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 $T_CONST_LBL = '^\$C(\d+):$'; # 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 = '\.quad';
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";
117 #--------------------------------------------------------#
118 } elsif ( $TargetPlatform =~ /^hppa/ ) {
120 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
121 $T_US = ''; # _ if symbols have an underscore on the front
122 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
123 $T_CONST_LBL = '^L\$C(\d+)$'; # regexp for what such a lbl looks like
126 $T_MOVE_DIRVS = '^((\s+\.(IMPORT|EXPORT|PARAM).*|\s+\.align\s+\d+|\s+\.(SPACE|SUBSPA)\s+\S+|\s*)\n)';
127 $T_COPY_DIRVS = '^\s+\.(IMPORT|EXPORT)';
129 $T_hsc_cc_PAT = '\.STRING.*\)(hsc|cc) (.*)\\\\x09(.*)\\\\x00"';
130 $T_DOT_WORD = '\.(blockz|word|half|byte)';
131 $T_DOT_GLOBAL = '^\s+\.EXPORT';
132 $T_HDR_literal = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
133 $T_HDR_misc = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
134 $T_HDR_data = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
135 $T_HDR_consist = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
136 $T_HDR_closure = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
137 $T_HDR_srt = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
138 $T_HDR_info = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
139 $T_HDR_entry = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
140 $T_HDR_fast = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
141 $T_HDR_vector = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
142 $T_HDR_direct = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
144 #--------------------------------------------------------#
145 } elsif ( $TargetPlatform =~ /^i386-.*-(linuxaout|freebsd2|netbsd_aout|openbsd|nextstep3|cygwin32|mingw32)$/ ) {
146 # NeXT added but not tested. CaS
148 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
149 $T_US = '_'; # _ if symbols have an underscore on the front
150 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
151 $T_CONST_LBL = '^LC(\d+):$';
153 $T_X86_PRE_LLBL_PAT = 'L';
154 $T_X86_PRE_LLBL = 'L';
155 $T_X86_BADJMP = '^\tjmp [^L\*]';
157 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s+\d+(,0x90)?|\.globl\s+\S+|\.text|\.data|\.stab[^n].*|\.type\s+.*|\.size\s+.*)\n)';
158 $T_COPY_DIRVS = '\.(globl|stab)';
159 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
160 $T_DOT_WORD = '\.(long|word|value|byte|space)';
161 $T_DOT_GLOBAL = '\.globl';
162 $T_HDR_literal = "\.text\n\t\.align 2\n";
163 $T_HDR_misc = "\.text\n\t\.align 2,0x90\n";
164 $T_HDR_data = "\.data\n\t\.align 2\n";
165 $T_HDR_consist = "\.text\n";
166 $T_HDR_closure = "\.data\n\t\.align 2\n";
167 $T_HDR_closure = "\.data\n\t\.align 2\n\t.long 0\n" if ( $TargetPlatform =~ /.*-mingw32$/ );
168 $T_HDR_srt = "\.text\n\t\.align 2\n";
169 $T_HDR_info = "\.text\n\t\.align 2\n"; # NB: requires padding
170 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
171 $T_HDR_fast = "\.text\n\t\.align 2,0x90\n";
172 $T_HDR_vector = "\.text\n\t\.align 2\n"; # NB: requires padding
173 $T_HDR_direct = "\.text\n\t\.align 2,0x90\n";
175 #--------------------------------------------------------#
176 } elsif ( $TargetPlatform =~ /^i386-.*-(solaris2|linux|freebsd|netbsd)$/ ) {
178 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
179 $T_US = ''; # _ if symbols have an underscore on the front
180 $T_PRE_APP = # regexp that says what comes before APP/NO_APP
181 ($TargetPlatform =~ /-(linux|freebsd|netbsd_elf)$/) ? '#' : '/' ;
182 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
184 $T_X86_PRE_LLBL_PAT = '\.L';
185 $T_X86_PRE_LLBL = '.L';
186 $T_X86_BADJMP = '^\tjmp\s+[^\.\*]';
188 $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)';
189 $T_COPY_DIRVS = '\.(globl)';
191 if ( $TargetPlatform =~ /freebsd|netbsd_elf/ ) {
192 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
194 $T_hsc_cc_PAT = '\.string.*\)(hsc|cc) (.*)\\\\t(.*)"';
197 $T_DOT_WORD = '\.(long|value|byte|zero)';
198 $T_DOT_GLOBAL = '\.globl';
199 $T_HDR_literal = "\.section\t\.rodata\n"; # or just use .text??? (WDP 95/11)
200 $T_HDR_misc = "\.text\n\t\.align 4\n";
201 $T_HDR_data = "\.data\n\t\.align 4\n"; # ToDo: change align??
202 $T_HDR_consist = "\.text\n";
203 $T_HDR_closure = "\.data\n\t\.align 4\n"; # ToDo: change align?
204 $T_HDR_srt = "\.text\n\t\.align 4\n"; # ToDo: change align?
205 $T_HDR_info = "\.text\n\t\.align 4\n"; # NB: requires padding
206 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
207 $T_HDR_fast = "\.text\n\t\.align 4\n";
208 $T_HDR_vector = "\.text\n\t\.align 4\n"; # NB: requires padding
209 $T_HDR_direct = "\.text\n\t\.align 4\n";
211 #--------------------------------------------------------#
212 } elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) {
214 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
215 $T_US = '_'; # _ if symbols have an underscore on the front
216 $T_PRE_APP = '^# MAY NOT APPLY'; # regexp that says what comes before APP/NO_APP
217 $T_CONST_LBL = '^LC(\d+):$';
220 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.proc\s+\d+|\.const|\.cstring|\.globl\s+\S+|\.text|\.data|\.even|\.stab[^n].*)\n)';
221 $T_COPY_DIRVS = '\.(globl|proc|stab)';
222 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
224 $T_DOT_WORD = '\.long';
225 $T_DOT_GLOBAL = '\.globl';
226 $T_HDR_literal = "\.text\n\t\.even\n";
227 $T_HDR_misc = "\.text\n\t\.even\n";
228 $T_HDR_data = "\.data\n\t\.even\n";
229 $T_HDR_consist = "\.text\n";
230 $T_HDR_closure = "\.data\n\t\.even\n";
231 $T_HDR_srt = "\.text\n\t\.even\n";
232 $T_HDR_info = "\.text\n\t\.even\n";
233 $T_HDR_entry = "\.text\n\t\.even\n";
234 $T_HDR_fast = "\.text\n\t\.even\n";
235 $T_HDR_vector = "\.text\n\t\.even\n";
236 $T_HDR_direct = "\.text\n\t\.even\n";
238 #--------------------------------------------------------#
239 } elsif ( $TargetPlatform =~ /^mips-.*/ ) {
241 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
242 $T_US = ''; # _ if symbols have an underscore on the front
243 $T_PRE_APP = '^\s*#'; # regexp that says what comes before APP/NO_APP
244 $T_CONST_LBL = '^\$LC(\d+):$'; # regexp for what such a lbl looks like
247 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.(globl|ent)\s+\S+|\.text|\.r?data)\n)';
248 $T_COPY_DIRVS = '\.(globl|ent)';
250 $T_hsc_cc_PAT = 'I WAS TOO LAZY TO DO THIS BIT (WDP 95/05)';
251 $T_DOT_WORD = '\.word';
252 $T_DOT_GLOBAL = '^\t\.globl';
253 $T_HDR_literal = "\t\.rdata\n\t\.align 2\n";
254 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
255 $T_HDR_data = "\t\.data\n\t\.align 2\n";
256 $T_HDR_consist = 'TOO LAZY TO DO THIS TOO';
257 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
258 $T_HDR_srt = "\t\.text\n\t\.align 2\n";
259 $T_HDR_info = "\t\.text\n\t\.align 2\n";
260 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
261 $T_HDR_fast = "\t\.text\n\t\.align 2\n";
262 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
263 $T_HDR_direct = "\t\.text\n\t\.align 2\n";
265 #--------------------------------------------------------#
266 } elsif ( $TargetPlatform =~ /^powerpc-.*|^rs6000-.*/ ) {
268 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
269 $T_US = ''; # _ if symbols have an underscore on the front
270 $T_PRE_APP = 'NOT APPLICABLE'; # regexp that says what comes before APP/NO_APP
271 $T_CONST_LBL = 'NOT APPLICABLE'; # regexp for what such a lbl looks like
274 $T_MOVE_DIRVS = '^(\s*(\.toc|\.align \d+|\.csect \S+|\t\.?l?globl \S+)\n)';
275 $T_COPY_DIRVS = '\.(l?globl)';
277 $T_hsc_cc_PAT = '\.byte.*\)(hsc|cc) (.*)"\n\t\.byte \d+\n\t\.byte "(.*)"\n\t\.byte \d+';
278 $T_DOT_WORD = '\.long';
279 $T_DOT_GLOBAL = '\.globl';
280 $T_HDR_toc = "\.toc\n";
281 $T_HDR_literal = "\.csect .data[RW]\n\t\.align 2\n"; #not RO!?
282 $T_HDR_misc = "# misc\n\.csect \.text[PR]\n\t\.align 2\n";
283 $T_HDR_data = "# data\n\.csect \.data[RW]\n\t\.align 2\n";
284 $T_HDR_consist = "# consist\n\.csect \.data[RW]\n\t\.align 2\n";
285 $T_HDR_closure = "# closure\n\.csect \.data[RW]\n\t\.align 2\n";
286 $T_HDR_srt = "# closure\n\.csect \.data[RW]\n\t\.align 2\n";
287 $T_HDR_info = "# info\n\.csect \.data[RW]\n\t\.align 2\n"; #not RO!?
288 $T_HDR_entry = "# entry\n\.csect \.text[PR]\n\t\.align 2\n";
289 $T_HDR_fast = "# fast\n\.csect \.text[PR]\n\t\.align 2\n";
290 $T_HDR_vector = "# vector\n\.csect \.data[RW]\n\t\.align 2\n"; #not RO!?
291 $T_HDR_direct = "# direct\n";
293 #--------------------------------------------------------#
294 } elsif ( $TargetPlatform =~ /^sparc-.*-solaris2/ ) {
296 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
297 $T_US = ''; # _ if symbols have an underscore on the front
298 $T_PRE_APP = 'DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
299 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
302 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*|\.section.*|\s+\.type.*|\s+\.size.*)\n)';
303 $T_COPY_DIRVS = '\.(global|proc|stab)';
305 $T_hsc_cc_PAT = '\.asciz.*\)(hsc|cc) (.*)\\\\t(.*)"';
306 $T_DOT_WORD = '\.(word|byte|half|skip|uahalf|uaword)';
307 $T_DOT_GLOBAL = '^\t\.global';
308 $T_HDR_literal = "\.text\n\t\.align 8\n";
309 $T_HDR_misc = "\.text\n\t\.align 4\n";
310 $T_HDR_data = "\.data\n\t\.align 8\n";
311 $T_HDR_consist = "\.text\n";
312 $T_HDR_closure = "\.data\n\t\.align 4\n";
313 $T_HDR_srt = "\.data\n\t\.align 4\n";
314 $T_HDR_info = "\.text\n\t\.align 4\n";
315 $T_HDR_entry = "\.text\n\t\.align 4\n";
316 $T_HDR_fast = "\.text\n\t\.align 4\n";
317 $T_HDR_vector = "\.text\n\t\.align 4\n";
318 $T_HDR_direct = "\.text\n\t\.align 4\n";
320 #--------------------------------------------------------#
321 } elsif ( $TargetPlatform =~ /^sparc-.*-sunos4/ ) {
323 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
324 $T_US = '_'; # _ if symbols have an underscore on the front
325 $T_PRE_APP = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
326 $T_CONST_LBL = '^LC(\d+):$';
329 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*)\n)';
330 $T_COPY_DIRVS = '\.(global|proc|stab)';
331 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
333 $T_DOT_WORD = '\.word';
334 $T_DOT_GLOBAL = '^\t\.global';
335 $T_HDR_literal = "\.text\n\t\.align 8\n";
336 $T_HDR_misc = "\.text\n\t\.align 4\n";
337 $T_HDR_data = "\.data\n\t\.align 8\n";
338 $T_HDR_consist = "\.text\n";
339 $T_HDR_closure = "\.data\n\t\.align 4\n";
340 $T_HDR_srt = "\.data\n\t\.align 4\n";
341 $T_HDR_info = "\.text\n\t\.align 4\n";
342 $T_HDR_entry = "\.text\n\t\.align 4\n";
343 $T_HDR_fast = "\.text\n\t\.align 4\n";
344 $T_HDR_vector = "\.text\n\t\.align 4\n";
345 $T_HDR_direct = "\.text\n\t\.align 4\n";
347 #--------------------------------------------------------#
349 print STDERR "$Pgm: don't know how to mangle assembly language for: $TargetPlatform\n";
354 print STDERR "T_STABBY: $T_STABBY\n";
355 print STDERR "T_US: $T_US\n";
356 print STDERR "T_PRE_APP: $T_PRE_APP\n";
357 print STDERR "T_CONST_LBL: $T_CONST_LBL\n";
358 print STDERR "T_POST_LBL: $T_POST_LBL\n";
359 if ( $TargetPlatform =~ /^i386-/ ) {
360 print STDERR "T_X86_PRE_LLBL_PAT: $T_X86_PRE_LLBL_PAT\n";
361 print STDERR "T_X86_PRE_LLBL: $T_X86_PRE_LLBL\n";
362 print STDERR "T_X86_BADJMP: $T_X86_BADJMP\n";
364 print STDERR "T_MOVE_DIRVS: $T_MOVE_DIRVS\n";
365 print STDERR "T_COPY_DIRVS: $T_COPY_DIRVS\n";
366 print STDERR "T_hsc_cc_PAT: $T_hsc_cc_PAT\n";
367 print STDERR "T_DOT_WORD: $T_DOT_WORD\n";
368 print STDERR "T_HDR_literal: $T_HDR_literal\n";
369 print STDERR "T_HDR_misc: $T_HDR_misc\n";
370 print STDERR "T_HDR_data: $T_HDR_data\n";
371 print STDERR "T_HDR_consist: $T_HDR_consist\n";
372 print STDERR "T_HDR_closure: $T_HDR_closure\n";
373 print STDERR "T_HDR_info: $T_HDR_info\n";
374 print STDERR "T_HDR_entry: $T_HDR_entry\n";
375 print STDERR "T_HDR_fast: $T_HDR_fast\n";
376 print STDERR "T_HDR_vector: $T_HDR_vector\n";
377 print STDERR "T_HDR_direct: $T_HDR_direct\n";
383 %************************************************************************
385 \subsection{Mangle away}
387 %************************************************************************
391 local($in_asmf, $out_asmf) = @_;
393 # multi-line regexp matching:
398 &init_TARGET_STUFF();
399 &init_FUNNY_THINGS();
401 open(INASM, "< $in_asmf")
402 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
403 open(OUTASM,"> $out_asmf")
404 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
406 # read whole file, divide into "chunks":
407 # record some info about what we've found...
409 @chk = (); # contents of the chunk
410 $numchks = 0; # number of them
411 @chkcat = (); # what category of thing in each chunk
412 @chksymb = (); # what symbol(base) is defined in this chunk
413 %slowchk = (); # ditto, its regular "slow" entry code
414 %fastchk = (); # ditto, fast entry code
415 %closurechk = (); # ditto, the (static) closure
416 %srtchk = (); # ditto, its SRT (for top-level things)
417 %infochk = (); # given a symbol base, say what chunk its info tbl is in
418 %vectorchk = (); # ditto, return vector table
419 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
421 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
424 tr/\r//d if $TargetPlatform =~ /-mingw32$/;
425 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
426 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
427 next if /^\t\.def.*endef$/;
428 next if /${T_PRE_APP}(NO_)?APP/o;
429 next if /^;/ && $TargetPlatform =~ /^hppa/;
431 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|powerpc|rs6000)-/;
433 last if /^_section_\.text:$/ && $TargetPlatform =~ /^powerpc-|^rs6000-/;
435 if ( $TargetPlatform =~ /^mips-/
436 && /^\t\.(globl \S+ \.text|comm\t)/ ) {
437 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
439 # As a temporary solution for compiling "foreign export" declarations,
440 # we use global variables to pass arguments from C to STG land.
441 # These declarations live in the .hc file and not in the generated C
442 # stub file, so we let them pass through here.
443 } elsif ( /^\t\.comm\t__fexp_.*$/ ) {
445 $chkcat[$i] = 'data';
448 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
449 # duplicated from the bottom
453 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
454 # Alphas: Local labels not to be confused with new chunks
457 # NB: all the rest start with a non-space
459 } elsif ( $TargetPlatform =~ /^mips-/
460 && /^\d+:/ ) { # a funny-looking very-local label
463 } elsif ( /$T_CONST_LBL/o ) {
465 $chkcat[$i] = 'literal';
468 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
470 $chkcat[$i] = 'splitmarker';
473 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
476 $chkcat[$i] = 'infotbl';
477 $chksymb[$i] = $symb;
479 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
481 $infochk{$symb} = $i;
483 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
485 $chkcat[$i] = 'slow';
490 } elsif ( /^${T_US}([A-Za-z0-9_]+)_fast\d*${T_POST_LBL}$/o ) {
492 $chkcat[$i] = 'fast';
497 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
499 $chkcat[$i] = 'closure';
502 $closurechk{$1} = $i;
504 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
511 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
513 $chkcat[$i] = 'data';
516 } elsif ( /^${T_US}ghc.*c_ID${T_POST_LBL}/o ) {
518 $chkcat[$i] = 'consist';
520 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
523 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
524 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
525 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
526 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
529 $chkcat[$i] = 'data';
532 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
537 } elsif ( $TargetPlatform =~ /^powerpc-|^rs6000-/ && /^LC\.\.([0-9]+)/ ) {
542 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
543 # all CC_ symbols go in the data section...
545 $chkcat[$i] = 'data';
548 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
550 $chkcat[$i] = 'misc';
553 #$chksymb[$i] = $symbtmp if ($TargetPlatform =~ /^powerpc-|^rs6000-/) ; #rm andre
555 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
557 $chkcat[$i] = 'vector';
562 # As a temporary solution for compiling "foreign export" declarations,
563 # we use global variables to pass arguments from C to STG land.
564 # These declarations live in the .hc file and not in the generated C
565 # stub file, so we let them pass through here.
566 } elsif ( /^[\t ]+\.comm[\t ]+__fexp_.*$/ ) {
568 $chkcat[$i] = 'data';
571 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
572 && /^(_uname|uname|stat|fstat):/ ) {
573 # for some utterly bizarre reason, this platform
574 # likes to drop little local C routines with these names
575 # into each and every .o file that #includes the
576 # relevant system .h file. Yuck. We just don't
577 # tolerate them in .hc files (which we are processing
578 # here). If you need to call one of these things from
579 # Haskell, make a call to your own C wrapper, then
580 # put that C wrapper (which calls one of these) in a
581 # plain .c file. WDP 95/12
583 $chkcat[$i] = 'toss';
586 } elsif ( /^${T_US}[A-Za-z0-9_]/o
587 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
589 && ( $TargetPlatform !~ /^powerpc|^rs6000/ # ditto
590 || ! /^(L\.\.\d+|LT\.\..*):$/ ) ) {
593 print "Funny global thing?: $_"
594 unless $KNOWN_FUNNY_THING{$thing}
595 || /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
596 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
597 || /^${T_US}__init.*${T_POST_LBL}$/o # __init<module>
598 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
599 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o; # closure tables
601 $chkcat[$i] = 'misc';
602 if ($TargetPlatform =~ /^powerpc-|^rs6000-/)
603 { $chksymb[$i] = $thing; }
604 else { $chksymb[$i] = ''; };
606 } else { # simple line (duplicated at the top)
611 $numchks = $#chk + 1;
613 # the division into chunks is imperfect;
614 # we throw some things over the fence into the next
617 # also, there are things we would like to know
618 # about the whole module before we start spitting
621 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
623 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
625 # Alphas: NB: we start meddling at chunk 1, not chunk 0
626 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
627 # spits a ".quad 0" in after the v first ".rdata"; we
628 # detect this special case (tossing the ".quad 0")!
629 local($magic_rdata_seen) = 0;
631 # HPPAs, MIPSen: also start medding at chunk 1
633 if ($TargetPlatform =~ /^powerpc|^rs6000/) {
634 print OUTASM $T_HDR_toc; # yes, we have to put a .toc
635 # in the beginning of every file!
636 %tocequiv = (); # maps toc symbol number to toc symbol
637 %revtocequiv = (); # maps toc symbol to toc symbol number
638 for ($i = 1; $i < $numchks; $i++) {
639 $chk[$i] =~ s/\[RW\]//g;
640 $chk[$i] =~ s/\[DS\]//g;
641 $chk[$i] =~ s/^\.csect .*\[DS\]$//g;
643 if ( $chkcat[$i] eq 'toc' && $chk[$i] !~ /\.byte/ )
644 #ToDo: instead of all these changes, just fix mangle_powerpc_tailjump and delete/ignore these tocs?
645 { $chk[$i] =~ s/$T_MOVE_DIRVS//g;
646 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(\S+_fast\d+)/\t\.tc \1\[TC\],\.\2/;
647 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(\S+_entry)\n/\t\.tc \1\[TC\],\.\2\n/;
648 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(ret_\S+)/\t\.tc \1\[TC\],\.\2/;
649 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(alt_\S+)/\t\.tc \1\[TC\],\.\2/;
650 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(vtbl_\S+)/\t\.tc \1\[TC\],\.\2/;
652 $tocnumber = $chksymb[$i];
654 $tocsymb =~ s/^LC\.\.\d+:\n//;
655 $tocsymb =~ s/^\t\.tc \S+,(\S+)\n/\1/;
656 $tocequiv{$tocnumber} = $tocsymb;
658 } elsif ( $chkcat[$i] eq 'toc' && $chk[$i] =~ /\.byte/ ) {
659 $chkcat[$i] = 'literal';
664 for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) {
665 $c = $chk[$i]; # convenience copy
667 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
669 # toss all prologue stuff; HPPA is pretty weird
671 $c = &mash_hppa_prologue($c) if $TargetPlatform =~ /^hppa/;
673 # be slightly paranoid to make sure there's
674 # nothing surprising in there
675 if ( $c =~ /--- BEGIN ---/ ) {
676 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
678 if ($TargetPlatform =~ /^i386-/) {
679 $p =~ s/^\tpushl\s+\%edi\n//;
680 $p =~ s/^\tpushl\s+\%esi\n//;
681 $p =~ s/^\tpushl\s+\%ebx\n//;
682 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
683 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-cygwin32/);
684 } elsif ($TargetPlatform =~ /^m68k-/) {
685 $p =~ s/^\tlink a6,#-?\d.*\n//;
686 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
687 # The above showed up in the asm code,
688 # so I added it here.
689 # I hope it's correct.
691 $p =~ s/^\tmovel d2,sp\@-\n//;
692 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
693 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
694 } elsif ($TargetPlatform =~ /^mips-/) {
695 # the .frame/.mask/.fmask that we use is the same
696 # as that produced by GCC for miniInterpret; this
697 # gives GDB some chance of figuring out what happened
698 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
699 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
700 $p =~ s/^\t\.(mask|fmask).*\n//g;
701 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
702 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
703 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
704 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
705 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
706 $p =~ s/__FRAME__/$FRAME/;
707 } elsif ($TargetPlatform =~ /^powerpc-|^rs6000/) {
708 $p =~ s/^\tmflr 0\n//;
709 $p =~ s/^\tstm \d+,-\d+\(1\)\n//;
710 $p =~ s/^\tstw? 0,\d+\(1\)\n//g;
711 $p =~ s/^\tstw? 1,\d+\(1\)\n//g; #mc
712 $p =~ s/^\tlw?z 0,0\(1\)\n//g; #mc
713 $p =~ s/^\tstw?u 1,-\d+\(1\)\n//;
714 $p =~ s/^\tstw? \d+,-\d+\(1\)\n//g;
715 $p =~ s/^\tstfd \d+,-\d+\(1\)\n//g;
717 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
720 # HWL HACK: dont die, just print a warning
721 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/
722 # && $TargetPlatform !~ /^powerpc-/; #ToDo: remove test
723 die "Prologue junk?: $p\n" if $p =~ /^\t[^\.]/
724 && $TargetPlatform !~ /^powerpc-/; #ToDo: remove test
726 # glue together what's left
728 $c =~ s/\n\t\n/\n/; # junk blank line
732 if ( $TargetPlatform =~ /^mips-/ ) {
733 # MIPS: first, this basic sequence may occur "--- END ---" or not
734 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
737 # toss all epilogue stuff; again, paranoidly
738 if ( $c =~ /--- END ---/ ) {
739 if (($r, $e) = split(/--- END ---/, $c)) {
740 if ($TargetPlatform =~ /^i386-/) {
742 $e =~ s/^\tpopl\s+\%edi\n//;
743 $e =~ s/^\tpopl\s+\%esi\n//;
744 $e =~ s/^\tpopl\s+\%edx\n//;
745 $e =~ s/^\tpopl\s+\%ecx\n//;
746 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
747 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
748 } elsif ($TargetPlatform =~ /^m68k-/) {
749 $e =~ s/^\tunlk a6\n//;
751 } elsif ($TargetPlatform =~ /^mips-/) {
752 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
753 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
754 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
755 $e =~ s/^\tj\t\$31\n//;
756 } elsif ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
757 $e =~ s/^\taddi 1,1,\d+\n//;
758 $e =~ s/^\tcal 1,\d+\(1\)\n//;
759 $e =~ s/^\tlw?z? \d+,\d+\(1\)\n//;
760 $e =~ s/^\tmtlr 0\n//;
763 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
766 print "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/
767 && $TargetPlatform !~ /^powerpc-/; #ToDo: remove test
769 # glue together what's left
771 $c =~ s/\n\t\n/\n/; # junk blank line
775 # On SPARCs, we don't do --- BEGIN/END ---, we just
776 # toss the register-windowing save/restore/ret* instructions
778 if ( $TargetPlatform =~ /^sparc-/ ) {
779 $c =~ s/^\t(save.*|restore|ret|retl)\n//g;
780 # throw away PROLOGUE comments
781 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
784 # On Alphas, the prologue mangling is done a little later (below)
786 # toss all calls to __DISCARD__
787 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
789 # MIPS: that may leave some gratuitous asm macros around
790 # (no harm done; but we get rid of them to be tidier)
791 $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/
792 if $TargetPlatform =~ /^mips-/;
794 # toss stack adjustment after DoSparks
795 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
796 if $TargetPlatform =~ /^m68k-/; # this looks old...
798 if ( $TargetPlatform =~ /^alpha-/ &&
799 ! $magic_rdata_seen &&
800 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
801 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
802 $magic_rdata_seen = 1;
805 # pick some end-things and move them to the next chunk
807 # pin a funny end-thing on (for easier matching):
808 $c .= 'FUNNY#END#THING';
810 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
814 # on x86 we try not to copy any directives into a literal
815 # chunk, rather we keep looking for the next real chunk. This
816 # is because we get things like
818 # .globl blah_closure
824 if ( $TargetPlatform =~ /^(i386|sparc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
826 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
829 if ( $j < $numchks ) {
830 $chk[$j] = $to_move . $chk[$j];
834 elsif ( $i < ($numchks - 1)
835 && ( $to_move =~ /${T_COPY_DIRVS}/
836 || ($TargetPlatform =~ /^hppa/ && $to_move =~ /align/ && $chkcat[$i+1] eq 'literal') )) {
837 $chk[$i + 1] = $to_move . $chk[$i + 1];
838 # otherwise they're tossed
841 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
844 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
846 # toss all prologue stuff, except for loading gp, and the ..ng address
847 if (($p, $r) = split(/^\t\.prologue/, $c)) {
848 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
849 $c = $keep . "..ng:\n";
851 print STDERR "malformed code block ($ent)?\n"
854 $c .= "\t.frame \$30,0,\$26,0\n\t.prologue" . $r;
857 $c =~ s/FUNNY#END#THING//;
859 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
861 $chk[$i] = $c; # update w/ convenience copy
864 if ( $TargetPlatform =~ /^alpha-/ ) {
865 # print out the header stuff first
866 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
867 print OUTASM $chk[0];
869 } elsif ( $TargetPlatform =~ /^hppa/ ) {
870 print OUTASM $chk[0];
872 } elsif ( $TargetPlatform =~ /^mips-/ ) {
873 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
875 # get rid of horrible "<dollar>Revision: .*$" strings
876 local(@lines0) = split(/\n/, $chk[0]);
878 while ( $z <= $#lines0 ) {
879 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
882 while ( $z <= $#lines0 ) {
884 last if $lines0[$z] =~ /[,\t]0x0$/;
890 $chk[0] = join("\n", @lines0);
891 $chk[0] =~ s/\n\n+/\n/;
892 print OUTASM $chk[0];
895 # print out all the literal strings next
896 for ($i = 0; $i < $numchks; $i++) {
897 if ( $chkcat[$i] eq 'literal' ) {
898 print OUTASM $T_HDR_literal, $chk[$i];
899 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
901 $chkcat[$i] = 'DONE ALREADY';
905 # on the HPPA, print out all the bss next
906 if ( $TargetPlatform =~ /^hppa/ ) {
907 for ($i = 1; $i < $numchks; $i++) {
908 if ( $chkcat[$i] eq 'bss' ) {
909 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
910 print OUTASM $chk[$i];
912 $chkcat[$i] = 'DONE ALREADY';
917 for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) {
918 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
920 next if $chkcat[$i] eq 'DONE ALREADY';
922 if ( $chkcat[$i] eq 'misc' ) {
923 if ($chk[$i] ne '') {
924 print OUTASM $T_HDR_misc;
925 if ($TargetPlatform =~ /^powerpc-|^rs6000/) {
926 $chksymb[$i] =~ s/://;
927 #new if ($chksymb[$i] =~ /ret.*upd/ || $KNOWN_FUNNY_THING{$chksymb[$i]}
928 #new || $chksymb[$i] =~ /^$.{T_US}_(PRIn|PRStart).*${T_POST_LBL}$/o )
929 #new { print OUTASM "\t\.globl $chksymb[$i]\n"; }
930 # if ($chksymb[$i] ne '' && $chksymb[$i] !~ /ret_[a-z]/ && $chksymb[$i] !~ /djn_[a-z]/)
931 if ($chksymb[$i] ne '')
932 { print OUTASM "\t\.globl \.$chksymb[$i]\n"; };
933 if ($chk[$i] =~ /TOC\[tc0\], 0\n/)
934 { ($p, $r) = split(/TOC\[tc0\], 0\n/, $chk[$i]); $printDS = 1;}
935 else { $r = $chk[$i]; $printDS = 0; };
936 $chk[$i] = &mangle_powerpc_tailjump($r);
938 &print_doctored($chk[$i], 0);
939 if ($TargetPlatform =~ /^powerpc-|^rs6000-/ && $printDS) {
940 #ok if ($chksymb[$i] !~ /\__init_Main/) {
941 print OUTASM "\.csect ${chksymb[$i]}[DS]\n";
942 print OUTASM "${p}TOC[tc0], 0\n";
947 } elsif ( $chkcat[$i] eq 'toss' ) {
948 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
950 } elsif ( $chkcat[$i] eq 'data' ) {
951 if ($chk[$i] ne '') {
952 print OUTASM $T_HDR_data;
953 print OUTASM $chk[$i];
956 } elsif ( $chkcat[$i] eq 'consist' ) {
957 if ( $chk[$i] =~ /$T_hsc_cc_PAT/o ) {
958 local($consist) = "$1.$2.$3";
960 $consist =~ s/\//./g;
962 $consist =~ s/[^A-Za-z0-9_.]/ZZ/g; # ToDo: properly?
964 # Using a cygnus-2.7-96q4 gcc build on hppas, the
965 # consistency chunk for ghc_cc_ID often (but not always!)
966 # gets lumped with a bunch of .IMPORT directives containing info on
967 # the code or data space nature of external symbols. We can't
968 # toss these, so once the consistency ID has been turned into
969 # a representable symbol, we substitute it for the symbol
970 # that the string was attached to in the first place (ghc_cc_ID.)
971 # (The original string is also substituted away.)
973 # This change may affect the code output on other platforms in
974 # adverse ways, hence we restrict this hack hppa targets only.
977 if ( $TargetPlatform =~ /^hppa/ ) {
978 $chk[$i] =~ s/^${T_US}ghc.*c_ID$TPOSTLBL/$consist/o;
979 $chk[$i] =~ s/\t$T_hsc_cc_PAT/$T_HDR_misc/o;
980 $consist = $chk[$i]; #clumsily
982 print OUTASM $T_HDR_consist, "${consist}${T_POST_LBL}\n";
984 } elsif ( $TargetPlatform !~ /^(mips)-/ ) { # we just don't try in those case (ToDo)
985 # on mips: consistency string is just a v
986 # horrible bunch of .bytes,
987 # which I am too lazy to sort out (WDP 95/05)
989 print STDERR "Couldn't grok consistency: ", $chk[$i];
992 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
993 # we can just re-constitute this one...
994 # NB: we emit _three_ underscores no matter what,
995 # so ghc-split doesn't have to care.
996 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
998 } elsif ( $chkcat[$i] eq 'closure'
999 || $chkcat[$i] eq 'srt'
1000 || $chkcat[$i] eq 'infotbl'
1001 || $chkcat[$i] eq 'slow'
1002 || $chkcat[$i] eq 'fast' ) { # do them in that order
1003 $symb = $chksymb[$i];
1006 if ( defined($closurechk{$symb}) ) {
1007 print OUTASM $T_HDR_closure;
1008 print OUTASM $chk[$closurechk{$symb}];
1009 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1013 if ( defined($srtchk{$symb}) ) {
1014 print OUTASM $T_HDR_srt;
1015 print OUTASM $chk[$srtchk{$symb}];
1016 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1020 if ( defined($infochk{$symb}) ) {
1022 print OUTASM $T_HDR_info;
1023 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1024 if ( !defined($slowchk{$symb}) && defined($fastchk{$symb}) ) {
1025 $fastname = $chk[$fastchk{$symb}];
1026 $fastname =~ s/([_A-Za-z]*_fast\d+):.*(.*\n)*/\1/;
1027 $chk[$infochk{$symb}] =~ s/\.long StdErrorCode/\.long $fastname/;
1029 $chk[$infochk{$symb}] =~ s/\.long ([_A-Za-z]\S+_entry)/\.long \.\1/;
1030 $chk[$infochk{$symb}] =~ s/\.long ([A-Za-z]\S+_upd)/\.long \.\1/;
1031 print OUTASM $chk[$infochk{$symb}];
1033 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1035 # entry code will be put here!
1037 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1041 if ( defined($slowchk{$symb}) ) {
1043 # teach it to drop through to the fast entry point:
1044 $c = $chk[$slowchk{$symb}];
1046 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1047 ($p, $r) = split(/TOC\[tc0\], 0\n/, $c);
1048 if ($symb =~ /^[_A-Z]/)
1050 print OUTASM "\t\.globl \.${chksymb[$i]}_entry\n";
1051 print OUTASM "\.csect ${symb}_entry[DS]\n";
1052 print OUTASM "${p}TOC[tc0], 0\n";
1054 $r =~ s/\.csect \.text\[PR\]\n//; # todo: properly - andre
1055 $c = &mangle_powerpc_tailjump($r);
1058 if ( defined($fastchk{$symb}) ) {
1059 if ( $TargetPlatform =~ /^alpha-/ ) {
1060 $c =~ s/^\tjmp \$31,\(\$27\),0\n\t\.align 4\n\t\.end/\t.align 4\n\t.end/;
1061 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1062 $c =~ s/^\s+ldil.*\n\s+ldo.*\n\s+bv.*\n(.*\n)?\s+\.EXIT/$1\t.EXIT/;
1063 } elsif ( $TargetPlatform =~ /^i386-/ ) {
1064 # Reg alloc depending, gcc generated code may jump to the fast entry point via
1065 # a number of registers.
1066 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%edx\n\tjmp\s+\*\%edx\n//;
1067 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%ecx\n\tjmp\s+\*\%ecx\n//;
1068 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%eax\n\tjmp\s+\*\%eax\n//;
1069 # The next two only apply if we're not stealing %esi or %edi.
1070 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%esi\n\tjmp\s+\*\%esi\n// if ($StolenX86Regs < 3);
1071 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%edi\n\tjmp\s+\*\%edi\n// if ($StolenX86Regs < 4);
1072 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1073 $c =~ s/^\tjmp \$31,\(\$27\),0\n\t\.align 4\n\t\.end/\t.align 4\n\t.end/;
1074 } elsif ( $TargetPlatform =~ /^m68k-/ ) {
1075 $c =~ s/^\tjmp ${T_US}${symb}_fast\d+.*\n\tnop\n//;
1076 $c =~ s/^\tjmp ${T_US}${symb}_fast\d+.*\n//;
1077 } elsif ( $TargetPlatform =~ /^powerpc-|^rs6000-/ ) {
1078 $c =~ s/^\tb \.${T_US}${symb}_fast\d+\n//;
1079 } elsif ( $TargetPlatform =~ /^sparc-/ ) {
1080 $c =~ s/^\tcall\s+${T_US}${symb}_fast\d+,.*\n\t\s*nop\n//;
1081 $c =~ s/^\tcall\s+${T_US}${symb}_fast\d+,.*\n(\t\s*[a-z].*\n)/$1/;
1083 print STDERR "$Pgm: mystery slow-fast dropthrough: $TargetPlatform\n";
1087 if ( $TargetPlatform !~ /^(alpha-|hppa|mips-)/ ) {
1088 # On alphas, hppa: no very good way to look for "dangling"
1089 # references to fast-entry point.
1090 # (questionable re hppa and mips...)
1091 print STDERR "still has jump to fast entry point:\n$c"
1092 if $c =~ /\b${T_US}${symb}_fast/;
1095 print OUTASM $T_HDR_entry;
1097 &print_doctored($c, 1); # NB: the 1!!!
1099 $chkcat[$slowchk{$symb}] = 'DONE ALREADY';
1103 if ( defined($fastchk{$symb}) ) {
1104 $c = $chk[$fastchk{$symb}];
1105 if ( ! defined($slowchk{$symb})
1106 # ToDo: the || clause can go once we're no longer
1107 # concerned about producing exactly the same output as before
1108 #OLD: || $TargetPlatform =~ /^(m68k|sparc|i386)-/
1110 print OUTASM $T_HDR_fast;
1112 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1113 local(@lbls) = split(/:/, $c);
1114 $fullname = $lbls[0];
1115 $fullname =~ s/$T_MOVE_DIRVS//g;
1116 if ( $fullname =~ /^[A-Z]/)
1117 { print OUTASM "\t\.globl \.${fullname}\n";
1119 # print OUTASM "\t\.lglobl \.${fullname}\n"; #todo: rm - andre
1121 $c =~ s/((.*\n)*)\t.long \S+, TOC\[tc0\], 0\n\.csect \.text\[PR\]\n((.*\n)*)/\1\3/;
1122 $c = &mangle_powerpc_tailjump($c);
1124 &print_doctored($c, 0);
1125 $chkcat[$fastchk{$symb}] = 'DONE ALREADY';
1128 } elsif ( $chkcat[$i] eq 'vector'
1129 || $chkcat[$i] eq 'direct' ) { # do them in that order
1130 $symb = $chksymb[$i];
1133 if ( defined($vectorchk{$symb}) ) {
1134 print OUTASM $T_HDR_vector;
1135 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1136 if ( $symb =~ /^[A-Z]/) {
1137 print OUTASM "\t\.globl \.${symb}_vtbl\n";
1138 print OUTASM "\t\.globl ${symb}_vtbl\n";
1140 $chk[$vectorchk{$symb}] =~ s/\.long (\S+)/\.long \.\1/g;
1141 print OUTASM ".${symb}_vtbl:\n";
1142 print OUTASM $chk[$vectorchk{$symb}];
1144 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1146 # direct return code will be put here!
1147 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1149 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1150 # Alphas: the commented nop is for the splitter, to ensure
1151 # that no module ends with a label as the very last
1152 # thing. (The linker will adjust the label to point
1153 # to the first code word of the next module linked in,
1154 # even if alignment constraints cause the label to move!)
1156 print OUTASM "\t# nop\n";
1159 } elsif ( $chkcat[$i] eq 'toc' ) {
1160 # silly optimisation to print tocs, since they come in groups...
1161 print OUTASM $T_HDR_toc;
1163 while ($chkcat[$j] eq 'toc')
1164 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1167 print OUTASM $chk[$j];
1169 $chkcat[$j] = 'DONE ALREADY';
1174 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1178 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1180 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1181 print OUTASM ".csect .text[PR]\n_section_.text:\n.csect .data[RW]\n\t.long _section_.text\n"
1185 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1186 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1191 sub mash_hppa_prologue { # OK, epilogue, too
1194 # toss all prologue stuff
1195 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1197 # Lie about our .CALLINFO
1198 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1205 # toss all epilogue stuff
1206 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1208 # Sorry; we moved the _info stuff to the code segment.
1209 s/_info,DATA/_info,CODE/g;
1216 sub print_doctored {
1217 local($_, $need_fallthru_patch) = @_;
1219 if ( $TargetPlatform !~ /^i386-/
1220 || ! /^\t[a-z]/ # no instructions in here, apparently
1221 || /^${T_US}__init_[A-Za-z0-9_]+${T_POST_LBL}/) {
1225 # OK, must do some x86 **HACKING**
1227 local($entry_patch) = '';
1228 local($exit_patch) = '';
1230 # gotta watch out for weird instructions that
1231 # invisibly smash various regs:
1232 # rep* %ecx used for counting
1233 # scas* %edi used for destination index
1234 # cmps* %e[sd]i used for indices
1235 # loop* %ecx used for counting
1240 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1242 # * GCC used an "STG reg" for its own purposes
1244 # * some secret uses of machine reg, requiring STG reg
1245 # to be saved/restored
1247 # The most dangerous "GCC uses" of an "STG reg" are when
1248 # the reg holds the target of a jmp -- it's tricky to
1249 # insert the patch-up code before we get to the target!
1250 # So here we change the jmps:
1252 # --------------------------------------------------------
1253 # it can happen that we have jumps of the form...
1254 # jmp *<something involving %esp>
1256 # jmp <something involving another naughty register...>
1258 # a reasonably-common case is:
1260 # movl $_blah,<bad-reg>
1264 s/^\tmovl\s+\$${T_US}(.*),(\%e[abcd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1266 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1267 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1268 s/^\tjmp\s+\*(-?\d*)\((.*\%esi.*)\)\n/\tmovl $2,\%eax\n\tjmp \*$1\(\%eax\)\n/g;
1269 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1270 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1271 if /(jmp|call)\s+.*\%esi/;
1273 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1274 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1275 s/^\tjmp\s+\*(-?\d*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1276 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1277 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1278 if /(jmp|call)\s+.*\%edi/;
1281 # OK, now we can decide what our patch-up code is going to
1284 # Offsets into register table - you'd better update these magic
1285 # numbers should you change its contents!
1286 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1287 local($OFFSET_Hp)=92;
1289 # Note funky ".=" stuff; we're *adding* to these _patch guys
1290 if ( $StolenX86Regs <= 2
1291 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1292 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1293 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1295 # nothing for call_{entry,exit} because %esi is callee-save
1297 if ( $StolenX86Regs <= 3
1298 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1299 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1300 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1302 # nothing for call_{entry,exit} because %edi is callee-save
1305 # --------------------------------------------------------
1306 # next, here we go with non-%esp patching!
1308 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1310 # Before calling GC we must set up the exit condition before the call
1311 # and entry condition when we come back
1313 # fix _all_ non-local jumps:
1315 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1316 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1318 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1320 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1321 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1323 if ($StolenX86Regs == 2 ) {
1324 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1325 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1326 } elsif ($StolenX86Regs == 3 ) {
1327 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1328 if /^\t(jmp|call)\s+.*\%edi/;
1331 # --------------------------------------------------------
1332 # that's it -- print it
1334 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1338 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1339 print OUTASM $exit_patch;
1340 # ToDo: make it not print if there is a "jmp" at the end
1346 sub init_FUNNY_THINGS {
1347 %KNOWN_FUNNY_THING = (
1349 # "${T_US}stg_.*{T_POST_LBL}", 1,
1354 The following table reversal is used for both info tables and return
1355 vectors. In both cases, we remove the first entry from the table,
1356 reverse the table, put the label at the end, and paste some code
1357 (that which is normally referred to by the first entry in the table)
1358 right after the table itself. (The code pasting is done elsewhere.)
1362 local($symb, $tbl, $discard1) = @_;
1364 local($before) = '';
1366 local(@imports) = (); # hppa only
1369 local(@lines) = split(/\n/, $tbl);
1370 local($i, $j); #local ($i, $extra, $words_to_pad, $j);
1372 # Deal with the header...
1373 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1374 $label .= $lines[$i] . "\n",
1375 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1376 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1377 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1379 $before .= $lines[$i] . "\n"; # otherwise...
1382 # Grab the table data...
1383 if ( $TargetPlatform !~ /^hppa/ ) {
1384 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1385 push(@words, $lines[$i]);
1387 } else { # hppa weirdness
1388 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1389 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1390 push(@imports, $lines[$i]);
1392 # We don't use HP's ``function pointers''
1393 # We just use labels in code space, like normal people
1394 $lines[$i] =~ s/P%//;
1395 push(@words, $lines[$i]);
1400 # now throw away the first word (SRT) iff it is empty.
1401 # The .zero business is for Linux/ELF.
1402 # The .skip business is for Sparc/Solaris/ELF.
1403 # The .blockz business is for HPPA.
1404 if ($discard1 && $words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1408 # Padding removed to reduce code size and improve performance on Pentiums.
1410 # for 486-cache-friendliness, we want our tables aligned
1411 # on 16-byte boundaries (.align 4). Let's pad:
1412 # $extra = ($#words + 1) % 4;
1413 # $words_to_pad = ($extra == 0) ? 0 : 4 - $extra;
1414 # for ($j = 0; $j < $words_to_pad; $j++) { push(@words, "\t${T_DOT_WORD} 0"); }
1416 for (; $i <= $#lines; $i++) {
1417 $after .= $lines[$i] . "\n";
1420 # Alphas:If we have anonymous text (not part of a procedure), the
1421 # linker may complain about missing exception information. Bleh.
1422 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1423 $before = "\t.ent $1\n" . $before;
1424 $after .= "\t.end $1\n";
1428 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
1429 . join("\n", @words) . "\n"
1432 # print STDERR "before=$before\n";
1433 # print STDERR "label=$label\n";
1434 # print STDERR "words=",(reverse @words),"\n";
1435 # print STDERR "after=$after\n";
1441 The HP is a major nuisance. The threaded code mangler moved info
1442 tables from data space to code space, but unthreaded code in the RTS
1443 still has references to info tables in data space. Since the HP
1444 linker is very precise about where symbols live, we need to patch the
1445 references in the unthreaded RTS as well.
1448 sub mini_mangle_asm_hppa {
1449 local($in_asmf, $out_asmf) = @_;
1451 open(INASM, "< $in_asmf")
1452 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1453 open(OUTASM,"> $out_asmf")
1454 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1457 s/_info,DATA/_info,CODE/; # Move _info references to code space
1463 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1464 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1470 sub mini_mangle_asm_powerpc {
1471 local($in_asmf, $out_asmf) = @_;
1473 open(INASM, "< $in_asmf")
1474 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1475 open(OUTASM,"> $out_asmf")
1476 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1479 s/long _PRStart/long ._PRStart/;
1480 s/long _PRIn_/long ._PRIn_/;
1481 s/long _Dummy_(\S+)_entry/long ._Dummy_\1_entry/;
1482 s/long _PRMarking_MarkNextRoot\[DS\]/long ._PRMarking_MarkNextRoot/;
1483 s/long _PRMarking_MarkNextCAF\[DS\]/long ._PRMarking_MarkNextCAF/;
1484 s/long _PRMarking_MarkNextAStack\[DS\]/long ._PRMarking_MarkNextAStack/;
1485 s/long _PRMarking_MarkNextBStack\[DS\]/long ._PRMarking_MarkNextBStack/;
1486 s/\.tc EnterNodeCode\[TC]\,EnterNodeCode\[DS\]/\.tc EnterNodeCode\[TC]\,.EnterNodeCode/; # CONC
1487 s/\.tc CheckHeapCode\[TC]\,CheckHeapCode\[DS\]/\.tc CheckHeapCode\[TC]\,.CheckHeapCode/; # CONC
1492 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1493 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1496 sub mangle_powerpc_tailjump {
1498 local($maybe_more) = 1;
1499 while (($c =~ /\tlw?z? \d+,LC\.\.\d+\(2\)\n\tmtctr \d+\n\tbctr\n/) && $maybe_more)
1502 $lcsymb =~ s/(.*\n)*\tlw?z? \d+,LC\.\.(\d+)\(2\)\n\tmtctr \d+\n\tbctr\n(.*\n)*/\2/;
1503 # the checks for r1 and r2 are mostly paranoia...
1505 $r1 =~ s/(.*\n)*\tlw?z? (\d+),LC\.\.\d+\(2\)\n\tmtctr \d+\n\tbctr\n(.*\n)*/\2/;
1507 $r2 =~ s/(.*\n)*\tlw?z? \d+,LC\.\.(\d+)\(2\)\n\tmtctr (\d+)\n\tbctr\n(.*\n)*/\3/;
1510 $c =~ s/((.*\n)*)\tlw?z? \d+,LC\.\.\d+\(2\)\n\tmtctr \d+\n\tbctr\n((.*\n)*)/\1\tb $tocequiv{$lcsymb}\n\3/;
1518 sub tidy_up_and_die {
1519 local($return_val, $msg) = @_;
1521 exit (($return_val == 0) ? 0 : 1);