Ticket #1421: ghc-asm.lprl

File ghc-asm.lprl, 74.4 KB (added by maeder@…, 8 years ago)

fix for Solaris 11 linker

Line 
1%************************************************************************
2%*                                                                      *
3\section[Driver-asm-fiddling]{Fiddling with assembler files}
4%*                                                                      *
5%************************************************************************
6
7Tasks:
8\begin{itemize}
9\item
10Utterly stomp out C functions' prologues and epilogues; i.e., the
11stuff to do with the C stack.
12\item
13Any other required tidying up.
14\end{itemize}
15
16General note [chak]: Many regexps are very fragile because they rely on white
17space being in the right place.  This caused trouble with gcc 2.95 (at least
18on Linux), where the use of white space in .s files generated by gcc suddenly
19changed.  To guarantee compatibility across different versions of gcc, make
20sure (at least on i386-.*-linux) that regexps tolerate varying amounts of white
21space between an assembler statement and its arguments as well as after a the
22comma separating multiple arguments. 
23
24\emph{For the time being, I have corrected the regexps for i386-.*-linux.  I
25didn't touch all the regexps for other i386 platforms, as I don't have
26a box to test these changes.}
27
28HPPA specific notes:
29\begin{itemize}
30\item
31The HP linker is very picky about symbols being in the appropriate
32space (code vs. data).  When we mangle the threaded code to put the
33info tables just prior to the code, they wind up in code space
34rather than data space.  This means that references to *_info from
35un-mangled parts of the RTS (e.g. unthreaded GC code) get
36unresolved symbols.  Solution:  mini-mangler for .c files on HP.  I
37think this should really be triggered in the driver by a new -rts
38option, so that user code doesn't get mangled inappropriately.
39\item
40With reversed tables, jumps are to the _info label rather than to
41the _entry label.  The _info label is just an address in code
42space, rather than an entry point with the descriptive blob we
43talked about yesterday.  As a result, you can't use the call-style
44JMP_ macro.  However, some JMP_ macros take _info labels as targets
45and some take code entry points within the RTS.  The latter won't
46work with the goto-style JMP_ macro.  Sigh.  Solution: Use the goto
47style JMP_ macro, and mangle some more assembly, changing all
48"RP'literal" and "LP'literal" references to "R'literal" and
49"L'literal," so that you get the real address of the code, rather
50than the descriptive blob.  Also change all ".word P%literal"
51entries in info tables and vector tables to just ".word literal,"
52for the same reason.  Advantage: No more ridiculous call sequences.
53\end{itemize}
54
55%************************************************************************
56%*                                                                      *
57\subsection{Top-level code}
58%*                                                                      *
59%************************************************************************
60
61\begin{code}
62$TargetPlatform = $TARGETPLATFORM;
63
64($Pgm = $0) =~ s|.*/||;
65$ifile = $ARGV[0];
66$ofile = $ARGV[1];
67
68if ( $TargetPlatform =~ /^i386-/ ) {
69    if ($ARGV[2] eq '') {
70        $StolenX86Regs = 4;
71    } else {
72        $StolenX86Regs = $ARGV[2];
73    }
74}
75
76&mangle_asm($ifile,$ofile);
77
78exit(0);
79\end{code}
80
81%************************************************************************
82%*                                                                      *
83\subsection{Constants for various architectures}
84%*                                                                      *
85%************************************************************************
86
87\begin{code}
88sub init_TARGET_STUFF {
89
90    #--------------------------------------------------------#
91    if ( $TargetPlatform =~ /^alpha-.*-.*/ ) {
92
93    $T_STABBY       = 0; # 1 iff .stab things (usually if a.out format)
94    $T_US           = ''; # _ if symbols have an underscore on the front
95    $T_PRE_APP      = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
96    $T_CONST_LBL    = '^\$L?C(\d+):$'; # regexp for what such a lbl looks like
97    $T_POST_LBL     = ':';
98
99    $T_MOVE_DIRVS   = '^(\s*(\$.*\.\.ng:|\.align\s+\d+|\.(globl|ent)\s+\S+|\#.*|\.(file|loc)\s+\S+\s+\S+|\.text|\.r?data)\n)';
100    $T_COPY_DIRVS   = '^\s*(\$.*\.\.ng:|\#|\.(file|globl|ent|loc))';
101
102    $T_DOT_WORD     = '\.(long|quad|byte|word)';
103    $T_DOT_GLOBAL   = '^\t\.globl';
104    $T_HDR_literal  = "\.rdata\n\t\.align 3\n";
105    $T_HDR_misc     = "\.text\n\t\.align 3\n";
106    $T_HDR_data     = "\.data\n\t\.align 3\n";
107    $T_HDR_rodata   = "\.rdata\n\t\.align 3\n";
108    $T_HDR_closure  = "\.data\n\t\.align 3\n";
109    $T_HDR_info     = "\.text\n\t\.align 3\n";
110    $T_HDR_entry    = "\.text\n\t\.align 3\n";
111    $T_HDR_vector   = "\.text\n\t\.align 3\n";
112
113    #--------------------------------------------------------#
114    } elsif ( $TargetPlatform =~ /^hppa/ ) {
115
116    $T_STABBY       = 0; # 1 iff .stab things (usually if a.out format)
117    $T_US           = ''; # _ if symbols have an underscore on the front
118    $T_PRE_APP      = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
119    $T_CONST_LBL    = '^L\$C(\d+)$'; # regexp for what such a lbl looks like
120    $T_POST_LBL     = '';
121
122    $T_MOVE_DIRVS   = '^((\s+\.(IMPORT|EXPORT|PARAM).*|\s+\.align\s+\d+|\s+\.(SPACE|SUBSPA)\s+\S+|\s*)\n)';
123    $T_COPY_DIRVS   = '^\s+\.(IMPORT|EXPORT)';
124
125    $T_DOT_WORD     = '\.(blockz|word|half|byte)';
126    $T_DOT_GLOBAL   = '^\s+\.EXPORT';
127    $T_HDR_literal  = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
128    $T_HDR_misc     = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
129    $T_HDR_data     = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
130    $T_HDR_rodata   = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
131    $T_HDR_closure  = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
132    $T_HDR_info     = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
133    $T_HDR_entry    = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
134    $T_HDR_vector   = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
135
136    #--------------------------------------------------------#
137    } elsif ( $TargetPlatform =~ /^i386-.*-(linuxaout|freebsd2|nextstep3|cygwin32|mingw32)$/ ) {
138                                # NeXT added but not tested. CaS
139
140    $T_STABBY       = 1; # 1 iff .stab things (usually if a.out format)
141    $T_US           = '_'; # _ if symbols have an underscore on the front
142    $T_PRE_APP      = '^#'; # regexp that says what comes before APP/NO_APP
143    $T_CONST_LBL    = '^LC(\d+):$';
144    $T_POST_LBL     = ':';
145    $T_X86_PRE_LLBL_PAT = 'L';
146    $T_X86_PRE_LLBL         = 'L';
147    $T_X86_BADJMP   = '^\tjmp [^L\*]';
148
149    $T_MOVE_DIRVS   = '^(\s*(\.(p2)?align\s.*|\.globl\s+\S+|\.text|\.data|\.stab[^n].*|\.type\s+.*|\.size\s+.*|\.lcomm.*)\n)';
150    $T_COPY_DIRVS   = '\.(globl|stab|lcomm)';
151    $T_DOT_WORD     = '\.(long|word|value|byte|space)';
152    $T_DOT_GLOBAL   = '\.globl';
153    $T_HDR_literal  = "\.text\n\t\.align 2\n";
154    $T_HDR_misc     = "\.text\n\t\.align 2,0x90\n";
155    $T_HDR_data     = "\.data\n\t\.align 2\n";
156    $T_HDR_rodata   = "\.text\n\t\.align 2\n";
157    $T_HDR_closure  = "\.data\n\t\.align 2\n";
158    $T_HDR_info     = "\.text\n\t\.align 2\n"; # NB: requires padding
159    $T_HDR_entry    = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
160    $T_HDR_vector   = "\.text\n\t\.align 2\n"; # NB: requires padding
161
162    #--------------------------------------------------------#
163    } elsif ( $TargetPlatform =~ /^i386-.*-(solaris2|linux|gnu|freebsd|netbsd|openbsd|kfreebsdgnu)$/ ) {
164
165    $T_STABBY       = 0; # 1 iff .stab things (usually if a.out format)
166    $T_US           = ''; # _ if symbols have an underscore on the front
167    $T_PRE_APP      = # regexp that says what comes before APP/NO_APP
168                      ($TargetPlatform =~ /-(linux|gnu|freebsd|netbsd|openbsd)$/) ? '#' : '/' ;
169    $T_CONST_LBL    = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
170    $T_POST_LBL     = ':';
171    $T_X86_PRE_LLBL_PAT = '\.L';
172    $T_X86_PRE_LLBL         = '.L';
173    $T_X86_BADJMP   = '^\tjmp\s+[^\.\*]';
174
175    $T_MOVE_DIRVS   = '^(\s*(\.(p2)?align\s.*|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.size\s+\S+\s*,\s*\d+|\.ident.*|\.local.*)\n)';
176    $T_COPY_DIRVS   = '^\s*\.(globl|local)';
177
178    $T_DOT_WORD     = '\.(long|value|word|byte|zero)';
179    $T_DOT_GLOBAL   = '\.globl';
180    $T_HDR_literal  = "\.section\t\.rodata\n"; # or just use .text??? (WDP 95/11)
181    $T_HDR_misc     = "\.text\n\t\.align 4\n";
182    $T_HDR_data     = "\.data\n\t\.align 4\n";
183    $T_HDR_rodata   = "\.section\t\.rodata\n\t\.align 4\n";
184    $T_HDR_closure  = "\.data\n\t\.align 4\n";
185    $T_HDR_info     = "\.text\n\t\.align 4\n";
186    $T_HDR_entry    = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
187    $T_HDR_vector   = "\.text\n\t\.align 4\n"; # NB: requires padding
188
189    #--------------------------------------------------------#
190    } elsif ( $TargetPlatform =~ /^ia64-.*-linux$/ ) {
191
192    $T_STABBY       = 0; # 1 iff .stab things (usually if a.out format)
193    $T_US           = ''; # _ if symbols have an underscore on the front
194    $T_PRE_APP      = '#';
195    $T_CONST_LBL    = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
196    $T_POST_LBL     = ':';
197
198    $T_MOVE_DIRVS   = '^(\s*\.(global|proc|pred\.safe_across_calls|text|data|section|subsection|align|size|type|ident)\s+.*\n)';
199    $T_COPY_DIRVS   = '\.(global|proc)';
200
201    $T_DOT_WORD     = '\.(long|value|byte|zero)';
202    $T_DOT_GLOBAL   = '\.global';
203    $T_HDR_literal  = "\.section\t\.rodata\n";
204    $T_HDR_misc     = "\.text\n\t\.align 16\n"; # May contain code; align like 'entry'
205    $T_HDR_data     = "\.data\n\t\.align 8\n";
206    $T_HDR_rodata   = "\.section\t\.rodata\n\t\.align 8\n";
207    $T_HDR_closure  = "\.data\n\t\.align 8\n";
208    $T_HDR_info     = "\.text\n\t\.align 8\n";
209    $T_HDR_entry    = "\.text\n\t\.align 16\n";
210    $T_HDR_vector   = "\.text\n\t\.align 8\n";
211
212    #--------------------------------------------------------#
213    } elsif ( $TargetPlatform =~ /^x86_64-.*-(linux|openbsd)$/ ) {
214
215    $T_STABBY       = 0; # 1 iff .stab things (usually if a.out format)
216    $T_US           = ''; # _ if symbols have an underscore on the front
217    $T_PRE_APP      = '#';
218    $T_CONST_LBL    = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
219    $T_POST_LBL     = ':';
220
221    $T_MOVE_DIRVS   = '^(\s*\.(globl|text|data|section|align|size|type|ident|local)\s+.*\n)';
222    $T_COPY_DIRVS   = '\.(globl|type|size|local)';
223
224    $T_DOT_WORD     = '\.(quad|long|value|byte|zero)';
225    $T_DOT_GLOBAL   = '\.global';
226
227    $T_HDR_literal16 = "\.section\t\.rodata.cst16\n\t.align 16\n";
228    $T_HDR_literal  = "\.section\t\.rodata\n";
229
230    $T_HDR_misc     = "\.text\n\t\.align 8\n";
231    $T_HDR_data     = "\.data\n\t\.align 8\n";
232    $T_HDR_rodata   = "\.section\t\.rodata\n\t\.align 8\n";
233
234        # the assembler on x86_64/Linux refuses to generate code for
235        #   .quad  x - y
236        # where x is in the text section and y in the rodata section.
237        # It works if y is in the text section, though.  This is probably
238        # going to cause difficulties for PIC, I imagine.
239    $T_HDR_relrodata= "\.text\n\t\.align 8\n";
240
241    $T_HDR_closure  = "\.data\n\t\.align 8\n";
242    $T_HDR_info     = "\.text\n\t\.align 8\n";
243    $T_HDR_entry    = "\.text\n\t\.align 8\n";
244    $T_HDR_vector   = "\.text\n\t\.align 8\n";
245
246    #--------------------------------------------------------#
247    } elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) {
248
249    $T_STABBY       = 1; # 1 iff .stab things (usually if a.out format)
250    $T_US           = '_'; # _ if symbols have an underscore on the front
251    $T_PRE_APP      = '^# MAY NOT APPLY'; # regexp that says what comes before APP/NO_APP
252    $T_CONST_LBL    = '^LC(\d+):$';
253    $T_POST_LBL     = ':';
254
255    $T_MOVE_DIRVS   = '^(\s*(\.align\s+\d+|\.proc\s+\d+|\.const|\.cstring|\.globl\s+\S+|\.text|\.data|\.even|\.stab[^n].*)\n)';
256    $T_COPY_DIRVS   = '\.(globl|proc|stab)';
257
258    $T_DOT_WORD     = '\.long';
259    $T_DOT_GLOBAL   = '\.globl';
260    $T_HDR_literal  = "\.text\n\t\.even\n";
261    $T_HDR_misc     = "\.text\n\t\.even\n";
262    $T_HDR_data     = "\.data\n\t\.even\n";
263    $T_HDR_rodata   = "\.text\n\t\.even\n";
264    $T_HDR_closure  = "\.data\n\t\.even\n";
265    $T_HDR_info     = "\.text\n\t\.even\n";
266    $T_HDR_entry    = "\.text\n\t\.even\n";
267    $T_HDR_vector   = "\.text\n\t\.even\n";
268
269    #--------------------------------------------------------#
270    } elsif ( $TargetPlatform =~ /^mips-.*/ ) {
271
272    $T_STABBY       = 0; # 1 iff .stab things (usually if a.out format)
273    $T_US           = ''; # _ if symbols have an underscore on the front
274    $T_PRE_APP      = '^\s*#'; # regexp that says what comes before APP/NO_APP
275    $T_CONST_LBL    = '^\$LC(\d+):$'; # regexp for what such a lbl looks like
276    $T_POST_LBL     = ':';
277
278    $T_MOVE_DIRVS   = '^(\s*(\.align\s+\d+|\.(globl|ent)\s+\S+|\.text|\.r?data)\n)';
279    $T_COPY_DIRVS   = '\.(globl|ent)';
280
281    $T_DOT_WORD     = '\.word';
282    $T_DOT_GLOBAL   = '^\t\.globl';
283    $T_HDR_literal  = "\t\.rdata\n\t\.align 2\n";
284    $T_HDR_misc     = "\t\.text\n\t\.align 2\n";
285    $T_HDR_data     = "\t\.data\n\t\.align 2\n";
286    $T_HDR_rodata   = "\t\.rdata\n\t\.align 2\n";
287    $T_HDR_closure  = "\t\.data\n\t\.align 2\n";
288    $T_HDR_info     = "\t\.text\n\t\.align 2\n";
289    $T_HDR_entry    = "\t\.text\n\t\.align 2\n";
290    $T_HDR_vector   = "\t\.text\n\t\.align 2\n";
291
292    #--------------------------------------------------------#
293    } elsif ( $TargetPlatform =~ /^powerpc-apple-darwin.*/ ) {
294                                # Apple PowerPC Darwin/MacOS X.
295    $T_STABBY       = 0; # 1 iff .stab things (usually if a.out format)
296    $T_US           = '_'; # _ if symbols have an underscore on the front
297    $T_PRE_APP      = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
298    $T_CONST_LBL    = '^\LC\d+:'; # regexp for what such a lbl looks like
299    $T_POST_LBL     = ':';
300
301    $T_MOVE_DIRVS   = '^(\s*(\.align \d+|\.text|\.data|\.const_data|\.cstring|\.non_lazy_symbol_pointer|\.const|\.static_const|\.literal4|\.literal8|\.static_data|\.globl \S+|\.section .*|\.lcomm.*)\n)';
302    $T_COPY_DIRVS   = '\.(globl|lcomm)';
303
304    $T_DOT_WORD     = '\.(long|short|byte|fill|space)';
305    $T_DOT_GLOBAL   = '\.globl';
306    $T_HDR_toc      = "\.toc\n";
307    $T_HDR_literal  = "\t\.const\n\t\.align 2\n";
308    $T_HDR_misc     = "\t\.text\n\t\.align 2\n";
309    $T_HDR_data     = "\t\.data\n\t\.align 2\n";
310    $T_HDR_rodata   = "\t\.const\n\t\.align 2\n";
311    $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
312    $T_HDR_closure  = "\t\.data\n\t\.align 2\n";
313    $T_HDR_info     = "\t\.text\n\t\.align 2\n";
314    $T_HDR_entry    = "\t\.text\n\t\.align 2\n";
315    $T_HDR_vector   = "\t\.text\n\t\.align 2\n";
316
317    #--------------------------------------------------------#
318    } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ ) {
319                                # Apple PowerPC Darwin/MacOS X.
320    $T_STABBY       = 0; # 1 iff .stab things (usually if a.out format)
321    $T_US           = '_'; # _ if symbols have an underscore on the front
322    $T_PRE_APP      = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
323    $T_CONST_LBL    = '^\LC\d+:'; # regexp for what such a lbl looks like
324    $T_POST_LBL     = ':';
325    $T_X86_PRE_LLBL_PAT = 'L';
326    $T_X86_PRE_LLBL         = 'L';
327    $T_X86_BADJMP   = '^\tjmp [^L\*]';
328
329    $T_MOVE_DIRVS   = '^(\s*(\.align \d+|\.text|\.data|\.const_data|\.cstring|\.non_lazy_symbol_pointer|\.const|\.static_const|\.literal4|\.literal8|\.static_data|\.globl \S+|\.section .*|\.lcomm.*)\n)';
330    $T_COPY_DIRVS   = '\.(globl|lcomm)';
331
332    $T_DOT_WORD     = '\.(long|short|byte|fill|space)';
333    $T_DOT_GLOBAL   = '\.globl';
334    $T_HDR_toc      = "\.toc\n";
335    $T_HDR_literal16= "\t\.literal8\n\t\.align 4\n";
336    $T_HDR_literal  = "\t\.const\n\t\.align 4\n";
337    $T_HDR_misc     = "\t\.text\n\t\.align 2\n";
338    $T_HDR_data     = "\t\.data\n\t\.align 2\n";
339    $T_HDR_rodata   = "\t\.const\n\t\.align 2\n";
340    $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
341    $T_HDR_closure  = "\t\.data\n\t\.align 2\n";
342    $T_HDR_info     = "\t\.text\n\t\.align 2\n";
343    $T_HDR_entry    = "\t\.text\n\t\.align 2\n";
344    $T_HDR_vector   = "\t\.text\n\t\.align 2\n";
345
346    #--------------------------------------------------------#
347    } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ ) {
348                                # PowerPC Linux
349    $T_STABBY       = 0; # 1 iff .stab things (usually if a.out format)
350    $T_US           = ''; # _ if symbols have an underscore on the front
351    $T_PRE_APP      = '^#'; # regexp that says what comes before APP/NO_APP
352    $T_CONST_LBL    = '^\.LC\d+:'; # regexp for what such a lbl looks like
353    $T_POST_LBL     = ':';
354
355    $T_MOVE_DIRVS   = '^(\s*(\.(p2)?align\s+\d+(,\s*0x90)?|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.size\s+\S+\s*,\s*\d+|\.ident.*|\.local.*)\n)';
356    $T_COPY_DIRVS   = '^\s*\.(globl|type|size|local)';
357
358    $T_DOT_WORD     = '\.(long|short|byte|fill|space)';
359    $T_DOT_GLOBAL   = '\.globl';
360    $T_HDR_toc      = "\.toc\n";
361    $T_HDR_literal  = "\t\.section\t.rodata\n\t\.align 2\n";
362    $T_HDR_misc     = "\t\.text\n\t\.align 2\n";
363    $T_HDR_data     = "\t\.data\n\t\.align 2\n";
364    $T_HDR_rodata   = "\t\.section\t.rodata\n\t\.align 2\n";
365    $T_HDR_closure  = "\t\.data\n\t\.align 2\n";
366    $T_HDR_info     = "\t\.text\n\t\.align 2\n";
367    $T_HDR_entry    = "\t\.text\n\t\.align 2\n";
368    $T_HDR_vector   = "\t\.text\n\t\.align 2\n";
369
370    #--------------------------------------------------------#
371    } elsif ( $TargetPlatform =~ /^powerpc64-.*-linux/ ) {
372                                # PowerPC 64 Linux
373    $T_STABBY       = 0; # 1 iff .stab things (usually if a.out format)
374    $T_US           = '\.'; # _ if symbols have an underscore on the front
375    $T_PRE_APP      = '^#'; # regexp that says what comes before APP/NO_APP
376    $T_CONST_LBL    = '^\.LC\d+:'; # regexp for what such a lbl looks like
377    $T_POST_LBL     = ':';
378
379    $T_MOVE_DIRVS   = '^(\s*(\.(p2)?align\s+\d+(,\s*0x90)?|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.size\s+\S+\s*,\s*\d+|\.ident.*|\.local.*)\n)';
380    $T_COPY_DIRVS   = '^\s*\.(globl|type|size|local)';
381
382    $T_DOT_WORD     = '\.(long|short|byte|fill|space)';
383    $T_DOT_GLOBAL   = '\.globl';
384    $T_HDR_toc      = "\.toc\n";
385    $T_HDR_literal  = "\t\.section\t\".toc\",\"aw\"\n";
386    $T_HDR_misc     = "\t\.text\n\t\.align 2\n";
387    $T_HDR_data     = "\t\.data\n\t\.align 2\n";
388    $T_HDR_rodata   = "\t\.section\t.rodata\n\t\.align 2\n";
389    $T_HDR_closure  = "\t\.data\n\t\.align 2\n";
390    $T_HDR_info     = "\t\.text\n\t\.align 2\n";
391    $T_HDR_entry    = "\t\.text\n\t\.align 2\n";
392    $T_HDR_vector   = "\t\.text\n\t\.align 2\n";
393
394    #--------------------------------------------------------#
395    } elsif ( $TargetPlatform =~ /^sparc-.*-(solaris2|openbsd)/ ) {
396
397    $T_STABBY       = 0; # 1 iff .stab things (usually if a.out format)
398    $T_US           = ''; # _ if symbols have an underscore on the front
399    $T_PRE_APP      = 'DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
400    $T_CONST_LBL    = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
401    $T_POST_LBL     = ':';
402
403    $T_MOVE_DIRVS   =  '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\s+\.local\s+\S+|\.text|\.data|\.stab.*|\s*\.section.*|\s+\.type.*|\s+\.size.*)\n)';
404    $T_COPY_DIRVS   = '\.(global|local|proc|stab)';
405
406    $T_DOT_WORD     = '\.(long|word|byte|half|skip|uahalf|uaword)';
407    $T_DOT_GLOBAL   = '^\t\.global';
408    $T_HDR_literal  = "\.text\n\t\.align 8\n";
409    $T_HDR_misc     = "\.text\n\t\.align 4\n";
410    $T_HDR_data     = "\.data\n\t\.align 8\n";
411    $T_HDR_rodata   = "\.text\n\t\.align 4\n";
412    $T_HDR_closure  = "\.data\n\t\.align 4\n";
413    $T_HDR_info     = "\.text\n\t\.align 4\n";
414    $T_HDR_entry    = "\.text\n\t\.align 4\n";
415    $T_HDR_vector   = "\.text\n\t\.align 4\n";
416
417    #--------------------------------------------------------#
418    } elsif ( $TargetPlatform =~ /^sparc-.*-sunos4/ ) {
419
420    $T_STABBY       = 1; # 1 iff .stab things (usually if a.out format)
421    $T_US           = '_'; # _ if symbols have an underscore on the front
422    $T_PRE_APP      = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
423    $T_CONST_LBL    = '^LC(\d+):$';
424    $T_POST_LBL     = ':';
425
426    $T_MOVE_DIRVS   = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*)\n)';
427    $T_COPY_DIRVS   = '\.(global|proc|stab)';
428
429    $T_DOT_WORD     = '\.word';
430    $T_DOT_GLOBAL   = '^\t\.global';
431    $T_HDR_literal  = "\.text\n\t\.align 8\n";
432    $T_HDR_misc     = "\.text\n\t\.align 4\n";
433    $T_HDR_data     = "\.data\n\t\.align 8\n";
434    $T_HDR_rodata   = "\.text\n\t\.align 4\n";
435    $T_HDR_closure  = "\.data\n\t\.align 4\n";
436    $T_HDR_info     = "\.text\n\t\.align 4\n";
437    $T_HDR_entry    = "\.text\n\t\.align 4\n";
438    $T_HDR_vector   = "\.text\n\t\.align 4\n";
439
440    #--------------------------------------------------------#
441    } elsif ( $TargetPlatform =~ /^sparc-.*-linux/ ) {
442    $T_STABBY       = 0; # 1 iff .stab things (usually if a.out format)
443    $T_US           = ''; # _ if symbols have an underscore on the front
444    $T_PRE_APP      = '#'; # regexp that says what comes before APP/NO_APP
445                           # Probably doesn't apply anyway
446    $T_CONST_LBL    = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
447    $T_POST_LBL     = ':';
448
449    $T_MOVE_DIRVS   = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\s+\.local\s+\S+|\.text|\.data|\.seg|\.stab.*|\s+?\.section.*|\s+\.type.*|\s+\.size.*)\n)';
450    $T_COPY_DIRVS   = '\.(global|local|globl|proc|stab)';
451
452    $T_DOT_WORD     = '\.(long|word|nword|xword|byte|half|short|skip|uahalf|uaword)';
453    $T_DOT_GLOBAL   = '^\t\.global';
454    $T_HDR_literal  = "\.text\n\t\.align 8\n";
455    $T_HDR_misc     = "\.text\n\t\.align 4\n";
456    $T_HDR_data     = "\.data\n\t\.align 8\n";
457    $T_HDR_rodata   = "\.text\n\t\.align 4\n";
458    $T_HDR_closure  = "\.data\n\t\.align 4\n";
459    $T_HDR_info     = "\.text\n\t\.align 4\n";
460    $T_HDR_entry    = "\.text\n\t\.align 4\n";
461    $T_HDR_vector   = "\.text\n\t\.align 4\n";
462
463    #--------------------------------------------------------#
464    } else {
465        print STDERR "$Pgm: don't know how to mangle assembly language for: $TargetPlatform\n";
466        exit 1;
467    }
468
469    if($T_HDR_relrodata eq "") {
470            # default values:
471            # relrodata defaults to rodata.
472        $T_HDR_relrodata = $T_HDR_rodata;
473    }
474
475if ( 0 ) {
476print STDERR "T_STABBY: $T_STABBY\n";
477print STDERR "T_US: $T_US\n";
478print STDERR "T_PRE_APP: $T_PRE_APP\n";
479print STDERR "T_CONST_LBL: $T_CONST_LBL\n";
480print STDERR "T_POST_LBL: $T_POST_LBL\n";
481if ( $TargetPlatform =~ /^i386-/ ) {
482    print STDERR "T_X86_PRE_LLBL_PAT: $T_X86_PRE_LLBL_PAT\n";
483    print STDERR "T_X86_PRE_LLBL: $T_X86_PRE_LLBL\n";
484    print STDERR "T_X86_BADJMP: $T_X86_BADJMP\n";
485}
486print STDERR "T_MOVE_DIRVS: $T_MOVE_DIRVS\n";
487print STDERR "T_COPY_DIRVS: $T_COPY_DIRVS\n";
488print STDERR "T_DOT_WORD: $T_DOT_WORD\n";
489print STDERR "T_HDR_literal: $T_HDR_literal\n";
490print STDERR "T_HDR_misc: $T_HDR_misc\n";
491print STDERR "T_HDR_data: $T_HDR_data\n";
492print STDERR "T_HDR_rodata: $T_HDR_rodata\n";
493print STDERR "T_HDR_closure: $T_HDR_closure\n";
494print STDERR "T_HDR_info: $T_HDR_info\n";
495print STDERR "T_HDR_entry: $T_HDR_entry\n";
496print STDERR "T_HDR_vector: $T_HDR_vector\n";
497}
498
499}
500\end{code}
501
502%************************************************************************
503%*                                                                      *
504\subsection{Mangle away}
505%*                                                                      *
506%************************************************************************
507
508\begin{code}
509sub mangle_asm {
510    local($in_asmf, $out_asmf) = @_;
511
512    # multi-line regexp matching:
513    local($*) = 1;
514    local($i, $c);
515
516    # ia64-specific information for code chunks
517    my $ia64_locnum;
518    my $ia64_outnum;
519
520    &init_TARGET_STUFF();
521    &init_FUNNY_THINGS();
522
523    open(INASM, "< $in_asmf")
524        || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
525    open(OUTASM,"> $out_asmf")
526        || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
527
528    # read whole file, divide into "chunks":
529    #   record some info about what we've found...
530
531    @chk = ();          # contents of the chunk
532    $numchks = 0;       # number of them
533    @chkcat = ();       # what category of thing in each chunk
534    @chksymb = ();      # what symbol(base) is defined in this chunk
535    %entrychk = ();     # ditto, its entry code
536    %closurechk = ();   # ditto, the (static) closure
537    %srtchk = ();       # ditto, its SRT (for top-level things)
538    %infochk = ();      # given a symbol base, say what chunk its info tbl is in
539    %vectorchk = ();    # ditto, return vector table
540    $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
541
542    $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
543
544    while (<INASM>) {
545        tr/\r//d if $TargetPlatform =~ /-mingw32$/; # In case Perl doesn't convert line endings
546        next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
547        next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
548        next if /^\t\.def.*endef$/;
549        next if /${T_PRE_APP}(NO_)?APP/o;
550        next if /^;/ && $TargetPlatform =~ /^hppa/;
551
552        next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|ia64)-/;
553
554        if ( $TargetPlatform =~ /^mips-/
555          && /^\t\.(globl\S+\.text|comm\t)/ ) {
556            $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
557        # Treat .comm variables as data.  These show up in two (known) places:
558        #
559        #    - the module_registered variable used in the __stginit fragment.
560        #      even though these are declared static and initialised, gcc 3.3
561        #      likes to make them .comm, presumably to save space in the
562        #      object file.
563        #
564        #    - global variables used to pass arguments from C to STG in
565        #      a foreign export.  (is this still true? --SDM)
566        #
567        } elsif ( /^\t\.comm.*$/ ) {
568            $chk[++$i]   = $_;
569            $chkcat[$i]  = 'data';
570            $chksymb[$i] = '';
571
572        # Labels ending "_str": these are literal strings.
573        } elsif ( /^${T_US}([A-Za-z0-9_]+)_str${T_POST_LBL}$/ ) {
574            $chk[++$i]   = $_;
575            $chkcat[$i]  = 'rodata';
576            $chksymb[$i] = '';
577        } elsif ( $TargetPlatform =~ /-darwin/
578                && (/^\s*\.subsections_via_symbols/
579                  ||/^\s*\.no_dead_strip.*/)) {
580            # Don't allow Apple's linker to do any dead-stripping of symbols
581            # in this file, because it will mess up info-tables in mangled
582            # code.
583            # The .no_dead_strip directives are actually put there by
584            # the gcc3 "used" attribute on entry points.
585       
586        } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && (
587                   /^\s*\.picsymbol_stub/
588                || /^\s*\.section __TEXT,__picsymbol_stub\d,.*/
589                || /^\s*\.section __TEXT,__picsymbolstub\d,.*/
590                || /^\s*\.symbol_stub/
591                || /^\s*\.section __TEXT,__symbol_stub\d,.*/
592                || /^\s*\.section __TEXT,__symbolstub\d,.*/
593                || /^\s*\.lazy_symbol_pointer/
594                || /^\s*\.non_lazy_symbol_pointer/
595                || /^\s*\.section __IMPORT.*/))
596        {
597            $chk[++$i]   = $_;
598            $chkcat[$i]  = 'dyld';
599            $chksymb[$i] = '';
600            $dyld_section = $_;
601
602        } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.data/)
603        {       # non_lazy_symbol_ptrs that point to local symbols
604            $chk[++$i]   = $_;
605            $chkcat[$i]  = 'dyld';
606            $chksymb[$i] = '';
607            $dyld_section = $_;
608        } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.align/)
609        {       # non_lazy_symbol_ptrs that point to local symbols
610            $dyld_section .= $_;
611        } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^L_.*:$/)
612        {       # non_lazy_symbol_ptrs that point to local symbols
613            $chk[++$i]   = $dyld_section . $_;
614            $chkcat[$i]  = 'dyld';
615            $chksymb[$i] = '';
616
617        } elsif ( /^\s+/ ) { # most common case first -- a simple line!
618            # duplicated from the bottom
619
620            $chk[$i] .= $_;
621
622        } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
623            # Alphas: Local labels not to be confused with new chunks
624            $chk[$i] .= $_;
625        # NB: all the rest start with a non-space
626
627        } elsif ( $TargetPlatform =~ /^mips-/
628               && /^\d+:/ ) { # a funny-looking very-local label
629            $chk[$i] .= $_;
630
631        } elsif ( /$T_CONST_LBL/o ) {
632            $chk[++$i]   = $_;
633            $chkcat[$i]  = 'literal';
634            $chksymb[$i] = $1;
635
636        } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
637            $chk[++$i]   = $_;
638            $chkcat[$i]  = 'splitmarker';
639            $chksymb[$i] = $1;
640
641        } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
642            $symb = $1;
643            $chk[++$i]   = $_;
644            $chkcat[$i]  = 'infotbl';
645            $chksymb[$i] = $symb;
646
647            die "Info table already? $symb; $i\n" if defined($infochk{$symb});
648
649            $infochk{$symb} = $i;
650
651        } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
652            $chk[++$i]   = $_;
653            $chkcat[$i]  = 'entry';
654            $chksymb[$i] = $1;
655
656            $entrychk{$1} = $i;
657
658        } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
659            $chk[++$i]   = $_;
660            $chkcat[$i]  = 'closure';
661            $chksymb[$i] = $1;
662
663            $closurechk{$1} = $i;
664
665        } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
666            $chk[++$i]   = $_;
667            $chkcat[$i]  = 'srt';
668            $chksymb[$i] = $1;
669
670            $srtchk{$1} = $i;
671
672        } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
673            $chk[++$i]   = $_;
674            $chkcat[$i]  = 'data';
675            $chksymb[$i] = '';
676
677        } elsif ( /^${T_US}(stg_ap_stack_entries|stg_stack_save_entries|stg_arg_bitmaps)${T_POST_LBL}$/o ) {
678            $chk[++$i]   = $_;
679            $chkcat[$i]  = 'data';
680            $chksymb[$i] = '';
681
682        } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
683            ; # toss it
684
685        } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
686               || /^${T_US}.*_CAT${T_POST_LBL}$/o               # PROF: _entryname_CAT
687               || /^${T_US}.*_done${T_POST_LBL}$/o              # PROF: _module_done
688               || /^${T_US}_module_registered${T_POST_LBL}$/o   # PROF: _module_registered
689               ) {
690            $chk[++$i]   = $_;
691            $chkcat[$i]  = 'data';
692            $chksymb[$i] = '';
693
694        } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
695            $chk[++$i]   = $_;
696            $chkcat[$i]  = 'bss';
697            $chksymb[$i] = '';
698
699        } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
700            # all CC_ symbols go in the data section...
701            $chk[++$i]   = $_;
702            $chkcat[$i]  = 'data';
703            $chksymb[$i] = '';
704
705        } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
706            $chk[++$i]   = $_;
707            $chkcat[$i]  = 'misc';
708            $chksymb[$i] = '';
709        } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
710            $chk[++$i]   = $_;
711            $chkcat[$i]  = 'vector';
712            $chksymb[$i] = $1;
713
714            $vectorchk{$1} = $i;
715
716        } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
717             &&   /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
718            # Some Solaris system headers contain function definitions (as
719            # opposed to mere prototypes), which end up in the .hc file when
720            # a Haskell module foreign imports the corresponding system
721            # functions (most notably stat()).  We put them into the text
722            # segment.  Note that this currently does not extend to function
723            # names starting with an underscore.
724            # - chak 7/2001
725            $chk[++$i]   = $_;
726            $chkcat[$i]  = 'misc';
727            $chksymb[$i] = $1;
728
729        } elsif ( $TargetPlatform =~ /^i386-apple-darwin/ && /^(___i686\.get_pc_thunk\.[abcd]x):/o) {
730                # To handle PIC on Darwin/x86, we need to appropriately pass through
731                # the get_pc_thunk functions. The need to be put into a special section
732                # marked as coalesced (otherwise the .weak_definition doesn't work
733                # on Darwin).
734            $chk[++$i]   = $_;
735            $chkcat[$i]  = 'get_pc_thunk';
736            $chksymb[$i] = $1;
737
738        } elsif ( /^${T_US}[A-Za-z0-9_]/o
739                && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
740                   || ! /^L\$\d+$/ )
741                && ( $TargetPlatform !~ /^powerpc64/ # we need to avoid local labels in this case
742                   || ! /^\.L\d+:$/ ) ) {
743            local($thing);
744            chop($thing = $_);
745            $thing =~ s/:$//;
746            $chk[++$i]   = $_;
747            $chksymb[$i] = '';
748            if (
749                       /^${T_US}stg_.*${T_POST_LBL}$/o          # RTS internals
750                    || /^${T_US}__stg_.*${T_POST_LBL}$/o        # more RTS internals
751                    || /^${T_US}__fexp_.*${T_POST_LBL}$/o       # foreign export
752                    || /^${T_US}.*_slow${T_POST_LBL}$/o         # slow entry
753                    || /^${T_US}__stginit.*${T_POST_LBL}$/o     # __stginit<module>
754                    || /^${T_US}.*_btm${T_POST_LBL}$/o          # large bitmaps
755                    || /^${T_US}.*_fast${T_POST_LBL}$/o         # primops
756                    || /^_uname:/o                              # x86/Solaris2
757                )
758            {
759                $chkcat[$i]  = 'misc';
760            } elsif (
761                       /^${T_US}.*_srtd${T_POST_LBL}$/o          # large bitmaps
762                    || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o  # closure tables
763                )
764            {
765                $chkcat[$i] = 'relrodata';
766            } else
767            {
768                print STDERR "Warning: retaining unknown function \`$thing' in output from C compiler\n";
769                $chkcat[$i]  = 'unknown';
770            }
771
772        } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ && /^\.LCTOC1 = /o ) {
773                # PowerPC Linux's large-model PIC (-fPIC) generates a gobal offset
774                # table "by hand". Be sure to copy it over.
775                # Note that this label and all entries in the table should actually
776                # go into the .got2 section, but it isn't easy to distinguish them
777                # from other constant literals (.LC\d+), so we just put everything
778                # in .rodata.
779            $chk[++$i]   = $_;
780            $chkcat[$i]  = 'literal';
781            $chksymb[$i] = 'LCTOC1';
782        } else { # simple line (duplicated at the top)
783
784            $chk[$i] .= $_;
785        }
786    }
787    $numchks = $#chk + 1;
788    $chk[$numchks] = ''; # We might push .note.GNU-stack into this
789    $chkcat[$numchks] = 'verbatim'; # If we do, write it straight back out
790
791    # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
792    # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
793    # close CHUNKS;
794
795    # the division into chunks is imperfect;
796    # we throw some things over the fence into the next
797    # chunk.
798    #
799    # also, there are things we would like to know
800    # about the whole module before we start spitting
801    # output.
802
803    local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
804    local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
805
806#   print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
807
808    # Alphas: NB: we start meddling at chunk 1, not chunk 0
809    # The first ".rdata" is quite magical; as of GCC 2.7.x, it
810    # spits a ".quad 0" in after the very first ".rdata"; we
811    # detect this special case (tossing the ".quad 0")!
812    local($magic_rdata_seen) = 0;
813 
814    # HPPAs, MIPSen: also start medding at chunk 1
815
816    for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
817        $c = $chk[$i]; # convenience copy
818
819#       print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
820
821        # toss all prologue stuff; HPPA is pretty weird
822        # (see elsewhere)
823        $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
824
825        undef $ia64_locnum;
826        undef $ia64_outnum;
827
828        # be slightly paranoid to make sure there's
829        # nothing surprising in there
830        if ( $c =~ /--- BEGIN ---/ ) {
831            if (($p, $r) = split(/--- BEGIN ---/, $c)) {
832
833                # remove junk whitespace around the split point
834                $p =~ s/\t+$//;
835                $r =~ s/^\s*\n//;
836
837                if ($TargetPlatform =~ /^i386-/) {
838                    if ($p =~ /^\tsubl\s+\$(\d+),\s*\%esp\n/) {
839                        if ($1 >= 8192) {
840                            die "Error: reserved stack space exceeded!\n  Possible workarounds: compile with -fasm, or try another version of gcc.\n"
841                        }
842                    }
843
844                # gcc 3.4.3 puts this kind of stuff in the prologue, eg.
845                # when compiling PrimOps.cmm with -optc-O2:
846                #        xorl    %ecx, %ecx
847                #        xorl    %edx, %edx
848                #        movl    %ecx, 16(%esp)
849                #        movl    %edx, 20(%esp)
850                # but then the code of the function doesn't assume
851                # anything about the contnets of these stack locations.
852                # I think it's to do with the use of inline functions for
853                # PK_Word64() and friends, where gcc is initialising the
854                # contents of the struct to zero, and failing to optimise
855                # away the initialisation.  Let's live dangerously and
856                # discard these initalisations.
857
858                    $p =~ s/^\tpushl\s+\%e(di|si|bx)\n//g;
859                    $p =~ s/^\txorl\s+\%e(ax|cx|dx),\s*\%e(ax|cx|dx)\n//g;
860                    $p =~ s/^\tmovl\s+\%e(ax|cx|dx|si|di),\s*\d*\(\%esp\)\n//g;
861                    $p =~ s/^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n//g;
862                    $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
863                    $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-(cygwin32|mingw32)/);
864
865                    if ($TargetPlatform =~ /^i386-apple-darwin/) {
866                        $pcrel_label = $p;
867                        $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
868                        $pcrel_reg = $p;
869                        $pcrel_reg =~ s/(.|\n)*.*___i686\.get_pc_thunk\.([abcd]x)\n(.|\n)*/$2/ or $pcrel_reg = "";
870                        $p =~ s/^\s+call\s+___i686\.get_pc_thunk\..x//;
871                        $p =~ s/^\"?L\d+\$pb\"?:\n//;
872
873                        if ($pcrel_reg eq "bx") {
874                            # Bad gcc. Goes and uses %ebx, our BaseReg, for PIC. Bad gcc.
875                            die "Darwin/x86: -fPIC -via-C doesn't work yet, use -fasm. Aborting."
876                        }
877                    }
878
879                } elsif ($TargetPlatform =~ /^x86_64-/) {
880                    $p =~ s/^\tpushq\s+\%r(bx|bp|12|13|14)\n//g;
881                    $p =~ s/^\tmovq\s+\%r(bx|bp|12|13|14),\s*\d*\(\%rsp\)\n//g;
882                    $p =~ s/^\tsubq\s+\$\d+,\s*\%rsp\n//;
883
884                } elsif ($TargetPlatform =~ /^ia64-/) {
885                    $p =~ s/^\t\.prologue .*\n//;
886
887                    # Record the number of local and out registers for register relocation later
888                    $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, (\d+), (\d+), 0\n//;
889                    $ia64_locnum = $1;
890                    $ia64_outnum = $2;
891
892                    $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
893                    $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
894
895                    # Ignore save/restore of these registers; they're taken
896                    # care of in StgRun()
897                    $p =~ s/^\t\.save ar\.lc, r\d+\n//;
898                    $p =~ s/^\t\.save pr, r\d+\n//;
899                    $p =~ s/^\tmov r\d+ = ar\.lc\n//;
900                    $p =~ s/^\tmov r\d+ = pr\n//;
901
902                    # Remove .proc and .body directives
903                    $p =~ s/^\t\.proc [a-zA-Z0-9_.]+#\n//;
904                    $p =~ s/^\t\.body\n//;
905
906                    # If there's a label, move it to the body
907                    if ($p =~ /^[a-zA-Z0-9.]+:\n/) {
908                        $p = $` . $';
909                        $r = $& . $r;
910                      }
911
912                    # Remove floating-point spill instructions.
913                    # Only fp registers 2-5 and 16-23 are saved by the runtime.
914                    if ($p =~ s/^\tstf\.spill \[r1[4-9]\] = f([2-5]|1[6-9]|2[0-3])(, [0-9]+)?\n//g) {
915                        # Being paranoid, only try to remove these if we saw a
916                        # spill operation.
917                        $p =~ s/^\tmov r1[4-9] = r12\n//;
918                        $p =~ s/^\tadds r1[4-9] = -[0-9]+, r12\n//g;
919                        $p =~ s/^\t\.save\.f 0x[0-9a-fA-F]\n//g;
920                        $p =~ s/^\t\.save\.gf 0x0, 0x[0-9a-fA-F]+\n//g;
921                    }
922
923                    $p =~ s/^\tnop(?:\.[mifb])?\s+\d+\n//g; # remove nop instructions
924                    $p =~ s/^\t\.(mii|mmi|mfi)\n//g;    # bundling is no longer sensible
925                    $p =~ s/^\t;;\n//g;         # discard stops
926                    $p =~ s/^\t\/\/.*\n//g;     # gcc inserts timings in // comments
927
928                    # GCC 3.3 saves r1 in the prologue, move this to the body
929                    # (Does this register get restored anywhere?)
930                    if ($p =~ /^\tmov r\d+ = r1\n/) {
931                      $p = $` . $';
932                      $r = $& . $r;
933                    }
934                } elsif ($TargetPlatform =~ /^m68k-/) {
935                    $p =~ s/^\tlink a6,#-?\d.*\n//;
936                    $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;   
937                                # The above showed up in the asm code,
938                                # so I added it here.
939                                # I hope it's correct.
940                                # CaS
941                    $p =~ s/^\tmovel d2,sp\@-\n//;
942                    $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
943                    $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
944                } elsif ($TargetPlatform =~ /^mips-/) {
945                    # the .frame/.mask/.fmask that we use is the same
946                    # as that produced by GCC for miniInterpret; this
947                    # gives GDB some chance of figuring out what happened
948                    $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
949                    $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
950                    $p =~ s/^\t\.(mask|fmask).*\n//g;
951                    $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
952                    $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
953                    $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
954                    $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
955                    $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
956                    $p =~ s/__FRAME__/$FRAME/;
957                } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
958                    $pcrel_label = $p;
959                    $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
960
961                    $p =~ s/^\tmflr r0\n//;
962                    $p =~ s/^\tbl saveFP # f\d+\n//;
963                    $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
964                    $p =~ s/^\"?L\d+\$pb\"?:\n//;
965                    $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
966                    $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
967                    $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
968                    $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
969                    $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
970                    $p =~ s/^\tbcl 20,31,L\d+\$pb\n//;
971                    $p =~ s/^L\d+\$pb:\n//;
972                    $p =~ s/^\tmflr r31\n//;
973
974                    # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
975                    # under some circumstances, only when generating position dependent code.
976                    # I have no idea why, and I don't think it is necessary, so let's toss it.
977                    $p =~ s/^\tli r\d+,0\n//g;
978                    $p =~ s/^\tstw r\d+,\d+\(r1\)\n//g;
979                } elsif ($TargetPlatform =~ /^powerpc-.*-linux/) {
980                    $p =~ s/^\tmflr 0\n//;
981                    $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
982                    $p =~ s/^\tstfd \d+,\d+\(1\)\n//g;
983                    $p =~ s/^\tstw r0,8\(1\)\n//;
984                    $p =~ s/^\tstwu 1,-\d+\(1\)\n//;
985                    $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
986                   
987                        # GCC's "large-model" PIC (-fPIC)
988                    $pcrel_label = $p;
989                    $pcrel_label =~ s/(.|\n)*^.LCF(\d+):\n(.|\n)*/$2/ or $pcrel_label = "";
990
991                    $p =~ s/^\tbcl 20,31,.LCF\d+\n//;
992                    $p =~ s/^.LCF\d+:\n//;
993                    $p =~ s/^\tmflr 30\n//;
994                    $p =~ s/^\tlwz 0,\.LCL\d+-\.LCF\d+\(30\)\n//;
995                    $p =~ s/^\tadd 30,0,30\n//;
996
997                    # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
998                    # under some circumstances, only when generating position dependent code.
999                    # I have no idea why, and I don't think it is necessary, so let's toss it.
1000                    $p =~ s/^\tli \d+,0\n//g;
1001                    $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1002                } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1003                    $p =~ s/^\tmr 31,1\n//;
1004                    $p =~ s/^\tmflr 0\n//;
1005                    $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
1006                    $p =~ s/^\tstfd \d+,-?\d+\(1\)\n//g;
1007                    $p =~ s/^\tstd r0,8\(1\)\n//;
1008                    $p =~ s/^\tstdu 1,-\d+\(1\)\n//;
1009                    $p =~ s/^\tstd \d+,-?\d+\(1\)\n//g;
1010                   
1011                    # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1012                    # under some circumstances, only when generating position dependent code.
1013                    # I have no idea why, and I don't think it is necessary, so let's toss it.
1014                    $p =~ s/^\tli \d+,0\n//g;
1015                    $p =~ s/^\tstd \d+,\d+\(1\)\n//g;
1016                } else {
1017                    print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
1018                }
1019               
1020                # HWL HACK: dont die, just print a warning
1021                #print stderr  "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
1022                die "Prologue junk?: $p\n" if $p =~ /^\s+[^\s\.]/;
1023               
1024                # For PIC, we want to keep part of the prologue
1025                if ($TargetPlatform =~ /^powerpc-apple-darwin.*/ && $pcrel_label ne "") {
1026                    # Darwin: load the current instruction pointer into register r31
1027                    $p .= "bcl 20,31,$pcrel_label\n";
1028                    $p .= "$pcrel_label:\n";
1029                    $p .= "\tmflr r31\n";
1030                } elsif ($TargetPlatform =~ /^powerpc-.*-linux/ && $pcrel_label ne "") {
1031                    # Linux: load the GOT pointer into register 30
1032                    $p .= "\tbcl 20,31,.LCF$pcrel_label\n";
1033                    $p .= ".LCF$pcrel_label:\n";
1034                    $p .= "\tmflr 30\n";
1035                    $p .= "\tlwz 0,.LCL$pcrel_label-.LCF$pcrel_label(30)\n";
1036                    $p .= "\tadd 30,0,30\n";
1037                } elsif ($TargetPlatform =~ /^i386-apple-darwin.*/ && $pcrel_label ne "") {
1038                    $p .= "\tcall ___i686.get_pc_thunk.$pcrel_reg\n";
1039                    $p .= "$pcrel_label:\n";
1040                }
1041               
1042                # glue together what's left
1043                $c = $p . $r;
1044            }
1045        }
1046
1047        if ( $TargetPlatform =~ /^mips-/ ) {
1048            # MIPS: first, this basic sequence may occur "--- END ---" or not
1049            $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
1050        }
1051
1052        # toss all epilogue stuff; again, paranoidly
1053        if ( $c =~ /--- END ---/ ) {
1054            # Gcc may decide to replicate the function epilogue.  We want
1055            # to process all epilogues, so we split the function and then
1056            # loop here.
1057            @fragments = split(/--- END ---/, $c);
1058            $r = shift(@fragments);
1059
1060            # Rebuild `c'; processed fragments will be appended to `c'
1061            $c = $r;
1062
1063            foreach $e (@fragments) {
1064                # etail holds code that is after the epilogue in the assembly-code
1065                # layout and should not be filtered as part of the epilogue.
1066                $etail = "";
1067                if ($TargetPlatform =~ /^i386-/) {
1068                    $e =~ s/^\tret\n//;
1069                    $e =~ s/^\tpopl\s+\%edi\n//;
1070                    $e =~ s/^\tpopl\s+\%esi\n//;
1071                    $e =~ s/^\tpopl\s+\%edx\n//;
1072                    $e =~ s/^\tpopl\s+\%ecx\n//;
1073                    $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
1074                    $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
1075                } elsif ($TargetPlatform =~ /^ia64-/) {
1076                    # The epilogue is first split into:
1077                    #     $e,    the epilogue code (up to the return instruction)
1078                    #     $etail, non-epilogue code (after the return instruction)
1079                    # The return instruction is stripped in the process.
1080                    if (!(($e, $etail) = split(/^\tbr\.ret\.sptk\.many b0\n/, $e))) {
1081                        die "Epilogue doesn't seem to have one return instruction: $e\n";
1082                    }
1083                    # Remove 'endp' directive from the tail
1084                    $etail =~ s/^\t\.endp [a-zA-Z0-9_.]+#\n//;
1085
1086                    # If a return value is saved here, discard it
1087                    $e =~ s/^\tmov r8 = r14\n//;
1088
1089                    # Remove floating-point fill instructions.
1090                    # Only fp registers 2-5 and 16-23 are saved by the runtime.
1091                    if ($e =~ s/^\tldf\.fill f([2-5]|1[6-9]|2[0-3]) = \[r1[4-9]\](, [0-9]+)?\n//g) {
1092                        # Being paranoid, only try to remove this if we saw a fill
1093                        # operation.
1094                        $e =~ s/^\tadds r1[4-9] = [0-9]+, r12//g;
1095                    }
1096
1097                    $e =~ s/^\tnop(?:\.[mifb])?\s+\d+\n//g; # remove nop instructions
1098                    $e =~ s/^\tmov ar\.pfs = r\d+\n//;
1099                    $e =~ s/^\tmov ar\.lc = r\d+\n//;
1100                    $e =~ s/^\tmov pr = r\d+, -1\n//;
1101                    $e =~ s/^\tmov b0 = r\d+\n//;
1102                    $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
1103                    #$e =~ s/^\tbr\.ret\.sptk\.many b0\n//; # already removed
1104                    $e =~ s/^\t\.(mii|mmi|mfi|mib)\n//g; # bundling is no longer sensible
1105                    $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
1106                    $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
1107                } elsif ($TargetPlatform =~ /^m68k-/) {
1108                    $e =~ s/^\tunlk a6\n//;
1109                    $e =~ s/^\trts\n//;
1110                } elsif ($TargetPlatform =~ /^mips-/) {
1111                    $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
1112                    $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
1113                    $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
1114                    $e =~ s/^\tj\t\$31\n//;
1115                } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
1116                    $e =~ s/^\taddi r1,r1,\d+\n//;
1117                    $e =~ s/^\tlwz r\d+,\d+\(r1\)\n//;
1118                    $e =~ s/^\tlmw r\d+,-\d+\(r1\)\n//;
1119                    $e =~ s/^\tmtlr r0\n//;
1120                    $e =~ s/^\tblr\n//;
1121                    $e =~ s/^\tb restFP ;.*\n//;
1122                } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1123                    $e =~ s/^\tmr 3,0\n//;
1124                    $e =~ s/^\taddi 1,1,\d+\n//;
1125                    $e =~ s/^\tld 0,16\(1\)\n//;
1126                    $e =~ s/^\tmtlr 0\n//;
1127
1128                    # callee-save registers
1129                    $e =~ s/^\tld \d+,-?\d+\(1\)\n//g;
1130                    $e =~ s/^\tlfd \d+,-?\d+\(1\)\n//g;
1131
1132                    # get rid of the debug junk along with the blr
1133                    $e =~ s/^\tblr\n\t.long .*\n\t.byte .*\n//;
1134
1135                    # incase we missed it with the last one get the blr alone
1136                    $e =~ s/^\tblr\n//;
1137                } else {
1138                    print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
1139                }
1140
1141                print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
1142
1143                # glue together what's left
1144                $c .= $e . $etail;
1145            }
1146            $c =~ s/\n\t\n/\n/; # junk blank line
1147        }
1148        else {
1149            if ($TargetPlatform =~ /^ia64-/) {
1150                # On IA64, remove an .endp directive even if no epilogue was found.
1151                # Code optimizations may have removed the "--- END ---" token.
1152                $c =~ s/^\t\.endp [a-zA-Z0-9_.]+#\n//;
1153            }
1154        }
1155
1156        # On SPARCs, we don't do --- BEGIN/END ---, we just
1157        # toss the register-windowing save/restore/ret* instructions
1158        # directly unless they've been generated by function definitions in header
1159        # files on Solaris:
1160        if ( $TargetPlatform =~ /^sparc-/ ) {
1161            if ( ! ( $TargetPlatform =~ /solaris2$/ && $chkcat[$i] eq 'unknown' )) {
1162                $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
1163            }
1164            # throw away PROLOGUE comments
1165            $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
1166        }
1167
1168        # On Alphas, the prologue mangling is done a little later (below)
1169
1170        # toss all calls to __DISCARD__
1171        $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
1172        $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
1173        $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-darwin.*/;
1174        $c =~ s/^\tbl\s+__DISCARD__(\@plt)?\n//go if $TargetPlatform =~ /^powerpc-.*-linux/;
1175        $c =~ s/^\tbl\s+\.__DISCARD__\n\s+nop\n//go if $TargetPlatform =~ /^powerpc64-.*-linux/;
1176        $c =~ s/^\tcall\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /i386-apple-darwin.*/;
1177
1178        # IA64: fix register allocation; mangle tailcalls into jumps
1179        if ($TargetPlatform =~ /^ia64-/) {
1180            ia64_rename_registers($ia64_locnum, $ia64_outnum) if (defined($ia64_locnum));
1181            ia64_mangle_tailcalls();
1182        }
1183
1184        # MIPS: that may leave some gratuitous asm macros around
1185        # (no harm done; but we get rid of them to be tidier)
1186        $c =~ s/^\t\.set\tnoreorder\n\t\.set\tnomacro\n\taddu\t(\S+)\n\t\.set\tmacro\n\t\.set\treorder\n/\taddu\t$1\n/
1187            if $TargetPlatform =~ /^mips-/;
1188
1189        # toss stack adjustment after DoSparks
1190        $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
1191                if $TargetPlatform =~ /^m68k-/; # this looks old...
1192
1193        if ( $TargetPlatform =~ /^alpha-/ &&
1194           ! $magic_rdata_seen &&
1195           $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
1196            $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
1197            $magic_rdata_seen = 1;
1198        }
1199
1200        # pick some end-things and move them to the next chunk
1201
1202        # pin a funny end-thing on (for easier matching):
1203        $c .= 'FUNNY#END#THING';
1204
1205        while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
1206
1207            $to_move = $1;
1208
1209            # on x86 we try not to copy any directives into a literal
1210            # chunk, rather we keep looking for the next real chunk.  This
1211            # is because we get things like
1212            #
1213            #    .globl blah_closure
1214            #    .LC32
1215            #           .string "..."
1216            #    blah_closure:
1217            #           ...
1218            #
1219            if ( $TargetPlatform =~ /^(i386|sparc|powerpc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
1220                $j = $i + 1;
1221                while ( $j < $numchks  && $chk[$j] =~ /$T_CONST_LBL/) {
1222                        $j++;
1223                }
1224                if ( $j < $numchks ) {
1225                        $chk[$j] = $to_move . $chk[$j];
1226                }
1227            }
1228
1229            elsif (   (    $i < ($numchks - 1)
1230                       && ( $to_move =~ /${T_COPY_DIRVS}/
1231                           || (   $TargetPlatform =~ /^hppa/
1232                               && $to_move =~ /align/
1233                               && $chkcat[$i+1] eq 'literal')
1234                          )
1235                      )
1236                   || ($to_move =~ /^[ \t]*\.section[ \t]+\.note\.GNU-stack,/)
1237                  ) {
1238                $chk[$i + 1] = $to_move . $chk[$i + 1];
1239                # otherwise they're tossed
1240            }
1241
1242            $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
1243        }
1244
1245        if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
1246            $ent = $1;
1247            # toss all prologue stuff, except for loading gp, and the ..ng address
1248            unless ($c =~ /\.ent.*\n\$.*\.\.ng:/) {
1249                if (($p, $r) = split(/^\t\.prologue/, $c)) {
1250                    if (($keep, $junk) = split(/\.\.ng:/, $p)) {
1251                        $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
1252                        $keep =~ s/^\t\.(mask|fmask).*\n//g;
1253                        $c = $keep . "..ng:\n";
1254                    } else {
1255                        print STDERR "malformed code block ($ent)?\n"
1256                    }
1257                }
1258                $c .= "\t.prologue" . $r;
1259            }
1260        }
1261 
1262        $c =~ s/FUNNY#END#THING//;
1263
1264#       print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
1265
1266        $chk[$i] = $c; # update w/ convenience copy
1267    }
1268
1269    # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
1270    # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
1271    # close CHUNKS;
1272
1273    if ( $TargetPlatform =~ /^alpha-/ ) {
1274        # print out the header stuff first
1275        $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
1276        print OUTASM $chk[0];
1277
1278    } elsif ( $TargetPlatform =~ /^hppa/ ) {
1279        print OUTASM $chk[0];
1280
1281    } elsif ( $TargetPlatform =~ /^mips-/ ) {
1282        $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
1283
1284        # get rid of horrible "<dollar>Revision: .*$" strings
1285        local(@lines0) = split(/\n/, $chk[0]);
1286        local($z) = 0;
1287        while ( $z <= $#lines0 ) {
1288            if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
1289                undef($lines0[$z]);
1290                $z++;
1291                while ( $z <= $#lines0 ) {
1292                    undef($lines0[$z]);
1293                    last if $lines0[$z] =~ /[,\t]0x0$/;
1294                    $z++;
1295                }
1296            }
1297            $z++;
1298        }
1299        $chk[0] = join("\n", @lines0);
1300        $chk[0] =~ s/\n\n+/\n/;
1301        print OUTASM $chk[0];
1302    }
1303
1304    # print out all the literal strings next
1305    for ($i = 0; $i < $numchks; $i++) {
1306        if ( $chkcat[$i] eq 'literal' ) {
1307
1308            # HACK: try to detect 16-byte constants and align them
1309            # on a 16-byte boundary.  x86_64 sometimes needs 128-bit
1310            # aligned constants, and so does Darwin/x86.
1311            if ( $TargetPlatform =~ /^x86_64/
1312                || $TargetPlatform =~ /^i386-apple-darwin/ ) {
1313                $z = $chk[$i];
1314                if ($z =~ /(\.long.*\n.*\.long.*\n.*\.long.*\n.*\.long|\.quad.*\n.*\.quad)/) {
1315                    print OUTASM $T_HDR_literal16;
1316                } else {
1317                    print OUTASM $T_HDR_literal;
1318                }
1319            } else {
1320                print OUTASM $T_HDR_literal;
1321            }
1322
1323            print OUTASM $chk[$i];
1324            print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
1325
1326            $chkcat[$i] = 'DONE ALREADY';
1327        }
1328    }
1329
1330    # on the HPPA, print out all the bss next
1331    if ( $TargetPlatform =~ /^hppa/ ) {
1332        for ($i = 1; $i < $numchks; $i++) {
1333            if ( $chkcat[$i] eq 'bss' ) {
1334                print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
1335                print OUTASM $chk[$i];
1336
1337                $chkcat[$i] = 'DONE ALREADY';
1338            }
1339        }
1340    }
1341
1342    # $numchks + 1 as we have the extra one for .note.GNU-stack
1343    for ($i = $FIRST_MANGLABLE; $i < $numchks + 1; $i++) {
1344#       print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
1345
1346        next if $chkcat[$i] eq 'DONE ALREADY';
1347
1348        if ( $chkcat[$i] eq 'misc' || $chkcat[$i] eq 'unknown' ) {
1349            if ($chk[$i] ne '') {
1350                print OUTASM $T_HDR_misc;
1351                &print_doctored($chk[$i], 0);
1352            }
1353
1354        } elsif ( $chkcat[$i] eq 'verbatim' ) {
1355            print OUTASM $chk[$i];
1356
1357        } elsif ( $chkcat[$i] eq 'toss' ) {
1358            print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
1359
1360        } elsif ( $chkcat[$i] eq 'data' ) {
1361            if ($chk[$i] ne '') {
1362                print OUTASM $T_HDR_data;
1363                print OUTASM $chk[$i];
1364            }
1365
1366        } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1367            # we can just re-constitute this one...
1368            # NB: we emit _three_ underscores no matter what,
1369            # so ghc-split doesn't have to care.
1370            print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1371
1372        } elsif ( $chkcat[$i] eq 'closure'
1373               || $chkcat[$i] eq 'srt'
1374               || $chkcat[$i] eq 'infotbl'
1375               || $chkcat[$i] eq 'entry') { # do them in that order
1376            $symb = $chksymb[$i];
1377
1378            # CLOSURE
1379            if ( defined($closurechk{$symb}) ) {
1380                print OUTASM $T_HDR_closure;
1381                print OUTASM $chk[$closurechk{$symb}];
1382                $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1383            }
1384
1385            # SRT
1386            if ( defined($srtchk{$symb}) ) {
1387                print OUTASM $T_HDR_relrodata;
1388                print OUTASM $chk[$srtchk{$symb}];
1389                $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1390            }
1391
1392            # INFO TABLE
1393            if ( defined($infochk{$symb}) ) {
1394
1395                print OUTASM $T_HDR_info;
1396                print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1397               
1398                # entry code will be put here!
1399
1400                $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1401            }
1402
1403            # ENTRY POINT
1404            if ( defined($entrychk{$symb}) ) {
1405
1406                $c = $chk[$entrychk{$symb}];
1407
1408                # If this is an entry point with an info table,
1409                # eliminate the entry symbol and all directives involving it.
1410                if (defined($infochk{$symb}) && $TargetPlatform !~ /^ia64-/) {
1411                        @o = ();
1412                        foreach $l (split(/\n/,$c)) {
1413                            next if $l =~ /^.*$symb_(entry|ret)${T_POST_LBL}/;
1414
1415                            # If we have .type/.size direrctives involving foo_entry,
1416                            # then make them refer to foo_info instead.  The information
1417                            # in these directives is used by the cachegrind annotator,
1418                            # so it is worthwhile keeping.
1419                            if ($l =~ /^\s*\.(type|size).*$symb_(entry|ret)/) {
1420                                $l =~ s/$symb(_entry|_ret)/${symb}_info/g;
1421                                push(@o,$l);
1422                                next;
1423                            }
1424                            next if $l =~ /^\s*\..*$symb.*\n?/;
1425                            push(@o,$l);
1426                        }
1427                        $c = join("\n",@o) . "\n";
1428                }
1429
1430                print OUTASM $T_HDR_entry;
1431
1432                &print_doctored($c, 1); # NB: the 1!!!
1433
1434                $chkcat[$entrychk{$symb}] = 'DONE ALREADY';
1435            }
1436           
1437        } elsif ( $chkcat[$i] eq 'vector' ) {
1438            $symb = $chksymb[$i];
1439
1440            # VECTOR TABLE
1441            if ( defined($vectorchk{$symb}) ) {
1442                print OUTASM $T_HDR_vector;
1443                print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1444
1445                # direct return code will be put here!
1446                $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1447
1448            } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1449                # Alphas: the commented nop is for the splitter, to ensure
1450                # that no module ends with a label as the very last
1451                # thing.  (The linker will adjust the label to point
1452                # to the first code word of the next module linked in,
1453                # even if alignment constraints cause the label to move!)
1454
1455                print OUTASM "\t# nop\n";
1456            }
1457           
1458        } elsif ( $chkcat[$i] eq 'rodata' ) {
1459                print OUTASM $T_HDR_rodata;
1460                print OUTASM $chk[$i];
1461                $chkcat[$i] = 'DONE ALREADY';
1462        } elsif ( $chkcat[$i] eq 'relrodata' ) {
1463                print OUTASM $T_HDR_relrodata;
1464                print OUTASM $chk[$i];
1465                $chkcat[$i] = 'DONE ALREADY';
1466        } elsif ( $chkcat[$i] eq 'toc' ) {
1467            # silly optimisation to print tocs, since they come in groups...
1468            print OUTASM $T_HDR_toc;
1469            local($j)   = $i;
1470            while ($chkcat[$j] eq 'toc')
1471              { if (   $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1472                   )
1473                {
1474                  print OUTASM $chk[$j];
1475                }
1476                $chkcat[$j] = 'DONE ALREADY';
1477                $j++;
1478            }
1479           
1480        } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' ) {
1481            # apple-darwin: dynamic linker stubs
1482            if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1483            {   # print them out unchanged, but remove the stubs for __DISCARD__
1484                print OUTASM $chk[$i];
1485            }
1486        } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ && $chkcat[$i] eq 'get_pc_thunk' ) {
1487            # i386-apple-darwin: __i686.get_pc_thunk.[abcd]x
1488            print OUTASM ".section __TEXT,__textcoal_nt,coalesced,no_toc\n";
1489            print OUTASM $chk[$i];
1490        } else {
1491            &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1492        }
1493    }
1494
1495    print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1496
1497    # finished
1498    close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1499    close(INASM)  || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1500}
1501\end{code}
1502
1503On IA64, tail calls are converted to branches at this point.  The mangler
1504searches for function calls immediately followed by a '--- TAILCALL ---'
1505token.  Since the compiler can put various combinations of labels, bundling
1506directives, nop instructions, stops, and a move of the return value
1507between the branch and the tail call, proper matching of the tail call
1508gets a little hairy.  This subroutine does the mangling.
1509
1510Here is an example of a tail call before mangling:
1511
1512\begin{verbatim}
1513        br.call.sptk.many b0 = b6
1514.L211
1515        ;;
1516        .mmi
1517        mov r1 = r32
1518        ;;
1519        nop.m 0
1520        nop.i 0
1521        ;;
1522        --- TAILCALL --
1523        ;;
1524.L123
1525\end{verbatim}
1526
1527\begin{code}
1528sub ia64_mangle_tailcalls {
1529    # Function input and output are in $c
1530
1531    # Construct the tailcall-mangling expression the first time this function
1532    # is called.
1533    if (!defined($IA64_MATCH_TAILCALL)) {
1534        # One-line pattern matching constructs.  None of these
1535        # should bind references; all parenthesized terms
1536        # should be (?:) terms.
1537        my $stop       = q/(?:\t;;\n)/;
1538        my $bundle     = q/(?:\t\.(?:mii|mib|mmi|mmb|mfi|mfb|mbb|bbb)\n)/;
1539        my $nop        = q/(?:\tnop(?:\.[mifb])?\s+\d+\n)/;
1540        my $movgp      = q/(?:\tmov r1 = r\d+\n)/;
1541        my $postbr     = q/(?:\tbr \.L\d+\n)/;
1542
1543        my $noeffect   = "(?:$stop$bundle?|$nop)*";
1544        my $postbundle = "(?:$bundle?$nop?$nop?$postbr)?";
1545
1546        # Important parts of the pattern match.  The branch target
1547        # and subsequent jump label are bound to $1 and $2
1548        # respectively.  Sometimes there is no label.
1549        my $callbr    = q/^\tbr\.call\.sptk\.many b0 = (.*)\n/;
1550        my $label     = q/(?:^\.L([0-9]*):\n)/;
1551        my $tailcall  = q/\t--- TAILCALL ---\n/;
1552
1553        $IA64_MATCH_TAILCALL =
1554          $callbr . $label . '?' . $noeffect . $movgp . '?' . $noeffect .
1555          $tailcall . $stop . '?' . '(?:' . $postbundle . ')?';
1556    }
1557
1558    # Find and mangle tailcalls
1559    while ($c =~ s/$IA64_MATCH_TAILCALL/\tbr\.few $1\n/o) {
1560        # Eek, the gcc optimiser is getting smarter... if we see a jump to the
1561        # --- TAILCALL --- marker then we reapply the substitution at the source sites
1562        $c =~ s/^\tbr \.L$2\n/\t--- TAILCALL ---\n/g if ($2);
1563    }
1564
1565    # Verify that all instances of TAILCALL were processed
1566    if ($c =~ /^\t--- TAILCALL ---\n/) {
1567        die "Unmangled TAILCALL tokens remain after mangling"
1568    }
1569}
1570\end{code}
1571
1572The number of registers allocated on the IA64 register stack is set
1573upon entry to the runtime with an `alloc' instruction at the entry
1574point of \verb+StgRun()+.  Gcc uses its own `alloc' to allocate
1575however many registers it likes in each function.  When we discard
1576gcc's alloc, we have to reconcile its register assignment with what
1577the STG uses.
1578
1579There are three stack areas: fixed registers, input/local registers,
1580and output registers.  We move the output registers to the output
1581register space and leave the other registers where they are.
1582
1583\begin{code}
1584sub ia64_rename_registers() {
1585    # The text to be mangled is in $c
1586    # Find number of registers in each stack area
1587    my ($loc, $out) = @_;
1588    my $cout;
1589    my $first_out_reg;
1590    my $regnum;
1591    my $fragment;
1592
1593    # These are the register numbers used in the STG runtime
1594    my $STG_FIRST_OUT_REG = 32 + 34;
1595    my $STG_LAST_OUT_REG = $STG_FIRST_OUT_REG + 7;
1596
1597    $first_out_reg = 32 + $loc;
1598
1599    if ($first_out_reg > $STG_FIRST_OUT_REG) {
1600        die "Too many local registers allocated by gcc";
1601    }
1602
1603    # Split the string into fragments containing one register name each.
1604    # Rename the register in each fragment and concatenate.
1605    $cout = "";
1606    foreach $fragment (split(/(?=r\d+[^a-zA-Z0-9_.])/s, $c)) {
1607        if ($fragment =~ /^r(\d+)((?:[^a-zA-Z0-9_.].*)?)$/s) {
1608            $regnum = $1;
1609
1610            if ($regnum < $first_out_reg) {
1611                # This is a local or fixed register
1612
1613                # Local registers 32 and 33 (r64 and r65) are
1614                # used to hold saved state; they shouldn't be touched
1615                if ($regnum == 64 || $regnum == 65) {
1616                   die "Reserved register $regnum is in use";
1617                }
1618            }
1619            else {
1620                # This is an output register
1621                $regnum = $regnum - $first_out_reg + $STG_FIRST_OUT_REG;
1622                if ($regnum > $STG_LAST_OUT_REG) {
1623                    die "Register number ($regnum) is out of expected range";
1624                }
1625            }
1626
1627            # Update this fragment
1628            $fragment = "r" . $regnum . $2;
1629        }
1630        $cout .= $fragment;
1631    }
1632
1633    $c = $cout;
1634}
1635
1636\end{code}
1637
1638\begin{code}
1639sub hppa_mash_prologue { # OK, epilogue, too
1640    local($_) = @_;
1641
1642    # toss all prologue stuff
1643    s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1644
1645    # Lie about our .CALLINFO
1646    s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1647
1648    # Get rid of P'
1649
1650    s/LP'/L'/g;
1651    s/RP'/R'/g;
1652
1653    # toss all epilogue stuff
1654    s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1655
1656    # Sorry; we moved the _info stuff to the code segment.
1657    s/_info,DATA/_info,CODE/g;
1658
1659    return($_);
1660}
1661\end{code}
1662
1663\begin{code}
1664sub print_doctored {
1665    local($_, $need_fallthru_patch) = @_;
1666
1667    if ( $TargetPlatform =~ /^x86_64-/ ) {
1668            # Catch things like
1669            #   
1670            #    movq -4(%ebp), %rax
1671            #    jmp  *%rax
1672            #
1673            # and optimise:
1674            #
1675            s/^\tmovq\s+(-?\d*\(\%r(bx|bp|13)\)),\s*(\%r(ax|cx|dx|10|11))\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1676            s/^\tmovl\s+\$${T_US}(.*),\s*(\%e(ax|cx|si|di))\n\tjmp\s+\*\%r\3/\tjmp\t$T_US$1/g;
1677    }
1678
1679    if ( $TargetPlatform !~ /^i386-/
1680      || ! /^\t[a-z]/  # no instructions in here, apparently
1681      || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1682        print OUTASM $_;
1683        return;
1684    }
1685
1686    # OK, must do some x86 **HACKING**
1687
1688    local($entry_patch) = '';
1689    local($exit_patch)  = '';
1690
1691    # gotta watch out for weird instructions that
1692    # invisibly smash various regs:
1693    #   rep*    %ecx used for counting
1694    #   scas*   %edi used for destination index
1695    #   cmps*   %e[sd]i used for indices
1696    #   loop*   %ecx used for counting
1697    #
1698    # SIGH.
1699
1700    # We cater for:
1701    #  * use of STG reg [ nn(%ebx) ] where no machine reg avail
1702    #
1703    #  * GCC used an "STG reg" for its own purposes
1704    #
1705    #  * some secret uses of machine reg, requiring STG reg
1706    #    to be saved/restored
1707
1708    # The most dangerous "GCC uses" of an "STG reg" are when
1709    # the reg holds the target of a jmp -- it's tricky to
1710    # insert the patch-up code before we get to the target!
1711    # So here we change the jmps:
1712
1713    # --------------------------------------------------------
1714    # it can happen that we have jumps of the form...
1715    #   jmp *<something involving %esp>
1716    # or
1717    #   jmp <something involving another naughty register...>
1718    #
1719    # a reasonably-common case is:
1720    #
1721    #   movl $_blah,<bad-reg>
1722    #   jmp  *<bad-reg>
1723    #
1724    s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[acd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1725
1726    # Catch things like
1727    #
1728    #    movl -4(%ebx), %eax
1729    #    jmp  *%eax
1730    #
1731    # and optimise:
1732    #
1733    s/^\tmovl\s+(-?\d*\(\%e(bx|si)\)),\s*(\%e[acd]x)\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1734
1735    if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1736        s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1737        s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1738        s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1739        die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1740            if /(jmp|call)\s+.*\%esi/;
1741    }
1742    if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1743        s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1744        s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1745        s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1746        die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1747            if /(jmp|call)\s+.*\%edi/;
1748    }
1749
1750    # OK, now we can decide what our patch-up code is going to
1751    # be:
1752
1753    # Offsets into register table - you'd better update these magic
1754    # numbers should you change its contents!
1755    # local($OFFSET_R1)=0;  No offset for R1 in new RTS.
1756    local($OFFSET_Hp)=88;
1757
1758        # Note funky ".=" stuff; we're *adding* to these _patch guys
1759    if ( $StolenX86Regs <= 2
1760         && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1761        $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1762        $exit_patch  .= "\tmovl (\%ebx),\%esi\n";
1763
1764        # nothing for call_{entry,exit} because %esi is callee-save
1765    }
1766    if ( $StolenX86Regs <= 3
1767         && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1768        $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1769        $exit_patch  .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1770
1771        # nothing for call_{entry,exit} because %edi is callee-save
1772    }
1773
1774    # --------------------------------------------------------
1775    # next, here we go with non-%esp patching!
1776    #
1777    s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1778
1779# Before calling GC we must set up the exit condition before the call
1780# and entry condition when we come back
1781
1782    # fix _all_ non-local jumps:
1783
1784    if ( $TargetPlatform =~ /^.*-apple-darwin.*/ ) {
1785        # On Darwin, we've got local-looking jumps that are
1786        # actually global (i.e. jumps to Lfoo$stub or via
1787        # Lfoo$non_lazy_ptr), so we fix those first.
1788        # In fact, we just fix everything that contains a dollar
1789        # because false positives don't hurt here.
1790
1791        s/^(\tjmp\s+\*?L.*\$.*\n)/$exit_patch$1/g;
1792    }
1793
1794    s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1795    s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1796
1797    s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1798
1799    s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1800    s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1801
1802    if ($StolenX86Regs == 2 ) {
1803        die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1804            if /^\t(jmp|call)\s+.*\%e(si|di)/;
1805    } elsif ($StolenX86Regs == 3 ) {
1806        die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1807            if /^\t(jmp|call)\s+.*\%edi/;
1808    }
1809
1810    # --------------------------------------------------------
1811    # that's it -- print it
1812    #
1813    #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1814
1815    print OUTASM $_;
1816
1817    if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1818        print OUTASM $exit_patch;
1819        # ToDo: make it not print if there is a "jmp" at the end
1820    }
1821}
1822\end{code}
1823
1824\begin{code}
1825sub init_FUNNY_THINGS {
1826    %KNOWN_FUNNY_THING = (
1827        # example
1828        # "${T_US}stg_.*{T_POST_LBL}", 1, 
1829    );
1830}
1831\end{code}
1832
1833The following table reversal is used for both info tables and return
1834vectors.  In both cases, we remove the first entry from the table,
1835reverse the table, put the label at the end, and paste some code
1836(that which is normally referred to by the first entry in the table)
1837right after the table itself.  (The code pasting is done elsewhere.)
1838
1839\begin{code}
1840sub rev_tbl {
1841    local($symb, $tbl, $discard1) = @_;
1842
1843    return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1844
1845    local($before) = '';
1846    local($label) = '';
1847    local(@imports) = (); # hppa only
1848    local(@words) = ();
1849    local($after) = '';
1850    local(@lines) = split(/\n/, $tbl);
1851    local($i, $j);
1852
1853    # Deal with the header...
1854    for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1855        $label .= $lines[$i] . "\n",
1856            next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1857                 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1858                 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1859
1860        $before .= $lines[$i] . "\n"; # otherwise...
1861    }
1862
1863    $infoname = $label;
1864    $infoname =~ s/(.|\n)*^([A-Za-z0-9_]+_info)${T_POST_LBL}$(.|\n)*/\2/;
1865   
1866    # Grab the table data...
1867    if ( $TargetPlatform !~ /^hppa/ ) {
1868        for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1869            $line = $lines[$i];
1870            # Convert addresses of SRTs, slow entrypoints and large bitmaps
1871            # to offsets (relative to the info label),
1872            # in order to support position independent code.
1873            $line =~ s/$infoname/0/
1874            || $line =~ s/([A-Za-z0-9_]+_srtd)$/\1 - $infoname/
1875            || $line =~ s/([A-Za-z0-9_]+_srt(\+\d+)?)$/\1 - $infoname/
1876            || $line =~ s/([A-Za-z0-9_]+_slow)$/\1 - $infoname/
1877            || $line =~ s/([A-Za-z0-9_]+_btm)$/\1 - $infoname/
1878            || $line =~ s/([A-Za-z0-9_]+_alt)$/\1 - $infoname/
1879            || $line =~ s/([A-Za-z0-9_]+_dflt)$/\1 - $infoname/
1880            || $line =~ s/([A-Za-z0-9_]+_ret)$/\1 - $infoname/;
1881            push(@words, $line);
1882        }
1883    } else { # hppa weirdness
1884        for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1885            # FIXME: the RTS now expects offsets instead of addresses
1886            # for all labels in info tables.
1887            if ($lines[$i] =~ /^\s+\.IMPORT/) {
1888                push(@imports, $lines[$i]);
1889            } else {
1890                # We don't use HP's ``function pointers''
1891                # We just use labels in code space, like normal people
1892                $lines[$i] =~ s/P%//;
1893                push(@words, $lines[$i]);
1894            }
1895        }
1896    }
1897
1898    # Now throw away any initial zero word from the table.  This is a hack
1899    # that lets us reduce the size of info tables when the SRT field is not
1900    # needed: see comments StgFunInfoTable in InfoTables.h.
1901    #
1902    # The .zero business is for Linux/ELF.
1903    # The .skip business is for Sparc/Solaris/ELF.
1904    # The .blockz business is for HPPA.
1905#    if ($discard1) {
1906#       if ($words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1907#               shift(@words);
1908#       }
1909#    }
1910
1911    for (; $i <= $#lines; $i++) {
1912        $after .= $lines[$i] . "\n";
1913    }
1914
1915    # Alphas: If we have anonymous text (not part of a procedure), the
1916    # linker may complain about missing exception information.  Bleh.
1917    # To suppress this, we place a .ent/.end pair around the code.
1918    # At the same time, we have to be careful and not enclose any leading
1919    # .file/.loc directives.
1920    if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1921        local ($ident) = $1;
1922        $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1923        $after .= "\t.end $ident\n";
1924    }
1925
1926    # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1927    # assembler (!) wherein .quad constants inside .text sections are
1928    # first narrowed to 32 bits then sign-extended back to 64 bits.
1929    # This obviously screws up our 64-bit bitmaps, so we work around
1930    # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1931    if ( $TargetPlatform =~ /^alpha-/ ) {
1932        foreach (@words) {
1933            if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1934                local ($number) = $1;
1935                if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1936                    local ($sign, $base, $digits) = ($1, $2, $3);
1937                    $base = (10, 8, 16)[length $base];
1938                    local ($hi, $lo) = (0, 0);
1939                    foreach $i (split(//, $digits)) {
1940                        $j = $lo * $base + $i;
1941                        $lo = $j % 4294967296;
1942                        $hi = $hi * $base + ($j - $lo) / 4294967296;
1943                    }
1944                    ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1945                        if $sign eq "-";
1946                    $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1947                    # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1948                } else {
1949                    print STDERR "Cannot handle \".quad $number\" in info table\n";
1950                    exit 1;
1951                }
1952            }
1953        }
1954    }
1955
1956    $tbl = $before
1957         . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
1958         . join("\n", @words) . "\n"
1959         . $label . $after;
1960
1961#   print STDERR "before=$before\n";
1962#   print STDERR "label=$label\n";
1963#   print STDERR "words=",(reverse @words),"\n";
1964#   print STDERR "after=$after\n";
1965
1966    $tbl;
1967}
1968\end{code}
1969
1970The HP is a major nuisance.  The threaded code mangler moved info
1971tables from data space to code space, but unthreaded code in the RTS
1972still has references to info tables in data space.  Since the HP
1973linker is very precise about where symbols live, we need to patch the
1974references in the unthreaded RTS as well.
1975
1976\begin{code}
1977sub mini_mangle_asm_hppa {
1978    local($in_asmf, $out_asmf) = @_;
1979
1980    open(INASM, "< $in_asmf")
1981        || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1982    open(OUTASM,"> $out_asmf")
1983        || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1984
1985    while (<INASM>) {
1986        s/_info,DATA/_info,CODE/;   # Move _info references to code space
1987        s/P%_PR/_PR/;
1988        print OUTASM;
1989    }
1990
1991    # finished:
1992    close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1993    close(INASM)  || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1994}
1995
1996\end{code}
1997
1998\begin{code}
1999sub tidy_up_and_die {
2000    local($return_val, $msg) = @_;
2001    print STDERR $msg;
2002    exit (($return_val == 0) ? 0 : 1);
2003}
2004\end{code}