Revert "codeGen: Remove binutils<2.17 hack, fixes T11758"
[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 #ifndef STG_H
27 #define STG_H
28
29 #if !(__STDC_VERSION__ >= 199901L)
30 # error __STDC_VERSION__ does not advertise C99 or later
31 #endif
32
33 /*
34 * If we are compiling a .hc file, then we want all the register
35 * variables. This is the what happens if you #include "Stg.h" first:
36 * we assume this is a .hc file, and set IN_STG_CODE==1, which later
37 * causes the register variables to be enabled in stg/Regs.h.
38 *
39 * If instead "Rts.h" is included first, then we are compiling a
40 * vanilla C file. Everything from Stg.h is provided, except that
41 * IN_STG_CODE is not defined, and the register variables will not be
42 * active.
43 */
44 #ifndef IN_STG_CODE
45 # define IN_STG_CODE 1
46
47 // Turn on C99 for .hc code. This gives us the INFINITY and NAN
48 // constants from math.h, which we occasionally need to use in .hc (#1861)
49 # define _ISOC99_SOURCE
50
51 // We need _BSD_SOURCE so that math.h defines things like gamma
52 // on Linux
53 # define _BSD_SOURCE
54
55 // On AIX we need _BSD defined, otherwise <math.h> includes <stdlib.h>
56 # if defined(_AIX)
57 # define _BSD 1
58 # endif
59
60 // '_BSD_SOURCE' is deprecated since glibc-2.20
61 // in favour of '_DEFAULT_SOURCE'
62 # define _DEFAULT_SOURCE
63 #endif
64
65 #if IN_STG_CODE == 0 || defined(llvm_CC_FLAVOR)
66 // C compilers that use an LLVM back end (clang or llvm-gcc) do not
67 // correctly support global register variables so we make sure that
68 // we do not declare them for these compilers.
69 # define NO_GLOBAL_REG_DECLS /* don't define fixed registers */
70 #endif
71
72 /* Configuration */
73 #include "ghcconfig.h"
74
75 /* The code generator calls the math functions directly in .hc code.
76 NB. after configuration stuff above, because this sets #defines
77 that depend on config info, such as __USE_FILE_OFFSET64 */
78 #include <math.h>
79
80 // On Solaris, we don't get the INFINITY and NAN constants unless we
81 // #define _STDC_C99, and we can't do that unless we also use -std=c99,
82 // because _STDC_C99 causes the headers to use C99 syntax (e.g. restrict).
83 // We aren't ready for -std=c99 yet, so define INFINITY/NAN by hand using
84 // the gcc builtins.
85 #if !defined(INFINITY)
86 #if defined(__GNUC__)
87 #define INFINITY __builtin_inf()
88 #else
89 #error No definition for INFINITY
90 #endif
91 #endif
92
93 #if !defined(NAN)
94 #if defined(__GNUC__)
95 #define NAN __builtin_nan("")
96 #else
97 #error No definition for NAN
98 #endif
99 #endif
100
101 /* -----------------------------------------------------------------------------
102 Useful definitions
103 -------------------------------------------------------------------------- */
104
105 /*
106 * The C backend likes to refer to labels by just mentioning their
107 * names. However, when a symbol is declared as a variable in C, the
108 * C compiler will implicitly dereference it when it occurs in source.
109 * So we must subvert this behaviour for .hc files by declaring
110 * variables as arrays, which eliminates the implicit dereference.
111 */
112 #if IN_STG_CODE
113 #define RTS_VAR(x) (x)[]
114 #define RTS_DEREF(x) (*(x))
115 #else
116 #define RTS_VAR(x) x
117 #define RTS_DEREF(x) x
118 #endif
119
120 /* bit macros
121 */
122 #define BITS_PER_BYTE 8
123 #define BITS_IN(x) (BITS_PER_BYTE * sizeof(x))
124
125 /* Compute offsets of struct fields
126 */
127 #define STG_FIELD_OFFSET(s_type, field) ((StgWord)&(((s_type*)0)->field))
128
129 /*
130 * 'Portable' inlining:
131 * INLINE_HEADER is for inline functions in header files (macros)
132 * STATIC_INLINE is for inline functions in source files
133 * EXTERN_INLINE is for functions that we want to inline sometimes
134 * (we also compile a static version of the function; see Inlines.c)
135 */
136
137 // We generally assume C99 semantics albeit these two definitions work fine even
138 // when gnu90 semantics are active (i.e. when __GNUC_GNU_INLINE__ is defined or
139 // when a GCC older than 4.2 is used)
140 //
141 // The problem, however, is with 'extern inline' whose semantics significantly
142 // differs between gnu90 and C99
143 #define INLINE_HEADER static inline
144 #define STATIC_INLINE static inline
145
146 // Figure out whether `__attributes__((gnu_inline))` is needed
147 // to force gnu90-style 'external inline' semantics.
148 #if defined(FORCE_GNU_INLINE)
149 // disable auto-detection since HAVE_GNU_INLINE has been defined externally
150 #elif __GNUC_GNU_INLINE__ && __GNUC__ == 4 && __GNUC_MINOR__ == 2
151 // GCC 4.2.x didn't properly support C99 inline semantics (GCC 4.3 was the first
152 // release to properly support C99 inline semantics), and therefore warned when
153 // using 'extern inline' while in C99 mode unless `__attributes__((gnu_inline))`
154 // was explicitly set.
155 # define FORCE_GNU_INLINE 1
156 #endif
157
158 #if FORCE_GNU_INLINE
159 // Force compiler into gnu90 semantics
160 # if defined(KEEP_INLINES)
161 # define EXTERN_INLINE inline __attribute__((gnu_inline))
162 # else
163 # define EXTERN_INLINE extern inline __attribute__((gnu_inline))
164 # endif
165 #elif __GNUC_GNU_INLINE__
166 // we're currently in gnu90 inline mode by default and
167 // __attribute__((gnu_inline)) may not be supported, so better leave it off
168 # if defined(KEEP_INLINES)
169 # define EXTERN_INLINE inline
170 # else
171 # define EXTERN_INLINE extern inline
172 # endif
173 #else
174 // Assume C99 semantics (yes, this curiously results in swapped definitions!)
175 // This is the preferred branch, and at some point we may drop support for
176 // compilers not supporting C99 semantics altogether.
177 # if defined(KEEP_INLINES)
178 # define EXTERN_INLINE extern inline
179 # else
180 # define EXTERN_INLINE inline
181 # endif
182 #endif
183
184
185 /*
186 * GCC attributes
187 */
188 #if defined(__GNUC__)
189 #define GNU_ATTRIBUTE(at) __attribute__((at))
190 #else
191 #define GNU_ATTRIBUTE(at)
192 #endif
193
194 #if __GNUC__ >= 3
195 #define GNUC3_ATTRIBUTE(at) __attribute__((at))
196 #else
197 #define GNUC3_ATTRIBUTE(at)
198 #endif
199
200 #if __GNUC__ > 4 || __GNUC__ == 4 && __GNUC_MINOR__ >= 3
201 #define GNUC_ATTR_HOT __attribute__((hot))
202 #else
203 #define GNUC_ATTR_HOT /* nothing */
204 #endif
205
206 #define STG_UNUSED GNUC3_ATTRIBUTE(__unused__)
207
208 /* -----------------------------------------------------------------------------
209 Global type definitions
210 -------------------------------------------------------------------------- */
211
212 #include "MachDeps.h"
213 #include "stg/Types.h"
214
215 /* -----------------------------------------------------------------------------
216 Shorthand forms
217 -------------------------------------------------------------------------- */
218
219 typedef StgChar C_;
220 typedef StgWord W_;
221 typedef StgWord* P_;
222 typedef StgInt I_;
223 typedef StgWord StgWordArray[];
224 typedef StgFunPtr F_;
225
226 #define EI_(X) extern StgWordArray (X) GNU_ATTRIBUTE(aligned (8))
227 #define II_(X) static StgWordArray (X) GNU_ATTRIBUTE(aligned (8))
228 #define IF_(f) static StgFunPtr GNUC3_ATTRIBUTE(used) f(void)
229 #define FN_(f) StgFunPtr f(void)
230 #define EF_(f) StgFunPtr f(void) /* External Cmm functions */
231 #define EFF_(f) void f() /* See Note [External function prototypes] */
232
233 /* Note [External function prototypes] See Trac #8965, #11395
234 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
235 In generated C code we need to distinct between two types
236 of external symbols:
237 1. Cmm functions declared by 'EF_' macro (External Functions)
238 2. C functions declared by 'EFF_' macro (External Foreign Functions)
239
240 Cmm functions are simple as they are internal to GHC.
241
242 C functions are trickier:
243
244 The external-function macro EFF_(F) used to be defined as
245 extern StgFunPtr f(void)
246 i.e a function of zero arguments. On most platforms this doesn't
247 matter very much: calls to these functions put the parameters in the
248 usual places anyway, and (with the exception of varargs) things just
249 work.
250
251 However, the ELFv2 ABI on ppc64 optimises stack allocation
252 (http://gcc.gnu.org/ml/gcc-patches/2013-11/msg01149.html): a call to a
253 function that has a prototype, is not varargs, and receives all parameters
254 in registers rather than on the stack does not require the caller to
255 allocate an argument save area. The incorrect prototypes cause GCC to
256 believe that all functions declared this way can be called without an
257 argument save area, but if the callee has sufficiently many arguments then
258 it will expect that area to be present, and will thus corrupt the caller's
259 stack. This happens in particular with calls to runInteractiveProcess in
260 libraries/process/cbits/runProcess.c, and led to Trac #8965.
261
262 The simplest fix appears to be to declare these external functions with an
263 unspecified argument list rather than a void argument list. This is no
264 worse for platforms that don't care either way, and allows a successful
265 bootstrap of GHC 7.8 on little-endian Linux ppc64 (which uses the ELFv2
266 ABI).
267
268 Another case is m68k ABI where 'void*' return type is returned by 'a0'
269 register while 'long' return type is returned by 'd0'. Thus we trick
270 external prototype return neither of these types to workaround #11395.
271 */
272
273
274 /* -----------------------------------------------------------------------------
275 Tail calls
276 -------------------------------------------------------------------------- */
277
278 #define JMP_(cont) return((StgFunPtr)(cont))
279
280 /* -----------------------------------------------------------------------------
281 Other Stg stuff...
282 -------------------------------------------------------------------------- */
283
284 #include "stg/DLL.h"
285 #include "stg/RtsMachRegs.h"
286 #include "stg/Regs.h"
287 #include "stg/Ticky.h"
288
289 #if IN_STG_CODE
290 /*
291 * This is included later for RTS sources, after definitions of
292 * StgInfoTable, StgClosure and so on.
293 */
294 #include "stg/MiscClosures.h"
295 #endif
296
297 #include "stg/Prim.h" /* ghc-prim fallbacks */
298 #include "stg/SMP.h" // write_barrier() inline is required
299
300 /* -----------------------------------------------------------------------------
301 Moving Floats and Doubles
302
303 ASSIGN_FLT is for assigning a float to memory (usually the
304 stack/heap). The memory address is guaranteed to be
305 StgWord aligned (currently == sizeof(void *)).
306
307 PK_FLT is for pulling a float out of memory. The memory is
308 guaranteed to be StgWord aligned.
309 -------------------------------------------------------------------------- */
310
311 INLINE_HEADER void ASSIGN_FLT (W_ [], StgFloat);
312 INLINE_HEADER StgFloat PK_FLT (W_ []);
313
314 #if ALIGNMENT_FLOAT <= ALIGNMENT_VOID_P
315
316 INLINE_HEADER void ASSIGN_FLT(W_ p_dest[], StgFloat src) { *(StgFloat *)p_dest = src; }
317 INLINE_HEADER StgFloat PK_FLT (W_ p_src[]) { return *(StgFloat *)p_src; }
318
319 #else /* ALIGNMENT_FLOAT > ALIGNMENT_UNSIGNED_INT */
320
321 INLINE_HEADER void ASSIGN_FLT(W_ p_dest[], StgFloat src)
322 {
323 float_thing y;
324 y.f = src;
325 *p_dest = y.fu;
326 }
327
328 INLINE_HEADER StgFloat PK_FLT(W_ p_src[])
329 {
330 float_thing y;
331 y.fu = *p_src;
332 return(y.f);
333 }
334
335 #endif /* ALIGNMENT_FLOAT > ALIGNMENT_VOID_P */
336
337 #if ALIGNMENT_DOUBLE <= ALIGNMENT_VOID_P
338
339 INLINE_HEADER void ASSIGN_DBL (W_ [], StgDouble);
340 INLINE_HEADER StgDouble PK_DBL (W_ []);
341
342 INLINE_HEADER void ASSIGN_DBL(W_ p_dest[], StgDouble src) { *(StgDouble *)p_dest = src; }
343 INLINE_HEADER StgDouble PK_DBL (W_ p_src[]) { return *(StgDouble *)p_src; }
344
345 #else /* ALIGNMENT_DOUBLE > ALIGNMENT_VOID_P */
346
347 /* Sparc uses two floating point registers to hold a double. We can
348 * write ASSIGN_DBL and PK_DBL by directly accessing the registers
349 * independently - unfortunately this code isn't writable in C, we
350 * have to use inline assembler.
351 */
352 #if sparc_HOST_ARCH
353
354 #define ASSIGN_DBL(dst0,src) \
355 { StgPtr dst = (StgPtr)(dst0); \
356 __asm__("st %2,%0\n\tst %R2,%1" : "=m" (((P_)(dst))[0]), \
357 "=m" (((P_)(dst))[1]) : "f" (src)); \
358 }
359
360 #define PK_DBL(src0) \
361 ( { StgPtr src = (StgPtr)(src0); \
362 register double d; \
363 __asm__("ld %1,%0\n\tld %2,%R0" : "=f" (d) : \
364 "m" (((P_)(src))[0]), "m" (((P_)(src))[1])); d; \
365 } )
366
367 #else /* ! sparc_HOST_ARCH */
368
369 INLINE_HEADER void ASSIGN_DBL (W_ [], StgDouble);
370 INLINE_HEADER StgDouble PK_DBL (W_ []);
371
372 typedef struct
373 { StgWord dhi;
374 StgWord dlo;
375 } unpacked_double;
376
377 typedef union
378 { StgDouble d;
379 unpacked_double du;
380 } double_thing;
381
382 INLINE_HEADER void ASSIGN_DBL(W_ p_dest[], StgDouble src)
383 {
384 double_thing y;
385 y.d = src;
386 p_dest[0] = y.du.dhi;
387 p_dest[1] = y.du.dlo;
388 }
389
390 /* GCC also works with this version, but it generates
391 the same code as the previous one, and is not ANSI
392
393 #define ASSIGN_DBL( p_dest, src ) \
394 *p_dest = ((double_thing) src).du.dhi; \
395 *(p_dest+1) = ((double_thing) src).du.dlo \
396 */
397
398 INLINE_HEADER StgDouble PK_DBL(W_ p_src[])
399 {
400 double_thing y;
401 y.du.dhi = p_src[0];
402 y.du.dlo = p_src[1];
403 return(y.d);
404 }
405
406 #endif /* ! sparc_HOST_ARCH */
407
408 #endif /* ALIGNMENT_DOUBLE > ALIGNMENT_UNSIGNED_INT */
409
410
411 /* -----------------------------------------------------------------------------
412 Moving 64-bit quantities around
413
414 ASSIGN_Word64 assign an StgWord64/StgInt64 to a memory location
415 PK_Word64 load an StgWord64/StgInt64 from a amemory location
416
417 In both cases the memory location might not be 64-bit aligned.
418 -------------------------------------------------------------------------- */
419
420 #if SIZEOF_HSWORD == 4
421
422 typedef struct
423 { StgWord dhi;
424 StgWord dlo;
425 } unpacked_double_word;
426
427 typedef union
428 { StgInt64 i;
429 unpacked_double_word iu;
430 } int64_thing;
431
432 typedef union
433 { StgWord64 w;
434 unpacked_double_word wu;
435 } word64_thing;
436
437 INLINE_HEADER void ASSIGN_Word64(W_ p_dest[], StgWord64 src)
438 {
439 word64_thing y;
440 y.w = src;
441 p_dest[0] = y.wu.dhi;
442 p_dest[1] = y.wu.dlo;
443 }
444
445 INLINE_HEADER StgWord64 PK_Word64(W_ p_src[])
446 {
447 word64_thing y;
448 y.wu.dhi = p_src[0];
449 y.wu.dlo = p_src[1];
450 return(y.w);
451 }
452
453 INLINE_HEADER void ASSIGN_Int64(W_ p_dest[], StgInt64 src)
454 {
455 int64_thing y;
456 y.i = src;
457 p_dest[0] = y.iu.dhi;
458 p_dest[1] = y.iu.dlo;
459 }
460
461 INLINE_HEADER StgInt64 PK_Int64(W_ p_src[])
462 {
463 int64_thing y;
464 y.iu.dhi = p_src[0];
465 y.iu.dlo = p_src[1];
466 return(y.i);
467 }
468
469 #elif SIZEOF_VOID_P == 8
470
471 INLINE_HEADER void ASSIGN_Word64(W_ p_dest[], StgWord64 src)
472 {
473 p_dest[0] = src;
474 }
475
476 INLINE_HEADER StgWord64 PK_Word64(W_ p_src[])
477 {
478 return p_src[0];
479 }
480
481 INLINE_HEADER void ASSIGN_Int64(W_ p_dest[], StgInt64 src)
482 {
483 p_dest[0] = src;
484 }
485
486 INLINE_HEADER StgInt64 PK_Int64(W_ p_src[])
487 {
488 return p_src[0];
489 }
490
491 #endif /* SIZEOF_HSWORD == 4 */
492
493 /* -----------------------------------------------------------------------------
494 Split markers
495 -------------------------------------------------------------------------- */
496
497 #if defined(USE_SPLIT_MARKERS)
498 #if defined(LEADING_UNDERSCORE)
499 #define __STG_SPLIT_MARKER __asm__("\n___stg_split_marker:");
500 #else
501 #define __STG_SPLIT_MARKER __asm__("\n__stg_split_marker:");
502 #endif
503 #else
504 #define __STG_SPLIT_MARKER /* nothing */
505 #endif
506
507 /* -----------------------------------------------------------------------------
508 Integer multiply with overflow
509 -------------------------------------------------------------------------- */
510
511 /* Multiply with overflow checking.
512 *
513 * This is tricky - the usual sign rules for add/subtract don't apply.
514 *
515 * On 32-bit machines we use gcc's 'long long' types, finding
516 * overflow with some careful bit-twiddling.
517 *
518 * On 64-bit machines where gcc's 'long long' type is also 64-bits,
519 * we use a crude approximation, testing whether either operand is
520 * larger than 32-bits; if neither is, then we go ahead with the
521 * multiplication.
522 *
523 * Return non-zero if there is any possibility that the signed multiply
524 * of a and b might overflow. Return zero only if you are absolutely sure
525 * that it won't overflow. If in doubt, return non-zero.
526 */
527
528 #if SIZEOF_VOID_P == 4
529
530 #ifdef WORDS_BIGENDIAN
531 #define RTS_CARRY_IDX__ 0
532 #define RTS_REM_IDX__ 1
533 #else
534 #define RTS_CARRY_IDX__ 1
535 #define RTS_REM_IDX__ 0
536 #endif
537
538 typedef union {
539 StgInt64 l;
540 StgInt32 i[2];
541 } long_long_u ;
542
543 #define mulIntMayOflo(a,b) \
544 ({ \
545 StgInt32 r, c; \
546 long_long_u z; \
547 z.l = (StgInt64)a * (StgInt64)b; \
548 r = z.i[RTS_REM_IDX__]; \
549 c = z.i[RTS_CARRY_IDX__]; \
550 if (c == 0 || c == -1) { \
551 c = ((StgWord)((a^b) ^ r)) \
552 >> (BITS_IN (I_) - 1); \
553 } \
554 c; \
555 })
556
557 /* Careful: the carry calculation above is extremely delicate. Make sure
558 * you test it thoroughly after changing it.
559 */
560
561 #else
562
563 /* Approximate version when we don't have long arithmetic (on 64-bit archs) */
564
565 /* If we have n-bit words then we have n-1 bits after accounting for the
566 * sign bit, so we can fit the result of multiplying 2 (n-1)/2-bit numbers */
567 #define HALF_POS_INT (((I_)1) << ((BITS_IN (I_) - 1) / 2))
568 #define HALF_NEG_INT (-HALF_POS_INT)
569
570 #define mulIntMayOflo(a,b) \
571 ({ \
572 I_ c; \
573 if ((I_)a <= HALF_NEG_INT || a >= HALF_POS_INT \
574 || (I_)b <= HALF_NEG_INT || b >= HALF_POS_INT) {\
575 c = 1; \
576 } else { \
577 c = 0; \
578 } \
579 c; \
580 })
581 #endif
582
583 #endif /* STG_H */