Fix Windows stack allocations.
[ghc.git] / includes / Stg.h
1 /* -----------------------------------------------------------------------------
2 *
3 * (c) The GHC Team, 1998-2009
4 *
5 * Top-level include file for everything required when compiling .hc
6 * code. NOTE: in .hc files, Stg.h must be included *before* any
7 * other headers, because we define some register variables which must
8 * be done before any inline functions are defined (some system
9 * headers have been known to define the odd inline function).
10 *
11 * We generally try to keep as little visible as possible when
12 * compiling .hc files. So for example the definitions of the
13 * InfoTable structs, closure structs and other RTS types are not
14 * visible here. The compiler knows enough about the representations
15 * of these types to generate code which manipulates them directly
16 * with pointer arithmetic.
17 *
18 * In ordinary C code, do not #include this file directly: #include
19 * "Rts.h" instead.
20 *
21 * To understand the structure of the RTS headers, see the wiki:
22 * http://ghc.haskell.org/trac/ghc/wiki/Commentary/SourceTree/Includes
23 *
24 * ---------------------------------------------------------------------------*/
25
26 #pragma once
27
28 #if !(__STDC_VERSION__ >= 199901L)
29 # error __STDC_VERSION__ does not advertise C99 or later
30 #endif
31
32 /*
33 * If we are compiling a .hc file, then we want all the register
34 * variables. This is the what happens if you #include "Stg.h" first:
35 * we assume this is a .hc file, and set IN_STG_CODE==1, which later
36 * causes the register variables to be enabled in stg/Regs.h.
37 *
38 * If instead "Rts.h" is included first, then we are compiling a
39 * vanilla C file. Everything from Stg.h is provided, except that
40 * IN_STG_CODE is not defined, and the register variables will not be
41 * active.
42 */
43 #if !defined(IN_STG_CODE)
44 # define IN_STG_CODE 1
45
46 // Turn on C99 for .hc code. This gives us the INFINITY and NAN
47 // constants from math.h, which we occasionally need to use in .hc (#1861)
48 # define _ISOC99_SOURCE
49
50 // We need _BSD_SOURCE so that math.h defines things like gamma
51 // on Linux
52 # define _BSD_SOURCE
53
54 // On AIX we need _BSD defined, otherwise <math.h> includes <stdlib.h>
55 # if defined(_AIX)
56 # define _BSD 1
57 # endif
58
59 // '_BSD_SOURCE' is deprecated since glibc-2.20
60 // in favour of '_DEFAULT_SOURCE'
61 # define _DEFAULT_SOURCE
62 #endif
63
64 #if IN_STG_CODE == 0 || defined(llvm_CC_FLAVOR)
65 // C compilers that use an LLVM back end (clang or llvm-gcc) do not
66 // correctly support global register variables so we make sure that
67 // we do not declare them for these compilers.
68 # define NO_GLOBAL_REG_DECLS /* don't define fixed registers */
69 #endif
70
71 /* Configuration */
72 #include "ghcconfig.h"
73
74 /* The code generator calls the math functions directly in .hc code.
75 NB. after configuration stuff above, because this sets #defines
76 that depend on config info, such as __USE_FILE_OFFSET64 */
77 #include <math.h>
78
79 // On Solaris, we don't get the INFINITY and NAN constants unless we
80 // #define _STDC_C99, and we can't do that unless we also use -std=c99,
81 // because _STDC_C99 causes the headers to use C99 syntax (e.g. restrict).
82 // We aren't ready for -std=c99 yet, so define INFINITY/NAN by hand using
83 // the gcc builtins.
84 #if !defined(INFINITY)
85 #if defined(__GNUC__)
86 #define INFINITY __builtin_inf()
87 #else
88 #error No definition for INFINITY
89 #endif
90 #endif
91
92 #if !defined(NAN)
93 #if defined(__GNUC__)
94 #define NAN __builtin_nan("")
95 #else
96 #error No definition for NAN
97 #endif
98 #endif
99
100 /* -----------------------------------------------------------------------------
101 Useful definitions
102 -------------------------------------------------------------------------- */
103
104 /*
105 * The C backend likes to refer to labels by just mentioning their
106 * names. However, when a symbol is declared as a variable in C, the
107 * C compiler will implicitly dereference it when it occurs in source.
108 * So we must subvert this behaviour for .hc files by declaring
109 * variables as arrays, which eliminates the implicit dereference.
110 */
111 #if IN_STG_CODE
112 #define RTS_VAR(x) (x)[]
113 #define RTS_DEREF(x) (*(x))
114 #else
115 #define RTS_VAR(x) x
116 #define RTS_DEREF(x) x
117 #endif
118
119 /* bit macros
120 */
121 #define BITS_PER_BYTE 8
122 #define BITS_IN(x) (BITS_PER_BYTE * sizeof(x))
123
124 /* Compute offsets of struct fields
125 */
126 #define STG_FIELD_OFFSET(s_type, field) ((StgWord)&(((s_type*)0)->field))
127
128 /*
129 * 'Portable' inlining:
130 * INLINE_HEADER is for inline functions in header files (macros)
131 * STATIC_INLINE is for inline functions in source files
132 * EXTERN_INLINE is for functions that we want to inline sometimes
133 * (we also compile a static version of the function; see Inlines.c)
134 */
135
136 // We generally assume C99 semantics albeit these two definitions work fine even
137 // when gnu90 semantics are active (i.e. when __GNUC_GNU_INLINE__ is defined or
138 // when a GCC older than 4.2 is used)
139 //
140 // The problem, however, is with 'extern inline' whose semantics significantly
141 // differs between gnu90 and C99
142 #define INLINE_HEADER static inline
143 #define STATIC_INLINE static inline
144
145 // Figure out whether `__attributes__((gnu_inline))` is needed
146 // to force gnu90-style 'external inline' semantics.
147 #if defined(FORCE_GNU_INLINE)
148 // disable auto-detection since HAVE_GNU_INLINE has been defined externally
149 #elif defined(__GNUC_GNU_INLINE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2
150 // GCC 4.2.x didn't properly support C99 inline semantics (GCC 4.3 was the first
151 // release to properly support C99 inline semantics), and therefore warned when
152 // using 'extern inline' while in C99 mode unless `__attributes__((gnu_inline))`
153 // was explicitly set.
154 # define FORCE_GNU_INLINE 1
155 #endif
156
157 #if defined(FORCE_GNU_INLINE)
158 // Force compiler into gnu90 semantics
159 # if defined(KEEP_INLINES)
160 # define EXTERN_INLINE inline __attribute__((gnu_inline))
161 # else
162 # define EXTERN_INLINE extern inline __attribute__((gnu_inline))
163 # endif
164 #elif defined(__GNUC_GNU_INLINE__)
165 // we're currently in gnu90 inline mode by default and
166 // __attribute__((gnu_inline)) may not be supported, so better leave it off
167 # if defined(KEEP_INLINES)
168 # define EXTERN_INLINE inline
169 # else
170 # define EXTERN_INLINE extern inline
171 # endif
172 #else
173 // Assume C99 semantics (yes, this curiously results in swapped definitions!)
174 // This is the preferred branch, and at some point we may drop support for
175 // compilers not supporting C99 semantics altogether.
176 # if defined(KEEP_INLINES)
177 # define EXTERN_INLINE extern inline
178 # else
179 # define EXTERN_INLINE inline
180 # endif
181 #endif
182
183
184 /*
185 * GCC attributes
186 */
187 #if defined(__GNUC__)
188 #define GNU_ATTRIBUTE(at) __attribute__((at))
189 #else
190 #define GNU_ATTRIBUTE(at)
191 #endif
192
193 #if __GNUC__ >= 3
194 #define GNUC3_ATTRIBUTE(at) __attribute__((at))
195 #else
196 #define GNUC3_ATTRIBUTE(at)
197 #endif
198
199 #if __GNUC__ > 4 || __GNUC__ == 4 && __GNUC_MINOR__ >= 3
200 #define GNUC_ATTR_HOT __attribute__((hot))
201 #else
202 #define GNUC_ATTR_HOT /* nothing */
203 #endif
204
205 #define STG_UNUSED GNUC3_ATTRIBUTE(__unused__)
206
207 /* Prevent functions from being optimized.
208 See Note [Windows Stack allocations] */
209 #if defined(__clang__)
210 #define STG_NO_OPTIMIZE __attribute__((optnone))
211 #elif defined(__GNUC__) || defined(__GNUG__)
212 #define STG_NO_OPTIMIZE __attribute__((optimize("O0")))
213 #else
214 #define STG_NO_OPTIMIZE /* nothing */
215 #endif
216
217 /* -----------------------------------------------------------------------------
218 Global type definitions
219 -------------------------------------------------------------------------- */
220
221 #include "MachDeps.h"
222 #include "stg/Types.h"
223
224 /* -----------------------------------------------------------------------------
225 Shorthand forms
226 -------------------------------------------------------------------------- */
227
228 typedef StgChar C_;
229 typedef StgWord W_;
230 typedef StgWord* P_;
231 typedef StgInt I_;
232 typedef StgWord StgWordArray[];
233 typedef StgFunPtr F_;
234
235 /* byte arrays (and strings): */
236 #define EB_(X) extern const char X[]
237 #define IB_(X) static const char X[]
238 /* static (non-heap) closures (requires alignment for pointer tagging): */
239 #define EC_(X) extern StgWordArray (X) GNU_ATTRIBUTE(aligned (8))
240 #define IC_(X) static StgWordArray (X) GNU_ATTRIBUTE(aligned (8))
241 /* writable data (does not require alignment): */
242 #define ERW_(X) extern StgWordArray (X)
243 #define IRW_(X) static StgWordArray (X)
244 /* read-only data (does not require alignment): */
245 #define ERO_(X) extern const StgWordArray (X)
246 #define IRO_(X) static const StgWordArray (X)
247 /* stg-native functions: */
248 #define IF_(f) static StgFunPtr GNUC3_ATTRIBUTE(used) f(void)
249 #define FN_(f) StgFunPtr f(void)
250 #define EF_(f) StgFunPtr f(void) /* External Cmm functions */
251 /* foreign functions: */
252 #define EFF_(f) void f() /* See Note [External function prototypes] */
253
254 /* Note [External function prototypes] See Trac #8965, #11395
255 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
256 In generated C code we need to distinct between two types
257 of external symbols:
258 1. Cmm functions declared by 'EF_' macro (External Functions)
259 2. C functions declared by 'EFF_' macro (External Foreign Functions)
260
261 Cmm functions are simple as they are internal to GHC.
262
263 C functions are trickier:
264
265 The external-function macro EFF_(F) used to be defined as
266 extern StgFunPtr f(void)
267 i.e a function of zero arguments. On most platforms this doesn't
268 matter very much: calls to these functions put the parameters in the
269 usual places anyway, and (with the exception of varargs) things just
270 work.
271
272 However, the ELFv2 ABI on ppc64 optimises stack allocation
273 (http://gcc.gnu.org/ml/gcc-patches/2013-11/msg01149.html): a call to a
274 function that has a prototype, is not varargs, and receives all parameters
275 in registers rather than on the stack does not require the caller to
276 allocate an argument save area. The incorrect prototypes cause GCC to
277 believe that all functions declared this way can be called without an
278 argument save area, but if the callee has sufficiently many arguments then
279 it will expect that area to be present, and will thus corrupt the caller's
280 stack. This happens in particular with calls to runInteractiveProcess in
281 libraries/process/cbits/runProcess.c, and led to Trac #8965.
282
283 The simplest fix appears to be to declare these external functions with an
284 unspecified argument list rather than a void argument list. This is no
285 worse for platforms that don't care either way, and allows a successful
286 bootstrap of GHC 7.8 on little-endian Linux ppc64 (which uses the ELFv2
287 ABI).
288
289 Another case is m68k ABI where 'void*' return type is returned by 'a0'
290 register while 'long' return type is returned by 'd0'. Thus we trick
291 external prototype return neither of these types to workaround #11395.
292 */
293
294
295 /* -----------------------------------------------------------------------------
296 Tail calls
297 -------------------------------------------------------------------------- */
298
299 #define JMP_(cont) return((StgFunPtr)(cont))
300
301 /* -----------------------------------------------------------------------------
302 Other Stg stuff...
303 -------------------------------------------------------------------------- */
304
305 #include "stg/DLL.h"
306 #include "stg/RtsMachRegs.h"
307 #include "stg/Regs.h"
308 #include "stg/Ticky.h"
309
310 #if IN_STG_CODE
311 /*
312 * This is included later for RTS sources, after definitions of
313 * StgInfoTable, StgClosure and so on.
314 */
315 #include "stg/MiscClosures.h"
316 #endif
317
318 #include "stg/Prim.h" /* ghc-prim fallbacks */
319 #include "stg/SMP.h" // write_barrier() inline is required
320
321 /* -----------------------------------------------------------------------------
322 Moving Floats and Doubles
323
324 ASSIGN_FLT is for assigning a float to memory (usually the
325 stack/heap). The memory address is guaranteed to be
326 StgWord aligned (currently == sizeof(void *)).
327
328 PK_FLT is for pulling a float out of memory. The memory is
329 guaranteed to be StgWord aligned.
330 -------------------------------------------------------------------------- */
331
332 INLINE_HEADER void ASSIGN_FLT (W_ [], StgFloat);
333 INLINE_HEADER StgFloat PK_FLT (W_ []);
334
335 #if ALIGNMENT_FLOAT <= ALIGNMENT_VOID_P
336
337 INLINE_HEADER void ASSIGN_FLT(W_ p_dest[], StgFloat src) { *(StgFloat *)p_dest = src; }
338 INLINE_HEADER StgFloat PK_FLT (W_ p_src[]) { return *(StgFloat *)p_src; }
339
340 #else /* ALIGNMENT_FLOAT > ALIGNMENT_UNSIGNED_INT */
341
342 INLINE_HEADER void ASSIGN_FLT(W_ p_dest[], StgFloat src)
343 {
344 float_thing y;
345 y.f = src;
346 *p_dest = y.fu;
347 }
348
349 INLINE_HEADER StgFloat PK_FLT(W_ p_src[])
350 {
351 float_thing y;
352 y.fu = *p_src;
353 return(y.f);
354 }
355
356 #endif /* ALIGNMENT_FLOAT > ALIGNMENT_VOID_P */
357
358 #if ALIGNMENT_DOUBLE <= ALIGNMENT_VOID_P
359
360 INLINE_HEADER void ASSIGN_DBL (W_ [], StgDouble);
361 INLINE_HEADER StgDouble PK_DBL (W_ []);
362
363 INLINE_HEADER void ASSIGN_DBL(W_ p_dest[], StgDouble src) { *(StgDouble *)p_dest = src; }
364 INLINE_HEADER StgDouble PK_DBL (W_ p_src[]) { return *(StgDouble *)p_src; }
365
366 #else /* ALIGNMENT_DOUBLE > ALIGNMENT_VOID_P */
367
368 /* Sparc uses two floating point registers to hold a double. We can
369 * write ASSIGN_DBL and PK_DBL by directly accessing the registers
370 * independently - unfortunately this code isn't writable in C, we
371 * have to use inline assembler.
372 */
373 #if defined(sparc_HOST_ARCH)
374
375 #define ASSIGN_DBL(dst0,src) \
376 { StgPtr dst = (StgPtr)(dst0); \
377 __asm__("st %2,%0\n\tst %R2,%1" : "=m" (((P_)(dst))[0]), \
378 "=m" (((P_)(dst))[1]) : "f" (src)); \
379 }
380
381 #define PK_DBL(src0) \
382 ( { StgPtr src = (StgPtr)(src0); \
383 register double d; \
384 __asm__("ld %1,%0\n\tld %2,%R0" : "=f" (d) : \
385 "m" (((P_)(src))[0]), "m" (((P_)(src))[1])); d; \
386 } )
387
388 #else /* ! sparc_HOST_ARCH */
389
390 INLINE_HEADER void ASSIGN_DBL (W_ [], StgDouble);
391 INLINE_HEADER StgDouble PK_DBL (W_ []);
392
393 typedef struct
394 { StgWord dhi;
395 StgWord dlo;
396 } unpacked_double;
397
398 typedef union
399 { StgDouble d;
400 unpacked_double du;
401 } double_thing;
402
403 INLINE_HEADER void ASSIGN_DBL(W_ p_dest[], StgDouble src)
404 {
405 double_thing y;
406 y.d = src;
407 p_dest[0] = y.du.dhi;
408 p_dest[1] = y.du.dlo;
409 }
410
411 /* GCC also works with this version, but it generates
412 the same code as the previous one, and is not ANSI
413
414 #define ASSIGN_DBL( p_dest, src ) \
415 *p_dest = ((double_thing) src).du.dhi; \
416 *(p_dest+1) = ((double_thing) src).du.dlo \
417 */
418
419 INLINE_HEADER StgDouble PK_DBL(W_ p_src[])
420 {
421 double_thing y;
422 y.du.dhi = p_src[0];
423 y.du.dlo = p_src[1];
424 return(y.d);
425 }
426
427 #endif /* ! sparc_HOST_ARCH */
428
429 #endif /* ALIGNMENT_DOUBLE > ALIGNMENT_UNSIGNED_INT */
430
431
432 /* -----------------------------------------------------------------------------
433 Moving 64-bit quantities around
434
435 ASSIGN_Word64 assign an StgWord64/StgInt64 to a memory location
436 PK_Word64 load an StgWord64/StgInt64 from a amemory location
437
438 In both cases the memory location might not be 64-bit aligned.
439 -------------------------------------------------------------------------- */
440
441 #if SIZEOF_HSWORD == 4
442
443 typedef struct
444 { StgWord dhi;
445 StgWord dlo;
446 } unpacked_double_word;
447
448 typedef union
449 { StgInt64 i;
450 unpacked_double_word iu;
451 } int64_thing;
452
453 typedef union
454 { StgWord64 w;
455 unpacked_double_word wu;
456 } word64_thing;
457
458 INLINE_HEADER void ASSIGN_Word64(W_ p_dest[], StgWord64 src)
459 {
460 word64_thing y;
461 y.w = src;
462 p_dest[0] = y.wu.dhi;
463 p_dest[1] = y.wu.dlo;
464 }
465
466 INLINE_HEADER StgWord64 PK_Word64(W_ p_src[])
467 {
468 word64_thing y;
469 y.wu.dhi = p_src[0];
470 y.wu.dlo = p_src[1];
471 return(y.w);
472 }
473
474 INLINE_HEADER void ASSIGN_Int64(W_ p_dest[], StgInt64 src)
475 {
476 int64_thing y;
477 y.i = src;
478 p_dest[0] = y.iu.dhi;
479 p_dest[1] = y.iu.dlo;
480 }
481
482 INLINE_HEADER StgInt64 PK_Int64(W_ p_src[])
483 {
484 int64_thing y;
485 y.iu.dhi = p_src[0];
486 y.iu.dlo = p_src[1];
487 return(y.i);
488 }
489
490 #elif SIZEOF_VOID_P == 8
491
492 INLINE_HEADER void ASSIGN_Word64(W_ p_dest[], StgWord64 src)
493 {
494 p_dest[0] = src;
495 }
496
497 INLINE_HEADER StgWord64 PK_Word64(W_ p_src[])
498 {
499 return p_src[0];
500 }
501
502 INLINE_HEADER void ASSIGN_Int64(W_ p_dest[], StgInt64 src)
503 {
504 p_dest[0] = src;
505 }
506
507 INLINE_HEADER StgInt64 PK_Int64(W_ p_src[])
508 {
509 return p_src[0];
510 }
511
512 #endif /* SIZEOF_HSWORD == 4 */
513
514 /* -----------------------------------------------------------------------------
515 Integer multiply with overflow
516 -------------------------------------------------------------------------- */
517
518 /* Multiply with overflow checking.
519 *
520 * This is tricky - the usual sign rules for add/subtract don't apply.
521 *
522 * On 32-bit machines we use gcc's 'long long' types, finding
523 * overflow with some careful bit-twiddling.
524 *
525 * On 64-bit machines where gcc's 'long long' type is also 64-bits,
526 * we use a crude approximation, testing whether either operand is
527 * larger than 32-bits; if neither is, then we go ahead with the
528 * multiplication.
529 *
530 * Return non-zero if there is any possibility that the signed multiply
531 * of a and b might overflow. Return zero only if you are absolutely sure
532 * that it won't overflow. If in doubt, return non-zero.
533 */
534
535 #if SIZEOF_VOID_P == 4
536
537 #if defined(WORDS_BIGENDIAN)
538 #define RTS_CARRY_IDX__ 0
539 #define RTS_REM_IDX__ 1
540 #else
541 #define RTS_CARRY_IDX__ 1
542 #define RTS_REM_IDX__ 0
543 #endif
544
545 typedef union {
546 StgInt64 l;
547 StgInt32 i[2];
548 } long_long_u ;
549
550 #define mulIntMayOflo(a,b) \
551 ({ \
552 StgInt32 r, c; \
553 long_long_u z; \
554 z.l = (StgInt64)a * (StgInt64)b; \
555 r = z.i[RTS_REM_IDX__]; \
556 c = z.i[RTS_CARRY_IDX__]; \
557 if (c == 0 || c == -1) { \
558 c = ((StgWord)((a^b) ^ r)) \
559 >> (BITS_IN (I_) - 1); \
560 } \
561 c; \
562 })
563
564 /* Careful: the carry calculation above is extremely delicate. Make sure
565 * you test it thoroughly after changing it.
566 */
567
568 #else
569
570 /* Approximate version when we don't have long arithmetic (on 64-bit archs) */
571
572 /* If we have n-bit words then we have n-1 bits after accounting for the
573 * sign bit, so we can fit the result of multiplying 2 (n-1)/2-bit numbers */
574 #define HALF_POS_INT (((I_)1) << ((BITS_IN (I_) - 1) / 2))
575 #define HALF_NEG_INT (-HALF_POS_INT)
576
577 #define mulIntMayOflo(a,b) \
578 ({ \
579 I_ c; \
580 if ((I_)a <= HALF_NEG_INT || a >= HALF_POS_INT \
581 || (I_)b <= HALF_NEG_INT || b >= HALF_POS_INT) {\
582 c = 1; \
583 } else { \
584 c = 0; \
585 } \
586 c; \
587 })
588 #endif