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+\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|\.const_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 || /^${T_US}.*_srtd${T_POST_LBL}$/o # large bitmaps
602 || /^${T_US}.*_fast${T_POST_LBL}$/o # primops
603 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
604 || /^_uname:/o; # x86/Solaris2
606 $chkcat[$i] = 'misc';
609 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.picsymbol_stub/ )
612 $chkcat[$i] = 'dyld';
614 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.symbol_stub/ )
617 $chkcat[$i] = 'dyld';
619 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.lazy_symbol_pointer/ )
622 $chkcat[$i] = 'dyld';
624 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.non_lazy_symbol_pointer/ )
627 $chkcat[$i] = 'dyld';
629 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.data/ && $chkcat[$i] eq 'dyld')
630 { # non_lazy_symbol_ptrs that point to local symbols
632 $chkcat[$i] = 'dyld';
634 } else { # simple line (duplicated at the top)
639 $numchks = $#chk + 1;
641 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
642 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
645 # the division into chunks is imperfect;
646 # we throw some things over the fence into the next
649 # also, there are things we would like to know
650 # about the whole module before we start spitting
653 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
654 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
656 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
658 # Alphas: NB: we start meddling at chunk 1, not chunk 0
659 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
660 # spits a ".quad 0" in after the very first ".rdata"; we
661 # detect this special case (tossing the ".quad 0")!
662 local($magic_rdata_seen) = 0;
664 # HPPAs, MIPSen: also start medding at chunk 1
666 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
667 $c = $chk[$i]; # convenience copy
669 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
671 # toss all prologue stuff; HPPA is pretty weird
673 $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
675 # be slightly paranoid to make sure there's
676 # nothing surprising in there
677 if ( $c =~ /--- BEGIN ---/ ) {
678 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
680 # remove junk whitespace around the split point
684 if ($TargetPlatform =~ /^i386-/) {
685 $p =~ s/^\tpushl\s+\%edi\n//;
686 $p =~ s/^\tpushl\s+\%esi\n//;
687 $p =~ s/^\tpushl\s+\%ebx\n//;
688 $p =~ s/^\tmovl\s+\%esi,\s*\d*\(\%esp\)\n//;
689 $p =~ s/^\tmovl\s+\%edi,\s*\d*\(\%esp\)\n//;
690 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
691 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-cygwin32/);
693 # GCC 3.1 is in the habit of adding spurious writes to the
694 # stack in the prologue. Just to be on the safe side,
695 # chuck these over the fence into the main code.
696 while ($p =~ /^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n/) {
697 # print "Spurious instruction: $&";
702 } elsif ($TargetPlatform =~ /^ia64-/) {
703 $p =~ s/^\t\.prologue .*\n//;
704 $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, 3[12], \d+, 0\n//;
705 $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
706 $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
707 $p =~ s/^\t\.(mii|mmi)\n//g; # bundling is no longer sensible
708 $p =~ s/^\t;;\n//g; # discard stops
709 $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
711 # GCC 3.3 saves r1 in the prologue, move this to the body
712 if ($p =~ /^\tmov r\d+ = r1\n/) {
716 } elsif ($TargetPlatform =~ /^m68k-/) {
717 $p =~ s/^\tlink a6,#-?\d.*\n//;
718 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
719 # The above showed up in the asm code,
720 # so I added it here.
721 # I hope it's correct.
723 $p =~ s/^\tmovel d2,sp\@-\n//;
724 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
725 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
726 } elsif ($TargetPlatform =~ /^mips-/) {
727 # the .frame/.mask/.fmask that we use is the same
728 # as that produced by GCC for miniInterpret; this
729 # gives GDB some chance of figuring out what happened
730 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
731 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
732 $p =~ s/^\t\.(mask|fmask).*\n//g;
733 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
734 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
735 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
736 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
737 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
738 $p =~ s/__FRAME__/$FRAME/;
739 } elsif ($TargetPlatform =~ /^powerpc-apple-.*/) {
741 $pcrel_label =~ s/(.|\n)*^(L\d+\$pb):\n(.|\n)*/$2/ or $pcrel_label = "";
743 $p =~ s/^\tmflr r0\n//;
744 $p =~ s/^\tbl saveFP # f\d+\n//;
745 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
746 $p =~ s/^L\d+\$pb:\n//;
747 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
748 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
749 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
750 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
751 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
752 $p =~ s/^\tbcl 20,31,L\d+\$pb\n//;
753 $p =~ s/^L\d+\$pb:\n//;
754 $p =~ s/^\tmflr r31\n//;
756 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
757 # under some circumstances, only when generating position dependent code.
758 # I have no idea why, and I don't think it is necessary, so let's toss it.
759 $p =~ s/^\tli r\d+,0\n//g;
760 $p =~ s/^\tstw r\d+,\d+\(r1\)\n//g;
762 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
765 # HWL HACK: dont die, just print a warning
766 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
767 die "Prologue junk?: $p\n" if $p =~ /^\s+[^\s\.]/;
769 if ($TargetPlatform =~ /^powerpc-apple-.*/ && $pcrel_label ne "") {
770 # on PowerPC, we have to keep a part of the prologue
771 # (which loads the current instruction pointer into register r31)
772 $p .= "bcl 20,31,$pcrel_label\n";
773 $p .= "$pcrel_label:\n";
774 $p .= "\tmflr r31\n";
777 # glue together what's left
782 if ( $TargetPlatform =~ /^mips-/ ) {
783 # MIPS: first, this basic sequence may occur "--- END ---" or not
784 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
787 # toss all epilogue stuff; again, paranoidly
788 if ( $c =~ /--- END ---/ ) {
789 if (($r, $e) = split(/--- END ---/, $c)) {
790 if ($TargetPlatform =~ /^i386-/) {
792 $e =~ s/^\tpopl\s+\%edi\n//;
793 $e =~ s/^\tpopl\s+\%esi\n//;
794 $e =~ s/^\tpopl\s+\%edx\n//;
795 $e =~ s/^\tpopl\s+\%ecx\n//;
796 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
797 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
798 } elsif ($TargetPlatform =~ /^ia64-/) {
799 $e =~ s/^\tmov ar\.pfs = r\d+\n//;
800 $e =~ s/^\tmov b0 = r\d+\n//;
801 $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
802 $e =~ s/^\tbr\.ret\.sptk\.many b0\n//;
803 $e =~ s/^\t\.(mii|mmi|mib)\n//g; # bundling is no longer sensible
804 $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
805 $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
806 } elsif ($TargetPlatform =~ /^m68k-/) {
807 $e =~ s/^\tunlk a6\n//;
809 } elsif ($TargetPlatform =~ /^mips-/) {
810 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
811 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
812 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
813 $e =~ s/^\tj\t\$31\n//;
814 } elsif ($TargetPlatform =~ /^powerpc-apple-.*/) {
815 $e =~ s/^\taddi r1,r1,\d+\n//;
816 $e =~ s/^\tcal r1,\d+\(r1\)\n//;
817 $e =~ s/^\tlw?z? r\d+,\d+\(r1\)\n//;
818 $e =~ s/^\tmtlr r0\n//;
821 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
824 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
826 # glue together what's left
828 $c =~ s/\n\t\n/\n/; # junk blank line
832 # On SPARCs, we don't do --- BEGIN/END ---, we just
833 # toss the register-windowing save/restore/ret* instructions
835 if ( $TargetPlatform =~ /^sparc-/ ) {
836 $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
837 # throw away PROLOGUE comments
838 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
841 # On Alphas, the prologue mangling is done a little later (below)
843 # toss all calls to __DISCARD__
844 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
845 $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
846 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-.*/;
848 # IA64: mangle tailcalls into jumps here
849 if ($TargetPlatform =~ /^ia64-/) {
850 while ($c =~ s/^\tbr\.call\.sptk\.many b0 = (.*)\n(?:^\.L([0-9]*):\n)?(?:\t;;\n)?(?:\tmov r1 = r\d+\n)?(?:\t;;\n)?\t--- TAILCALL ---\n(?:\t;;\n\tbr \.L\d+\n)?/\tbr\.few $1\n/) {
851 # Eek, the gcc optimiser is getting smarter... if we see a jump to the --- TAILCALL ---
852 # marker then we reapply the substitution at the source sites
853 $c =~ s/^\tbr \.L$2\n/\t--- TAILCALL ---\n/g if ($2);
857 # MIPS: that may leave some gratuitous asm macros around
858 # (no harm done; but we get rid of them to be tidier)
859 $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/
860 if $TargetPlatform =~ /^mips-/;
862 # toss stack adjustment after DoSparks
863 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
864 if $TargetPlatform =~ /^m68k-/; # this looks old...
866 if ( $TargetPlatform =~ /^alpha-/ &&
867 ! $magic_rdata_seen &&
868 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
869 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
870 $magic_rdata_seen = 1;
873 # pick some end-things and move them to the next chunk
875 # pin a funny end-thing on (for easier matching):
876 $c .= 'FUNNY#END#THING';
878 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
882 # on x86 we try not to copy any directives into a literal
883 # chunk, rather we keep looking for the next real chunk. This
884 # is because we get things like
886 # .globl blah_closure
892 if ( $TargetPlatform =~ /^(i386|sparc|powerpc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
894 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
897 if ( $j < $numchks ) {
898 $chk[$j] = $to_move . $chk[$j];
902 elsif ( $i < ($numchks - 1)
903 && ( $to_move =~ /${T_COPY_DIRVS}/
904 || ($TargetPlatform =~ /^hppa/ && $to_move =~ /align/ && $chkcat[$i+1] eq 'literal') )) {
905 $chk[$i + 1] = $to_move . $chk[$i + 1];
906 # otherwise they're tossed
909 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
912 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
914 # toss all prologue stuff, except for loading gp, and the ..ng address
915 unless ($c =~ /\.ent.*\n\$.*\.\.ng:/) {
916 if (($p, $r) = split(/^\t\.prologue/, $c)) {
917 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
918 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
919 $keep =~ s/^\t\.(mask|fmask).*\n//g;
920 $c = $keep . "..ng:\n";
922 print STDERR "malformed code block ($ent)?\n"
925 $c .= "\t.prologue" . $r;
929 $c =~ s/FUNNY#END#THING//;
931 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
933 $chk[$i] = $c; # update w/ convenience copy
936 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
937 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
940 if ( $TargetPlatform =~ /^alpha-/ ) {
941 # print out the header stuff first
942 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
943 print OUTASM $chk[0];
945 } elsif ( $TargetPlatform =~ /^hppa/ ) {
946 print OUTASM $chk[0];
948 } elsif ( $TargetPlatform =~ /^mips-/ ) {
949 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
951 # get rid of horrible "<dollar>Revision: .*$" strings
952 local(@lines0) = split(/\n/, $chk[0]);
954 while ( $z <= $#lines0 ) {
955 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
958 while ( $z <= $#lines0 ) {
960 last if $lines0[$z] =~ /[,\t]0x0$/;
966 $chk[0] = join("\n", @lines0);
967 $chk[0] =~ s/\n\n+/\n/;
968 print OUTASM $chk[0];
971 # print out all the literal strings next
972 for ($i = 0; $i < $numchks; $i++) {
973 if ( $chkcat[$i] eq 'literal' ) {
974 print OUTASM $T_HDR_literal, $chk[$i];
975 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
977 $chkcat[$i] = 'DONE ALREADY';
981 # on the HPPA, print out all the bss next
982 if ( $TargetPlatform =~ /^hppa/ ) {
983 for ($i = 1; $i < $numchks; $i++) {
984 if ( $chkcat[$i] eq 'bss' ) {
985 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
986 print OUTASM $chk[$i];
988 $chkcat[$i] = 'DONE ALREADY';
993 for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) {
994 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
996 next if $chkcat[$i] eq 'DONE ALREADY';
998 if ( $chkcat[$i] eq 'misc' ) {
999 if ($chk[$i] ne '') {
1000 print OUTASM $T_HDR_misc;
1001 &print_doctored($chk[$i], 0);
1004 } elsif ( $chkcat[$i] eq 'toss' ) {
1005 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
1007 } elsif ( $chkcat[$i] eq 'data' ) {
1008 if ($chk[$i] ne '') {
1009 print OUTASM $T_HDR_data;
1010 print OUTASM $chk[$i];
1013 } elsif ( $chkcat[$i] eq 'consist' ) {
1014 if ( $chk[$i] =~ /$T_hsc_cc_PAT/o ) {
1015 local($consist) = "$1.$2.$3";
1016 $consist =~ s/,/./g;
1017 $consist =~ s/\//./g;
1018 $consist =~ s/-/_/g;
1019 $consist =~ s/[^A-Za-z0-9_.]/ZZ/g; # ToDo: properly?
1021 # Using a cygnus-2.7-96q4 gcc build on hppas, the
1022 # consistency chunk for ghc_cc_ID often (but not always!)
1023 # gets lumped with a bunch of .IMPORT directives containing info on
1024 # the code or data space nature of external symbols. We can't
1025 # toss these, so once the consistency ID has been turned into
1026 # a representable symbol, we substitute it for the symbol
1027 # that the string was attached to in the first place (ghc_cc_ID.)
1028 # (The original string is also substituted away.)
1030 # This change may affect the code output on other platforms in
1031 # adverse ways, hence we restrict this hack hppa targets only.
1034 if ( $TargetPlatform =~ /^hppa/ ) {
1035 $chk[$i] =~ s/^${T_US}ghc.*c_ID$TPOSTLBL/$consist/o;
1036 $chk[$i] =~ s/\t$T_hsc_cc_PAT/$T_HDR_misc/o;
1037 $consist = $chk[$i]; #clumsily
1039 print OUTASM $T_HDR_consist, "${consist}${T_POST_LBL}\n";
1041 } elsif ( $TargetPlatform !~ /^(mips)-/ ) { # we just don't try in those case (ToDo)
1042 # on mips: consistency string is just a v
1043 # horrible bunch of .bytes,
1044 # which I am too lazy to sort out (WDP 95/05)
1046 print STDERR "Couldn't grok consistency: ", $chk[$i];
1049 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1050 # we can just re-constitute this one...
1051 # NB: we emit _three_ underscores no matter what,
1052 # so ghc-split doesn't have to care.
1053 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1055 } elsif ( $chkcat[$i] eq 'closure'
1056 || $chkcat[$i] eq 'srt'
1057 || $chkcat[$i] eq 'infotbl'
1058 || $chkcat[$i] eq 'entry') { # do them in that order
1059 $symb = $chksymb[$i];
1062 if ( defined($closurechk{$symb}) ) {
1063 print OUTASM $T_HDR_closure;
1064 print OUTASM $chk[$closurechk{$symb}];
1065 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1069 if ( defined($srtchk{$symb}) ) {
1070 print OUTASM $T_HDR_srt;
1071 print OUTASM $chk[$srtchk{$symb}];
1072 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1076 if ( defined($infochk{$symb}) ) {
1078 print OUTASM $T_HDR_info;
1079 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1081 # entry code will be put here!
1083 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1087 if ( defined($entrychk{$symb}) ) {
1089 $c = $chk[$entrychk{$symb}];
1091 print OUTASM $T_HDR_entry;
1093 &print_doctored($c, 1); # NB: the 1!!!
1095 $chkcat[$entrychk{$symb}] = 'DONE ALREADY';
1098 } elsif ( $chkcat[$i] eq 'vector'
1099 || $chkcat[$i] eq 'direct' ) { # do them in that order
1100 $symb = $chksymb[$i];
1103 if ( defined($vectorchk{$symb}) ) {
1104 print OUTASM $T_HDR_vector;
1105 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1107 # direct return code will be put here!
1108 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1110 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1111 # Alphas: the commented nop is for the splitter, to ensure
1112 # that no module ends with a label as the very last
1113 # thing. (The linker will adjust the label to point
1114 # to the first code word of the next module linked in,
1115 # even if alignment constraints cause the label to move!)
1117 print OUTASM "\t# nop\n";
1120 } elsif ( $chkcat[$i] eq 'toc' ) {
1121 # silly optimisation to print tocs, since they come in groups...
1122 print OUTASM $T_HDR_toc;
1124 while ($chkcat[$j] eq 'toc')
1125 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1128 print OUTASM $chk[$j];
1130 $chkcat[$j] = 'DONE ALREADY';
1134 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && $chkcat[$i] eq 'dyld' ) {
1135 # powerpc-apple: dynamic linker stubs
1136 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1137 { # print them out unchanged, but remove the stubs for __DISCARD__
1138 print OUTASM $chk[$i];
1141 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1145 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1148 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1149 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1154 sub hppa_mash_prologue { # OK, epilogue, too
1157 # toss all prologue stuff
1158 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1160 # Lie about our .CALLINFO
1161 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1168 # toss all epilogue stuff
1169 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1171 # Sorry; we moved the _info stuff to the code segment.
1172 s/_info,DATA/_info,CODE/g;
1179 sub print_doctored {
1180 local($_, $need_fallthru_patch) = @_;
1182 if ( $TargetPlatform !~ /^i386-/
1183 || ! /^\t[a-z]/ # no instructions in here, apparently
1184 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1188 # OK, must do some x86 **HACKING**
1190 local($entry_patch) = '';
1191 local($exit_patch) = '';
1193 # gotta watch out for weird instructions that
1194 # invisibly smash various regs:
1195 # rep* %ecx used for counting
1196 # scas* %edi used for destination index
1197 # cmps* %e[sd]i used for indices
1198 # loop* %ecx used for counting
1203 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1205 # * GCC used an "STG reg" for its own purposes
1207 # * some secret uses of machine reg, requiring STG reg
1208 # to be saved/restored
1210 # The most dangerous "GCC uses" of an "STG reg" are when
1211 # the reg holds the target of a jmp -- it's tricky to
1212 # insert the patch-up code before we get to the target!
1213 # So here we change the jmps:
1215 # --------------------------------------------------------
1216 # it can happen that we have jumps of the form...
1217 # jmp *<something involving %esp>
1219 # jmp <something involving another naughty register...>
1221 # a reasonably-common case is:
1223 # movl $_blah,<bad-reg>
1227 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[abcd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1229 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1230 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1231 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1232 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1233 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1234 if /(jmp|call)\s+.*\%esi/;
1236 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1237 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1238 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1239 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1240 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1241 if /(jmp|call)\s+.*\%edi/;
1244 # OK, now we can decide what our patch-up code is going to
1247 # Offsets into register table - you'd better update these magic
1248 # numbers should you change its contents!
1249 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1250 local($OFFSET_Hp)=88;
1252 # Note funky ".=" stuff; we're *adding* to these _patch guys
1253 if ( $StolenX86Regs <= 2
1254 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1255 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1256 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1258 # nothing for call_{entry,exit} because %esi is callee-save
1260 if ( $StolenX86Regs <= 3
1261 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1262 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1263 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1265 # nothing for call_{entry,exit} because %edi is callee-save
1268 # --------------------------------------------------------
1269 # next, here we go with non-%esp patching!
1271 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1273 # Before calling GC we must set up the exit condition before the call
1274 # and entry condition when we come back
1276 # fix _all_ non-local jumps:
1278 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1279 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1281 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1283 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1284 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1286 if ($StolenX86Regs == 2 ) {
1287 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1288 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1289 } elsif ($StolenX86Regs == 3 ) {
1290 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1291 if /^\t(jmp|call)\s+.*\%edi/;
1294 # --------------------------------------------------------
1295 # that's it -- print it
1297 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1301 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1302 print OUTASM $exit_patch;
1303 # ToDo: make it not print if there is a "jmp" at the end
1309 sub init_FUNNY_THINGS {
1310 %KNOWN_FUNNY_THING = (
1312 # "${T_US}stg_.*{T_POST_LBL}", 1,
1317 The following table reversal is used for both info tables and return
1318 vectors. In both cases, we remove the first entry from the table,
1319 reverse the table, put the label at the end, and paste some code
1320 (that which is normally referred to by the first entry in the table)
1321 right after the table itself. (The code pasting is done elsewhere.)
1325 local($symb, $tbl, $discard1) = @_;
1327 return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1329 local($before) = '';
1331 local(@imports) = (); # hppa only
1334 local(@lines) = split(/\n/, $tbl);
1337 # Deal with the header...
1338 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1339 $label .= $lines[$i] . "\n",
1340 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1341 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1342 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1344 $before .= $lines[$i] . "\n"; # otherwise...
1347 # Grab the table data...
1348 if ( $TargetPlatform !~ /^hppa/ ) {
1349 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1350 push(@words, $lines[$i]);
1352 } else { # hppa weirdness
1353 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1354 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1355 push(@imports, $lines[$i]);
1357 # We don't use HP's ``function pointers''
1358 # We just use labels in code space, like normal people
1359 $lines[$i] =~ s/P%//;
1360 push(@words, $lines[$i]);
1365 # Now throw away any initial zero word from the table. This is a hack
1366 # that lets us reduce the size of info tables when the SRT field is not
1367 # needed: see comments StgFunInfoTable in InfoTables.h.
1369 # The .zero business is for Linux/ELF.
1370 # The .skip business is for Sparc/Solaris/ELF.
1371 # The .blockz business is for HPPA.
1373 # if ($words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1378 for (; $i <= $#lines; $i++) {
1379 $after .= $lines[$i] . "\n";
1382 # Alphas: If we have anonymous text (not part of a procedure), the
1383 # linker may complain about missing exception information. Bleh.
1384 # To suppress this, we place a .ent/.end pair around the code.
1385 # At the same time, we have to be careful and not enclose any leading
1386 # .file/.loc directives.
1387 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1388 local ($ident) = $1;
1389 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1390 $after .= "\t.end $ident\n";
1393 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1394 # assembler (!) wherein .quad constants inside .text sections are
1395 # first narrowed to 32 bits then sign-extended back to 64 bits.
1396 # This obviously screws up our 64-bit bitmaps, so we work around
1397 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1398 if ( $TargetPlatform =~ /^alpha-/ ) {
1400 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1401 local ($number) = $1;
1402 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1403 local ($sign, $base, $digits) = ($1, $2, $3);
1404 $base = (10, 8, 16)[length $base];
1405 local ($hi, $lo) = (0, 0);
1406 foreach $i (split(//, $digits)) {
1407 $j = $lo * $base + $i;
1408 $lo = $j % 4294967296;
1409 $hi = $hi * $base + ($j - $lo) / 4294967296;
1411 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1413 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1414 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1416 print STDERR "Cannot handle \".quad $number\" in info table\n";
1424 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
1425 . join("\n", @words) . "\n"
1428 # print STDERR "before=$before\n";
1429 # print STDERR "label=$label\n";
1430 # print STDERR "words=",(reverse @words),"\n";
1431 # print STDERR "after=$after\n";
1437 The HP is a major nuisance. The threaded code mangler moved info
1438 tables from data space to code space, but unthreaded code in the RTS
1439 still has references to info tables in data space. Since the HP
1440 linker is very precise about where symbols live, we need to patch the
1441 references in the unthreaded RTS as well.
1444 sub mini_mangle_asm_hppa {
1445 local($in_asmf, $out_asmf) = @_;
1447 open(INASM, "< $in_asmf")
1448 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1449 open(OUTASM,"> $out_asmf")
1450 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1453 s/_info,DATA/_info,CODE/; # Move _info references to code space
1459 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1460 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1466 sub tidy_up_and_die {
1467 local($return_val, $msg) = @_;
1469 exit (($return_val == 0) ? 0 : 1);