1 /* -----------------------------------------------------------------------------
2 * $Id: TailCalls.h,v 1.19 2005/03/08 08:59:58 simonmar Exp $
4 * (c) The GHC Team, 1998-1999
6 * Stuff for implementing proper tail jumps.
8 * ---------------------------------------------------------------------------*/
13 /* -----------------------------------------------------------------------------
14 Unmangled tail-jumping: use the mini interpretter.
15 -------------------------------------------------------------------------- */
17 #ifdef USE_MINIINTERPRETER
19 #define JMP_(cont) return((StgFunPtr)(cont))
25 extern void __DISCARD__(void);
27 /* -----------------------------------------------------------------------------
29 -------------------------------------------------------------------------- */
33 /* Note about discard: possibly there to fool GCC into clearing up
34 before we do the jump eg. if there are some arguments left on the C
35 stack that GCC hasn't popped yet. Also possibly to fool any
36 optimisations (a function call often acts as a barrier). Not sure
37 if any of this is necessary now -- SDM
39 Comment to above note: I don't think the __DISCARD__() in JMP_ is
40 necessary. Arguments should be popped from the C stack immediately
41 after returning from a function, as long as we pass -fno-defer-pop
42 to gcc. Moreover, a goto to a first-class label acts as a barrier
43 for optimisations in the same way a function call does.
47 /* The goto here seems to cause gcc -O2 to delete all the code after
48 it - including the FE_ marker and the epilogue code - exactly what
56 __target = (void *)(cont); \
60 #endif /* i386_HOST_ARCH */
62 /* -----------------------------------------------------------------------------
63 Tail calling on x86_64
64 -------------------------------------------------------------------------- */
69 NOTE about __DISCARD__():
71 On x86_64 this is necessary to work around bugs in the register
72 variable support in gcc. Without the __DISCARD__() call, gcc will
73 silently throw away assignements to global register variables that
74 happen before the jump.
85 without the dummy function call, gcc throws away the assignment to R1
94 __target = (void *)(cont); \
98 #endif /* x86_64_HOST_ARCH */
100 /* -----------------------------------------------------------------------------
101 Tail calling on Sparc
102 -------------------------------------------------------------------------- */
104 #ifdef sparc_HOST_ARCH
106 #define JMP_(cont) ((F_) (cont))()
107 /* Oh so happily, the above turns into a "call" instruction,
108 which, on a SPARC, is nothing but a "jmpl" with the
109 return address in %o7 [which we don't care about].
112 /* Don't need these for sparc mangling */
116 #endif /* sparc_HOST_ARCH */
118 /* -----------------------------------------------------------------------------
119 Tail calling on Alpha
120 -------------------------------------------------------------------------- */
122 #ifdef alpha_HOST_ARCH
125 register void *_procedure __asm__("$27");
129 do { _procedure = (void *)(cont); \
134 /* Don't need these for alpha mangling */
138 #endif /* alpha_HOST_ARCH */
140 /* -----------------------------------------------------------------------------
143 Description of HP's weird procedure linkage, many thanks to Andy Bennet
144 <andy_bennett@hp.com>:
146 I've been digging a little further into the problem of how HP-UX does
147 dynamic procedure calls. My solution in the last e-mail inserting an extra
148 'if' statement into the JMP_ I think is probably the best general solution I
149 can come up with. There are still a few problems with it however: It wont
150 work, if JMP_ ever has to call anything in a shared library, if this is
151 likely to be required it'll need something more elaborate. It also wont work
152 with PA-RISC 2.0 wide mode (64-bit) which uses a different format PLT.
154 I had some feedback from someone in HP's compiler lab and the problem
155 relates to the linker on HP-UX, not gcc as I first suspected. The reason the
156 'hsc' executable works is most likely due to a change in 'ld's behaviour for
157 performance reasons between your revision and mine.
159 The major issue relating to this is shared libraries and how they are
160 implented under HP-UX. The whole point of the Procedure Label Table (PLT) is
161 to allow a function pointer to hold the address of the function and a
162 pointer to the library's global data lookup table (DLT) used by position
163 independent code (PIC). This makes the PLT absolutely essential for shared
164 library calls. HP has two linker introduced assembly functions for dealing
165 with dynamic calls, $$dyncall and $$dyncall_external. The former does a
166 check to see if the address is a PLT pointer and dereferences if necessary
167 or just calls the address otherwise; the latter skips the check and just
168 does the indirect jump no matter what.
170 Since $$dyncall_external runs faster due to its not having the test, the
171 linker nowadays prefers to generate calls to that, rather than $$dyncall. It
172 makes this decision based on the presence of any shared library. If it even
173 smells an sl's existence at link time, it rigs the runtime system to
174 generate PLT references for everything on the assumption that the result
175 will be slightly more efficient. This is what is crashing GHC since the
176 calls it is generating have no understanding of the procedure label proper.
177 The only way to get real addresses is to link everything archive, including
178 system libraries, at which point it assumes you probably are going to be
179 using calls similar to GHC's (its rigged for HP's +ESfic compiler option)
180 but uses $$dyncall if necessary to cope, just in case you aren't.
182 -------------------------------------------------------------------------- */
184 #ifdef hppa1_1_hp_hpux_TARGET
187 do { void *_procedure = (void *)(cont); \
188 if (((int) _procedure) & 2) \
189 _procedure = (void *)(*((int *) (_procedure - 2))); \
193 #endif /* hppa1_1_hp_hpux_TARGET */
195 /* -----------------------------------------------------------------------------
196 Tail calling on PowerPC
197 -------------------------------------------------------------------------- */
199 #ifdef powerpc_HOST_ARCH
204 target = (void *)(cont); \
210 The __DISCARD__ is there because Apple's April 2002 Beta of GCC 3.1
211 sometimes generates incorrect code otherwise.
212 It tends to "forget" to update global register variables in the presence
213 of decrement/increment operators:
214 JMP_(*(--Sp)) is wrongly compiled as JMP_(Sp[-1]).
215 Calling __DISCARD__ in between works around this problem.
219 I would _love_ to use the following instead,
220 but some versions of Apple's GCC fail to generate code for it
221 if it is called for a casted data pointer - which is exactly what
222 we are going to do...
224 #define JMP_(cont) ((F_) (cont))()
227 #endif /* powerpc_HOST_ARCH */
229 #ifdef powerpc64_HOST_ARCH
230 #define JMP_(cont) ((F_) (cont))()
233 /* -----------------------------------------------------------------------------
235 -------------------------------------------------------------------------- */
237 #ifdef ia64_HOST_ARCH
239 /* The compiler can more intelligently decide how to do this. We therefore
240 * implement it as a call and optimise to a jump at mangle time. */
241 #define JMP_(cont) ((F_) (cont))(); __asm__ volatile ("--- TAILCALL ---");
243 /* Don't emit calls to __DISCARD__ as this causes hassles */
244 #define __DISCARD__()
248 /* -----------------------------------------------------------------------------
251 These are markers indicating the start and end of Real Code in a
252 function. All instructions between the actual start and end of the
253 function and these markers is shredded by the mangler.
254 -------------------------------------------------------------------------- */
256 /* The following __DISCARD__() has become necessary with gcc 2.96 on x86.
257 * It prevents gcc from moving stack manipulation code from the function
258 * body (aka the Real Code) into the function prologue, ie, from moving it
259 * over the --- BEGIN --- marker. It should be noted that (like some
260 * other black magic in GHC's code), there is no essential reason why gcc
261 * could not move some stack manipulation code across the __DISCARD__() -
262 * it just doesn't choose to do it at the moment.
267 #define FB_ __asm__ volatile ("--- BEGIN ---"); __DISCARD__ ();
271 #define FE_ __asm__ volatile ("--- END ---");
274 #endif /* !USE_MINIINTERPRETER */
276 #endif /* TAILCALLS_H */