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*(\.align\s+\d+|\.(globl|ent)\s+\S+|\#.*|\.(file|loc)\s+\S+\s+\S+|\.text|\.r?data)\n)';
100 $T_COPY_DIRVS = '^\s*(\#|\.(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_fast = "\.text\n\t\.align 3\n";
114 $T_HDR_vector = "\.text\n\t\.align 3\n";
115 $T_HDR_direct = "\.text\n\t\.align 3\n";
116 $T_create_word = "\t.quad";
118 #--------------------------------------------------------#
119 } elsif ( $TargetPlatform =~ /^hppa/ ) {
121 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
122 $T_US = ''; # _ if symbols have an underscore on the front
123 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
124 $T_CONST_LBL = '^L\$C(\d+)$'; # regexp for what such a lbl looks like
127 $T_MOVE_DIRVS = '^((\s+\.(IMPORT|EXPORT|PARAM).*|\s+\.align\s+\d+|\s+\.(SPACE|SUBSPA)\s+\S+|\s*)\n)';
128 $T_COPY_DIRVS = '^\s+\.(IMPORT|EXPORT)';
130 $T_hsc_cc_PAT = '\.STRING.*\)(hsc|cc) (.*)\\\\x09(.*)\\\\x00"';
131 $T_DOT_WORD = '\.(blockz|word|half|byte)';
132 $T_DOT_GLOBAL = '^\s+\.EXPORT';
133 $T_HDR_literal = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
134 $T_HDR_misc = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
135 $T_HDR_data = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
136 $T_HDR_consist = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
137 $T_HDR_closure = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
138 $T_HDR_srt = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
139 $T_HDR_info = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
140 $T_HDR_entry = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
141 $T_HDR_fast = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
142 $T_HDR_vector = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
143 $T_HDR_direct = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
144 $T_create_word = "\t.word";
146 #--------------------------------------------------------#
147 } elsif ( $TargetPlatform =~ /^i386-.*-(linuxaout|freebsd2|openbsd|nextstep3|cygwin32|mingw32)$/ ) {
148 # NeXT added but not tested. CaS
150 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
151 $T_US = '_'; # _ if symbols have an underscore on the front
152 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
153 $T_CONST_LBL = '^LC(\d+):$';
155 $T_X86_PRE_LLBL_PAT = 'L';
156 $T_X86_PRE_LLBL = 'L';
157 $T_X86_BADJMP = '^\tjmp [^L\*]';
159 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s+\d+(,0x90)?|\.globl\s+\S+|\.text|\.data|\.stab[^n].*|\.type\s+.*|\.size\s+.*)\n)';
160 $T_COPY_DIRVS = '\.(globl|stab)';
161 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
162 $T_DOT_WORD = '\.(long|word|value|byte|space)';
163 $T_DOT_GLOBAL = '\.globl';
164 $T_HDR_literal = "\.text\n\t\.align 2\n";
165 $T_HDR_misc = "\.text\n\t\.align 2,0x90\n";
166 $T_HDR_data = "\.data\n\t\.align 2\n";
167 $T_HDR_consist = "\.text\n";
168 $T_HDR_closure = "\.data\n\t\.align 2\n";
169 $T_HDR_closure = "\.data\n\t\.align 2\n\t.long 0\n" if ( $TargetPlatform =~ /.*-mingw32$/ );
170 $T_HDR_srt = "\.text\n\t\.align 2\n";
171 $T_HDR_info = "\.text\n\t\.align 2\n"; # NB: requires padding
172 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
173 $T_HDR_fast = "\.text\n\t\.align 2,0x90\n";
174 $T_HDR_vector = "\.text\n\t\.align 2\n"; # NB: requires padding
175 $T_HDR_direct = "\.text\n\t\.align 2,0x90\n";
176 $T_create_word = "\t.word";
178 #--------------------------------------------------------#
179 } elsif ( $TargetPlatform =~ /^i386-.*-(solaris2|linux|freebsd|netbsd)$/ ) {
181 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
182 $T_US = ''; # _ if symbols have an underscore on the front
183 $T_PRE_APP = # regexp that says what comes before APP/NO_APP
184 ($TargetPlatform =~ /-(linux|freebsd|netbsd)$/) ? '#' : '/' ;
185 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
187 $T_X86_PRE_LLBL_PAT = '\.L';
188 $T_X86_PRE_LLBL = '.L';
189 $T_X86_BADJMP = '^\tjmp\s+[^\.\*]';
191 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s+\d+(,\s*0x90)?|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.Lfe.*\n\s*\.size\s+.*|\.size\s+.*|\.ident.*)\n)';
192 $T_COPY_DIRVS = '\.(globl)';
194 if ( $TargetPlatform =~ /freebsd|netbsd/ ) {
195 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
197 $T_hsc_cc_PAT = '\.string.*\)(hsc|cc) (.*)\\\\t(.*)"';
200 $T_DOT_WORD = '\.(long|value|byte|zero)';
201 $T_DOT_GLOBAL = '\.globl';
202 $T_HDR_literal = "\.section\t\.rodata\n"; # or just use .text??? (WDP 95/11)
203 $T_HDR_misc = "\.text\n\t\.align 4\n";
204 $T_HDR_data = "\.data\n\t\.align 4\n"; # ToDo: change align??
205 $T_HDR_consist = "\.text\n";
206 $T_HDR_closure = "\.data\n\t\.align 4\n"; # ToDo: change align?
207 $T_HDR_srt = "\.text\n\t\.align 4\n"; # ToDo: change align?
208 $T_HDR_info = "\.text\n\t\.align 4\n"; # NB: requires padding
209 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
210 $T_HDR_fast = "\.text\n\t\.align 4\n";
211 $T_HDR_vector = "\.text\n\t\.align 4\n"; # NB: requires padding
212 $T_HDR_direct = "\.text\n\t\.align 4\n";
213 $T_create_word = "\t.word";
215 #--------------------------------------------------------#
216 } elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) {
218 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
219 $T_US = '_'; # _ if symbols have an underscore on the front
220 $T_PRE_APP = '^# MAY NOT APPLY'; # regexp that says what comes before APP/NO_APP
221 $T_CONST_LBL = '^LC(\d+):$';
224 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.proc\s+\d+|\.const|\.cstring|\.globl\s+\S+|\.text|\.data|\.even|\.stab[^n].*)\n)';
225 $T_COPY_DIRVS = '\.(globl|proc|stab)';
226 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
228 $T_DOT_WORD = '\.long';
229 $T_DOT_GLOBAL = '\.globl';
230 $T_HDR_literal = "\.text\n\t\.even\n";
231 $T_HDR_misc = "\.text\n\t\.even\n";
232 $T_HDR_data = "\.data\n\t\.even\n";
233 $T_HDR_consist = "\.text\n";
234 $T_HDR_closure = "\.data\n\t\.even\n";
235 $T_HDR_srt = "\.text\n\t\.even\n";
236 $T_HDR_info = "\.text\n\t\.even\n";
237 $T_HDR_entry = "\.text\n\t\.even\n";
238 $T_HDR_fast = "\.text\n\t\.even\n";
239 $T_HDR_vector = "\.text\n\t\.even\n";
240 $T_HDR_direct = "\.text\n\t\.even\n";
241 $T_create_word = "\t.long";
243 #--------------------------------------------------------#
244 } elsif ( $TargetPlatform =~ /^mips-.*/ ) {
246 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
247 $T_US = ''; # _ if symbols have an underscore on the front
248 $T_PRE_APP = '^\s*#'; # regexp that says what comes before APP/NO_APP
249 $T_CONST_LBL = '^\$LC(\d+):$'; # regexp for what such a lbl looks like
252 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.(globl|ent)\s+\S+|\.text|\.r?data)\n)';
253 $T_COPY_DIRVS = '\.(globl|ent)';
255 $T_hsc_cc_PAT = 'I WAS TOO LAZY TO DO THIS BIT (WDP 95/05)';
256 $T_DOT_WORD = '\.word';
257 $T_DOT_GLOBAL = '^\t\.globl';
258 $T_HDR_literal = "\t\.rdata\n\t\.align 2\n";
259 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
260 $T_HDR_data = "\t\.data\n\t\.align 2\n";
261 $T_HDR_consist = 'TOO LAZY TO DO THIS TOO';
262 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
263 $T_HDR_srt = "\t\.text\n\t\.align 2\n";
264 $T_HDR_info = "\t\.text\n\t\.align 2\n";
265 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
266 $T_HDR_fast = "\t\.text\n\t\.align 2\n";
267 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
268 $T_HDR_direct = "\t\.text\n\t\.align 2\n";
269 $T_create_word = "\t.word";
271 #--------------------------------------------------------#
272 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ ) {
273 # Apple PowerPC Darwin/MacOS X.
274 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
275 $T_US = '_'; # _ if symbols have an underscore on the front
276 $T_PRE_APP = 'WHAT IS THIS'; # regexp that says what comes before APP/NO_APP
277 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
280 $T_MOVE_DIRVS = '^(\s*(\.align \d+|\.text|\.data|\.cstring|\.non_lazy_symbol_pointer|\.const|\.static_const|\.literal4|\.literal8|\.static_data|\.globl \S+)\n)';
281 $T_COPY_DIRVS = '\.(globl)';
283 $T_hsc_cc_PAT = '\.byte.*\)(hsc|cc) (.*)"\n\t\.byte \d+\n\t\.byte "(.*)"\n\t\.byte \d+';
284 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
285 $T_DOT_GLOBAL = '\.globl';
286 $T_HDR_toc = "\.toc\n";
287 $T_HDR_literal = "\t\.const_data\n\t\.align 2\n";
288 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
289 $T_HDR_data = "\t\.data\n\t\.align 2\n";
290 $T_HDR_consist = "\t\.text\n\t\.align 2\n";
291 $T_HDR_closure = "\t\.const_data\n\t\.align 2\n";
292 $T_HDR_srt = "\t\.text\n\t\.align 2\n";
293 $T_HDR_info = "\t\.text\n\t\.align 2\n";
294 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
295 $T_HDR_fast = "\t\.text\n\t\.align 2\n";
296 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
297 $T_HDR_direct = "\t\.text\n\t\.align 2\n";
298 $T_create_word = "\t.long";
300 #--------------------------------------------------------#
301 } elsif ( $TargetPlatform =~ /^sparc-.*-solaris2/ ) {
303 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
304 $T_US = ''; # _ if symbols have an underscore on the front
305 $T_PRE_APP = 'DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
306 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
309 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*|\.section.*|\s+\.type.*|\s+\.size.*)\n)';
310 $T_COPY_DIRVS = '\.(global|proc|stab)';
312 $T_hsc_cc_PAT = '\.asciz.*\)(hsc|cc) (.*)\\\\t(.*)"';
313 $T_DOT_WORD = '\.(word|byte|half|skip|uahalf|uaword)';
314 $T_DOT_GLOBAL = '^\t\.global';
315 $T_HDR_literal = "\.text\n\t\.align 8\n";
316 $T_HDR_misc = "\.text\n\t\.align 4\n";
317 $T_HDR_data = "\.data\n\t\.align 8\n";
318 $T_HDR_consist = "\.text\n";
319 $T_HDR_closure = "\.data\n\t\.align 4\n";
320 $T_HDR_srt = "\.data\n\t\.align 4\n";
321 $T_HDR_info = "\.text\n\t\.align 4\n";
322 $T_HDR_entry = "\.text\n\t\.align 4\n";
323 $T_HDR_fast = "\.text\n\t\.align 4\n";
324 $T_HDR_vector = "\.text\n\t\.align 4\n";
325 $T_HDR_direct = "\.text\n\t\.align 4\n";
326 $T_create_word = "\t.word";
328 #--------------------------------------------------------#
329 } elsif ( $TargetPlatform =~ /^sparc-.*-sunos4/ ) {
331 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
332 $T_US = '_'; # _ if symbols have an underscore on the front
333 $T_PRE_APP = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
334 $T_CONST_LBL = '^LC(\d+):$';
337 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*)\n)';
338 $T_COPY_DIRVS = '\.(global|proc|stab)';
339 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
341 $T_DOT_WORD = '\.word';
342 $T_DOT_GLOBAL = '^\t\.global';
343 $T_HDR_literal = "\.text\n\t\.align 8\n";
344 $T_HDR_misc = "\.text\n\t\.align 4\n";
345 $T_HDR_data = "\.data\n\t\.align 8\n";
346 $T_HDR_consist = "\.text\n";
347 $T_HDR_closure = "\.data\n\t\.align 4\n";
348 $T_HDR_srt = "\.data\n\t\.align 4\n";
349 $T_HDR_info = "\.text\n\t\.align 4\n";
350 $T_HDR_entry = "\.text\n\t\.align 4\n";
351 $T_HDR_fast = "\.text\n\t\.align 4\n";
352 $T_HDR_vector = "\.text\n\t\.align 4\n";
353 $T_HDR_direct = "\.text\n\t\.align 4\n";
354 $T_create_word = "\t.word";
356 #--------------------------------------------------------#
358 print STDERR "$Pgm: don't know how to mangle assembly language for: $TargetPlatform\n";
363 print STDERR "T_STABBY: $T_STABBY\n";
364 print STDERR "T_US: $T_US\n";
365 print STDERR "T_PRE_APP: $T_PRE_APP\n";
366 print STDERR "T_CONST_LBL: $T_CONST_LBL\n";
367 print STDERR "T_POST_LBL: $T_POST_LBL\n";
368 if ( $TargetPlatform =~ /^i386-/ ) {
369 print STDERR "T_X86_PRE_LLBL_PAT: $T_X86_PRE_LLBL_PAT\n";
370 print STDERR "T_X86_PRE_LLBL: $T_X86_PRE_LLBL\n";
371 print STDERR "T_X86_BADJMP: $T_X86_BADJMP\n";
373 print STDERR "T_MOVE_DIRVS: $T_MOVE_DIRVS\n";
374 print STDERR "T_COPY_DIRVS: $T_COPY_DIRVS\n";
375 print STDERR "T_hsc_cc_PAT: $T_hsc_cc_PAT\n";
376 print STDERR "T_DOT_WORD: $T_DOT_WORD\n";
377 print STDERR "T_HDR_literal: $T_HDR_literal\n";
378 print STDERR "T_HDR_misc: $T_HDR_misc\n";
379 print STDERR "T_HDR_data: $T_HDR_data\n";
380 print STDERR "T_HDR_consist: $T_HDR_consist\n";
381 print STDERR "T_HDR_closure: $T_HDR_closure\n";
382 print STDERR "T_HDR_info: $T_HDR_info\n";
383 print STDERR "T_HDR_entry: $T_HDR_entry\n";
384 print STDERR "T_HDR_fast: $T_HDR_fast\n";
385 print STDERR "T_HDR_vector: $T_HDR_vector\n";
386 print STDERR "T_HDR_direct: $T_HDR_direct\n";
392 %************************************************************************
394 \subsection{Mangle away}
396 %************************************************************************
400 local($in_asmf, $out_asmf) = @_;
402 # multi-line regexp matching:
407 &init_TARGET_STUFF();
408 &init_FUNNY_THINGS();
410 open(INASM, "< $in_asmf")
411 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
412 open(OUTASM,"> $out_asmf")
413 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
415 # read whole file, divide into "chunks":
416 # record some info about what we've found...
418 @chk = (); # contents of the chunk
419 $numchks = 0; # number of them
420 @chkcat = (); # what category of thing in each chunk
421 @chksymb = (); # what symbol(base) is defined in this chunk
422 %slowchk = (); # ditto, its regular "slow" entry code
423 %fastchk = (); # ditto, fast entry code
424 %closurechk = (); # ditto, the (static) closure
425 %srtchk = (); # ditto, its SRT (for top-level things)
426 %infochk = (); # given a symbol base, say what chunk its info tbl is in
427 %vectorchk = (); # ditto, return vector table
428 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
430 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
433 tr/\r//d if $TargetPlatform =~ /-mingw32$/; # In case Perl doesn't convert line endings
434 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
435 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
436 next if /^\t\.def.*endef$/;
437 next if /${T_PRE_APP}(NO_)?APP/o;
438 next if /^;/ && $TargetPlatform =~ /^hppa/;
440 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips)-/;
442 if ( $TargetPlatform =~ /^mips-/
443 && /^\t\.(globl\S+\.text|comm\t)/ ) {
444 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
446 # As a temporary solution for compiling "foreign export" declarations,
447 # we use global variables to pass arguments from C to STG land.
448 # These declarations live in the .hc file and not in the generated C
449 # stub file, so we let them pass through here.
450 } elsif ( /^\t\.comm\t__fexp_.*$/ ) {
452 $chkcat[$i] = 'data';
455 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
456 # duplicated from the bottom
460 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
461 # Alphas: Local labels not to be confused with new chunks
464 # NB: all the rest start with a non-space
466 } elsif ( $TargetPlatform =~ /^mips-/
467 && /^\d+:/ ) { # a funny-looking very-local label
470 } elsif ( /$T_CONST_LBL/o ) {
472 $chkcat[$i] = 'literal';
475 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
477 $chkcat[$i] = 'splitmarker';
480 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
483 $chkcat[$i] = 'infotbl';
484 $chksymb[$i] = $symb;
486 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
488 $infochk{$symb} = $i;
490 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
492 $chkcat[$i] = 'slow';
497 } elsif ( /^${T_US}([A-Za-z0-9_]+)_fast\d*${T_POST_LBL}$/o ) {
499 $chkcat[$i] = 'fast';
504 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
506 $chkcat[$i] = 'closure';
509 $closurechk{$1} = $i;
511 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
518 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
520 $chkcat[$i] = 'data';
523 } elsif ( /^${T_US}ghc.*c_ID${T_POST_LBL}/o ) {
525 $chkcat[$i] = 'consist';
527 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
530 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
531 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
532 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
533 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
536 $chkcat[$i] = 'data';
539 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
544 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
545 # all CC_ symbols go in the data section...
547 $chkcat[$i] = 'data';
550 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
552 $chkcat[$i] = 'misc';
554 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
556 $chkcat[$i] = 'vector';
561 # As a temporary solution for compiling "foreign export" declarations,
562 # we use global variables to pass arguments from C to STG land.
563 # These declarations live in the .hc file and not in the generated C
564 # stub file, so we let them pass through here.
565 } elsif ( /^[\t ]+\.comm[\t ]+__fexp_.*$/ ) {
567 $chkcat[$i] = 'data';
570 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
571 && /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
572 # Some Solaris system headers contain function definitions (as
573 # opposed to mere prototypes), which end up in the .hc file when
574 # a Haskell module foreign imports the corresponding system
575 # functions (most notably stat()). We put them into the text
576 # segment. Note that this currently does not extend to function
577 # names starting with an underscore.
580 $chkcat[$i] = 'misc';
583 } elsif ( /^${T_US}[A-Za-z0-9_]/o
584 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
585 || ! /^L\$\d+$/ ) ) {
588 print STDERR "Funny global thing?: $_"
589 unless # $KNOWN_FUNNY_THING{$thing}
590 /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
591 || /^${T_US}__stg_.*${T_POST_LBL}$/o # more RTS internals
592 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
593 || /^${T_US}__stginit.*${T_POST_LBL}$/o # __stginit<module>
594 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
595 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
596 || /^_uname:/o; # x86/Solaris2
598 $chkcat[$i] = 'misc';
601 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.picsymbol_stub/ )
604 $chkcat[$i] = 'dyld';
606 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.lazy_symbol_pointer/ )
609 $chkcat[$i] = 'dyld';
611 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.non_lazy_symbol_pointer/ )
614 $chkcat[$i] = 'dyld';
616 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.data/ && $chkcat[$i] == 'dyld')
617 { # non_lazy_symbol_ptrs that point to local symbols
619 $chkcat[$i] = 'dyld';
621 } else { # simple line (duplicated at the top)
626 $numchks = $#chk + 1;
628 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
629 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
632 # the division into chunks is imperfect;
633 # we throw some things over the fence into the next
636 # also, there are things we would like to know
637 # about the whole module before we start spitting
640 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
641 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
643 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
645 # Alphas: NB: we start meddling at chunk 1, not chunk 0
646 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
647 # spits a ".quad 0" in after the very first ".rdata"; we
648 # detect this special case (tossing the ".quad 0")!
649 local($magic_rdata_seen) = 0;
651 # HPPAs, MIPSen: also start medding at chunk 1
653 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
654 $c = $chk[$i]; # convenience copy
656 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
658 # toss all prologue stuff; HPPA is pretty weird
660 $c = &mash_hppa_prologue($c) if $TargetPlatform =~ /^hppa/;
662 # be slightly paranoid to make sure there's
663 # nothing surprising in there
664 if ( $c =~ /--- BEGIN ---/ ) {
665 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
667 if ($TargetPlatform =~ /^i386-/) {
668 $p =~ s/^\tpushl\s+\%edi\n//;
669 $p =~ s/^\tpushl\s+\%esi\n//;
670 $p =~ s/^\tpushl\s+\%ebx\n//;
671 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
672 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-cygwin32/);
673 } elsif ($TargetPlatform =~ /^m68k-/) {
674 $p =~ s/^\tlink a6,#-?\d.*\n//;
675 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
676 # The above showed up in the asm code,
677 # so I added it here.
678 # I hope it's correct.
680 $p =~ s/^\tmovel d2,sp\@-\n//;
681 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
682 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
683 } elsif ($TargetPlatform =~ /^mips-/) {
684 # the .frame/.mask/.fmask that we use is the same
685 # as that produced by GCC for miniInterpret; this
686 # gives GDB some chance of figuring out what happened
687 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
688 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
689 $p =~ s/^\t\.(mask|fmask).*\n//g;
690 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
691 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
692 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
693 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
694 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
695 $p =~ s/__FRAME__/$FRAME/;
696 } elsif ($TargetPlatform =~ /^powerpc-apple-.*/) {
698 $pcrel_label =~ s/(.|\n)*^(L\d+\$pb):\n(.|\n)*/$2/ or $pcrel_label = "";
700 $p =~ s/^\tmflr r0\n//;
701 $p =~ s/^\tbl saveFP # f\d+\n//;
702 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
703 $p =~ s/^L\d+\$pb:\n//;
704 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
705 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
706 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
707 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
708 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
709 $p =~ s/^\tbcl 20,31,L\d+\$pb\n//;
710 $p =~ s/^L\d+\$pb:\n//;
711 $p =~ s/^\tmflr r31\n//;
713 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
716 # HWL HACK: dont die, just print a warning
717 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
718 die "Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
720 if ($TargetPlatform =~ /^powerpc-apple-.*/ && $pcrel_label ne "") {
721 # on PowerPC, we have to keep a part of the prologue
722 # (which loads the current instruction pointer into register r31)
723 $p .= "bcl 20,31,$pcrel_label\n";
724 $p .= "$pcrel_label:\n";
728 # glue together what's left
730 $c =~ s/\n\t\n/\n/; # junk blank line
734 if ( $TargetPlatform =~ /^mips-/ ) {
735 # MIPS: first, this basic sequence may occur "--- END ---" or not
736 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
739 # toss all epilogue stuff; again, paranoidly
740 if ( $c =~ /--- END ---/ ) {
741 if (($r, $e) = split(/--- END ---/, $c)) {
742 if ($TargetPlatform =~ /^i386-/) {
744 $e =~ s/^\tpopl\s+\%edi\n//;
745 $e =~ s/^\tpopl\s+\%esi\n//;
746 $e =~ s/^\tpopl\s+\%edx\n//;
747 $e =~ s/^\tpopl\s+\%ecx\n//;
748 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
749 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
750 } elsif ($TargetPlatform =~ /^m68k-/) {
751 $e =~ s/^\tunlk a6\n//;
753 } elsif ($TargetPlatform =~ /^mips-/) {
754 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
755 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
756 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
757 $e =~ s/^\tj\t\$31\n//;
758 } elsif ($TargetPlatform =~ /^powerpc-apple-.*/) {
759 $e =~ s/^\taddi r1,r1,\d+\n//;
760 $e =~ s/^\tcal r1,\d+\(r1\)\n//;
761 $e =~ s/^\tlw?z? r\d+,\d+\(r1\)\n//;
762 $e =~ s/^\tmtlr r0\n//;
765 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
768 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
770 # glue together what's left
772 $c =~ s/\n\t\n/\n/; # junk blank line
776 # On SPARCs, we don't do --- BEGIN/END ---, we just
777 # toss the register-windowing save/restore/ret* instructions
779 if ( $TargetPlatform =~ /^sparc-/ ) {
780 $c =~ s/^\t(save.*|restore|ret|retl)\n//g;
781 # throw away PROLOGUE comments
782 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
785 # On Alphas, the prologue mangling is done a little later (below)
787 # toss all calls to __DISCARD__
788 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
789 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-.*/;
791 # MIPS: that may leave some gratuitous asm macros around
792 # (no harm done; but we get rid of them to be tidier)
793 $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/
794 if $TargetPlatform =~ /^mips-/;
796 # toss stack adjustment after DoSparks
797 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
798 if $TargetPlatform =~ /^m68k-/; # this looks old...
800 if ( $TargetPlatform =~ /^alpha-/ &&
801 ! $magic_rdata_seen &&
802 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
803 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
804 $magic_rdata_seen = 1;
807 # pick some end-things and move them to the next chunk
809 # pin a funny end-thing on (for easier matching):
810 $c .= 'FUNNY#END#THING';
812 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
816 # on x86 we try not to copy any directives into a literal
817 # chunk, rather we keep looking for the next real chunk. This
818 # is because we get things like
820 # .globl blah_closure
826 if ( $TargetPlatform =~ /^(i386|sparc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
828 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
831 if ( $j < $numchks ) {
832 $chk[$j] = $to_move . $chk[$j];
836 elsif ( $i < ($numchks - 1)
837 && ( $to_move =~ /${T_COPY_DIRVS}/
838 || ($TargetPlatform =~ /^hppa/ && $to_move =~ /align/ && $chkcat[$i+1] eq 'literal') )) {
839 $chk[$i + 1] = $to_move . $chk[$i + 1];
840 # otherwise they're tossed
843 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
846 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
848 # toss all prologue stuff, except for loading gp, and the ..ng address
849 if (($p, $r) = split(/^\t\.prologue/, $c)) {
850 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
851 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
852 $keep =~ s/^\t\.(mask|fmask).*\n//g;
853 $c = $keep . "..ng:\n";
855 print STDERR "malformed code block ($ent)?\n"
858 $c .= "\t.prologue" . $r;
861 $c =~ s/FUNNY#END#THING//;
863 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
865 $chk[$i] = $c; # update w/ convenience copy
868 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
869 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
872 if ( $TargetPlatform =~ /^alpha-/ ) {
873 # print out the header stuff first
874 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
875 print OUTASM $chk[0];
877 } elsif ( $TargetPlatform =~ /^hppa/ ) {
878 print OUTASM $chk[0];
880 } elsif ( $TargetPlatform =~ /^mips-/ ) {
881 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
883 # get rid of horrible "<dollar>Revision: .*$" strings
884 local(@lines0) = split(/\n/, $chk[0]);
886 while ( $z <= $#lines0 ) {
887 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
890 while ( $z <= $#lines0 ) {
892 last if $lines0[$z] =~ /[,\t]0x0$/;
898 $chk[0] = join("\n", @lines0);
899 $chk[0] =~ s/\n\n+/\n/;
900 print OUTASM $chk[0];
903 # print out all the literal strings next
904 for ($i = 0; $i < $numchks; $i++) {
905 if ( $chkcat[$i] eq 'literal' ) {
906 print OUTASM $T_HDR_literal, $chk[$i];
907 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
909 $chkcat[$i] = 'DONE ALREADY';
913 # on the HPPA, print out all the bss next
914 if ( $TargetPlatform =~ /^hppa/ ) {
915 for ($i = 1; $i < $numchks; $i++) {
916 if ( $chkcat[$i] eq 'bss' ) {
917 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
918 print OUTASM $chk[$i];
920 $chkcat[$i] = 'DONE ALREADY';
925 for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) {
926 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
928 next if $chkcat[$i] eq 'DONE ALREADY';
930 if ( $chkcat[$i] eq 'misc' ) {
931 if ($chk[$i] ne '') {
932 print OUTASM $T_HDR_misc;
933 &print_doctored($chk[$i], 0);
936 } elsif ( $chkcat[$i] eq 'toss' ) {
937 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
939 } elsif ( $chkcat[$i] eq 'data' ) {
940 if ($chk[$i] ne '') {
941 print OUTASM $T_HDR_data;
942 print OUTASM $chk[$i];
945 } elsif ( $chkcat[$i] eq 'consist' ) {
946 if ( $chk[$i] =~ /$T_hsc_cc_PAT/o ) {
947 local($consist) = "$1.$2.$3";
949 $consist =~ s/\//./g;
951 $consist =~ s/[^A-Za-z0-9_.]/ZZ/g; # ToDo: properly?
953 # Using a cygnus-2.7-96q4 gcc build on hppas, the
954 # consistency chunk for ghc_cc_ID often (but not always!)
955 # gets lumped with a bunch of .IMPORT directives containing info on
956 # the code or data space nature of external symbols. We can't
957 # toss these, so once the consistency ID has been turned into
958 # a representable symbol, we substitute it for the symbol
959 # that the string was attached to in the first place (ghc_cc_ID.)
960 # (The original string is also substituted away.)
962 # This change may affect the code output on other platforms in
963 # adverse ways, hence we restrict this hack hppa targets only.
966 if ( $TargetPlatform =~ /^hppa/ ) {
967 $chk[$i] =~ s/^${T_US}ghc.*c_ID$TPOSTLBL/$consist/o;
968 $chk[$i] =~ s/\t$T_hsc_cc_PAT/$T_HDR_misc/o;
969 $consist = $chk[$i]; #clumsily
971 print OUTASM $T_HDR_consist, "${consist}${T_POST_LBL}\n";
973 } elsif ( $TargetPlatform !~ /^(mips)-/ ) { # we just don't try in those case (ToDo)
974 # on mips: consistency string is just a v
975 # horrible bunch of .bytes,
976 # which I am too lazy to sort out (WDP 95/05)
978 print STDERR "Couldn't grok consistency: ", $chk[$i];
981 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
982 # we can just re-constitute this one...
983 # NB: we emit _three_ underscores no matter what,
984 # so ghc-split doesn't have to care.
985 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
987 } elsif ( $chkcat[$i] eq 'closure'
988 || $chkcat[$i] eq 'srt'
989 || $chkcat[$i] eq 'infotbl'
990 || $chkcat[$i] eq 'slow'
991 || $chkcat[$i] eq 'fast' ) { # do them in that order
992 $symb = $chksymb[$i];
995 if ( defined($closurechk{$symb}) ) {
996 print OUTASM $T_HDR_closure;
997 print OUTASM $chk[$closurechk{$symb}];
998 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1002 if ( defined($srtchk{$symb}) ) {
1003 print OUTASM $T_HDR_srt;
1004 print OUTASM $chk[$srtchk{$symb}];
1005 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1009 if ( defined($infochk{$symb}) ) {
1011 print OUTASM $T_HDR_info;
1012 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1014 # entry code will be put here!
1016 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1020 if ( defined($slowchk{$symb}) ) {
1022 # teach it to drop through to the fast entry point:
1023 $c = $chk[$slowchk{$symb}];
1025 if ( defined($fastchk{$symb}) ) {
1026 if ( $TargetPlatform =~ /^alpha-/ ) {
1027 $c =~ s/^\tjmp \$31,\(\$27\),0\n\t\.align 4\n\t\.end/\t.align 4\n\t.end/;
1028 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1029 $c =~ s/^\s+ldil.*\n\s+ldo.*\n\s+bv.*\n(.*\n)?\s+\.EXIT/$1\t.EXIT/;
1030 } elsif ( $TargetPlatform =~ /^i386-/ ) {
1031 # Reg alloc depending, gcc generated code may jump to the fast entry point via
1032 # a number of registers.
1033 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%edx\n\tjmp\s+\*\%edx\n//;
1034 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%ecx\n\tjmp\s+\*\%ecx\n//;
1035 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%eax\n\tjmp\s+\*\%eax\n//;
1036 # The next two only apply if we're not stealing %esi or %edi.
1037 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%esi\n\tjmp\s+\*\%esi\n// if ($StolenX86Regs < 3);
1038 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%edi\n\tjmp\s+\*\%edi\n// if ($StolenX86Regs < 4);
1039 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1040 $c =~ s/^\tjmp \$31,\(\$27\),0\n\t\.align 4\n\t\.end/\t.align 4\n\t.end/;
1041 } elsif ( $TargetPlatform =~ /^m68k-/ ) {
1042 $c =~ s/^\tjmp ${T_US}${symb}_fast\d+.*\n\tnop\n//;
1043 $c =~ s/^\tjmp ${T_US}${symb}_fast\d+.*\n//;
1044 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ ) {
1045 $c =~ s/^\taddis r9,r31,ha16\(L_${symb}_fast\d*\$non_lazy_ptr-L\d+\$pb\)\n\tlwz r9,lo16\(L_${symb}_fast\d*\$non_lazy_ptr-L\d+\$pb\)\(r9\)\n\tmtctr r9\n\tbctr\n//;
1046 # addis r9,r31,ha16(L_XXXX_fast3$non_lazy_ptr-L32$pb)
1047 # lwz r9,lo16(L_XXXX_fast3$non_lazy_ptr-L32$pb)(r9)
1050 } elsif ( $TargetPlatform =~ /^sparc-/ ) {
1051 $c =~ s/^\tcall\s+${T_US}${symb}_fast\d+,.*\n\t\s*nop\n//;
1052 $c =~ s/^\tcall\s+${T_US}${symb}_fast\d+,.*\n(\t\s*[a-z].*\n)/$1/;
1054 print STDERR "$Pgm: mystery slow-fast dropthrough: $TargetPlatform\n";
1058 if ( $TargetPlatform !~ /^(alpha-|hppa|mips-|powerpc-apple-)/) {
1059 # On alphas, hppa: no very good way to look for "dangling"
1060 # references to fast-entry point.
1061 # (questionable re hppa and mips...)
1062 print STDERR "still has jump to fast entry point:\n$c"
1063 if $c =~ /\b${T_US}${symb}_fast/;
1066 print OUTASM $T_HDR_entry;
1068 &print_doctored($c, 1); # NB: the 1!!!
1070 $chkcat[$slowchk{$symb}] = 'DONE ALREADY';
1074 if ( defined($fastchk{$symb}) ) {
1075 $c = $chk[$fastchk{$symb}];
1076 if ( ! defined($slowchk{$symb})
1077 # ToDo: the || clause can go once we're no longer
1078 # concerned about producing exactly the same output as before
1079 #OLD: || $TargetPlatform =~ /^(m68k|sparc|i386)-/
1081 print OUTASM $T_HDR_fast;
1084 &print_doctored($c, 0);
1085 $chkcat[$fastchk{$symb}] = 'DONE ALREADY';
1088 } elsif ( $chkcat[$i] eq 'vector'
1089 || $chkcat[$i] eq 'direct' ) { # do them in that order
1090 $symb = $chksymb[$i];
1093 if ( defined($vectorchk{$symb}) ) {
1094 print OUTASM $T_HDR_vector;
1096 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1097 # DO NOT DELETE THE NEXT LINE. It fixes a rather subtle GC bug
1098 # which showed up as a segfault reported by Ryszard Kubiak.
1099 # Problem is with vector tables. They wind up as follows:
1101 # .word some-other-word
1103 # Problem is that we want the label fooble_vtbl to be considered
1104 # in the same section as the vtbl itself, but the label actually
1105 # lives at the next word along. If a data segment should happen
1106 # to immediately follow the vtbl, as it can in GHCi, the label will
1107 # be malclassified as in the data rather than text segment (during
1108 # GC), and so we will regard references to it as static closure
1109 # pointers rather than as code pointers, which is an error which
1110 # usually crashes the garbage collectors.
1111 # To fix this, we place a dummy word after the label, so as to
1112 # ensure that the label is in the same segment as the vtbl proper.
1113 # The native code generator has an analogous fix; see
1114 # ghc/compiler/nativeGen/AbsCStixGen.lhs line 107.
1115 print OUTASM "${T_create_word} 0\n";
1117 # direct return code will be put here!
1118 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1120 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1121 # Alphas: the commented nop is for the splitter, to ensure
1122 # that no module ends with a label as the very last
1123 # thing. (The linker will adjust the label to point
1124 # to the first code word of the next module linked in,
1125 # even if alignment constraints cause the label to move!)
1127 print OUTASM "\t# nop\n";
1130 } elsif ( $chkcat[$i] eq 'toc' ) {
1131 # silly optimisation to print tocs, since they come in groups...
1132 print OUTASM $T_HDR_toc;
1134 while ($chkcat[$j] eq 'toc')
1135 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1138 print OUTASM $chk[$j];
1140 $chkcat[$j] = 'DONE ALREADY';
1144 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && $chkcat[$i] eq 'dyld' ) {
1145 # powerpc-apple: dynamic linker stubs
1146 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1147 { # print them out unchanged, but remove the stubs for __DISCARD__
1148 print OUTASM $chk[$i];
1151 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1155 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1158 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1159 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1164 sub mash_hppa_prologue { # OK, epilogue, too
1167 # toss all prologue stuff
1168 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1170 # Lie about our .CALLINFO
1171 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1178 # toss all epilogue stuff
1179 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1181 # Sorry; we moved the _info stuff to the code segment.
1182 s/_info,DATA/_info,CODE/g;
1189 sub print_doctored {
1190 local($_, $need_fallthru_patch) = @_;
1192 if ( $TargetPlatform !~ /^i386-/
1193 || ! /^\t[a-z]/ # no instructions in here, apparently
1194 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1198 # OK, must do some x86 **HACKING**
1200 local($entry_patch) = '';
1201 local($exit_patch) = '';
1203 # gotta watch out for weird instructions that
1204 # invisibly smash various regs:
1205 # rep* %ecx used for counting
1206 # scas* %edi used for destination index
1207 # cmps* %e[sd]i used for indices
1208 # loop* %ecx used for counting
1213 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1215 # * GCC used an "STG reg" for its own purposes
1217 # * some secret uses of machine reg, requiring STG reg
1218 # to be saved/restored
1220 # The most dangerous "GCC uses" of an "STG reg" are when
1221 # the reg holds the target of a jmp -- it's tricky to
1222 # insert the patch-up code before we get to the target!
1223 # So here we change the jmps:
1225 # --------------------------------------------------------
1226 # it can happen that we have jumps of the form...
1227 # jmp *<something involving %esp>
1229 # jmp <something involving another naughty register...>
1231 # a reasonably-common case is:
1233 # movl $_blah,<bad-reg>
1237 s/^\tmovl\s+\$${T_US}(.*),(\%e[abcd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1239 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1240 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1241 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1242 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1243 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1244 if /(jmp|call)\s+.*\%esi/;
1246 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1247 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1248 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1249 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1250 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1251 if /(jmp|call)\s+.*\%edi/;
1254 # OK, now we can decide what our patch-up code is going to
1257 # Offsets into register table - you'd better update these magic
1258 # numbers should you change its contents!
1259 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1260 local($OFFSET_Hp)=92;
1262 # Note funky ".=" stuff; we're *adding* to these _patch guys
1263 if ( $StolenX86Regs <= 2
1264 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1265 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1266 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1268 # nothing for call_{entry,exit} because %esi is callee-save
1270 if ( $StolenX86Regs <= 3
1271 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1272 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1273 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1275 # nothing for call_{entry,exit} because %edi is callee-save
1278 # --------------------------------------------------------
1279 # next, here we go with non-%esp patching!
1281 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1283 # Before calling GC we must set up the exit condition before the call
1284 # and entry condition when we come back
1286 # fix _all_ non-local jumps:
1288 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1289 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1291 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1293 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1294 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1296 if ($StolenX86Regs == 2 ) {
1297 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1298 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1299 } elsif ($StolenX86Regs == 3 ) {
1300 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1301 if /^\t(jmp|call)\s+.*\%edi/;
1304 # --------------------------------------------------------
1305 # that's it -- print it
1307 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1311 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1312 print OUTASM $exit_patch;
1313 # ToDo: make it not print if there is a "jmp" at the end
1319 sub init_FUNNY_THINGS {
1320 %KNOWN_FUNNY_THING = (
1322 # "${T_US}stg_.*{T_POST_LBL}", 1,
1327 The following table reversal is used for both info tables and return
1328 vectors. In both cases, we remove the first entry from the table,
1329 reverse the table, put the label at the end, and paste some code
1330 (that which is normally referred to by the first entry in the table)
1331 right after the table itself. (The code pasting is done elsewhere.)
1335 local($symb, $tbl, $discard1) = @_;
1337 local($before) = '';
1339 local(@imports) = (); # hppa only
1342 local(@lines) = split(/\n/, $tbl);
1345 # Deal with the header...
1346 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1347 $label .= $lines[$i] . "\n",
1348 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1349 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1350 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1352 $before .= $lines[$i] . "\n"; # otherwise...
1355 # Grab the table data...
1356 if ( $TargetPlatform !~ /^hppa/ ) {
1357 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1358 push(@words, $lines[$i]);
1360 } else { # hppa weirdness
1361 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1362 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1363 push(@imports, $lines[$i]);
1365 # We don't use HP's ``function pointers''
1366 # We just use labels in code space, like normal people
1367 $lines[$i] =~ s/P%//;
1368 push(@words, $lines[$i]);
1373 # now throw away the first word (SRT) iff it is empty.
1374 # The .zero business is for Linux/ELF.
1375 # The .skip business is for Sparc/Solaris/ELF.
1376 # The .blockz business is for HPPA.
1377 if ($discard1 && $words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1381 for (; $i <= $#lines; $i++) {
1382 $after .= $lines[$i] . "\n";
1385 # Alphas: If we have anonymous text (not part of a procedure), the
1386 # linker may complain about missing exception information. Bleh.
1387 # To suppress this, we place a .ent/.end pair around the code.
1388 # At the same time, we have to be careful and not enclose any leading
1389 # .file/.loc directives.
1390 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1391 local ($ident) = $1;
1392 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1393 $after .= "\t.end $ident\n";
1396 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1397 # assembler (!) wherein .quad constants inside .text sections are
1398 # first narrowed to 32 bits then sign-extended back to 64 bits.
1399 # This obviously screws up our 64-bit bitmaps, so we work around
1400 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1401 if ( $TargetPlatform =~ /^alpha-/ ) {
1403 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1404 local ($number) = $1;
1405 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1406 local ($sign, $base, $digits) = ($1, $2, $3);
1407 $base = (10, 8, 16)[length $base];
1408 local ($hi, $lo) = (0, 0);
1409 foreach $i (split(//, $digits)) {
1410 $j = $lo * $base + $i;
1411 $lo = $j % 4294967296;
1412 $hi = $hi * $base + ($j - $lo) / 4294967296;
1414 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1416 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1417 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1419 print STDERR "Cannot handle \".quad $number\" in info table\n";
1427 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
1428 . join("\n", @words) . "\n"
1431 # print STDERR "before=$before\n";
1432 # print STDERR "label=$label\n";
1433 # print STDERR "words=",(reverse @words),"\n";
1434 # print STDERR "after=$after\n";
1440 The HP is a major nuisance. The threaded code mangler moved info
1441 tables from data space to code space, but unthreaded code in the RTS
1442 still has references to info tables in data space. Since the HP
1443 linker is very precise about where symbols live, we need to patch the
1444 references in the unthreaded RTS as well.
1447 sub mini_mangle_asm_hppa {
1448 local($in_asmf, $out_asmf) = @_;
1450 open(INASM, "< $in_asmf")
1451 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1452 open(OUTASM,"> $out_asmf")
1453 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1456 s/_info,DATA/_info,CODE/; # Move _info references to code space
1462 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1463 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1469 sub tidy_up_and_die {
1470 local($return_val, $msg) = @_;
1472 exit (($return_val == 0) ? 0 : 1);