1 %************************************************************************
3 \section[Driver-asm-fiddling]{Fiddling with assembler files}
5 %************************************************************************
10 Utterly stomp out C functions' prologues and epilogues; i.e., the
11 stuff to do with the C stack.
13 Any other required tidying up.
16 General note [chak]: Many regexps are very fragile because they rely on white
17 space being in the right place. This caused trouble with gcc 2.95 (at least
18 on Linux), where the use of white space in .s files generated by gcc suddenly
19 changed. To guarantee compatibility across different versions of gcc, make
20 sure (at least on i386-.*-linux) that regexps tolerate varying amounts of white
21 space between an assembler statement and its arguments as well as after a the
22 comma separating multiple arguments.
24 \emph{For the time being, I have corrected the regexps for i386-.*-linux. I
25 didn't touch all the regexps for other i386 platforms, as I don't have
26 a box to test these changes.}
31 The HP linker is very picky about symbols being in the appropriate
32 space (code vs. data). When we mangle the threaded code to put the
33 info tables just prior to the code, they wind up in code space
34 rather than data space. This means that references to *_info from
35 un-mangled parts of the RTS (e.g. unthreaded GC code) get
36 unresolved symbols. Solution: mini-mangler for .c files on HP. I
37 think this should really be triggered in the driver by a new -rts
38 option, so that user code doesn't get mangled inappropriately.
40 With reversed tables, jumps are to the _info label rather than to
41 the _entry label. The _info label is just an address in code
42 space, rather than an entry point with the descriptive blob we
43 talked about yesterday. As a result, you can't use the call-style
44 JMP_ macro. However, some JMP_ macros take _info labels as targets
45 and some take code entry points within the RTS. The latter won't
46 work with the goto-style JMP_ macro. Sigh. Solution: Use the goto
47 style 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
50 than the descriptive blob. Also change all ".word P%literal"
51 entries in info tables and vector tables to just ".word literal,"
52 for the same reason. Advantage: No more ridiculous call sequences.
55 %************************************************************************
57 \subsection{Top-level code}
59 %************************************************************************
62 $TargetPlatform = $TARGETPLATFORM;
64 ($Pgm = $0) =~ s|.*/||;
68 if ( $TargetPlatform =~ /^i386-/ ) {
72 $StolenX86Regs = $ARGV[2];
76 &mangle_asm($ifile,$ofile);
81 %************************************************************************
83 \subsection{Constants for various architectures}
85 %************************************************************************
88 sub init_TARGET_STUFF {
90 #--------------------------------------------------------#
91 if ( $TargetPlatform =~ /^alpha-.*-.*/ ) {
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
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))';
102 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
103 $T_DOT_WORD = '\.(long|quad|byte|word)';
104 $T_DOT_GLOBAL = '^\t\.globl';
105 $T_HDR_literal = "\.rdata\n\t\.align 3\n";
106 $T_HDR_misc = "\.text\n\t\.align 3\n";
107 $T_HDR_data = "\.data\n\t\.align 3\n";
108 $T_HDR_consist = "\.text\n";
109 $T_HDR_closure = "\.data\n\t\.align 3\n";
110 $T_HDR_srt = "\.text\n\t\.align 3\n";
111 $T_HDR_info = "\.text\n\t\.align 3\n";
112 $T_HDR_entry = "\.text\n\t\.align 3\n";
113 $T_HDR_vector = "\.text\n\t\.align 3\n";
114 $T_HDR_direct = "\.text\n\t\.align 3\n";
115 $T_create_word = "\t.quad";
117 #--------------------------------------------------------#
118 } elsif ( $TargetPlatform =~ /^hppa/ ) {
120 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
121 $T_US = ''; # _ if symbols have an underscore on the front
122 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
123 $T_CONST_LBL = '^L\$C(\d+)$'; # regexp for what such a lbl looks like
126 $T_MOVE_DIRVS = '^((\s+\.(IMPORT|EXPORT|PARAM).*|\s+\.align\s+\d+|\s+\.(SPACE|SUBSPA)\s+\S+|\s*)\n)';
127 $T_COPY_DIRVS = '^\s+\.(IMPORT|EXPORT)';
129 $T_hsc_cc_PAT = '\.STRING.*\)(hsc|cc) (.*)\\\\x09(.*)\\\\x00"';
130 $T_DOT_WORD = '\.(blockz|word|half|byte)';
131 $T_DOT_GLOBAL = '^\s+\.EXPORT';
132 $T_HDR_literal = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
133 $T_HDR_misc = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
134 $T_HDR_data = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
135 $T_HDR_consist = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
136 $T_HDR_closure = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
137 $T_HDR_srt = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
138 $T_HDR_info = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
139 $T_HDR_entry = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
140 $T_HDR_vector = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
141 $T_HDR_direct = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
142 $T_create_word = "\t.word";
144 #--------------------------------------------------------#
145 } elsif ( $TargetPlatform =~ /^i386-.*-(linuxaout|freebsd2|openbsd|nextstep3|cygwin32|mingw32)$/ ) {
146 # NeXT added but not tested. CaS
148 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
149 $T_US = '_'; # _ if symbols have an underscore on the front
150 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
151 $T_CONST_LBL = '^LC(\d+):$';
153 $T_X86_PRE_LLBL_PAT = 'L';
154 $T_X86_PRE_LLBL = 'L';
155 $T_X86_BADJMP = '^\tjmp [^L\*]';
157 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s+\d+(,0x90)?|\.globl\s+\S+|\.text|\.data|\.stab[^n].*|\.type\s+.*|\.size\s+.*)\n)';
158 $T_COPY_DIRVS = '\.(globl|stab)';
159 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
160 $T_DOT_WORD = '\.(long|word|value|byte|space)';
161 $T_DOT_GLOBAL = '\.globl';
162 $T_HDR_literal = "\.text\n\t\.align 2\n";
163 $T_HDR_misc = "\.text\n\t\.align 2,0x90\n";
164 $T_HDR_data = "\.data\n\t\.align 2\n";
165 $T_HDR_consist = "\.text\n";
166 $T_HDR_closure = "\.data\n\t\.align 2\n";
167 $T_HDR_srt = "\.text\n\t\.align 2\n";
168 $T_HDR_info = "\.text\n\t\.align 2\n"; # NB: requires padding
169 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
170 $T_HDR_vector = "\.text\n\t\.align 2\n"; # NB: requires padding
171 $T_HDR_direct = "\.text\n\t\.align 2,0x90\n";
172 $T_create_word = "\t.word";
174 #--------------------------------------------------------#
175 } elsif ( $TargetPlatform =~ /^i386-.*-(solaris2|linux|freebsd|netbsd)$/ ) {
177 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
178 $T_US = ''; # _ if symbols have an underscore on the front
179 $T_PRE_APP = # regexp that says what comes before APP/NO_APP
180 ($TargetPlatform =~ /-(linux|freebsd|netbsd)$/) ? '#' : '/' ;
181 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
183 $T_X86_PRE_LLBL_PAT = '\.L';
184 $T_X86_PRE_LLBL = '.L';
185 $T_X86_BADJMP = '^\tjmp\s+[^\.\*]';
187 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s+\d+(,\s*0x90)?|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.size\s+\S+,\d+|\.ident.*)\n)';
188 $T_COPY_DIRVS = '^\s*\.(globl|type|size)';
190 if ( $TargetPlatform =~ /freebsd|netbsd/ ) {
191 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
193 $T_hsc_cc_PAT = '\.string.*\)(hsc|cc) (.*)\\\\t(.*)"';
196 $T_DOT_WORD = '\.(long|value|word|byte|zero)';
197 $T_DOT_GLOBAL = '\.globl';
198 $T_HDR_literal = "\.section\t\.rodata\n"; # or just use .text??? (WDP 95/11)
199 $T_HDR_misc = "\.text\n\t\.align 4\n";
200 $T_HDR_data = "\.data\n\t\.align 4\n"; # ToDo: change align??
201 $T_HDR_consist = "\.text\n";
202 $T_HDR_closure = "\.data\n\t\.align 4\n"; # ToDo: change align?
203 $T_HDR_srt = "\.text\n\t\.align 4\n"; # ToDo: change align?
204 $T_HDR_info = "\.text\n\t\.align 4\n"; # NB: requires padding
205 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
206 $T_HDR_vector = "\.text\n\t\.align 4\n"; # NB: requires padding
207 $T_HDR_direct = "\.text\n\t\.align 4\n";
208 $T_create_word = "\t.word";
210 #--------------------------------------------------------#
211 } elsif ( $TargetPlatform =~ /^ia64-.*-linux$/ ) {
213 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
214 $T_US = ''; # _ if symbols have an underscore on the front
216 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
219 $T_MOVE_DIRVS = '^(\s*\.(global|proc|pred\.safe_across_calls|text|data|section|subsection|align|size|type|ident)\s+.*\n)';
220 $T_COPY_DIRVS = '\.(global|proc)';
222 $T_hsc_cc_PAT = '\.string.*\)(hsc|cc) (.*)\\\\t(.*)"';
223 $T_DOT_WORD = '\.(long|value|byte|zero)';
224 $T_DOT_GLOBAL = '\.global';
225 $T_HDR_literal = "\.section\t\.rodata\n";
226 $T_HDR_misc = "\.text\n\t\.align 8\n";
227 $T_HDR_data = "\.data\n\t\.align 8\n";
228 $T_HDR_consist = "\.text\n";
229 $T_HDR_closure = "\.data\n\t\.align 8\n";
230 $T_HDR_srt = "\.text\n\t\.align 8\n";
231 $T_HDR_info = "\.text\n\t\.align 8\n";
232 $T_HDR_entry = "\.text\n\t\.align 16\n";
233 $T_HDR_vector = "\.text\n\t\.align 8\n";
234 $T_HDR_direct = "\.text\n\t\.align 8\n";
235 $T_create_word = "\t.word";
237 #--------------------------------------------------------#
238 } elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) {
240 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
241 $T_US = '_'; # _ if symbols have an underscore on the front
242 $T_PRE_APP = '^# MAY NOT APPLY'; # regexp that says what comes before APP/NO_APP
243 $T_CONST_LBL = '^LC(\d+):$';
246 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.proc\s+\d+|\.const|\.cstring|\.globl\s+\S+|\.text|\.data|\.even|\.stab[^n].*)\n)';
247 $T_COPY_DIRVS = '\.(globl|proc|stab)';
248 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
250 $T_DOT_WORD = '\.long';
251 $T_DOT_GLOBAL = '\.globl';
252 $T_HDR_literal = "\.text\n\t\.even\n";
253 $T_HDR_misc = "\.text\n\t\.even\n";
254 $T_HDR_data = "\.data\n\t\.even\n";
255 $T_HDR_consist = "\.text\n";
256 $T_HDR_closure = "\.data\n\t\.even\n";
257 $T_HDR_srt = "\.text\n\t\.even\n";
258 $T_HDR_info = "\.text\n\t\.even\n";
259 $T_HDR_entry = "\.text\n\t\.even\n";
260 $T_HDR_vector = "\.text\n\t\.even\n";
261 $T_HDR_direct = "\.text\n\t\.even\n";
262 $T_create_word = "\t.long";
264 #--------------------------------------------------------#
265 } elsif ( $TargetPlatform =~ /^mips-.*/ ) {
267 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
268 $T_US = ''; # _ if symbols have an underscore on the front
269 $T_PRE_APP = '^\s*#'; # regexp that says what comes before APP/NO_APP
270 $T_CONST_LBL = '^\$LC(\d+):$'; # regexp for what such a lbl looks like
273 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.(globl|ent)\s+\S+|\.text|\.r?data)\n)';
274 $T_COPY_DIRVS = '\.(globl|ent)';
276 $T_hsc_cc_PAT = 'I WAS TOO LAZY TO DO THIS BIT (WDP 95/05)';
277 $T_DOT_WORD = '\.word';
278 $T_DOT_GLOBAL = '^\t\.globl';
279 $T_HDR_literal = "\t\.rdata\n\t\.align 2\n";
280 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
281 $T_HDR_data = "\t\.data\n\t\.align 2\n";
282 $T_HDR_consist = 'TOO LAZY TO DO THIS TOO';
283 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
284 $T_HDR_srt = "\t\.text\n\t\.align 2\n";
285 $T_HDR_info = "\t\.text\n\t\.align 2\n";
286 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
287 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
288 $T_HDR_direct = "\t\.text\n\t\.align 2\n";
289 $T_create_word = "\t.word";
291 #--------------------------------------------------------#
292 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ ) {
293 # Apple PowerPC Darwin/MacOS X.
294 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
295 $T_US = '_'; # _ if symbols have an underscore on the front
296 $T_PRE_APP = 'WHAT IS THIS'; # regexp that says what comes before APP/NO_APP
297 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
300 $T_MOVE_DIRVS = '^(\s*(\.align \d+|\.text|\.data|\.cstring|\.non_lazy_symbol_pointer|\.const|\.static_const|\.literal4|\.literal8|\.static_data|\.globl \S+)\n)';
301 $T_COPY_DIRVS = '\.(globl)';
303 $T_hsc_cc_PAT = '\.byte.*\)(hsc|cc) (.*)"\n\t\.byte \d+\n\t\.byte "(.*)"\n\t\.byte \d+';
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_data\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_consist = "\t\.text\n\t\.align 2\n";
311 $T_HDR_closure = "\t\.const_data\n\t\.align 2\n";
312 $T_HDR_srt = "\t\.text\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 $T_HDR_direct = "\t\.text\n\t\.align 2\n";
317 $T_create_word = "\t.long";
319 #--------------------------------------------------------#
320 } elsif ( $TargetPlatform =~ /^sparc-.*-solaris2/ ) {
322 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
323 $T_US = ''; # _ if symbols have an underscore on the front
324 $T_PRE_APP = 'DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
325 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
328 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*|\.section.*|\s+\.type.*|\s+\.size.*)\n)';
329 $T_COPY_DIRVS = '\.(global|proc|stab)';
331 $T_hsc_cc_PAT = '\.asciz.*\)(hsc|cc) (.*)\\\\t(.*)"';
332 $T_DOT_WORD = '\.(long|word|byte|half|skip|uahalf|uaword)';
333 $T_DOT_GLOBAL = '^\t\.global';
334 $T_HDR_literal = "\.text\n\t\.align 8\n";
335 $T_HDR_misc = "\.text\n\t\.align 4\n";
336 $T_HDR_data = "\.data\n\t\.align 8\n";
337 $T_HDR_consist = "\.text\n";
338 $T_HDR_closure = "\.data\n\t\.align 4\n";
339 $T_HDR_srt = "\.data\n\t\.align 4\n";
340 $T_HDR_info = "\.text\n\t\.align 4\n";
341 $T_HDR_entry = "\.text\n\t\.align 4\n";
342 $T_HDR_vector = "\.text\n\t\.align 4\n";
343 $T_HDR_direct = "\.text\n\t\.align 4\n";
344 $T_create_word = "\t.word";
346 #--------------------------------------------------------#
347 } elsif ( $TargetPlatform =~ /^sparc-.*-sunos4/ ) {
349 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
350 $T_US = '_'; # _ if symbols have an underscore on the front
351 $T_PRE_APP = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
352 $T_CONST_LBL = '^LC(\d+):$';
355 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*)\n)';
356 $T_COPY_DIRVS = '\.(global|proc|stab)';
357 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
359 $T_DOT_WORD = '\.word';
360 $T_DOT_GLOBAL = '^\t\.global';
361 $T_HDR_literal = "\.text\n\t\.align 8\n";
362 $T_HDR_misc = "\.text\n\t\.align 4\n";
363 $T_HDR_data = "\.data\n\t\.align 8\n";
364 $T_HDR_consist = "\.text\n";
365 $T_HDR_closure = "\.data\n\t\.align 4\n";
366 $T_HDR_srt = "\.data\n\t\.align 4\n";
367 $T_HDR_info = "\.text\n\t\.align 4\n";
368 $T_HDR_entry = "\.text\n\t\.align 4\n";
369 $T_HDR_vector = "\.text\n\t\.align 4\n";
370 $T_HDR_direct = "\.text\n\t\.align 4\n";
371 $T_create_word = "\t.word";
373 #--------------------------------------------------------#
375 print STDERR "$Pgm: don't know how to mangle assembly language for: $TargetPlatform\n";
380 print STDERR "T_STABBY: $T_STABBY\n";
381 print STDERR "T_US: $T_US\n";
382 print STDERR "T_PRE_APP: $T_PRE_APP\n";
383 print STDERR "T_CONST_LBL: $T_CONST_LBL\n";
384 print STDERR "T_POST_LBL: $T_POST_LBL\n";
385 if ( $TargetPlatform =~ /^i386-/ ) {
386 print STDERR "T_X86_PRE_LLBL_PAT: $T_X86_PRE_LLBL_PAT\n";
387 print STDERR "T_X86_PRE_LLBL: $T_X86_PRE_LLBL\n";
388 print STDERR "T_X86_BADJMP: $T_X86_BADJMP\n";
390 print STDERR "T_MOVE_DIRVS: $T_MOVE_DIRVS\n";
391 print STDERR "T_COPY_DIRVS: $T_COPY_DIRVS\n";
392 print STDERR "T_hsc_cc_PAT: $T_hsc_cc_PAT\n";
393 print STDERR "T_DOT_WORD: $T_DOT_WORD\n";
394 print STDERR "T_HDR_literal: $T_HDR_literal\n";
395 print STDERR "T_HDR_misc: $T_HDR_misc\n";
396 print STDERR "T_HDR_data: $T_HDR_data\n";
397 print STDERR "T_HDR_consist: $T_HDR_consist\n";
398 print STDERR "T_HDR_closure: $T_HDR_closure\n";
399 print STDERR "T_HDR_info: $T_HDR_info\n";
400 print STDERR "T_HDR_entry: $T_HDR_entry\n";
401 print STDERR "T_HDR_vector: $T_HDR_vector\n";
402 print STDERR "T_HDR_direct: $T_HDR_direct\n";
408 %************************************************************************
410 \subsection{Mangle away}
412 %************************************************************************
416 local($in_asmf, $out_asmf) = @_;
418 # multi-line regexp matching:
423 &init_TARGET_STUFF();
424 &init_FUNNY_THINGS();
426 open(INASM, "< $in_asmf")
427 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
428 open(OUTASM,"> $out_asmf")
429 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
431 # read whole file, divide into "chunks":
432 # record some info about what we've found...
434 @chk = (); # contents of the chunk
435 $numchks = 0; # number of them
436 @chkcat = (); # what category of thing in each chunk
437 @chksymb = (); # what symbol(base) is defined in this chunk
438 %entrychk = (); # ditto, its entry code
439 %closurechk = (); # ditto, the (static) closure
440 %srtchk = (); # ditto, its SRT (for top-level things)
441 %infochk = (); # given a symbol base, say what chunk its info tbl is in
442 %vectorchk = (); # ditto, return vector table
443 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
445 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
448 tr/\r//d if $TargetPlatform =~ /-mingw32$/; # In case Perl doesn't convert line endings
449 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
450 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
451 next if /^\t\.def.*endef$/;
452 next if /${T_PRE_APP}(NO_)?APP/o;
453 next if /^;/ && $TargetPlatform =~ /^hppa/;
455 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|ia64)-/;
457 if ( $TargetPlatform =~ /^mips-/
458 && /^\t\.(globl\S+\.text|comm\t)/ ) {
459 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
461 # As a temporary solution for compiling "foreign export" declarations,
462 # we use global variables to pass arguments from C to STG land.
463 # These declarations live in the .hc file and not in the generated C
464 # stub file, so we let them pass through here.
465 } elsif ( /^\t\.comm\t__fexp_.*$/ ) {
467 $chkcat[$i] = 'data';
470 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
471 # duplicated from the bottom
475 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
476 # Alphas: Local labels not to be confused with new chunks
479 # NB: all the rest start with a non-space
481 } elsif ( $TargetPlatform =~ /^mips-/
482 && /^\d+:/ ) { # a funny-looking very-local label
485 } elsif ( /$T_CONST_LBL/o ) {
487 $chkcat[$i] = 'literal';
490 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
492 $chkcat[$i] = 'splitmarker';
495 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
498 $chkcat[$i] = 'infotbl';
499 $chksymb[$i] = $symb;
501 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
503 $infochk{$symb} = $i;
505 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
507 $chkcat[$i] = 'entry';
512 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
514 $chkcat[$i] = 'closure';
517 $closurechk{$1} = $i;
519 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
526 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
528 $chkcat[$i] = 'data';
531 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
534 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
535 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
536 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
537 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
540 $chkcat[$i] = 'data';
543 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
548 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
549 # all CC_ symbols go in the data section...
551 $chkcat[$i] = 'data';
554 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
556 $chkcat[$i] = 'misc';
558 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
560 $chkcat[$i] = 'vector';
565 # As a temporary solution for compiling "foreign export" declarations,
566 # we use global variables to pass arguments from C to STG land.
567 # These declarations live in the .hc file and not in the generated C
568 # stub file, so we let them pass through here.
569 } elsif ( /^[\t ]+\.comm[\t ]+__fexp_.*$/ ) {
571 $chkcat[$i] = 'data';
574 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
575 && /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
576 # Some Solaris system headers contain function definitions (as
577 # opposed to mere prototypes), which end up in the .hc file when
578 # a Haskell module foreign imports the corresponding system
579 # functions (most notably stat()). We put them into the text
580 # segment. Note that this currently does not extend to function
581 # names starting with an underscore.
584 $chkcat[$i] = 'misc';
587 } elsif ( /^${T_US}[A-Za-z0-9_]/o
588 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
589 || ! /^L\$\d+$/ ) ) {
593 print STDERR "Warning: retaining unknown function \`$thing' in output from C compiler\n"
594 unless # $KNOWN_FUNNY_THING{$thing}
595 /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
596 || /^${T_US}__stg_.*${T_POST_LBL}$/o # more RTS internals
597 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
598 || /^${T_US}.*_slow${T_POST_LBL}$/o # slow entry
599 || /^${T_US}__stginit.*${T_POST_LBL}$/o # __stginit<module>
600 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
601 || /^${T_US}.*_fast${T_POST_LBL}$/o # primops
602 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
603 || /^_uname:/o; # x86/Solaris2
605 $chkcat[$i] = 'misc';
608 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.picsymbol_stub/ )
611 $chkcat[$i] = 'dyld';
613 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.symbol_stub/ )
616 $chkcat[$i] = 'dyld';
618 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.lazy_symbol_pointer/ )
621 $chkcat[$i] = 'dyld';
623 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.non_lazy_symbol_pointer/ )
626 $chkcat[$i] = 'dyld';
628 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.data/ && $chkcat[$i] == 'dyld')
629 { # non_lazy_symbol_ptrs that point to local symbols
631 $chkcat[$i] = 'dyld';
633 } else { # simple line (duplicated at the top)
638 $numchks = $#chk + 1;
640 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
641 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
644 # the division into chunks is imperfect;
645 # we throw some things over the fence into the next
648 # also, there are things we would like to know
649 # about the whole module before we start spitting
652 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
653 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
655 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
657 # Alphas: NB: we start meddling at chunk 1, not chunk 0
658 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
659 # spits a ".quad 0" in after the very first ".rdata"; we
660 # detect this special case (tossing the ".quad 0")!
661 local($magic_rdata_seen) = 0;
663 # HPPAs, MIPSen: also start medding at chunk 1
665 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
666 $c = $chk[$i]; # convenience copy
668 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
670 # toss all prologue stuff; HPPA is pretty weird
672 $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
674 # be slightly paranoid to make sure there's
675 # nothing surprising in there
676 if ( $c =~ /--- BEGIN ---/ ) {
677 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
679 # remove junk whitespace around the split point
683 if ($TargetPlatform =~ /^i386-/) {
684 $p =~ s/^\tpushl\s+\%edi\n//;
685 $p =~ s/^\tpushl\s+\%esi\n//;
686 $p =~ s/^\tpushl\s+\%ebx\n//;
687 $p =~ s/^\tmovl\s+\%esi,\s*\d*\(\%esp\)\n//;
688 $p =~ s/^\tmovl\s+\%edi,\s*\d*\(\%esp\)\n//;
689 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
690 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-cygwin32/);
692 # GCC 3.1 is in the habit of adding spurious writes to the
693 # stack in the prologue. Just to be on the safe side,
694 # chuck these over the fence into the main code.
695 while ($p =~ /^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n/) {
696 # print "Spurious instruction: $&";
701 } elsif ($TargetPlatform =~ /^ia64-/) {
702 $p =~ s/^\t\.prologue .*\n//;
703 $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, 31, \d+, 0\n//;
704 $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
705 $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
706 $p =~ s/^\t\.(mii|mmi)\n//; # bundling is no longer sensible
707 $p =~ s/^\t;;\n//g; # discard stops
708 $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
709 } elsif ($TargetPlatform =~ /^m68k-/) {
710 $p =~ s/^\tlink a6,#-?\d.*\n//;
711 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
712 # The above showed up in the asm code,
713 # so I added it here.
714 # I hope it's correct.
716 $p =~ s/^\tmovel d2,sp\@-\n//;
717 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
718 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
719 } elsif ($TargetPlatform =~ /^mips-/) {
720 # the .frame/.mask/.fmask that we use is the same
721 # as that produced by GCC for miniInterpret; this
722 # gives GDB some chance of figuring out what happened
723 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
724 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
725 $p =~ s/^\t\.(mask|fmask).*\n//g;
726 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
727 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
728 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
729 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
730 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
731 $p =~ s/__FRAME__/$FRAME/;
732 } elsif ($TargetPlatform =~ /^powerpc-apple-.*/) {
734 $pcrel_label =~ s/(.|\n)*^(L\d+\$pb):\n(.|\n)*/$2/ or $pcrel_label = "";
736 $p =~ s/^\tmflr r0\n//;
737 $p =~ s/^\tbl saveFP # f\d+\n//;
738 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
739 $p =~ s/^L\d+\$pb:\n//;
740 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
741 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
742 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
743 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
744 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
745 $p =~ s/^\tbcl 20,31,L\d+\$pb\n//;
746 $p =~ s/^L\d+\$pb:\n//;
747 $p =~ s/^\tmflr r31\n//;
749 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
750 # under some circumstances, only when generating position dependent code.
751 # I have no idea why, and I don't think it is necessary, so let's toss it.
752 $p =~ s/^\tli r\d+,0\n//g;
753 $p =~ s/^\tstw r\d+,\d+\(r1\)\n//g;
755 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
758 # HWL HACK: dont die, just print a warning
759 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
760 die "Prologue junk?: $p\n" if $p =~ /^\s+[^\s\.]/;
762 if ($TargetPlatform =~ /^powerpc-apple-.*/ && $pcrel_label ne "") {
763 # on PowerPC, we have to keep a part of the prologue
764 # (which loads the current instruction pointer into register r31)
765 $p .= "bcl 20,31,$pcrel_label\n";
766 $p .= "$pcrel_label:\n";
767 $p .= "\tmflr r31\n";
770 # glue together what's left
775 if ( $TargetPlatform =~ /^mips-/ ) {
776 # MIPS: first, this basic sequence may occur "--- END ---" or not
777 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
780 # toss all epilogue stuff; again, paranoidly
781 if ( $c =~ /--- END ---/ ) {
782 if (($r, $e) = split(/--- END ---/, $c)) {
783 if ($TargetPlatform =~ /^i386-/) {
785 $e =~ s/^\tpopl\s+\%edi\n//;
786 $e =~ s/^\tpopl\s+\%esi\n//;
787 $e =~ s/^\tpopl\s+\%edx\n//;
788 $e =~ s/^\tpopl\s+\%ecx\n//;
789 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
790 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
791 } elsif ($TargetPlatform =~ /^ia64-/) {
792 $e =~ s/^\tmov ar\.pfs = r\d+\n//;
793 $e =~ s/^\tmov b0 = r\d+\n//;
794 $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
795 $e =~ s/^\tbr\.ret\.sptk\.many b0\n//;
796 $e =~ s/^\t\.(mii|mmi|mib)\n//g; # bundling is no longer sensible
797 $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
798 $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
799 } elsif ($TargetPlatform =~ /^m68k-/) {
800 $e =~ s/^\tunlk a6\n//;
802 } elsif ($TargetPlatform =~ /^mips-/) {
803 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
804 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
805 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
806 $e =~ s/^\tj\t\$31\n//;
807 } elsif ($TargetPlatform =~ /^powerpc-apple-.*/) {
808 $e =~ s/^\taddi r1,r1,\d+\n//;
809 $e =~ s/^\tcal r1,\d+\(r1\)\n//;
810 $e =~ s/^\tlw?z? r\d+,\d+\(r1\)\n//;
811 $e =~ s/^\tmtlr r0\n//;
814 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
817 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
819 # glue together what's left
821 $c =~ s/\n\t\n/\n/; # junk blank line
825 # On SPARCs, we don't do --- BEGIN/END ---, we just
826 # toss the register-windowing save/restore/ret* instructions
828 if ( $TargetPlatform =~ /^sparc-/ ) {
829 $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
830 # throw away PROLOGUE comments
831 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
834 # On Alphas, the prologue mangling is done a little later (below)
836 # toss all calls to __DISCARD__
837 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
838 $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
839 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-.*/;
841 # IA64: mangle tailcalls into jumps here
842 $c =~ s/^\tbr\.call\.sptk\.many b0 = (.*)\n\t;;\n(\tmov r1 = r\d+\n)?\t;;\n\t--- TAILCALL ---\n/\tbr\.few $1\n/g
843 if $TargetPlatform =~ /^ia64-/;
845 # MIPS: that may leave some gratuitous asm macros around
846 # (no harm done; but we get rid of them to be tidier)
847 $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/
848 if $TargetPlatform =~ /^mips-/;
850 # toss stack adjustment after DoSparks
851 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
852 if $TargetPlatform =~ /^m68k-/; # this looks old...
854 if ( $TargetPlatform =~ /^alpha-/ &&
855 ! $magic_rdata_seen &&
856 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
857 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
858 $magic_rdata_seen = 1;
861 # pick some end-things and move them to the next chunk
863 # pin a funny end-thing on (for easier matching):
864 $c .= 'FUNNY#END#THING';
866 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
870 # on x86 we try not to copy any directives into a literal
871 # chunk, rather we keep looking for the next real chunk. This
872 # is because we get things like
874 # .globl blah_closure
880 if ( $TargetPlatform =~ /^(i386|sparc|powerpc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
882 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
885 if ( $j < $numchks ) {
886 $chk[$j] = $to_move . $chk[$j];
890 elsif ( $i < ($numchks - 1)
891 && ( $to_move =~ /${T_COPY_DIRVS}/
892 || ($TargetPlatform =~ /^hppa/ && $to_move =~ /align/ && $chkcat[$i+1] eq 'literal') )) {
893 $chk[$i + 1] = $to_move . $chk[$i + 1];
894 # otherwise they're tossed
897 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
900 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
902 # toss all prologue stuff, except for loading gp, and the ..ng address
903 unless ($c =~ /\.ent.*\n\$.*\.\.ng:/) {
904 if (($p, $r) = split(/^\t\.prologue/, $c)) {
905 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
906 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
907 $keep =~ s/^\t\.(mask|fmask).*\n//g;
908 $c = $keep . "..ng:\n";
910 print STDERR "malformed code block ($ent)?\n"
913 $c .= "\t.prologue" . $r;
917 $c =~ s/FUNNY#END#THING//;
919 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
921 $chk[$i] = $c; # update w/ convenience copy
924 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
925 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
928 if ( $TargetPlatform =~ /^alpha-/ ) {
929 # print out the header stuff first
930 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
931 print OUTASM $chk[0];
933 } elsif ( $TargetPlatform =~ /^hppa/ ) {
934 print OUTASM $chk[0];
936 } elsif ( $TargetPlatform =~ /^mips-/ ) {
937 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
939 # get rid of horrible "<dollar>Revision: .*$" strings
940 local(@lines0) = split(/\n/, $chk[0]);
942 while ( $z <= $#lines0 ) {
943 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
946 while ( $z <= $#lines0 ) {
948 last if $lines0[$z] =~ /[,\t]0x0$/;
954 $chk[0] = join("\n", @lines0);
955 $chk[0] =~ s/\n\n+/\n/;
956 print OUTASM $chk[0];
959 # print out all the literal strings next
960 for ($i = 0; $i < $numchks; $i++) {
961 if ( $chkcat[$i] eq 'literal' ) {
962 print OUTASM $T_HDR_literal, $chk[$i];
963 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
965 $chkcat[$i] = 'DONE ALREADY';
969 # on the HPPA, print out all the bss next
970 if ( $TargetPlatform =~ /^hppa/ ) {
971 for ($i = 1; $i < $numchks; $i++) {
972 if ( $chkcat[$i] eq 'bss' ) {
973 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
974 print OUTASM $chk[$i];
976 $chkcat[$i] = 'DONE ALREADY';
981 for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) {
982 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
984 next if $chkcat[$i] eq 'DONE ALREADY';
986 if ( $chkcat[$i] eq 'misc' ) {
987 if ($chk[$i] ne '') {
988 print OUTASM $T_HDR_misc;
989 &print_doctored($chk[$i], 0);
992 } elsif ( $chkcat[$i] eq 'toss' ) {
993 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
995 } elsif ( $chkcat[$i] eq 'data' ) {
996 if ($chk[$i] ne '') {
997 print OUTASM $T_HDR_data;
998 print OUTASM $chk[$i];
1001 } elsif ( $chkcat[$i] eq 'consist' ) {
1002 if ( $chk[$i] =~ /$T_hsc_cc_PAT/o ) {
1003 local($consist) = "$1.$2.$3";
1004 $consist =~ s/,/./g;
1005 $consist =~ s/\//./g;
1006 $consist =~ s/-/_/g;
1007 $consist =~ s/[^A-Za-z0-9_.]/ZZ/g; # ToDo: properly?
1009 # Using a cygnus-2.7-96q4 gcc build on hppas, the
1010 # consistency chunk for ghc_cc_ID often (but not always!)
1011 # gets lumped with a bunch of .IMPORT directives containing info on
1012 # the code or data space nature of external symbols. We can't
1013 # toss these, so once the consistency ID has been turned into
1014 # a representable symbol, we substitute it for the symbol
1015 # that the string was attached to in the first place (ghc_cc_ID.)
1016 # (The original string is also substituted away.)
1018 # This change may affect the code output on other platforms in
1019 # adverse ways, hence we restrict this hack hppa targets only.
1022 if ( $TargetPlatform =~ /^hppa/ ) {
1023 $chk[$i] =~ s/^${T_US}ghc.*c_ID$TPOSTLBL/$consist/o;
1024 $chk[$i] =~ s/\t$T_hsc_cc_PAT/$T_HDR_misc/o;
1025 $consist = $chk[$i]; #clumsily
1027 print OUTASM $T_HDR_consist, "${consist}${T_POST_LBL}\n";
1029 } elsif ( $TargetPlatform !~ /^(mips)-/ ) { # we just don't try in those case (ToDo)
1030 # on mips: consistency string is just a v
1031 # horrible bunch of .bytes,
1032 # which I am too lazy to sort out (WDP 95/05)
1034 print STDERR "Couldn't grok consistency: ", $chk[$i];
1037 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1038 # we can just re-constitute this one...
1039 # NB: we emit _three_ underscores no matter what,
1040 # so ghc-split doesn't have to care.
1041 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1043 } elsif ( $chkcat[$i] eq 'closure'
1044 || $chkcat[$i] eq 'srt'
1045 || $chkcat[$i] eq 'infotbl'
1046 || $chkcat[$i] eq 'entry') { # do them in that order
1047 $symb = $chksymb[$i];
1050 if ( defined($closurechk{$symb}) ) {
1051 print OUTASM $T_HDR_closure;
1052 print OUTASM $chk[$closurechk{$symb}];
1053 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1057 if ( defined($srtchk{$symb}) ) {
1058 print OUTASM $T_HDR_srt;
1059 print OUTASM $chk[$srtchk{$symb}];
1060 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1064 if ( defined($infochk{$symb}) ) {
1066 print OUTASM $T_HDR_info;
1067 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1069 # entry code will be put here!
1071 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1075 if ( defined($entrychk{$symb}) ) {
1077 $c = $chk[$entrychk{$symb}];
1079 print OUTASM $T_HDR_entry;
1081 &print_doctored($c, 1); # NB: the 1!!!
1083 $chkcat[$entrychk{$symb}] = 'DONE ALREADY';
1086 } elsif ( $chkcat[$i] eq 'vector'
1087 || $chkcat[$i] eq 'direct' ) { # do them in that order
1088 $symb = $chksymb[$i];
1091 if ( defined($vectorchk{$symb}) ) {
1092 print OUTASM $T_HDR_vector;
1093 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1095 # direct return code will be put here!
1096 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1098 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1099 # Alphas: the commented nop is for the splitter, to ensure
1100 # that no module ends with a label as the very last
1101 # thing. (The linker will adjust the label to point
1102 # to the first code word of the next module linked in,
1103 # even if alignment constraints cause the label to move!)
1105 print OUTASM "\t# nop\n";
1108 } elsif ( $chkcat[$i] eq 'toc' ) {
1109 # silly optimisation to print tocs, since they come in groups...
1110 print OUTASM $T_HDR_toc;
1112 while ($chkcat[$j] eq 'toc')
1113 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1116 print OUTASM $chk[$j];
1118 $chkcat[$j] = 'DONE ALREADY';
1122 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && $chkcat[$i] eq 'dyld' ) {
1123 # powerpc-apple: dynamic linker stubs
1124 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1125 { # print them out unchanged, but remove the stubs for __DISCARD__
1126 print OUTASM $chk[$i];
1129 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1133 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1136 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1137 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1142 sub hppa_mash_prologue { # OK, epilogue, too
1145 # toss all prologue stuff
1146 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1148 # Lie about our .CALLINFO
1149 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1156 # toss all epilogue stuff
1157 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1159 # Sorry; we moved the _info stuff to the code segment.
1160 s/_info,DATA/_info,CODE/g;
1167 sub print_doctored {
1168 local($_, $need_fallthru_patch) = @_;
1170 if ( $TargetPlatform !~ /^i386-/
1171 || ! /^\t[a-z]/ # no instructions in here, apparently
1172 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1176 # OK, must do some x86 **HACKING**
1178 local($entry_patch) = '';
1179 local($exit_patch) = '';
1181 # gotta watch out for weird instructions that
1182 # invisibly smash various regs:
1183 # rep* %ecx used for counting
1184 # scas* %edi used for destination index
1185 # cmps* %e[sd]i used for indices
1186 # loop* %ecx used for counting
1191 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1193 # * GCC used an "STG reg" for its own purposes
1195 # * some secret uses of machine reg, requiring STG reg
1196 # to be saved/restored
1198 # The most dangerous "GCC uses" of an "STG reg" are when
1199 # the reg holds the target of a jmp -- it's tricky to
1200 # insert the patch-up code before we get to the target!
1201 # So here we change the jmps:
1203 # --------------------------------------------------------
1204 # it can happen that we have jumps of the form...
1205 # jmp *<something involving %esp>
1207 # jmp <something involving another naughty register...>
1209 # a reasonably-common case is:
1211 # movl $_blah,<bad-reg>
1215 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[abcd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1217 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1218 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1219 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1220 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1221 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1222 if /(jmp|call)\s+.*\%esi/;
1224 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1225 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1226 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1227 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1228 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1229 if /(jmp|call)\s+.*\%edi/;
1232 # OK, now we can decide what our patch-up code is going to
1235 # Offsets into register table - you'd better update these magic
1236 # numbers should you change its contents!
1237 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1238 local($OFFSET_Hp)=88;
1240 # Note funky ".=" stuff; we're *adding* to these _patch guys
1241 if ( $StolenX86Regs <= 2
1242 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1243 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1244 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1246 # nothing for call_{entry,exit} because %esi is callee-save
1248 if ( $StolenX86Regs <= 3
1249 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1250 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1251 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1253 # nothing for call_{entry,exit} because %edi is callee-save
1256 # --------------------------------------------------------
1257 # next, here we go with non-%esp patching!
1259 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1261 # Before calling GC we must set up the exit condition before the call
1262 # and entry condition when we come back
1264 # fix _all_ non-local jumps:
1266 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1267 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1269 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1271 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1272 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1274 if ($StolenX86Regs == 2 ) {
1275 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1276 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1277 } elsif ($StolenX86Regs == 3 ) {
1278 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1279 if /^\t(jmp|call)\s+.*\%edi/;
1282 # --------------------------------------------------------
1283 # that's it -- print it
1285 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1289 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1290 print OUTASM $exit_patch;
1291 # ToDo: make it not print if there is a "jmp" at the end
1297 sub init_FUNNY_THINGS {
1298 %KNOWN_FUNNY_THING = (
1300 # "${T_US}stg_.*{T_POST_LBL}", 1,
1305 The following table reversal is used for both info tables and return
1306 vectors. In both cases, we remove the first entry from the table,
1307 reverse the table, put the label at the end, and paste some code
1308 (that which is normally referred to by the first entry in the table)
1309 right after the table itself. (The code pasting is done elsewhere.)
1313 local($symb, $tbl, $discard1) = @_;
1315 return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1317 local($before) = '';
1319 local(@imports) = (); # hppa only
1322 local(@lines) = split(/\n/, $tbl);
1325 # Deal with the header...
1326 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1327 $label .= $lines[$i] . "\n",
1328 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1329 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1330 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1332 $before .= $lines[$i] . "\n"; # otherwise...
1335 # Grab the table data...
1336 if ( $TargetPlatform !~ /^hppa/ ) {
1337 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1338 push(@words, $lines[$i]);
1340 } else { # hppa weirdness
1341 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1342 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1343 push(@imports, $lines[$i]);
1345 # We don't use HP's ``function pointers''
1346 # We just use labels in code space, like normal people
1347 $lines[$i] =~ s/P%//;
1348 push(@words, $lines[$i]);
1353 # Now throw away any initial zero word from the table. This is a hack
1354 # that lets us reduce the size of info tables when the SRT field is not
1355 # needed: see comments StgFunInfoTable in InfoTables.h.
1357 # The .zero business is for Linux/ELF.
1358 # The .skip business is for Sparc/Solaris/ELF.
1359 # The .blockz business is for HPPA.
1361 # if ($words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1366 for (; $i <= $#lines; $i++) {
1367 $after .= $lines[$i] . "\n";
1370 # Alphas: If we have anonymous text (not part of a procedure), the
1371 # linker may complain about missing exception information. Bleh.
1372 # To suppress this, we place a .ent/.end pair around the code.
1373 # At the same time, we have to be careful and not enclose any leading
1374 # .file/.loc directives.
1375 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1376 local ($ident) = $1;
1377 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1378 $after .= "\t.end $ident\n";
1381 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1382 # assembler (!) wherein .quad constants inside .text sections are
1383 # first narrowed to 32 bits then sign-extended back to 64 bits.
1384 # This obviously screws up our 64-bit bitmaps, so we work around
1385 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1386 if ( $TargetPlatform =~ /^alpha-/ ) {
1388 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1389 local ($number) = $1;
1390 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1391 local ($sign, $base, $digits) = ($1, $2, $3);
1392 $base = (10, 8, 16)[length $base];
1393 local ($hi, $lo) = (0, 0);
1394 foreach $i (split(//, $digits)) {
1395 $j = $lo * $base + $i;
1396 $lo = $j % 4294967296;
1397 $hi = $hi * $base + ($j - $lo) / 4294967296;
1399 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1401 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1402 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1404 print STDERR "Cannot handle \".quad $number\" in info table\n";
1412 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
1413 . join("\n", @words) . "\n"
1416 # print STDERR "before=$before\n";
1417 # print STDERR "label=$label\n";
1418 # print STDERR "words=",(reverse @words),"\n";
1419 # print STDERR "after=$after\n";
1425 The HP is a major nuisance. The threaded code mangler moved info
1426 tables from data space to code space, but unthreaded code in the RTS
1427 still has references to info tables in data space. Since the HP
1428 linker is very precise about where symbols live, we need to patch the
1429 references in the unthreaded RTS as well.
1432 sub mini_mangle_asm_hppa {
1433 local($in_asmf, $out_asmf) = @_;
1435 open(INASM, "< $in_asmf")
1436 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1437 open(OUTASM,"> $out_asmf")
1438 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1441 s/_info,DATA/_info,CODE/; # Move _info references to code space
1447 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1448 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1454 sub tidy_up_and_die {
1455 local($return_val, $msg) = @_;
1457 exit (($return_val == 0) ? 0 : 1);