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 ( $TargetPlatform =~ /^ia64-.*-linux$/ ) {
218 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
219 $T_US = ''; # _ if symbols have an underscore on the front
221 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
224 $T_MOVE_DIRVS = '^(\s*\.(global|proc|pred\.safe_across_calls|text|data|section|subsection|align|size|type|ident)\s+.*\n)';
225 $T_COPY_DIRVS = '\.(global|proc)';
227 $T_hsc_cc_PAT = '\.string.*\)(hsc|cc) (.*)\\\\t(.*)"';
228 $T_DOT_WORD = '\.(long|value|byte|zero)';
229 $T_DOT_GLOBAL = '\.global';
230 $T_HDR_literal = "\.section\t\.rodata\n";
231 $T_HDR_misc = "\.text\n\t\.align 8\n";
232 $T_HDR_data = "\.data\n\t\.align 8\n";
233 $T_HDR_consist = "\.text\n";
234 $T_HDR_closure = "\.data\n\t\.align 8\n";
235 $T_HDR_srt = "\.text\n\t\.align 8\n";
236 $T_HDR_info = "\.text\n\t\.align 8\n";
237 $T_HDR_entry = "\.text\n\t\.align 16\n";
238 $T_HDR_fast = "\.text\n\t\.align 16\n";
239 $T_HDR_vector = "\.text\n\t\.align 8\n";
240 $T_HDR_direct = "\.text\n\t\.align 8\n";
241 $T_create_word = "\t.word";
243 #--------------------------------------------------------#
244 } elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) {
246 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
247 $T_US = '_'; # _ if symbols have an underscore on the front
248 $T_PRE_APP = '^# MAY NOT APPLY'; # regexp that says what comes before APP/NO_APP
249 $T_CONST_LBL = '^LC(\d+):$';
252 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.proc\s+\d+|\.const|\.cstring|\.globl\s+\S+|\.text|\.data|\.even|\.stab[^n].*)\n)';
253 $T_COPY_DIRVS = '\.(globl|proc|stab)';
254 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
256 $T_DOT_WORD = '\.long';
257 $T_DOT_GLOBAL = '\.globl';
258 $T_HDR_literal = "\.text\n\t\.even\n";
259 $T_HDR_misc = "\.text\n\t\.even\n";
260 $T_HDR_data = "\.data\n\t\.even\n";
261 $T_HDR_consist = "\.text\n";
262 $T_HDR_closure = "\.data\n\t\.even\n";
263 $T_HDR_srt = "\.text\n\t\.even\n";
264 $T_HDR_info = "\.text\n\t\.even\n";
265 $T_HDR_entry = "\.text\n\t\.even\n";
266 $T_HDR_fast = "\.text\n\t\.even\n";
267 $T_HDR_vector = "\.text\n\t\.even\n";
268 $T_HDR_direct = "\.text\n\t\.even\n";
269 $T_create_word = "\t.long";
271 #--------------------------------------------------------#
272 } elsif ( $TargetPlatform =~ /^mips-.*/ ) {
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 = '^\s*#'; # 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\s+\d+|\.(globl|ent)\s+\S+|\.text|\.r?data)\n)';
281 $T_COPY_DIRVS = '\.(globl|ent)';
283 $T_hsc_cc_PAT = 'I WAS TOO LAZY TO DO THIS BIT (WDP 95/05)';
284 $T_DOT_WORD = '\.word';
285 $T_DOT_GLOBAL = '^\t\.globl';
286 $T_HDR_literal = "\t\.rdata\n\t\.align 2\n";
287 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
288 $T_HDR_data = "\t\.data\n\t\.align 2\n";
289 $T_HDR_consist = 'TOO LAZY TO DO THIS TOO';
290 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
291 $T_HDR_srt = "\t\.text\n\t\.align 2\n";
292 $T_HDR_info = "\t\.text\n\t\.align 2\n";
293 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
294 $T_HDR_fast = "\t\.text\n\t\.align 2\n";
295 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
296 $T_HDR_direct = "\t\.text\n\t\.align 2\n";
297 $T_create_word = "\t.word";
299 #--------------------------------------------------------#
300 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ ) {
301 # Apple PowerPC Darwin/MacOS X.
302 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
303 $T_US = '_'; # _ if symbols have an underscore on the front
304 $T_PRE_APP = 'WHAT IS THIS'; # regexp that says what comes before APP/NO_APP
305 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
308 $T_MOVE_DIRVS = '^(\s*(\.align \d+|\.text|\.data|\.cstring|\.non_lazy_symbol_pointer|\.const|\.static_const|\.literal4|\.literal8|\.static_data|\.globl \S+)\n)';
309 $T_COPY_DIRVS = '\.(globl)';
311 $T_hsc_cc_PAT = '\.byte.*\)(hsc|cc) (.*)"\n\t\.byte \d+\n\t\.byte "(.*)"\n\t\.byte \d+';
312 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
313 $T_DOT_GLOBAL = '\.globl';
314 $T_HDR_toc = "\.toc\n";
315 $T_HDR_literal = "\t\.const_data\n\t\.align 2\n";
316 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
317 $T_HDR_data = "\t\.data\n\t\.align 2\n";
318 $T_HDR_consist = "\t\.text\n\t\.align 2\n";
319 $T_HDR_closure = "\t\.const_data\n\t\.align 2\n";
320 $T_HDR_srt = "\t\.text\n\t\.align 2\n";
321 $T_HDR_info = "\t\.text\n\t\.align 2\n";
322 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
323 $T_HDR_fast = "\t\.text\n\t\.align 2\n";
324 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
325 $T_HDR_direct = "\t\.text\n\t\.align 2\n";
326 $T_create_word = "\t.long";
328 #--------------------------------------------------------#
329 } elsif ( $TargetPlatform =~ /^sparc-.*-solaris2/ ) {
331 $T_STABBY = 0; # 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 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
337 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*|\.section.*|\s+\.type.*|\s+\.size.*)\n)';
338 $T_COPY_DIRVS = '\.(global|proc|stab)';
340 $T_hsc_cc_PAT = '\.asciz.*\)(hsc|cc) (.*)\\\\t(.*)"';
341 $T_DOT_WORD = '\.(long|word|byte|half|skip|uahalf|uaword)';
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 #--------------------------------------------------------#
357 } elsif ( $TargetPlatform =~ /^sparc-.*-sunos4/ ) {
359 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
360 $T_US = '_'; # _ if symbols have an underscore on the front
361 $T_PRE_APP = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
362 $T_CONST_LBL = '^LC(\d+):$';
365 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*)\n)';
366 $T_COPY_DIRVS = '\.(global|proc|stab)';
367 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
369 $T_DOT_WORD = '\.word';
370 $T_DOT_GLOBAL = '^\t\.global';
371 $T_HDR_literal = "\.text\n\t\.align 8\n";
372 $T_HDR_misc = "\.text\n\t\.align 4\n";
373 $T_HDR_data = "\.data\n\t\.align 8\n";
374 $T_HDR_consist = "\.text\n";
375 $T_HDR_closure = "\.data\n\t\.align 4\n";
376 $T_HDR_srt = "\.data\n\t\.align 4\n";
377 $T_HDR_info = "\.text\n\t\.align 4\n";
378 $T_HDR_entry = "\.text\n\t\.align 4\n";
379 $T_HDR_fast = "\.text\n\t\.align 4\n";
380 $T_HDR_vector = "\.text\n\t\.align 4\n";
381 $T_HDR_direct = "\.text\n\t\.align 4\n";
382 $T_create_word = "\t.word";
384 #--------------------------------------------------------#
386 print STDERR "$Pgm: don't know how to mangle assembly language for: $TargetPlatform\n";
391 print STDERR "T_STABBY: $T_STABBY\n";
392 print STDERR "T_US: $T_US\n";
393 print STDERR "T_PRE_APP: $T_PRE_APP\n";
394 print STDERR "T_CONST_LBL: $T_CONST_LBL\n";
395 print STDERR "T_POST_LBL: $T_POST_LBL\n";
396 if ( $TargetPlatform =~ /^i386-/ ) {
397 print STDERR "T_X86_PRE_LLBL_PAT: $T_X86_PRE_LLBL_PAT\n";
398 print STDERR "T_X86_PRE_LLBL: $T_X86_PRE_LLBL\n";
399 print STDERR "T_X86_BADJMP: $T_X86_BADJMP\n";
401 print STDERR "T_MOVE_DIRVS: $T_MOVE_DIRVS\n";
402 print STDERR "T_COPY_DIRVS: $T_COPY_DIRVS\n";
403 print STDERR "T_hsc_cc_PAT: $T_hsc_cc_PAT\n";
404 print STDERR "T_DOT_WORD: $T_DOT_WORD\n";
405 print STDERR "T_HDR_literal: $T_HDR_literal\n";
406 print STDERR "T_HDR_misc: $T_HDR_misc\n";
407 print STDERR "T_HDR_data: $T_HDR_data\n";
408 print STDERR "T_HDR_consist: $T_HDR_consist\n";
409 print STDERR "T_HDR_closure: $T_HDR_closure\n";
410 print STDERR "T_HDR_info: $T_HDR_info\n";
411 print STDERR "T_HDR_entry: $T_HDR_entry\n";
412 print STDERR "T_HDR_fast: $T_HDR_fast\n";
413 print STDERR "T_HDR_vector: $T_HDR_vector\n";
414 print STDERR "T_HDR_direct: $T_HDR_direct\n";
420 %************************************************************************
422 \subsection{Mangle away}
424 %************************************************************************
428 local($in_asmf, $out_asmf) = @_;
430 # multi-line regexp matching:
435 &init_TARGET_STUFF();
436 &init_FUNNY_THINGS();
438 open(INASM, "< $in_asmf")
439 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
440 open(OUTASM,"> $out_asmf")
441 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
443 # read whole file, divide into "chunks":
444 # record some info about what we've found...
446 @chk = (); # contents of the chunk
447 $numchks = 0; # number of them
448 @chkcat = (); # what category of thing in each chunk
449 @chksymb = (); # what symbol(base) is defined in this chunk
450 %slowchk = (); # ditto, its regular "slow" entry code
451 %fastchk = (); # ditto, fast entry code
452 %closurechk = (); # ditto, the (static) closure
453 %srtchk = (); # ditto, its SRT (for top-level things)
454 %infochk = (); # given a symbol base, say what chunk its info tbl is in
455 %vectorchk = (); # ditto, return vector table
456 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
458 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
461 tr/\r//d if $TargetPlatform =~ /-mingw32$/; # In case Perl doesn't convert line endings
462 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
463 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
464 next if /^\t\.def.*endef$/;
465 next if /${T_PRE_APP}(NO_)?APP/o;
466 next if /^;/ && $TargetPlatform =~ /^hppa/;
468 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|ia64)-/;
470 if ( $TargetPlatform =~ /^mips-/
471 && /^\t\.(globl\S+\.text|comm\t)/ ) {
472 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
474 # As a temporary solution for compiling "foreign export" declarations,
475 # we use global variables to pass arguments from C to STG land.
476 # These declarations live in the .hc file and not in the generated C
477 # stub file, so we let them pass through here.
478 } elsif ( /^\t\.comm\t__fexp_.*$/ ) {
480 $chkcat[$i] = 'data';
483 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
484 # duplicated from the bottom
488 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
489 # Alphas: Local labels not to be confused with new chunks
492 # NB: all the rest start with a non-space
494 } elsif ( $TargetPlatform =~ /^mips-/
495 && /^\d+:/ ) { # a funny-looking very-local label
498 } elsif ( /$T_CONST_LBL/o ) {
500 $chkcat[$i] = 'literal';
503 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
505 $chkcat[$i] = 'splitmarker';
508 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
511 $chkcat[$i] = 'infotbl';
512 $chksymb[$i] = $symb;
514 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
516 $infochk{$symb} = $i;
518 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
520 $chkcat[$i] = 'slow';
525 } elsif ( /^${T_US}([A-Za-z0-9_]+)_fast\d*${T_POST_LBL}$/o ) {
527 $chkcat[$i] = 'fast';
532 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
534 $chkcat[$i] = 'closure';
537 $closurechk{$1} = $i;
539 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
546 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
548 $chkcat[$i] = 'data';
551 } elsif ( /^${T_US}ghc.*c_ID${T_POST_LBL}/o ) {
553 $chkcat[$i] = 'consist';
555 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
558 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
559 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
560 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
561 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
564 $chkcat[$i] = 'data';
567 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
572 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
573 # all CC_ symbols go in the data section...
575 $chkcat[$i] = 'data';
578 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
580 $chkcat[$i] = 'misc';
582 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
584 $chkcat[$i] = 'vector';
589 # As a temporary solution for compiling "foreign export" declarations,
590 # we use global variables to pass arguments from C to STG land.
591 # These declarations live in the .hc file and not in the generated C
592 # stub file, so we let them pass through here.
593 } elsif ( /^[\t ]+\.comm[\t ]+__fexp_.*$/ ) {
595 $chkcat[$i] = 'data';
598 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
599 && /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
600 # Some Solaris system headers contain function definitions (as
601 # opposed to mere prototypes), which end up in the .hc file when
602 # a Haskell module foreign imports the corresponding system
603 # functions (most notably stat()). We put them into the text
604 # segment. Note that this currently does not extend to function
605 # names starting with an underscore.
608 $chkcat[$i] = 'misc';
611 } elsif ( /^${T_US}[A-Za-z0-9_]/o
612 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
613 || ! /^L\$\d+$/ ) ) {
616 print STDERR "Warning: retaining unknown function `$_' in output from C compiler"
617 unless # $KNOWN_FUNNY_THING{$thing}
618 /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
619 || /^${T_US}__stg_.*${T_POST_LBL}$/o # more RTS internals
620 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
621 || /^${T_US}__stginit.*${T_POST_LBL}$/o # __stginit<module>
622 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
623 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
624 || /^_uname:/o; # x86/Solaris2
626 $chkcat[$i] = 'misc';
629 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.picsymbol_stub/ )
632 $chkcat[$i] = 'dyld';
634 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.lazy_symbol_pointer/ )
637 $chkcat[$i] = 'dyld';
639 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.non_lazy_symbol_pointer/ )
642 $chkcat[$i] = 'dyld';
644 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.data/ && $chkcat[$i] == 'dyld')
645 { # non_lazy_symbol_ptrs that point to local symbols
647 $chkcat[$i] = 'dyld';
649 } else { # simple line (duplicated at the top)
654 $numchks = $#chk + 1;
656 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
657 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
660 # the division into chunks is imperfect;
661 # we throw some things over the fence into the next
664 # also, there are things we would like to know
665 # about the whole module before we start spitting
668 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
669 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
671 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
673 # Alphas: NB: we start meddling at chunk 1, not chunk 0
674 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
675 # spits a ".quad 0" in after the very first ".rdata"; we
676 # detect this special case (tossing the ".quad 0")!
677 local($magic_rdata_seen) = 0;
679 # HPPAs, MIPSen: also start medding at chunk 1
681 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
682 $c = $chk[$i]; # convenience copy
684 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
686 # toss all prologue stuff; HPPA is pretty weird
688 $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
690 # be slightly paranoid to make sure there's
691 # nothing surprising in there
692 if ( $c =~ /--- BEGIN ---/ ) {
693 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
695 # remove junk whitespace around the split point
699 if ($TargetPlatform =~ /^i386-/) {
700 $p =~ s/^\tpushl\s+\%edi\n//;
701 $p =~ s/^\tpushl\s+\%esi\n//;
702 $p =~ s/^\tpushl\s+\%ebx\n//;
703 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
704 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-cygwin32/);
706 # GCC 3.1 is in the habit of adding spurious writes to the
707 # stack in the prologue. Just to be on the safe side,
708 # chuck these over the fence into the main code.
709 while ($p =~ /^\tmovl\s+(\$\d+|\%edi),\s*\d*\(\%esp\)\n/) {
710 # print "Spurious instruction: $&";
715 } elsif ($TargetPlatform =~ /^ia64-/) {
716 $p =~ s/^\t\.prologue .*\n//;
717 $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, 31, \d+, 0\n//;
718 $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
719 $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
720 $p =~ s/^\t\.(mii|mmi)\n//; # bundling is no longer sensible
721 $p =~ s/^\t;;\n//g; # discard stops
722 $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
723 } elsif ($TargetPlatform =~ /^m68k-/) {
724 $p =~ s/^\tlink a6,#-?\d.*\n//;
725 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
726 # The above showed up in the asm code,
727 # so I added it here.
728 # I hope it's correct.
730 $p =~ s/^\tmovel d2,sp\@-\n//;
731 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
732 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
733 } elsif ($TargetPlatform =~ /^mips-/) {
734 # the .frame/.mask/.fmask that we use is the same
735 # as that produced by GCC for miniInterpret; this
736 # gives GDB some chance of figuring out what happened
737 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
738 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
739 $p =~ s/^\t\.(mask|fmask).*\n//g;
740 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
741 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
742 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
743 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
744 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
745 $p =~ s/__FRAME__/$FRAME/;
746 } elsif ($TargetPlatform =~ /^powerpc-apple-.*/) {
748 $pcrel_label =~ s/(.|\n)*^(L\d+\$pb):\n(.|\n)*/$2/ or $pcrel_label = "";
750 $p =~ s/^\tmflr r0\n//;
751 $p =~ s/^\tbl saveFP # f\d+\n//;
752 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
753 $p =~ s/^L\d+\$pb:\n//;
754 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
755 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
756 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
757 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
758 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
759 $p =~ s/^\tbcl 20,31,L\d+\$pb\n//;
760 $p =~ s/^L\d+\$pb:\n//;
761 $p =~ s/^\tmflr r31\n//;
763 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
766 # HWL HACK: dont die, just print a warning
767 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
768 die "Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
770 if ($TargetPlatform =~ /^powerpc-apple-.*/ && $pcrel_label ne "") {
771 # on PowerPC, we have to keep a part of the prologue
772 # (which loads the current instruction pointer into register r31)
773 $p .= "bcl 20,31,$pcrel_label\n";
774 $p .= "$pcrel_label:\n";
775 $p .= "\tmflr r31\n";
778 # glue together what's left
783 if ( $TargetPlatform =~ /^mips-/ ) {
784 # MIPS: first, this basic sequence may occur "--- END ---" or not
785 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
788 # toss all epilogue stuff; again, paranoidly
789 if ( $c =~ /--- END ---/ ) {
790 if (($r, $e) = split(/--- END ---/, $c)) {
791 if ($TargetPlatform =~ /^i386-/) {
793 $e =~ s/^\tpopl\s+\%edi\n//;
794 $e =~ s/^\tpopl\s+\%esi\n//;
795 $e =~ s/^\tpopl\s+\%edx\n//;
796 $e =~ s/^\tpopl\s+\%ecx\n//;
797 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
798 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
799 } elsif ($TargetPlatform =~ /^ia64-/) {
800 $e =~ s/^\tmov ar\.pfs = r\d+\n//;
801 $e =~ s/^\tmov b0 = r\d+\n//;
802 $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
803 $e =~ s/^\tbr\.ret\.sptk\.many b0\n//;
804 $e =~ s/^\t\.(mii|mmi|mib)\n//g; # bundling is no longer sensible
805 $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
806 $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
807 } elsif ($TargetPlatform =~ /^m68k-/) {
808 $e =~ s/^\tunlk a6\n//;
810 } elsif ($TargetPlatform =~ /^mips-/) {
811 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
812 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
813 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
814 $e =~ s/^\tj\t\$31\n//;
815 } elsif ($TargetPlatform =~ /^powerpc-apple-.*/) {
816 $e =~ s/^\taddi r1,r1,\d+\n//;
817 $e =~ s/^\tcal r1,\d+\(r1\)\n//;
818 $e =~ s/^\tlw?z? r\d+,\d+\(r1\)\n//;
819 $e =~ s/^\tmtlr r0\n//;
822 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
825 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
827 # glue together what's left
829 $c =~ s/\n\t\n/\n/; # junk blank line
833 # On SPARCs, we don't do --- BEGIN/END ---, we just
834 # toss the register-windowing save/restore/ret* instructions
836 if ( $TargetPlatform =~ /^sparc-/ ) {
837 $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
838 # throw away PROLOGUE comments
839 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
842 # On Alphas, the prologue mangling is done a little later (below)
844 # toss all calls to __DISCARD__
845 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
846 $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
847 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-.*/;
849 # IA64: mangle tailcalls into jumps here
850 $c =~ s/^\tbr\.call\.sptk\.many b0 = (.*)\n\t;;\n(\tmov r1 = r\d+\n)?\t;;\n\t--- TAILCALL ---\n/\tbr\.few $1\n/g
851 if $TargetPlatform =~ /^ia64-/;
853 # MIPS: that may leave some gratuitous asm macros around
854 # (no harm done; but we get rid of them to be tidier)
855 $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/
856 if $TargetPlatform =~ /^mips-/;
858 # toss stack adjustment after DoSparks
859 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
860 if $TargetPlatform =~ /^m68k-/; # this looks old...
862 if ( $TargetPlatform =~ /^alpha-/ &&
863 ! $magic_rdata_seen &&
864 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
865 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
866 $magic_rdata_seen = 1;
869 # pick some end-things and move them to the next chunk
871 # pin a funny end-thing on (for easier matching):
872 $c .= 'FUNNY#END#THING';
874 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
878 # on x86 we try not to copy any directives into a literal
879 # chunk, rather we keep looking for the next real chunk. This
880 # is because we get things like
882 # .globl blah_closure
888 if ( $TargetPlatform =~ /^(i386|sparc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
890 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
893 if ( $j < $numchks ) {
894 $chk[$j] = $to_move . $chk[$j];
898 elsif ( $i < ($numchks - 1)
899 && ( $to_move =~ /${T_COPY_DIRVS}/
900 || ($TargetPlatform =~ /^hppa/ && $to_move =~ /align/ && $chkcat[$i+1] eq 'literal') )) {
901 $chk[$i + 1] = $to_move . $chk[$i + 1];
902 # otherwise they're tossed
905 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
908 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
910 # toss all prologue stuff, except for loading gp, and the ..ng address
911 if (($p, $r) = split(/^\t\.prologue/, $c)) {
912 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
913 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
914 $keep =~ s/^\t\.(mask|fmask).*\n//g;
915 $c = $keep . "..ng:\n";
917 print STDERR "malformed code block ($ent)?\n"
920 $c .= "\t.prologue" . $r;
923 $c =~ s/FUNNY#END#THING//;
925 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
927 $chk[$i] = $c; # update w/ convenience copy
930 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
931 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
934 if ( $TargetPlatform =~ /^alpha-/ ) {
935 # print out the header stuff first
936 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
937 print OUTASM $chk[0];
939 } elsif ( $TargetPlatform =~ /^hppa/ ) {
940 print OUTASM $chk[0];
942 } elsif ( $TargetPlatform =~ /^mips-/ ) {
943 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
945 # get rid of horrible "<dollar>Revision: .*$" strings
946 local(@lines0) = split(/\n/, $chk[0]);
948 while ( $z <= $#lines0 ) {
949 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
952 while ( $z <= $#lines0 ) {
954 last if $lines0[$z] =~ /[,\t]0x0$/;
960 $chk[0] = join("\n", @lines0);
961 $chk[0] =~ s/\n\n+/\n/;
962 print OUTASM $chk[0];
965 # print out all the literal strings next
966 for ($i = 0; $i < $numchks; $i++) {
967 if ( $chkcat[$i] eq 'literal' ) {
968 print OUTASM $T_HDR_literal, $chk[$i];
969 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
971 $chkcat[$i] = 'DONE ALREADY';
975 # on the HPPA, print out all the bss next
976 if ( $TargetPlatform =~ /^hppa/ ) {
977 for ($i = 1; $i < $numchks; $i++) {
978 if ( $chkcat[$i] eq 'bss' ) {
979 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
980 print OUTASM $chk[$i];
982 $chkcat[$i] = 'DONE ALREADY';
987 for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) {
988 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
990 next if $chkcat[$i] eq 'DONE ALREADY';
992 if ( $chkcat[$i] eq 'misc' ) {
993 if ($chk[$i] ne '') {
994 print OUTASM $T_HDR_misc;
995 &print_doctored($chk[$i], 0);
998 } elsif ( $chkcat[$i] eq 'toss' ) {
999 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
1001 } elsif ( $chkcat[$i] eq 'data' ) {
1002 if ($chk[$i] ne '') {
1003 print OUTASM $T_HDR_data;
1004 print OUTASM $chk[$i];
1007 } elsif ( $chkcat[$i] eq 'consist' ) {
1008 if ( $chk[$i] =~ /$T_hsc_cc_PAT/o ) {
1009 local($consist) = "$1.$2.$3";
1010 $consist =~ s/,/./g;
1011 $consist =~ s/\//./g;
1012 $consist =~ s/-/_/g;
1013 $consist =~ s/[^A-Za-z0-9_.]/ZZ/g; # ToDo: properly?
1015 # Using a cygnus-2.7-96q4 gcc build on hppas, the
1016 # consistency chunk for ghc_cc_ID often (but not always!)
1017 # gets lumped with a bunch of .IMPORT directives containing info on
1018 # the code or data space nature of external symbols. We can't
1019 # toss these, so once the consistency ID has been turned into
1020 # a representable symbol, we substitute it for the symbol
1021 # that the string was attached to in the first place (ghc_cc_ID.)
1022 # (The original string is also substituted away.)
1024 # This change may affect the code output on other platforms in
1025 # adverse ways, hence we restrict this hack hppa targets only.
1028 if ( $TargetPlatform =~ /^hppa/ ) {
1029 $chk[$i] =~ s/^${T_US}ghc.*c_ID$TPOSTLBL/$consist/o;
1030 $chk[$i] =~ s/\t$T_hsc_cc_PAT/$T_HDR_misc/o;
1031 $consist = $chk[$i]; #clumsily
1033 print OUTASM $T_HDR_consist, "${consist}${T_POST_LBL}\n";
1035 } elsif ( $TargetPlatform !~ /^(mips)-/ ) { # we just don't try in those case (ToDo)
1036 # on mips: consistency string is just a v
1037 # horrible bunch of .bytes,
1038 # which I am too lazy to sort out (WDP 95/05)
1040 print STDERR "Couldn't grok consistency: ", $chk[$i];
1043 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1044 # we can just re-constitute this one...
1045 # NB: we emit _three_ underscores no matter what,
1046 # so ghc-split doesn't have to care.
1047 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1049 } elsif ( $chkcat[$i] eq 'closure'
1050 || $chkcat[$i] eq 'srt'
1051 || $chkcat[$i] eq 'infotbl'
1052 || $chkcat[$i] eq 'slow'
1053 || $chkcat[$i] eq 'fast' ) { # do them in that order
1054 $symb = $chksymb[$i];
1057 if ( defined($closurechk{$symb}) ) {
1058 print OUTASM $T_HDR_closure;
1059 print OUTASM $chk[$closurechk{$symb}];
1060 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1064 if ( defined($srtchk{$symb}) ) {
1065 print OUTASM $T_HDR_srt;
1066 print OUTASM $chk[$srtchk{$symb}];
1067 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1071 if ( defined($infochk{$symb}) ) {
1073 print OUTASM $T_HDR_info;
1074 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1076 # entry code will be put here!
1078 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1082 if ( defined($slowchk{$symb}) ) {
1084 # teach it to drop through to the fast entry point:
1085 $c = $chk[$slowchk{$symb}];
1087 if ( defined($fastchk{$symb}) ) {
1088 if ( $TargetPlatform =~ /^alpha-/ ) {
1089 $c =~ s/^\tjmp \$31,\(\$27\),0\n\t\.align 4\n\t\.end/\t.align 4\n\t.end/;
1090 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1091 $c =~ s/^\s+ldil.*\n\s+ldo.*\n\s+bv.*\n(.*\n)?\s+\.EXIT/$1\t.EXIT/;
1092 } elsif ( $TargetPlatform =~ /^i386-/ ) {
1093 # Reg alloc depending, gcc generated code may jump to the fast entry point via
1094 # a number of registers.
1095 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%edx\n\tjmp\s+\*\%edx\n//;
1096 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%ecx\n\tjmp\s+\*\%ecx\n//;
1097 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%eax\n\tjmp\s+\*\%eax\n//;
1098 # The next two only apply if we're not stealing %esi or %edi.
1099 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%esi\n\tjmp\s+\*\%esi\n// if ($StolenX86Regs < 3);
1100 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%edi\n\tjmp\s+\*\%edi\n// if ($StolenX86Regs < 4);
1101 } elsif ( $TargetPlatform =~ /^ia64-/ ) {
1102 #$c =~ s/^\tbr\.few ${symb}_fast\d*#\n\t;;\n(\t;;\n\t\.endp ${symb}_entry#\n)/$1/;
1103 $c =~ s/^\tbr\.few ${symb}_fast\d*#\n(\t;;\n\t\.endp ${symb}_entry#\n)/$1/;
1104 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1105 $c =~ s/^\tjmp \$31,\(\$27\),0\n\t\.align 4\n\t\.end/\t.align 4\n\t.end/;
1106 } elsif ( $TargetPlatform =~ /^m68k-/ ) {
1107 $c =~ s/^\tjmp ${T_US}${symb}_fast\d+.*\n\tnop\n//;
1108 $c =~ s/^\tjmp ${T_US}${symb}_fast\d+.*\n//;
1109 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ ) {
1110 $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//;
1111 # addis r9,r31,ha16(L_XXXX_fast3$non_lazy_ptr-L32$pb)
1112 # lwz r9,lo16(L_XXXX_fast3$non_lazy_ptr-L32$pb)(r9)
1115 } elsif ( $TargetPlatform =~ /^sparc-/ ) {
1116 $c =~ s/^\tcall\s+${T_US}${symb}_fast\d+,.*\n\t\s*nop\n//;
1117 $c =~ s/^\tcall\s+${T_US}${symb}_fast\d+,.*\n(\t\s*[a-z].*\n)/$1/;
1119 print STDERR "$Pgm: mystery slow-fast dropthrough: $TargetPlatform\n";
1123 if ( $TargetPlatform !~ /^(alpha-|hppa|mips-|powerpc-apple-)/) {
1124 # On alphas, hppa: no very good way to look for "dangling"
1125 # references to fast-entry point.
1126 # (questionable re hppa and mips...)
1127 print STDERR "still has jump to fast entry point:\n$c"
1128 if $c =~ /\b${T_US}${symb}_fast/;
1131 print OUTASM $T_HDR_entry;
1133 &print_doctored($c, 1); # NB: the 1!!!
1135 $chkcat[$slowchk{$symb}] = 'DONE ALREADY';
1139 if ( defined($fastchk{$symb}) ) {
1140 $c = $chk[$fastchk{$symb}];
1141 if ( ! defined($slowchk{$symb})
1142 # ToDo: the || clause can go once we're no longer
1143 # concerned about producing exactly the same output as before
1144 #OLD: || $TargetPlatform =~ /^(m68k|sparc|i386)-/
1146 print OUTASM $T_HDR_fast;
1149 &print_doctored($c, 0);
1150 $chkcat[$fastchk{$symb}] = 'DONE ALREADY';
1153 } elsif ( $chkcat[$i] eq 'vector'
1154 || $chkcat[$i] eq 'direct' ) { # do them in that order
1155 $symb = $chksymb[$i];
1158 if ( defined($vectorchk{$symb}) ) {
1159 print OUTASM $T_HDR_vector;
1161 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1162 # DO NOT DELETE THE NEXT LINE. It fixes a rather subtle GC bug
1163 # which showed up as a segfault reported by Ryszard Kubiak.
1164 # Problem is with vector tables. They wind up as follows:
1166 # .word some-other-word
1168 # Problem is that we want the label fooble_vtbl to be considered
1169 # in the same section as the vtbl itself, but the label actually
1170 # lives at the next word along. If a data segment should happen
1171 # to immediately follow the vtbl, as it can in GHCi, the label will
1172 # be malclassified as in the data rather than text segment (during
1173 # GC), and so we will regard references to it as static closure
1174 # pointers rather than as code pointers, which is an error which
1175 # usually crashes the garbage collectors.
1176 # To fix this, we place a dummy word after the label, so as to
1177 # ensure that the label is in the same segment as the vtbl proper.
1178 # The native code generator has an analogous fix; see
1179 # ghc/compiler/nativeGen/AbsCStixGen.lhs line 107.
1180 print OUTASM "${T_create_word} 0\n";
1182 # direct return code will be put here!
1183 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1185 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1186 # Alphas: the commented nop is for the splitter, to ensure
1187 # that no module ends with a label as the very last
1188 # thing. (The linker will adjust the label to point
1189 # to the first code word of the next module linked in,
1190 # even if alignment constraints cause the label to move!)
1192 print OUTASM "\t# nop\n";
1195 } elsif ( $chkcat[$i] eq 'toc' ) {
1196 # silly optimisation to print tocs, since they come in groups...
1197 print OUTASM $T_HDR_toc;
1199 while ($chkcat[$j] eq 'toc')
1200 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1203 print OUTASM $chk[$j];
1205 $chkcat[$j] = 'DONE ALREADY';
1209 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && $chkcat[$i] eq 'dyld' ) {
1210 # powerpc-apple: dynamic linker stubs
1211 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1212 { # print them out unchanged, but remove the stubs for __DISCARD__
1213 print OUTASM $chk[$i];
1216 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1220 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1223 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1224 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1229 sub hppa_mash_prologue { # OK, epilogue, too
1232 # toss all prologue stuff
1233 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1235 # Lie about our .CALLINFO
1236 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1243 # toss all epilogue stuff
1244 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1246 # Sorry; we moved the _info stuff to the code segment.
1247 s/_info,DATA/_info,CODE/g;
1254 sub print_doctored {
1255 local($_, $need_fallthru_patch) = @_;
1257 if ( $TargetPlatform !~ /^i386-/
1258 || ! /^\t[a-z]/ # no instructions in here, apparently
1259 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1263 # OK, must do some x86 **HACKING**
1265 local($entry_patch) = '';
1266 local($exit_patch) = '';
1268 # gotta watch out for weird instructions that
1269 # invisibly smash various regs:
1270 # rep* %ecx used for counting
1271 # scas* %edi used for destination index
1272 # cmps* %e[sd]i used for indices
1273 # loop* %ecx used for counting
1278 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1280 # * GCC used an "STG reg" for its own purposes
1282 # * some secret uses of machine reg, requiring STG reg
1283 # to be saved/restored
1285 # The most dangerous "GCC uses" of an "STG reg" are when
1286 # the reg holds the target of a jmp -- it's tricky to
1287 # insert the patch-up code before we get to the target!
1288 # So here we change the jmps:
1290 # --------------------------------------------------------
1291 # it can happen that we have jumps of the form...
1292 # jmp *<something involving %esp>
1294 # jmp <something involving another naughty register...>
1296 # a reasonably-common case is:
1298 # movl $_blah,<bad-reg>
1302 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[abcd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1304 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1305 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1306 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1307 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1308 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1309 if /(jmp|call)\s+.*\%esi/;
1311 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1312 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1313 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1314 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1315 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1316 if /(jmp|call)\s+.*\%edi/;
1319 # OK, now we can decide what our patch-up code is going to
1322 # Offsets into register table - you'd better update these magic
1323 # numbers should you change its contents!
1324 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1325 local($OFFSET_Hp)=92;
1327 # Note funky ".=" stuff; we're *adding* to these _patch guys
1328 if ( $StolenX86Regs <= 2
1329 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1330 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1331 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1333 # nothing for call_{entry,exit} because %esi is callee-save
1335 if ( $StolenX86Regs <= 3
1336 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1337 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1338 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1340 # nothing for call_{entry,exit} because %edi is callee-save
1343 # --------------------------------------------------------
1344 # next, here we go with non-%esp patching!
1346 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1348 # Before calling GC we must set up the exit condition before the call
1349 # and entry condition when we come back
1351 # fix _all_ non-local jumps:
1353 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1354 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1356 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1358 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1359 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1361 if ($StolenX86Regs == 2 ) {
1362 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1363 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1364 } elsif ($StolenX86Regs == 3 ) {
1365 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1366 if /^\t(jmp|call)\s+.*\%edi/;
1369 # --------------------------------------------------------
1370 # that's it -- print it
1372 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1376 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1377 print OUTASM $exit_patch;
1378 # ToDo: make it not print if there is a "jmp" at the end
1384 sub init_FUNNY_THINGS {
1385 %KNOWN_FUNNY_THING = (
1387 # "${T_US}stg_.*{T_POST_LBL}", 1,
1392 The following table reversal is used for both info tables and return
1393 vectors. In both cases, we remove the first entry from the table,
1394 reverse the table, put the label at the end, and paste some code
1395 (that which is normally referred to by the first entry in the table)
1396 right after the table itself. (The code pasting is done elsewhere.)
1400 local($symb, $tbl, $discard1) = @_;
1402 return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1404 local($before) = '';
1406 local(@imports) = (); # hppa only
1409 local(@lines) = split(/\n/, $tbl);
1412 # Deal with the header...
1413 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1414 $label .= $lines[$i] . "\n",
1415 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1416 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1417 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1419 $before .= $lines[$i] . "\n"; # otherwise...
1422 # Grab the table data...
1423 if ( $TargetPlatform !~ /^hppa/ ) {
1424 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1425 push(@words, $lines[$i]);
1427 } else { # hppa weirdness
1428 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1429 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1430 push(@imports, $lines[$i]);
1432 # We don't use HP's ``function pointers''
1433 # We just use labels in code space, like normal people
1434 $lines[$i] =~ s/P%//;
1435 push(@words, $lines[$i]);
1440 # now throw away the first word (SRT) iff it is empty.
1441 # The .zero business is for Linux/ELF.
1442 # The .skip business is for Sparc/Solaris/ELF.
1443 # The .blockz business is for HPPA.
1444 if ($discard1 && $words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1448 for (; $i <= $#lines; $i++) {
1449 $after .= $lines[$i] . "\n";
1452 # Alphas: If we have anonymous text (not part of a procedure), the
1453 # linker may complain about missing exception information. Bleh.
1454 # To suppress this, we place a .ent/.end pair around the code.
1455 # At the same time, we have to be careful and not enclose any leading
1456 # .file/.loc directives.
1457 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1458 local ($ident) = $1;
1459 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1460 $after .= "\t.end $ident\n";
1463 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1464 # assembler (!) wherein .quad constants inside .text sections are
1465 # first narrowed to 32 bits then sign-extended back to 64 bits.
1466 # This obviously screws up our 64-bit bitmaps, so we work around
1467 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1468 if ( $TargetPlatform =~ /^alpha-/ ) {
1470 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1471 local ($number) = $1;
1472 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1473 local ($sign, $base, $digits) = ($1, $2, $3);
1474 $base = (10, 8, 16)[length $base];
1475 local ($hi, $lo) = (0, 0);
1476 foreach $i (split(//, $digits)) {
1477 $j = $lo * $base + $i;
1478 $lo = $j % 4294967296;
1479 $hi = $hi * $base + ($j - $lo) / 4294967296;
1481 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1483 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1484 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1486 print STDERR "Cannot handle \".quad $number\" in info table\n";
1494 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
1495 . join("\n", @words) . "\n"
1498 # print STDERR "before=$before\n";
1499 # print STDERR "label=$label\n";
1500 # print STDERR "words=",(reverse @words),"\n";
1501 # print STDERR "after=$after\n";
1507 The HP is a major nuisance. The threaded code mangler moved info
1508 tables from data space to code space, but unthreaded code in the RTS
1509 still has references to info tables in data space. Since the HP
1510 linker is very precise about where symbols live, we need to patch the
1511 references in the unthreaded RTS as well.
1514 sub mini_mangle_asm_hppa {
1515 local($in_asmf, $out_asmf) = @_;
1517 open(INASM, "< $in_asmf")
1518 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1519 open(OUTASM,"> $out_asmf")
1520 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1523 s/_info,DATA/_info,CODE/; # Move _info references to code space
1529 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1530 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1536 sub tidy_up_and_die {
1537 local($return_val, $msg) = @_;
1539 exit (($return_val == 0) ? 0 : 1);