess
 278
 279        And run it with +RTS -sstderr.  The stats code in the RTS, in
 280        particular statsPrintf(), relies on the stack alignment because
 281        it saves the %xmm regs on the stack, so it'll fall over if the
 282        stack isn't aligned, and calling exitWith from Haskell invokes
 283        shutdownHaskellAndExit using a C call.
 284
 285        Future gcc releases will almost certainly break this hack...
 286     */
 287 }
 288
 289 #endif /* x86-64 */
 290
 291 /* -----------------------------------------------------------------------------
 292    Sparc architecture
 293
 294    --
 295    OLD COMMENT from GHC-3.02:
 296
 297    We want tailjumps to be calls, because `call xxx' is the only Sparc
 298    branch that allows an arbitrary label as a target.  (Gcc's ``goto
 299    *target'' construct ends up loading the label into a register and
 300    then jumping, at the cost of two extra instructions for the 32-bit
 301    load.)
 302
 303    When entering the threaded world, we stash our return address in a
 304    known location so that \tr{%i7} is available as an extra
 305    callee-saves register.  Of course, we have to restore this when
 306    coming out of the threaded world.
 307
 308    I hate this god-forsaken architecture.  Since the top of the
 309    reserved stack space is used for globals and the bottom is reserved
 310    for outgoing arguments, we have to stick our return address
 311    somewhere in the middle.  Currently, I'm allowing 100 extra
 312    outgoing arguments beyond the first 6.  --JSM
 313
 314    Updated info (GHC 4.06): we don't appear to use %i7 any more, so
 315    I'm not sure whether we still need to save it.  Incedentally, what
 316    does the last paragraph above mean when it says "the top of the
 317    stack is used for globals"?  What globals?  --SDM
 318
 319    Updated info (GHC 4.08.2): not saving %i7 any more (see below).
 320    -------------------------------------------------------------------------- */
 321
 322 #ifdef sparc_HOST_ARCH
 323
 324 StgThreadReturnCode
 325 StgRun(StgFunPtr f, StgRegTable *basereg) {
 326
 327     unsigned char space[RESERVED_C_STACK_BYTES];
 328 #if 0
 329     register void *i7 __asm__("%i7");
 330     ((void **)(space))[100] = i7;
 331 #endif
 332     f();
 333     __asm__ volatile (
 334             ".align 4\n"
 335             ".global " STG_RETURN "\n"
 336             STG_RETURN ":"
 337             : : : "l0","l1","l2","l3","l4","l5","l6","l7");
 338     /* we tell the C compiler that l0-l7 are clobbered on return to
 339      * StgReturn, otherwise it tries to use these to save eg. the
 340      * address of space[100] across the call.  The correct thing
 341      * to do would be to save all the callee-saves regs, but we
 342      * can't be bothered to do that.
 343      *
 344      * The code that gcc generates for this little fragment is now
 345      * terrible.  We could do much better by coding it directly in
 346      * assembler.
 347      */
 348 #if 0
 349     /* updated 4.08.2: we don't save %i7 in the middle of the reserved
 350      * space any more, since gcc tries to save its address across the
 351      * call to f(), this gets clobbered in STG land and we end up
 352      * dereferencing a bogus pointer in StgReturn.
 353      */
 354     __asm__ volatile ("ld %1,%0"
 355                       : "=r" (i7) : "m" (((void **)(space))[100]));
 356 #endif
 357     return (StgThreadReturnCode)R1.i;
 358 }
 359
 360 #endif
 361
 362 /* -----------------------------------------------------------------------------
 363    alpha architecture
 364
 365    "The stack pointer (SP) must at all times denote an address that has octaword
 366     alignment. (This restriction has the side effect that the in-memory portion
 367     of the argument list, if any, will start on an octaword boundary.) Note that
 368     the stack grows toward lower addresses. During a procedure invocation, SP
 369     can never be set to a value that is higher than the value of SP at entry to
 370     that procedure invocation.
 371
 372    "The contents of the stack, located above the portion of the argument list
 373     (if any) that is passed in memory, belong to the calling procedure. Because
 374     they are part of the calling procedure, they should not be read or written
 375     by the called procedure, except as specified by indirect arguments or
 376     language-controlled up-level references.
 377
 378    "The SP value might be used by the hardware when raising exceptions and
 379     asynchronous interrupts. It must be assumed that the contents of the stack
 380     below the current SP value and within the stack for the current thread are
 381     continually and unpredictably modified, as specified in the _Alpha
 382     Architecture Reference Manual_, and as a result of asynchronous software
 383     actions."
 384
 385    -- Compaq Computer Corporation, Houston. Tru64 UNIX Calling Standard for
 386       Alpha Systems, 5.1 edition, August 2000, section 3.2.1.  http://www.
 387       tru64unix.compaq.com/docs/base_doc/DOCUMENTATION/V51_PDF/ARH9MBTE.PDF
 388    -------------------------------------------------------------------------- */
 389
 390 #ifdef alpha_HOST_ARCH
 391
 392 StgThreadReturnCode
 393 StgRun(StgFunPtr f, StgRegTable *basereg)
 394 {
 395     register long   real_ra __asm__("$26"); volatile long   save_ra;
 396     register long   real_gp __asm__("$29"); volatile long   save_gp;
 397
 398     register long   real_s0 __asm__("$9" ); volatile long   save_s0;
 399     register long   real_s1 __asm__("$10"); volatile long   save_s1;
 400     register long   real_s2 __asm__("$11"); volatile long   save_s2;
 401     register long   real_s3 __asm__("$12"); volatile long   save_s3;
 402     register long   real_s4 __asm__("$13"); volatile long   save_s4;
 403     register long   real_s5 __asm__("$14"); volatile long   save_s5;
 404 #ifdef alpha_EXTRA_CAREFUL
 405     register long   real_s6 __asm__("$15"); volatile long   save_s6;
 406 #endif
 407
 408     register double real_f2 __asm__("$f2"); volatile double save_f2;
 409     register double real_f3 __asm__("$f3"); volatile double save_f3;
 410     register double real_f4 __asm__("$f4"); volatile double save_f4;
 411     register double real_f5 __asm__("$f5"); volatile double save_f5;
 412     register double real_f6 __asm__("$f6"); volatile double save_f6;
 413     register double real_f7 __asm__("$f7"); volatile double save_f7;
 414 #ifdef alpha_EXTRA_CAREFUL
 415     register double real_f8 __asm__("$f8"); volatile double save_f8;
 416     register double real_f9 __asm__("$f9"); volatile double save_f9;
 417 #endif
 418
 419     register StgFunPtr real_pv __asm__("$27");
 420
 421     StgThreadReturnCode ret;
 422
 423     save_ra = real_ra;
 424     save_gp = real_gp;
 425
 426     save_s0 = real_s0;
 427     save_s1 = real_s1;
 428     save_s2 = real_s2;
 429     save_s3 = real_s3;
 430     save_s4 = real_s4;
 431     save_s5 = real_s5;
 432 #ifdef alpha_EXTRA_CAREFUL
 433     save_s6 = real_s6;
 434 #endif
 435
 436     save_f2 = real_f2;
 437     save_f3 = real_f3;
 438     save_f4 = real_f4;
 439     save_f5 = real_f5;
 440     save_f6 = real_f6;
 441     save_f7 = real_f7;
 442 #ifdef alpha_EXTRA_CAREFUL
 443     save_f8 = real_f8;
 444     save_f9 = real_f9;
 445 #endif
 446
 447     real_pv = f;
 448
 449     __asm__ volatile(   "lda $30,-%0($30)"      "\n"
 450                 "\t"    "jmp ($27)"             "\n"
 451                 "\t"    ".align 3"              "\n"
 452                 ".globl " STG_RETURN            "\n"
 453                 STG_RETURN ":"                  "\n"
 454                 "\t"    "lda $30,%0($30)"       "\n"
 455                 : : "K" (RESERVED_C_STACK_BYTES));
 456
 457     ret = real_s5;
 458
 459     real_s0 = save_s0;
 460     real_s1 = save_s1;
 461     real_s2 = save_s2;
 462     real_s3 = save_s3;
 463     real_s4 = save_s4;
 464     real_s5 = save_s5;
 465 #ifdef alpha_EXTRA_CAREFUL
 466     real_s6 = save_s6;
 467 #endif
 468
 469     real_f2 = save_f2;
 470     real_f3 = save_f3;
 471     real_f4 = save_f4;
 472     real_f5 = save_f5;
 473     real_f6 = save_f6;
 474     real_f7 = save_f7;
 475 #ifdef alpha_EXTRA_CAREFUL
 476     real_f8 = save_f8;
 477     real_f9 = save_f9;
 478 #endif
 479
 480     real_ra = save_ra;
 481     real_gp = save_gp;
 482
 483     return ret;
 484 }
 485
 486 #endif /* alpha_HOST_ARCH */
 487
 488 /* -----------------------------------------------------------------------------
 489    HP-PA architecture
 490    -------------------------------------------------------------------------- */
 491
 492 #ifdef hppa1_1_HOST_ARCH
 493
 494 StgThreadReturnCode
 495 StgRun(StgFunPtr f, StgRegTable *basereg)
 496 {
 497     StgChar space[RESERVED_C_STACK_BYTES+16*sizeof(long)+10*sizeof(double)];
 498     StgThreadReturnCode ret;
 499
 500     __asm__ volatile ("ldo %0(%%r30),%%r19\n"
 501                       "\tstw %%r3, 0(0,%%r19)\n"
 502                       "\tstw %%r4, 4(0,%%r19)\n"
 503                       "\tstw %%r5, 8(0,%%r19)\n"
 504                       "\tstw %%r6,12(0,%%r19)\n"
 505                       "\tstw %%r7,16(0,%%r19)\n"
 506                       "\tstw %%r8,20(0,%%r19)\n"
 507                       "\tstw %%r9,24(0,%%r19)\n"
 508                       "\tstw %%r10,28(0,%%r19)\n"
 509                       "\tstw %%r11,32(0,%%r19)\n"
 510                       "\tstw %%r12,36(0,%%r19)\n"
 511                       "\tstw %%r13,40(0,%%r19)\n"
 512                       "\tstw %%r14,44(0,%%r19)\n"
 513                       "\tstw %%r15,48(0,%%r19)\n"
 514                       "\tstw %%r16,52(0,%%r19)\n"
 515                       "\tstw %%r17,56(0,%%r19)\n"
 516                       "\tstw %%r18,60(0,%%r19)\n"
 517                       "\tldo 80(%%r19),%%r19\n"
 518                       "\tfstds %%fr12,-16(0,%%r19)\n"
 519                       "\tfstds %%fr13, -8(0,%%r19)\n"
 520                       "\tfstds %%fr14,  0(0,%%r19)\n"
 521                       "\tfstds %%fr15,  8(0,%%r19)\n"
 522                       "\tldo 32(%%r19),%%r19\n"
 523                       "\tfstds %%fr16,-16(0,%%r19)\n"
 524                       "\tfstds %%fr17, -8(0,%%r19)\n"
 525                       "\tfstds %%fr18,  0(0,%%r19)\n"
 526                       "\tfstds %%fr19,  8(0,%%r19)\n"
 527                       "\tldo 32(%%r19),%%r19\n"
 528                       "\tfstds %%fr20,-16(0,%%r19)\n"
 529                       "\tfstds %%fr21, -8(0,%%r19)\n" : :
 530                       "n" (-(116 * sizeof(long) + 10 * sizeof(double))) : "%r19"
 531                       );
 532
 533     f();
 534
 535     __asm__ volatile (".align 4\n"
 536                       "\t.EXPORT " STG_RETURN ",CODE\n"
 537                       "\t.EXPORT " STG_RETURN ",ENTRY,PRIV_LEV=3\n"
 538                       STG_RETURN "\n"
 539                       /* "\tldo %0(%%r3),%%r19\n" */
 540                       "\tldo %1(%%r30),%%r19\n"
 541                       "\tcopy %%r11, %0\n"  /* save R1 */
 542                       "\tldw  0(0,%%r19),%%r3\n"
 543                       "\tldw  4(0,%%r19),%%r4\n"
 544                       "\tldw  8(0,%%r19),%%r5\n"
 545                       "\tldw 12(0,%%r19),%%r6\n"
 546                       "\tldw 16(0,%%r19),%%r7\n"
 547                       "\tldw 20(0,%%r19),%%r8\n"
 548                       "\tldw 24(0,%%r19),%%r9\n"
 549                       "\tldw 28(0,%%r19),%%r10\n"
 550                       "\tldw 32(0,%%r19),%%r11\n"
 551                       "\tldw 36(0,%%r19),%%r12\n"
 552                       "\tldw 40(0,%%r19),%%r13\n"
 553                       "\tldw 44(0,%%r19),%%r14\n"
 554                       "\tldw 48(0,%%r19),%%r15\n"
 555                       "\tldw 52(0,%%r19),%%r16\n"
 556                       "\tldw 56(0,%%r19),%%r17\n"
 557                       "\tldw 60(0,%%r19),%%r18\n"
 558                       "\tldo 80(%%r19),%%r19\n"
 559                       "\tfldds -16(0,%%r19),%%fr12\n"
 560                       "\tfldds  -8(0,%%r19),%%fr13\n"
 561                       "\tfldds   0(0,%%r19),%%fr14\n"
 562                       "\tfldds   8(0,%%r19),%%fr15\n"
 563                       "\tldo 32(%%r19),%%r19\n"
 564                       "\tfldds -16(0,%%r19),%%fr16\n"
 565                       "\tfldds  -8(0,%%r19),%%fr17\n"
 566                       "\tfldds   0(0,%%r19),%%fr18\n"
 567                       "\tfldds   8(0,%%r19),%%fr19\n"
 568                       "\tldo 32(%%r19),%%r19\n"
 569                       "\tfldds -16(0,%%r19),%%fr20\n"
 570                       "\tfldds  -8(0,%%r19),%%fr21\n"
 571                          : "=r" (ret)
 572                          : "n" (-(116 * sizeof(long) + 10 * sizeof(double)))
 573                          : "%r19"
 574                       );
 575
 576     return ret;
 577 }
 578
 579 #endif /* hppa1_1_HOST_ARCH */
 580
 581 /* -----------------------------------------------------------------------------
 582    PowerPC architecture
 583
 584    Everything is in assembler, so we don't have to deal with GCC...
 585
 586    -------------------------------------------------------------------------- */
 587
 588 #ifdef powerpc_HOST_ARCH
 589
 590 extern StgThreadReturnCode StgRun(StgFunPtr f, StgRegTable *basereg);
 591
 592 #ifdef darwin_HOST_OS
 593 static void StgRunIsImplementedInAssembler(void)
 594 {
 595 #if HAVE_SUBSECTIONS_VIA_SYMBOLS
 596             // if the toolchain supports deadstripping, we have to
 597             // prevent it here (it tends to get confused here).
 598         __asm__ volatile (".no_dead_strip _StgRunIsImplementedInAssembler");
 599 #endif
 600         __asm__ volatile (
 601                 "\n.globl _StgRun\n"
 602                 "_StgRun:\n"
 603                 "\tmflr r0\n"
 604                 "\tbl saveFP # f14\n"
 605                 "\tstmw r13,-220(r1)\n"
 606                 "\tstwu r1,-%0(r1)\n"
 607                 "\tmr r27,r4\n" // BaseReg == r27
 608                 "\tmtctr r3\n"
 609                 "\tmr r12,r3\n"
 610                 "\tbctr\n"
 611                 ".globl _StgReturn\n"
 612                 "_StgReturn:\n"
 613                 "\tmr r3,r14\n"
 614                 "\tla r1,%0(r1)\n"
 615                 "\tlmw r13,-220(r1)\n"
 616                 "\tb restFP # f14\n"
 617         : : "i"(RESERVED_C_STACK_BYTES+224 /*stack frame size*/));
 618 }
 619 #else
 620
 621 // This version is for PowerPC Linux.
 622
 623 // Differences from the Darwin/Mac OS X version:
 624 // *) Different Assembler Syntax
 625 // *) Doesn't use Register Saving Helper Functions (although they exist somewhere)
 626 // *) We may not access positive stack offsets
 627 //    (no "Red Zone" as in the Darwin ABI)
 628 // *) The Link Register is saved to a different offset in the caller's stack frame
 629 //    (Linux: 4(r1), Darwin 8(r1))
 630
 631 static void StgRunIsImplementedInAssembler(void)
 632 {
 633         __asm__ volatile (
 634                 "\t.globl StgRun\n"
 635                 "\t.type StgRun,@function\n"
 636                 "StgRun:\n"
 637                 "\tmflr 0\n"
 638                 "\tstw 0,4(1)\n"
 639                 "\tmr 5,1\n"
 640                 "\tstwu 1,-%0(1)\n"
 641                 "\tstmw 13,-220(5)\n"
 642                 "\tstfd 14,-144(5)\n"
 643                 "\tstfd 15,-136(5)\n"
 644                 "\tstfd 16,-128(5)\n"
 645                 "\tstfd 17,-120(5)\n"
 646                 "\tstfd 18,-112(5)\n"
 647                 "\tstfd 19,-104(5)\n"
 648                 "\tstfd 20,-96(5)\n"
 649                 "\tstfd 21,-88(5)\n"
 650                 "\tstfd 22,-80(5)\n"
 651                 "\tstfd 23,-72(5)\n"
 652                 "\tstfd 24,-64(5)\n"
 653                 "\tstfd 25,-56(5)\n"
 654                 "\tstfd 26,-48(5)\n"
 655                 "\tstfd 27,-40(5)\n"
 656                 "\tstfd 28,-32(5)\n"
 657                 "\tstfd 29,-24(5)\n"
 658                 "\tstfd 30,-16(5)\n"
 659                 "\tstfd 31,-8(5)\n"
 660                 "\tmr 27,4\n"  // BaseReg == r27
 661                 "\tmtctr 3\n"
 662                 "\tmr 12,3\n"
 663                 "\tbctr\n"
 664                 ".globl StgReturn\n"
 665                 "\t.type StgReturn,@function\n"
 666                 "StgReturn:\n"
 667                 "\tmr 3,14\n"
 668                 "\tla 5,%0(1)\n"
 669                 "\tlmw 13,-220(5)\n"
 670                 "\tlfd 14,-144(5)\n"
 671                 "\tlfd 15,-136(5)\n"
 672                 "\tlfd 16,-128(5)\n"
 673                 "\tlfd 17,-120(5)\n"
 674                 "\tlfd 18,-112(5)\n"
 675                 "\tlfd 19,-104(5)\n"
 676                 "\tlfd 20,-96(5)\n"
 677                 "\tlfd 21,-88(5)\n"
 678                 "\tlfd 22,-80(5)\n"
 679                 "\tlfd 23,-72(5)\n"
 680                 "\tlfd 24,-64(5)\n"
 681                 "\tlfd 25,-56(5)\n"
 682                 "\tlfd 26,-48(5)\n"
 683                 "\tlfd 27,-40(5)\n"
 684                 "\tlfd 28,-32(5)\n"
 685                 "\tlfd 29,-24(5)\n"
 686                 "\tlfd 30,-16(5)\n"
 687                 "\tlfd 31,-8(5)\n"
 688                 "\tmr 1,5\n"
 689                 "\tlwz 0,4(1)\n"
 690                 "\tmtlr 0\n"
 691                 "\tblr\n"
 692         : : "i"(RESERVED_C_STACK_BYTES+224 /*stack frame size*/));
 693 }
 694 #endif
 695
 696 #endif
 697
 698 /* -----------------------------------------------------------------------------
 699    PowerPC 64 architecture
 700
 701    Everything is in assembler, so we don't have to deal with GCC...
 702
 703    -------------------------------------------------------------------------- */
 704
 705 #ifdef powerpc64_HOST_ARCH
 706
 707 #ifdef linux_HOST_OS
 708 extern StgThreadReturnCode StgRun(StgFunPtr f, StgRegTable *basereg);
 709
 710 static void StgRunIsImplementedInAssembler(void)
 711 {
 712         // r0 volatile
 713         // r1 stack pointer
 714         // r2 toc - needs to be saved
 715         // r3-r10 argument passing, volatile
 716         // r11, r12 very volatile (not saved across cross-module calls)
 717         // r13 thread local state (never modified, don't need to save)
 718         // r14-r31 callee-save
 719         __asm__ volatile (
 720                 ".section \".opd\",\"aw\"\n"
 721                 ".align 3\n"
 722                 ".globl StgRun\n"
 723                 "StgRun:\n"
 724                         "\t.quad\t.StgRun,.TOC.@tocbase,0\n"
 725                         "\t.size StgRun,24\n"
 726                 ".globl StgReturn\n"
 727                 "StgReturn:\n"
 728                         "\t.quad\t.StgReturn,.TOC.@tocbase,0\n"
 729                         "\t.size StgReturn,24\n"
 730                 ".previous\n"
 731                 ".globl .StgRun\n"
 732                 ".type .StgRun,@function\n"
 733                 ".StgRun:\n"
 734                         "\tmflr 0\n"
 735                         "\tmr 5, 1\n"
 736                         "\tstd 0, 16(1)\n"
 737                         "\tstdu 1, -%0(1)\n"
 738                         "\tstd 2, -296(5)\n"
 739                         "\tstd 14, -288(5)\n"
 740                         "\tstd 15, -280(5)\n"
 741                         "\tstd 16, -272(5)\n"
 742                         "\tstd 17, -264(5)\n"
 743                         "\tstd 18, -256(5)\n"
 744                         "\tstd 19, -248(5)\n"
 745                         "\tstd 20, -240(5)\n"
 746                         "\tstd 21, -232(5)\n"
 747                         "\tstd 22, -224(5)\n"
 748                         "\tstd 23, -216(5)\n"
 749                         "\tstd 24, -208(5)\n"
 750                         "\tstd 25, -200(5)\n"
 751                         "\tstd 26, -192(5)\n"
 752                         "\tstd 27, -184(5)\n"
 753                         "\tstd 28, -176(5)\n"
 754                         "\tstd 29, -168(5)\n"
 755                         "\tstd 30, -160(5)\n"
 756                         "\tstd 31, -152(5)\n"
 757                         "\tstfd 14, -144(5)\n"
 758                         "\tstfd 15, -136(5)\n"
 759                         "\tstfd 16, -128(5)\n"
 760                         "\tstfd 17, -120(5)\n"
 761                         "\tstfd 18, -112(5)\n"
 762                         "\tstfd 19, -104(5)\n"
 763                         "\tstfd 20, -96(5)\n"
 764                         "\tstfd 21, -88(5)\n"
 765                         "\tstfd 22, -80(5)\n"
 766                         "\tstfd 23, -72(5)\n"
 767                         "\tstfd 24, -64(5)\n"
 768                         "\tstfd 25, -56(5)\n"
 769                         "\tstfd 26, -48(5)\n"
 770                         "\tstfd 27, -40(5)\n"
 771                         "\tstfd 28, -32(5)\n"
 772                         "\tstfd 29, -24(5)\n"
 773                         "\tstfd 30, -16(5)\n"
 774                         "\tstfd 31, -8(5)\n"
 775                         "\tmr 27, 4\n"  // BaseReg == r27
 776                         "\tld 2, 8(3)\n"
 777                         "\tld 3, 0(3)\n"
 778                         "\tmtctr 3\n"
 779                         "\tbctr\n"
 780                 ".globl .StgReturn\n"
 781                 ".type .StgReturn,@function\n"
 782                 ".StgReturn:\n"
 783                         "\tmr 3,14\n"
 784                         "\tla 5, %0(1)\n" // load address == addi r5, r1, %0
 785                         "\tld 2, -296(5)\n"
 786                         "\tld 14, -288(5)\n"
 787                         "\tld 15, -280(5)\n"
 788                         "\tld 16, -272(5)\n"
 789                         "\tld 17, -264(5)\n"
 790                         "\tld 18, -256(5)\n"
 791                         "\tld 19, -248(5)\n"
 792                         "\tld 20, -240(5)\n"
 793                         "\tld 21, -232(5)\n"
 794                         "\tld 22, -224(5)\n"
 795                         "\tld 23, -216(5)\n"
 796                         "\tld 24, -208(5)\n"
 797                         "\tld 25, -200(5)\n"
 798                         "\tld 26, -192(5)\n"
 799                         "\tld 27, -184(5)\n"
 800                         "\tld 28, -176(5)\n"
 801                         "\tld 29, -168(5)\n"
 802                         "\tld 30, -160(5)\n"
 803                         "\tld 31, -152(5)\n"
 804                         "\tlfd 14, -144(5)\n"
 805                         "\tlfd 15, -136(5)\n"
 806                         "\tlfd 16, -128(5)\n"
 807                         "\tlfd 17, -120(5)\n"
 808                         "\tlfd 18, -112(5)\n"
 809                         "\tlfd 19, -104(5)\n"
 810                         "\tlfd 20, -96(5)\n"
 811                         "\tlfd 21, -88(5)\n"
 812                         "\tlfd 22, -80(5)\n"
 813                         "\tlfd 23, -72(5)\n"
 814                         "\tlfd 24, -64(5)\n"
 815                         "\tlfd 25, -56(5)\n"
 816                         "\tlfd 26, -48(5)\n"
 817                         "\tlfd 27, -40(5)\n"
 818                         "\tlfd 28, -32(5)\n"
 819                         "\tlfd 29, -24(5)\n"
 820                         "\tlfd 30, -16(5)\n"
 821                         "\tlfd 31, -8(5)\n"
 822                         "\tmr 1, 5\n"
 823                         "\tld 0, 16(1)\n"
 824                         "\tmtlr 0\n"
 825                         "\tblr\n"
 826         : : "i"(RESERVED_C_STACK_BYTES+304 /*stack frame size*/));
 827 }
 828 #else // linux_HOST_OS
 829 #error Only linux support for power64 right now.
 830 #endif
 831
 832 #endif
 833
 834 /* -----------------------------------------------------------------------------
 835    IA64 architecture
 836
 837    Again, in assembler - so we can fiddle with the register stack, and because
 838    gcc doesn't handle asm-clobbered callee-saves correctly.
 839
 840    loc0  - loc15: preserved locals
 841    loc16 - loc28: STG registers
 842            loc29: saved ar.pfs
 843            loc30: saved b0
 844            loc31: saved gp (gcc 3.3 uses this slot)
 845    -------------------------------------------------------------------------- */
 846
 847 #ifdef ia64_HOST_ARCH
 848
 849 /* the memory stack is rarely used, so 16K is excessive */
 850 #undef RESERVED_C_STACK_BYTES
 851 #define RESERVED_C_STACK_BYTES 1024
 852
 853 #if ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 3)) || (__GNUC__ > 3)
 854 /* gcc 3.3+: leave an extra slot for gp saves */
 855 #define LOCALS 32
 856 #else
 857 #define LOCALS 31
 858 #endif
 859
 860 static void StgRunIsImplementedInAssembler(void)
 861 {
 862     __asm__ volatile(
 863                 ".global StgRun\n"
 864                 "StgRun:\n"
 865                 "\talloc loc29 = ar.pfs, 0, %1, 8, 0\n" /* setup register frame */
 866                 "\tld8 r18 = [r32],8\n"                 /* get procedure address */
 867                 "\tadds sp = -%0, sp ;;\n"              /* setup stack */
 868                 "\tld8 gp = [r32]\n"                    /* get procedure GP */
 869                 "\tadds r16 = %0-(6*16), sp\n"
 870                 "\tadds r17 = %0-(5*16), sp ;;\n"
 871                 "\tstf.spill [r16] = f16,32\n"          /* spill callee-saved fp regs */
 872                 "\tstf.spill [r17] = f17,32\n"
 873                 "\tmov b6 = r18 ;;\n"                   /* set target address */
 874                 "\tstf.spill [r16] = f18,32\n"
 875                 "\tstf.spill [r17] = f19,32\n"
 876                 "\tmov loc30 = b0 ;;\n"                 /* save return address */
 877                 "\tstf.spill [r16] = f20,32\n"
 878                 "\tstf.spill [r17] = f21,32\n"
 879                 "\tbr.few b6 ;;\n"                      /* branch to function */
 880                 ".global StgReturn\n"
 881                 "StgReturn:\n"
 882                 "\tmov r8 = loc16\n"            /* return value in r8 */
 883                 "\tadds r16 = %0-(6*16), sp\n"
 884                 "\tadds r17 = %0-(5*16), sp ;;\n"
 885                 "\tldf.fill f16 = [r16],32\n"   /* start restoring fp regs */
 886                 "\tldf.fill f17 = [r17],32\n"
 887                 "\tmov ar.pfs = loc29 ;;\n"     /* restore register frame */
 888                 "\tldf.fill f18 = [r16],32\n"
 889                 "\tldf.fill f19 = [r17],32\n"
 890                 "\tmov b0 = loc30 ;;\n"         /* restore return address */
 891                 "\tldf.fill f20 = [r16],32\n"
 892                 "\tldf.fill f21 = [r17],32\n"
 893                 "\tadds sp = %0, sp\n"          /* restore stack */
 894                 "\tbr.ret.sptk.many b0 ;;\n"    /* return */
 895         : : "i"(RESERVED_C_STACK_BYTES + 6*16), "i"(LOCALS));
 896 }
 897
 898 #endif
 899
 900 #endif /* !USE_MINIINTERPRETER */