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-/ ) {
69 $StolenX86Regs = $ARGV[2];
72 &mangle_asm($ifile,$ofile);
77 %************************************************************************
79 \subsection{Constants for various architectures}
81 %************************************************************************
84 sub init_TARGET_STUFF {
86 #--------------------------------------------------------#
87 if ( $TargetPlatform =~ /^alpha-.*-.*/ ) {
89 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
90 $T_US = ''; # _ if symbols have an underscore on the front
91 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
92 $T_CONST_LBL = '^\$C(\d+):$'; # regexp for what such a lbl looks like
95 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.(globl|ent)\s+\S+|\#.*|\.(file|loc)\s+\S+\s+\S+|\.text|\.r?data)\n)';
96 $T_COPY_DIRVS = '^\s*(\#|\.(file|globl|ent|loc))';
98 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
99 $T_DOT_WORD = '\.quad';
100 $T_DOT_GLOBAL = '^\t\.globl';
101 $T_HDR_literal = "\.rdata\n\t\.align 3\n";
102 $T_HDR_misc = "\.text\n\t\.align 3\n";
103 $T_HDR_data = "\.data\n\t\.align 3\n";
104 $T_HDR_consist = "\.text\n";
105 $T_HDR_closure = "\.data\n\t\.align 3\n";
106 $T_HDR_srt = "\.text\n\t\.align 3\n";
107 $T_HDR_info = "\.text\n\t\.align 3\n";
108 $T_HDR_entry = "\.text\n\t\.align 3\n";
109 $T_HDR_fast = "\.text\n\t\.align 3\n";
110 $T_HDR_vector = "\.text\n\t\.align 3\n";
111 $T_HDR_direct = "\.text\n\t\.align 3\n";
113 #--------------------------------------------------------#
114 } elsif ( $TargetPlatform =~ /^hppa/ ) {
116 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
117 $T_US = ''; # _ if symbols have an underscore on the front
118 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
119 $T_CONST_LBL = '^L\$C(\d+)$'; # regexp for what such a lbl looks like
122 $T_MOVE_DIRVS = '^((\s+\.(IMPORT|EXPORT|PARAM).*|\s+\.align\s+\d+|\s+\.(SPACE|SUBSPA)\s+\S+|\s*)\n)';
123 $T_COPY_DIRVS = '^\s+\.(IMPORT|EXPORT)';
125 $T_hsc_cc_PAT = '\.STRING.*\)(hsc|cc) (.*)\\\\x09(.*)\\\\x00"';
126 $T_DOT_WORD = '\.(blockz|word|half|byte)';
127 $T_DOT_GLOBAL = '^\s+\.EXPORT';
128 $T_HDR_literal = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
129 $T_HDR_misc = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
130 $T_HDR_data = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
131 $T_HDR_consist = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
132 $T_HDR_closure = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
133 $T_HDR_srt = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
134 $T_HDR_info = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
135 $T_HDR_entry = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
136 $T_HDR_fast = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
137 $T_HDR_vector = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
138 $T_HDR_direct = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
140 #--------------------------------------------------------#
141 } elsif ( $TargetPlatform =~ /^i386-.*-(linuxaout|freebsd2|netbsd|openbsd|nextstep3|cygwin32|mingw32)$/ ) {
142 # NeXT added but not tested. CaS
144 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
145 $T_US = '_'; # _ if symbols have an underscore on the front
146 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
147 $T_CONST_LBL = '^LC(\d+):$';
149 $T_X86_PRE_LLBL_PAT = 'L';
150 $T_X86_PRE_LLBL = 'L';
151 $T_X86_BADJMP = '^\tjmp [^L\*]';
153 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s+\d+(,0x90)?|\.globl\s+\S+|\.text|\.data|\.stab[^n].*|\.type\s+.*|\.size\s+.*)\n)';
154 $T_COPY_DIRVS = '\.(globl|stab)';
155 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
156 $T_DOT_WORD = '\.(long|word|value|byte|space)';
157 $T_DOT_GLOBAL = '\.globl';
158 $T_HDR_literal = "\.text\n\t\.align 2\n";
159 $T_HDR_misc = "\.text\n\t\.align 2,0x90\n";
160 $T_HDR_data = "\.data\n\t\.align 2\n";
161 $T_HDR_consist = "\.text\n";
162 $T_HDR_closure = "\.data\n\t\.align 2\n";
163 $T_HDR_closure = "\.data\n\t\.align 2\n\t.long 0\n" if ( $TargetPlatform =~ /.*-mingw32$/ );
164 $T_HDR_srt = "\.text\n\t\.align 2\n";
165 $T_HDR_info = "\.text\n\t\.align 2\n"; # NB: requires padding
166 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
167 $T_HDR_fast = "\.text\n\t\.align 2,0x90\n";
168 $T_HDR_vector = "\.text\n\t\.align 2\n"; # NB: requires padding
169 $T_HDR_direct = "\.text\n\t\.align 2,0x90\n";
171 #--------------------------------------------------------#
172 } elsif ( $TargetPlatform =~ /^i386-.*-(solaris2|linux|freebsd|netbsd_elf)$/ ) {
174 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
175 $T_US = ''; # _ if symbols have an underscore on the front
176 $T_PRE_APP = # regexp that says what comes before APP/NO_APP
177 ($TargetPlatform =~ /-(linux|freebsd|netbsd_elf)$/) ? '#' : '/' ;
178 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
180 $T_X86_PRE_LLBL_PAT = '\.L';
181 $T_X86_PRE_LLBL = '.L';
182 $T_X86_BADJMP = '^\tjmp\s+[^\.\*]';
184 $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)';
185 $T_COPY_DIRVS = '\.(globl)';
187 if ( $TargetPlatform =~ /freebsd|netbsd_elf/ ) {
188 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
190 $T_hsc_cc_PAT = '\.string.*\)(hsc|cc) (.*)\\\\t(.*)"';
193 $T_DOT_WORD = '\.(long|value|byte|zero)';
194 $T_DOT_GLOBAL = '\.globl';
195 $T_HDR_literal = "\.section\t\.rodata\n"; # or just use .text??? (WDP 95/11)
196 $T_HDR_misc = "\.text\n\t\.align 4\n";
197 $T_HDR_data = "\.data\n\t\.align 4\n"; # ToDo: change align??
198 $T_HDR_consist = "\.text\n";
199 $T_HDR_closure = "\.data\n\t\.align 4\n"; # ToDo: change align?
200 $T_HDR_srt = "\.text\n\t\.align 4\n"; # ToDo: change align?
201 $T_HDR_info = "\.text\n\t\.align 4\n"; # NB: requires padding
202 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
203 $T_HDR_fast = "\.text\n\t\.align 4\n";
204 $T_HDR_vector = "\.text\n\t\.align 4\n"; # NB: requires padding
205 $T_HDR_direct = "\.text\n\t\.align 4\n";
207 #--------------------------------------------------------#
208 } elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) {
210 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
211 $T_US = '_'; # _ if symbols have an underscore on the front
212 $T_PRE_APP = '^# MAY NOT APPLY'; # regexp that says what comes before APP/NO_APP
213 $T_CONST_LBL = '^LC(\d+):$';
216 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.proc\s+\d+|\.const|\.cstring|\.globl\s+\S+|\.text|\.data|\.even|\.stab[^n].*)\n)';
217 $T_COPY_DIRVS = '\.(globl|proc|stab)';
218 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
220 $T_DOT_WORD = '\.long';
221 $T_DOT_GLOBAL = '\.globl';
222 $T_HDR_literal = "\.text\n\t\.even\n";
223 $T_HDR_misc = "\.text\n\t\.even\n";
224 $T_HDR_data = "\.data\n\t\.even\n";
225 $T_HDR_consist = "\.text\n";
226 $T_HDR_closure = "\.data\n\t\.even\n";
227 $T_HDR_srt = "\.text\n\t\.even\n";
228 $T_HDR_info = "\.text\n\t\.even\n";
229 $T_HDR_entry = "\.text\n\t\.even\n";
230 $T_HDR_fast = "\.text\n\t\.even\n";
231 $T_HDR_vector = "\.text\n\t\.even\n";
232 $T_HDR_direct = "\.text\n\t\.even\n";
234 #--------------------------------------------------------#
235 } elsif ( $TargetPlatform =~ /^mips-.*/ ) {
237 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
238 $T_US = ''; # _ if symbols have an underscore on the front
239 $T_PRE_APP = '^\s*#'; # regexp that says what comes before APP/NO_APP
240 $T_CONST_LBL = '^\$LC(\d+):$'; # regexp for what such a lbl looks like
243 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.(globl|ent)\s+\S+|\.text|\.r?data)\n)';
244 $T_COPY_DIRVS = '\.(globl|ent)';
246 $T_hsc_cc_PAT = 'I WAS TOO LAZY TO DO THIS BIT (WDP 95/05)';
247 $T_DOT_WORD = '\.word';
248 $T_DOT_GLOBAL = '^\t\.globl';
249 $T_HDR_literal = "\t\.rdata\n\t\.align 2\n";
250 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
251 $T_HDR_data = "\t\.data\n\t\.align 2\n";
252 $T_HDR_consist = 'TOO LAZY TO DO THIS TOO';
253 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
254 $T_HDR_srt = "\t\.text\n\t\.align 2\n";
255 $T_HDR_info = "\t\.text\n\t\.align 2\n";
256 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
257 $T_HDR_fast = "\t\.text\n\t\.align 2\n";
258 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
259 $T_HDR_direct = "\t\.text\n\t\.align 2\n";
261 #--------------------------------------------------------#
262 } elsif ( $TargetPlatform =~ /^powerpc-.*|^rs6000-.*/ ) {
264 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
265 $T_US = ''; # _ if symbols have an underscore on the front
266 $T_PRE_APP = 'NOT APPLICABLE'; # regexp that says what comes before APP/NO_APP
267 $T_CONST_LBL = 'NOT APPLICABLE'; # regexp for what such a lbl looks like
270 $T_MOVE_DIRVS = '^(\s*(\.toc|\.align \d+|\.csect \S+|\t\.?l?globl \S+)\n)';
271 $T_COPY_DIRVS = '\.(l?globl)';
273 $T_hsc_cc_PAT = '\.byte.*\)(hsc|cc) (.*)"\n\t\.byte \d+\n\t\.byte "(.*)"\n\t\.byte \d+';
274 $T_DOT_WORD = '\.long';
275 $T_DOT_GLOBAL = '\.globl';
276 $T_HDR_toc = "\.toc\n";
277 $T_HDR_literal = "\.csect .data[RW]\n\t\.align 2\n"; #not RO!?
278 $T_HDR_misc = "# misc\n\.csect \.text[PR]\n\t\.align 2\n";
279 $T_HDR_data = "# data\n\.csect \.data[RW]\n\t\.align 2\n";
280 $T_HDR_consist = "# consist\n\.csect \.data[RW]\n\t\.align 2\n";
281 $T_HDR_closure = "# closure\n\.csect \.data[RW]\n\t\.align 2\n";
282 $T_HDR_srt = "# closure\n\.csect \.data[RW]\n\t\.align 2\n";
283 $T_HDR_info = "# info\n\.csect \.data[RW]\n\t\.align 2\n"; #not RO!?
284 $T_HDR_entry = "# entry\n\.csect \.text[PR]\n\t\.align 2\n";
285 $T_HDR_fast = "# fast\n\.csect \.text[PR]\n\t\.align 2\n";
286 $T_HDR_vector = "# vector\n\.csect \.data[RW]\n\t\.align 2\n"; #not RO!?
287 $T_HDR_direct = "# direct\n";
289 #--------------------------------------------------------#
290 } elsif ( $TargetPlatform =~ /^sparc-.*-solaris2/ ) {
292 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
293 $T_US = ''; # _ if symbols have an underscore on the front
294 $T_PRE_APP = 'DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
295 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
298 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*|\.section.*|\s+\.type.*|\s+\.size.*)\n)';
299 $T_COPY_DIRVS = '\.(global|proc|stab)';
301 $T_hsc_cc_PAT = '\.asciz.*\)(hsc|cc) (.*)\\\\t(.*)"';
302 $T_DOT_WORD = '\.(word|byte|half|skip|uahalf|uaword)';
303 $T_DOT_GLOBAL = '^\t\.global';
304 $T_HDR_literal = "\.text\n\t\.align 8\n";
305 $T_HDR_misc = "\.text\n\t\.align 4\n";
306 $T_HDR_data = "\.data\n\t\.align 8\n";
307 $T_HDR_consist = "\.text\n";
308 $T_HDR_closure = "\.data\n\t\.align 4\n";
309 $T_HDR_srt = "\.data\n\t\.align 4\n";
310 $T_HDR_info = "\.text\n\t\.align 4\n";
311 $T_HDR_entry = "\.text\n\t\.align 4\n";
312 $T_HDR_fast = "\.text\n\t\.align 4\n";
313 $T_HDR_vector = "\.text\n\t\.align 4\n";
314 $T_HDR_direct = "\.text\n\t\.align 4\n";
316 #--------------------------------------------------------#
317 } elsif ( $TargetPlatform =~ /^sparc-.*-sunos4/ ) {
319 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
320 $T_US = '_'; # _ if symbols have an underscore on the front
321 $T_PRE_APP = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
322 $T_CONST_LBL = '^LC(\d+):$';
325 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*)\n)';
326 $T_COPY_DIRVS = '\.(global|proc|stab)';
327 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
329 $T_DOT_WORD = '\.word';
330 $T_DOT_GLOBAL = '^\t\.global';
331 $T_HDR_literal = "\.text\n\t\.align 8\n";
332 $T_HDR_misc = "\.text\n\t\.align 4\n";
333 $T_HDR_data = "\.data\n\t\.align 8\n";
334 $T_HDR_consist = "\.text\n";
335 $T_HDR_closure = "\.data\n\t\.align 4\n";
336 $T_HDR_srt = "\.data\n\t\.align 4\n";
337 $T_HDR_info = "\.text\n\t\.align 4\n";
338 $T_HDR_entry = "\.text\n\t\.align 4\n";
339 $T_HDR_fast = "\.text\n\t\.align 4\n";
340 $T_HDR_vector = "\.text\n\t\.align 4\n";
341 $T_HDR_direct = "\.text\n\t\.align 4\n";
343 #--------------------------------------------------------#
345 print STDERR "$Pgm: don't know how to mangle assembly language for: $TargetPlatform\n";
350 print STDERR "T_STABBY: $T_STABBY\n";
351 print STDERR "T_US: $T_US\n";
352 print STDERR "T_PRE_APP: $T_PRE_APP\n";
353 print STDERR "T_CONST_LBL: $T_CONST_LBL\n";
354 print STDERR "T_POST_LBL: $T_POST_LBL\n";
355 if ( $TargetPlatform =~ /^i386-/ ) {
356 print STDERR "T_X86_PRE_LLBL_PAT: $T_X86_PRE_LLBL_PAT\n";
357 print STDERR "T_X86_PRE_LLBL: $T_X86_PRE_LLBL\n";
358 print STDERR "T_X86_BADJMP: $T_X86_BADJMP\n";
360 print STDERR "T_MOVE_DIRVS: $T_MOVE_DIRVS\n";
361 print STDERR "T_COPY_DIRVS: $T_COPY_DIRVS\n";
362 print STDERR "T_hsc_cc_PAT: $T_hsc_cc_PAT\n";
363 print STDERR "T_DOT_WORD: $T_DOT_WORD\n";
364 print STDERR "T_HDR_literal: $T_HDR_literal\n";
365 print STDERR "T_HDR_misc: $T_HDR_misc\n";
366 print STDERR "T_HDR_data: $T_HDR_data\n";
367 print STDERR "T_HDR_consist: $T_HDR_consist\n";
368 print STDERR "T_HDR_closure: $T_HDR_closure\n";
369 print STDERR "T_HDR_info: $T_HDR_info\n";
370 print STDERR "T_HDR_entry: $T_HDR_entry\n";
371 print STDERR "T_HDR_fast: $T_HDR_fast\n";
372 print STDERR "T_HDR_vector: $T_HDR_vector\n";
373 print STDERR "T_HDR_direct: $T_HDR_direct\n";
379 %************************************************************************
381 \subsection{Mangle away}
383 %************************************************************************
387 local($in_asmf, $out_asmf) = @_;
389 # multi-line regexp matching:
394 &init_TARGET_STUFF();
395 &init_FUNNY_THINGS();
397 open(INASM, "< $in_asmf")
398 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
399 open(OUTASM,"> $out_asmf")
400 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
402 # read whole file, divide into "chunks":
403 # record some info about what we've found...
405 @chk = (); # contents of the chunk
406 $numchks = 0; # number of them
407 @chkcat = (); # what category of thing in each chunk
408 @chksymb = (); # what symbol(base) is defined in this chunk
409 %slowchk = (); # ditto, its regular "slow" entry code
410 %fastchk = (); # ditto, fast entry code
411 %closurechk = (); # ditto, the (static) closure
412 %srtchk = (); # ditto, its SRT (for top-level things)
413 %infochk = (); # given a symbol base, say what chunk its info tbl is in
414 %vectorchk = (); # ditto, return vector table
415 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
417 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
420 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
421 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
422 next if /^\t\.def.*endef$/;
423 next if /${T_PRE_APP}(NO_)?APP/o;
424 next if /^;/ && $TargetPlatform =~ /^hppa/;
426 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|powerpc|rs6000)-/;
428 last if /^_section_\.text:$/ && $TargetPlatform =~ /^powerpc-|^rs6000-/;
430 if ( $TargetPlatform =~ /^mips-/
431 && /^\t\.(globl \S+ \.text|comm\t)/ ) {
432 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
434 # As a temporary solution for compiling "foreign export" declarations,
435 # we use global variables to pass arguments from C to STG land.
436 # These declarations live in the .hc file and not in the generated C
437 # stub file, so we let them pass through here.
438 } elsif ( /^\t\.comm\t__fexp_.*$/ ) {
440 $chkcat[$i] = 'data';
443 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
444 # duplicated from the bottom
448 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
449 # Alphas: Local labels not to be confused with new chunks
452 # NB: all the rest start with a non-space
454 } elsif ( $TargetPlatform =~ /^mips-/
455 && /^\d+:/ ) { # a funny-looking very-local label
458 } elsif ( /$T_CONST_LBL/o ) {
460 $chkcat[$i] = 'literal';
463 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
465 $chkcat[$i] = 'splitmarker';
468 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
471 $chkcat[$i] = 'infotbl';
472 $chksymb[$i] = $symb;
474 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
476 $infochk{$symb} = $i;
478 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
480 $chkcat[$i] = 'slow';
485 } elsif ( /^${T_US}([A-Za-z0-9_]+)_fast\d*${T_POST_LBL}$/o ) {
487 $chkcat[$i] = 'fast';
492 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
494 $chkcat[$i] = 'closure';
497 $closurechk{$1} = $i;
499 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
506 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
508 $chkcat[$i] = 'data';
511 } elsif ( /^${T_US}ghc.*c_ID${T_POST_LBL}/o ) {
513 $chkcat[$i] = 'consist';
515 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
518 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
519 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
520 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
521 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
524 $chkcat[$i] = 'data';
527 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
532 } elsif ( $TargetPlatform =~ /^powerpc-|^rs6000-/ && /^LC\.\.([0-9]+)/ ) {
537 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
538 # all CC_ symbols go in the data section...
540 $chkcat[$i] = 'data';
543 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
545 $chkcat[$i] = 'misc';
548 #$chksymb[$i] = $symbtmp if ($TargetPlatform =~ /^powerpc-|^rs6000-/) ; #rm andre
550 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
552 $chkcat[$i] = 'vector';
557 # As a temporary solution for compiling "foreign export" declarations,
558 # we use global variables to pass arguments from C to STG land.
559 # These declarations live in the .hc file and not in the generated C
560 # stub file, so we let them pass through here.
561 } elsif ( /^[\t ]+\.comm[\t ]+__fexp_.*$/ ) {
563 $chkcat[$i] = 'data';
566 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
567 && /^(_uname|uname|stat|fstat):/ ) {
568 # for some utterly bizarre reason, this platform
569 # likes to drop little local C routines with these names
570 # into each and every .o file that #includes the
571 # relevant system .h file. Yuck. We just don't
572 # tolerate them in .hc files (which we are processing
573 # here). If you need to call one of these things from
574 # Haskell, make a call to your own C wrapper, then
575 # put that C wrapper (which calls one of these) in a
576 # plain .c file. WDP 95/12
578 $chkcat[$i] = 'toss';
581 } elsif ( /^${T_US}[A-Za-z0-9_]/o
582 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
584 && ( $TargetPlatform !~ /^powerpc|^rs6000/ # ditto
585 || ! /^(L\.\.\d+|LT\.\..*):$/ ) ) {
588 print "Funny global thing?: $_"
589 unless $KNOWN_FUNNY_THING{$thing}
590 || /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
591 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
592 || /^${T_US}__init.*${T_POST_LBL}$/o # __init<module>
593 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
594 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o; # closure tables
596 $chkcat[$i] = 'misc';
597 if ($TargetPlatform =~ /^powerpc-|^rs6000-/)
598 { $chksymb[$i] = $thing; }
599 else { $chksymb[$i] = ''; };
601 } else { # simple line (duplicated at the top)
606 $numchks = $#chk + 1;
608 # the division into chunks is imperfect;
609 # we throw some things over the fence into the next
612 # also, there are things we would like to know
613 # about the whole module before we start spitting
616 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
618 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
620 # Alphas: NB: we start meddling at chunk 1, not chunk 0
621 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
622 # spits a ".quad 0" in after the v first ".rdata"; we
623 # detect this special case (tossing the ".quad 0")!
624 local($magic_rdata_seen) = 0;
626 # HPPAs, MIPSen: also start medding at chunk 1
628 if ($TargetPlatform =~ /^powerpc|^rs6000/) {
629 print OUTASM $T_HDR_toc; # yes, we have to put a .toc
630 # in the beginning of every file!
631 %tocequiv = (); # maps toc symbol number to toc symbol
632 %revtocequiv = (); # maps toc symbol to toc symbol number
633 for ($i = 1; $i < $numchks; $i++) {
634 $chk[$i] =~ s/\[RW\]//g;
635 $chk[$i] =~ s/\[DS\]//g;
636 $chk[$i] =~ s/^\.csect .*\[DS\]$//g;
638 if ( $chkcat[$i] eq 'toc' && $chk[$i] !~ /\.byte/ )
639 #ToDo: instead of all these changes, just fix mangle_powerpc_tailjump and delete/ignore these tocs?
640 { $chk[$i] =~ s/$T_MOVE_DIRVS//g;
641 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(\S+_fast\d+)/\t\.tc \1\[TC\],\.\2/;
642 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(\S+_entry)\n/\t\.tc \1\[TC\],\.\2\n/;
643 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(ret_\S+)/\t\.tc \1\[TC\],\.\2/;
644 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(alt_\S+)/\t\.tc \1\[TC\],\.\2/;
645 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(vtbl_\S+)/\t\.tc \1\[TC\],\.\2/;
647 $tocnumber = $chksymb[$i];
649 $tocsymb =~ s/^LC\.\.\d+:\n//;
650 $tocsymb =~ s/^\t\.tc \S+,(\S+)\n/\1/;
651 $tocequiv{$tocnumber} = $tocsymb;
653 } elsif ( $chkcat[$i] eq 'toc' && $chk[$i] =~ /\.byte/ ) {
654 $chkcat[$i] = 'literal';
659 for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) {
660 $c = $chk[$i]; # convenience copy
662 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
664 # toss all prologue stuff; HPPA is pretty weird
666 $c = &mash_hppa_prologue($c) if $TargetPlatform =~ /^hppa/;
668 # be slightly paranoid to make sure there's
669 # nothing surprising in there
670 if ( $c =~ /--- BEGIN ---/ ) {
671 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
673 if ($TargetPlatform =~ /^i386-/) {
674 $p =~ s/^\tpushl\s+\%edi\n//;
675 $p =~ s/^\tpushl\s+\%esi\n//;
676 $p =~ s/^\tpushl\s+\%ebx\n//;
677 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
678 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-cygwin32/);
679 } elsif ($TargetPlatform =~ /^m68k-/) {
680 $p =~ s/^\tlink a6,#-?\d.*\n//;
681 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
682 # The above showed up in the asm code,
683 # so I added it here.
684 # I hope it's correct.
686 $p =~ s/^\tmovel d2,sp\@-\n//;
687 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
688 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
689 } elsif ($TargetPlatform =~ /^mips-/) {
690 # the .frame/.mask/.fmask that we use is the same
691 # as that produced by GCC for miniInterpret; this
692 # gives GDB some chance of figuring out what happened
693 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
694 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
695 $p =~ s/^\t\.(mask|fmask).*\n//g;
696 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
697 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
698 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
699 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
700 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
701 $p =~ s/__FRAME__/$FRAME/;
702 } elsif ($TargetPlatform =~ /^powerpc-|^rs6000/) {
703 $p =~ s/^\tmflr 0\n//;
704 $p =~ s/^\tstm \d+,-\d+\(1\)\n//;
705 $p =~ s/^\tstw? 0,\d+\(1\)\n//g;
706 $p =~ s/^\tstw? 1,\d+\(1\)\n//g; #mc
707 $p =~ s/^\tlw?z 0,0\(1\)\n//g; #mc
708 $p =~ s/^\tstw?u 1,-\d+\(1\)\n//;
709 $p =~ s/^\tstw? \d+,-\d+\(1\)\n//g;
710 $p =~ s/^\tstfd \d+,-\d+\(1\)\n//g;
712 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
715 # HWL HACK: dont die, just print a warning
716 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/
717 # && $TargetPlatform !~ /^powerpc-/; #ToDo: remove test
718 die "Prologue junk?: $p\n" if $p =~ /^\t[^\.]/
719 && $TargetPlatform !~ /^powerpc-/; #ToDo: remove test
721 # glue together what's left
723 $c =~ s/\n\t\n/\n/; # junk blank line
727 if ( $TargetPlatform =~ /^mips-/ ) {
728 # MIPS: first, this basic sequence may occur "--- END ---" or not
729 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
732 # toss all epilogue stuff; again, paranoidly
733 if ( $c =~ /--- END ---/ ) {
734 if (($r, $e) = split(/--- END ---/, $c)) {
735 if ($TargetPlatform =~ /^i386-/) {
737 $e =~ s/^\tpopl\s+\%edi\n//;
738 $e =~ s/^\tpopl\s+\%esi\n//;
739 $e =~ s/^\tpopl\s+\%edx\n//;
740 $e =~ s/^\tpopl\s+\%ecx\n//;
741 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
742 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
743 } elsif ($TargetPlatform =~ /^m68k-/) {
744 $e =~ s/^\tunlk a6\n//;
746 } elsif ($TargetPlatform =~ /^mips-/) {
747 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
748 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
749 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
750 $e =~ s/^\tj\t\$31\n//;
751 } elsif ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
752 $e =~ s/^\taddi 1,1,\d+\n//;
753 $e =~ s/^\tcal 1,\d+\(1\)\n//;
754 $e =~ s/^\tlw?z? \d+,\d+\(1\)\n//;
755 $e =~ s/^\tmtlr 0\n//;
758 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
761 print "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/
762 && $TargetPlatform !~ /^powerpc-/; #ToDo: remove test
764 # glue together what's left
766 $c =~ s/\n\t\n/\n/; # junk blank line
770 # On SPARCs, we don't do --- BEGIN/END ---, we just
771 # toss the register-windowing save/restore/ret* instructions
773 if ( $TargetPlatform =~ /^sparc-/ ) {
774 $c =~ s/^\t(save.*|restore|ret|retl)\n//g;
775 # throw away PROLOGUE comments
776 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
779 # On Alphas, the prologue mangling is done a little later (below)
781 # toss all calls to __DISCARD__
782 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
784 # MIPS: that may leave some gratuitous asm macros around
785 # (no harm done; but we get rid of them to be tidier)
786 $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/
787 if $TargetPlatform =~ /^mips-/;
789 # toss stack adjustment after DoSparks
790 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
791 if $TargetPlatform =~ /^m68k-/; # this looks old...
793 if ( $TargetPlatform =~ /^alpha-/ &&
794 ! $magic_rdata_seen &&
795 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
796 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
797 $magic_rdata_seen = 1;
800 # pick some end-things and move them to the next chunk
802 # pin a funny end-thing on (for easier matching):
803 $c .= 'FUNNY#END#THING';
805 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
809 # on x86 we try not to copy any directives into a literal
810 # chunk, rather we keep looking for the next real chunk. This
811 # is because we get things like
813 # .globl blah_closure
819 if ( $TargetPlatform =~ /^(i386|sparc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
821 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
824 if ( $j < $numchks ) {
825 $chk[$j] = $to_move . $chk[$j];
829 elsif ( $i < ($numchks - 1)
830 && ( $to_move =~ /${T_COPY_DIRVS}/
831 || ($TargetPlatform =~ /^hppa/ && $to_move =~ /align/ && $chkcat[$i+1] eq 'literal') )) {
832 $chk[$i + 1] = $to_move . $chk[$i + 1];
833 # otherwise they're tossed
836 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
839 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
841 # toss all prologue stuff, except for loading gp, and the ..ng address
842 if (($p, $r) = split(/^\t\.prologue/, $c)) {
843 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
844 $c = $keep . "..ng:\n";
846 print STDERR "malformed code block ($ent)?\n"
849 $c .= "\t.frame \$30,0,\$26,0\n\t.prologue" . $r;
852 $c =~ s/FUNNY#END#THING//;
854 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
856 $chk[$i] = $c; # update w/ convenience copy
859 if ( $TargetPlatform =~ /^alpha-/ ) {
860 # print out the header stuff first
861 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
862 print OUTASM $chk[0];
864 } elsif ( $TargetPlatform =~ /^hppa/ ) {
865 print OUTASM $chk[0];
867 } elsif ( $TargetPlatform =~ /^mips-/ ) {
868 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
870 # get rid of horrible "<dollar>Revision: .*$" strings
871 local(@lines0) = split(/\n/, $chk[0]);
873 while ( $z <= $#lines0 ) {
874 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
877 while ( $z <= $#lines0 ) {
879 last if $lines0[$z] =~ /[,\t]0x0$/;
885 $chk[0] = join("\n", @lines0);
886 $chk[0] =~ s/\n\n+/\n/;
887 print OUTASM $chk[0];
890 # print out all the literal strings next
891 for ($i = 0; $i < $numchks; $i++) {
892 if ( $chkcat[$i] eq 'literal' ) {
893 print OUTASM $T_HDR_literal, $chk[$i];
894 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
896 $chkcat[$i] = 'DONE ALREADY';
900 # on the HPPA, print out all the bss next
901 if ( $TargetPlatform =~ /^hppa/ ) {
902 for ($i = 1; $i < $numchks; $i++) {
903 if ( $chkcat[$i] eq 'bss' ) {
904 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
905 print OUTASM $chk[$i];
907 $chkcat[$i] = 'DONE ALREADY';
912 for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) {
913 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
915 next if $chkcat[$i] eq 'DONE ALREADY';
917 if ( $chkcat[$i] eq 'misc' ) {
918 if ($chk[$i] ne '') {
919 print OUTASM $T_HDR_misc;
920 if ($TargetPlatform =~ /^powerpc-|^rs6000/) {
921 $chksymb[$i] =~ s/://;
922 #new if ($chksymb[$i] =~ /ret.*upd/ || $KNOWN_FUNNY_THING{$chksymb[$i]}
923 #new || $chksymb[$i] =~ /^$.{T_US}_(PRIn|PRStart).*${T_POST_LBL}$/o )
924 #new { print OUTASM "\t\.globl $chksymb[$i]\n"; }
925 # if ($chksymb[$i] ne '' && $chksymb[$i] !~ /ret_[a-z]/ && $chksymb[$i] !~ /djn_[a-z]/)
926 if ($chksymb[$i] ne '')
927 { print OUTASM "\t\.globl \.$chksymb[$i]\n"; };
928 if ($chk[$i] =~ /TOC\[tc0\], 0\n/)
929 { ($p, $r) = split(/TOC\[tc0\], 0\n/, $chk[$i]); $printDS = 1;}
930 else { $r = $chk[$i]; $printDS = 0; };
931 $chk[$i] = &mangle_powerpc_tailjump($r);
933 &print_doctored($chk[$i], 0);
934 if ($TargetPlatform =~ /^powerpc-|^rs6000-/ && $printDS) {
935 #ok if ($chksymb[$i] !~ /\__init_Main/) {
936 print OUTASM "\.csect ${chksymb[$i]}[DS]\n";
937 print OUTASM "${p}TOC[tc0], 0\n";
942 } elsif ( $chkcat[$i] eq 'toss' ) {
943 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
945 } elsif ( $chkcat[$i] eq 'data' ) {
946 if ($chk[$i] ne '') {
947 print OUTASM $T_HDR_data;
948 print OUTASM $chk[$i];
951 } elsif ( $chkcat[$i] eq 'consist' ) {
952 if ( $chk[$i] =~ /$T_hsc_cc_PAT/o ) {
953 local($consist) = "$1.$2.$3";
955 $consist =~ s/\//./g;
957 $consist =~ s/[^A-Za-z0-9_.]/ZZ/g; # ToDo: properly?
959 # Using a cygnus-2.7-96q4 gcc build on hppas, the
960 # consistency chunk for ghc_cc_ID often (but not always!)
961 # gets lumped with a bunch of .IMPORT directives containing info on
962 # the code or data space nature of external symbols. We can't
963 # toss these, so once the consistency ID has been turned into
964 # a representable symbol, we substitute it for the symbol
965 # that the string was attached to in the first place (ghc_cc_ID.)
966 # (The original string is also substituted away.)
968 # This change may affect the code output on other platforms in
969 # adverse ways, hence we restrict this hack hppa targets only.
972 if ( $TargetPlatform =~ /^hppa/ ) {
973 $chk[$i] =~ s/^${T_US}ghc.*c_ID$TPOSTLBL/$consist/o;
974 $chk[$i] =~ s/\t$T_hsc_cc_PAT/$T_HDR_misc/o;
975 $consist = $chk[$i]; #clumsily
977 print OUTASM $T_HDR_consist, "${consist}${T_POST_LBL}\n";
979 } elsif ( $TargetPlatform !~ /^(mips)-/ ) { # we just don't try in those case (ToDo)
980 # on mips: consistency string is just a v
981 # horrible bunch of .bytes,
982 # which I am too lazy to sort out (WDP 95/05)
984 print STDERR "Couldn't grok consistency: ", $chk[$i];
987 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
988 # we can just re-constitute this one...
989 # NB: we emit _three_ underscores no matter what,
990 # so ghc-split doesn't have to care.
991 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
993 } elsif ( $chkcat[$i] eq 'closure'
994 || $chkcat[$i] eq 'srt'
995 || $chkcat[$i] eq 'infotbl'
996 || $chkcat[$i] eq 'slow'
997 || $chkcat[$i] eq 'fast' ) { # do them in that order
998 $symb = $chksymb[$i];
1001 if ( defined($closurechk{$symb}) ) {
1002 print OUTASM $T_HDR_closure;
1003 print OUTASM $chk[$closurechk{$symb}];
1004 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1008 if ( defined($srtchk{$symb}) ) {
1009 print OUTASM $T_HDR_srt;
1010 print OUTASM $chk[$srtchk{$symb}];
1011 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1015 if ( defined($infochk{$symb}) ) {
1017 print OUTASM $T_HDR_info;
1018 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1019 if ( !defined($slowchk{$symb}) && defined($fastchk{$symb}) ) {
1020 $fastname = $chk[$fastchk{$symb}];
1021 $fastname =~ s/([_A-Za-z]*_fast\d+):.*(.*\n)*/\1/;
1022 $chk[$infochk{$symb}] =~ s/\.long StdErrorCode/\.long $fastname/;
1024 $chk[$infochk{$symb}] =~ s/\.long ([_A-Za-z]\S+_entry)/\.long \.\1/;
1025 $chk[$infochk{$symb}] =~ s/\.long ([A-Za-z]\S+_upd)/\.long \.\1/;
1026 print OUTASM $chk[$infochk{$symb}];
1028 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1030 # entry code will be put here!
1032 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1036 if ( defined($slowchk{$symb}) ) {
1038 # teach it to drop through to the fast entry point:
1039 $c = $chk[$slowchk{$symb}];
1041 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1042 ($p, $r) = split(/TOC\[tc0\], 0\n/, $c);
1043 if ($symb =~ /^[_A-Z]/)
1045 print OUTASM "\t\.globl \.${chksymb[$i]}_entry\n";
1046 print OUTASM "\.csect ${symb}_entry[DS]\n";
1047 print OUTASM "${p}TOC[tc0], 0\n";
1049 $r =~ s/\.csect \.text\[PR\]\n//; # todo: properly - andre
1050 $c = &mangle_powerpc_tailjump($r);
1053 if ( defined($fastchk{$symb}) ) {
1054 if ( $TargetPlatform =~ /^alpha-/ ) {
1055 $c =~ s/^\tjmp \$31,\(\$27\),0\n\t\.align 4\n\t\.end/\t.align 4\n\t.end/;
1056 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1057 $c =~ s/^\s+ldil.*\n\s+ldo.*\n\s+bv.*\n(.*\n)?\s+\.EXIT/$1\t.EXIT/;
1058 } elsif ( $TargetPlatform =~ /^i386-/ ) {
1059 # Reg alloc depending, gcc generated code may jump to the fast entry point via
1060 # a number of registers.
1061 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%edx\n\tjmp\s+\*\%edx\n//;
1062 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%ecx\n\tjmp\s+\*\%ecx\n//;
1063 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%eax\n\tjmp\s+\*\%eax\n//;
1064 # The next two only apply if we're not stealing %esi or %edi.
1065 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%esi\n\tjmp\s+\*\%esi\n// if ($StolenX86Regs < 3);
1066 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%edi\n\tjmp\s+\*\%edi\n// if ($StolenX86Regs < 4);
1067 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1068 $c =~ s/^\tjmp \$31,\(\$27\),0\n\t\.align 4\n\t\.end/\t.align 4\n\t.end/;
1069 } elsif ( $TargetPlatform =~ /^m68k-/ ) {
1070 $c =~ s/^\tjmp ${T_US}${symb}_fast\d+.*\n\tnop\n//;
1071 $c =~ s/^\tjmp ${T_US}${symb}_fast\d+.*\n//;
1072 } elsif ( $TargetPlatform =~ /^powerpc-|^rs6000-/ ) {
1073 $c =~ s/^\tb \.${T_US}${symb}_fast\d+\n//;
1074 } elsif ( $TargetPlatform =~ /^sparc-/ ) {
1075 $c =~ s/^\tcall\s+${T_US}${symb}_fast\d+,.*\n\t\s*nop\n//;
1076 $c =~ s/^\tcall\s+${T_US}${symb}_fast\d+,.*\n(\t\s*[a-z].*\n)/$1/;
1078 print STDERR "$Pgm: mystery slow-fast dropthrough: $TargetPlatform\n";
1082 if ( $TargetPlatform !~ /^(alpha-|hppa|mips-)/ ) {
1083 # On alphas, hppa: no very good way to look for "dangling"
1084 # references to fast-entry point.
1085 # (questionable re hppa and mips...)
1086 print STDERR "still has jump to fast entry point:\n$c"
1087 if $c =~ /\b${T_US}${symb}_fast/;
1090 print OUTASM $T_HDR_entry;
1092 &print_doctored($c, 1); # NB: the 1!!!
1094 $chkcat[$slowchk{$symb}] = 'DONE ALREADY';
1098 if ( defined($fastchk{$symb}) ) {
1099 $c = $chk[$fastchk{$symb}];
1100 if ( ! defined($slowchk{$symb})
1101 # ToDo: the || clause can go once we're no longer
1102 # concerned about producing exactly the same output as before
1103 #OLD: || $TargetPlatform =~ /^(m68k|sparc|i386)-/
1105 print OUTASM $T_HDR_fast;
1107 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1108 local(@lbls) = split(/:/, $c);
1109 $fullname = $lbls[0];
1110 $fullname =~ s/$T_MOVE_DIRVS//g;
1111 if ( $fullname =~ /^[A-Z]/)
1112 { print OUTASM "\t\.globl \.${fullname}\n";
1114 # print OUTASM "\t\.lglobl \.${fullname}\n"; #todo: rm - andre
1116 $c =~ s/((.*\n)*)\t.long \S+, TOC\[tc0\], 0\n\.csect \.text\[PR\]\n((.*\n)*)/\1\3/;
1117 $c = &mangle_powerpc_tailjump($c);
1119 &print_doctored($c, 0);
1120 $chkcat[$fastchk{$symb}] = 'DONE ALREADY';
1123 } elsif ( $chkcat[$i] eq 'vector'
1124 || $chkcat[$i] eq 'direct' ) { # do them in that order
1125 $symb = $chksymb[$i];
1128 if ( defined($vectorchk{$symb}) ) {
1129 print OUTASM $T_HDR_vector;
1130 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1131 if ( $symb =~ /^[A-Z]/) {
1132 print OUTASM "\t\.globl \.${symb}_vtbl\n";
1133 print OUTASM "\t\.globl ${symb}_vtbl\n";
1135 $chk[$vectorchk{$symb}] =~ s/\.long (\S+)/\.long \.\1/g;
1136 print OUTASM ".${symb}_vtbl:\n";
1137 print OUTASM $chk[$vectorchk{$symb}];
1139 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1141 # direct return code will be put here!
1142 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1144 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1145 # Alphas: the commented nop is for the splitter, to ensure
1146 # that no module ends with a label as the very last
1147 # thing. (The linker will adjust the label to point
1148 # to the first code word of the next module linked in,
1149 # even if alignment constraints cause the label to move!)
1151 print OUTASM "\t# nop\n";
1154 } elsif ( $chkcat[$i] eq 'toc' ) {
1155 # silly optimisation to print tocs, since they come in groups...
1156 print OUTASM $T_HDR_toc;
1158 while ($chkcat[$j] eq 'toc')
1159 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1162 print OUTASM $chk[$j];
1164 $chkcat[$j] = 'DONE ALREADY';
1169 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1173 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1175 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1176 print OUTASM ".csect .text[PR]\n_section_.text:\n.csect .data[RW]\n\t.long _section_.text\n"
1180 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1181 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1186 sub mash_hppa_prologue { # OK, epilogue, too
1189 # toss all prologue stuff
1190 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1192 # Lie about our .CALLINFO
1193 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1200 # toss all epilogue stuff
1201 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1203 # Sorry; we moved the _info stuff to the code segment.
1204 s/_info,DATA/_info,CODE/g;
1211 sub print_doctored {
1212 local($_, $need_fallthru_patch) = @_;
1214 if ( $TargetPlatform !~ /^i386-/
1215 || ! /^\t[a-z]/ # no instructions in here, apparently
1216 || /^${T_US}__init_[A-Za-z0-9_]+${T_POST_LBL}/) {
1220 # OK, must do some x86 **HACKING**
1222 local($entry_patch) = '';
1223 local($exit_patch) = '';
1225 # gotta watch out for weird instructions that
1226 # invisibly smash various regs:
1227 # rep* %ecx used for counting
1228 # scas* %edi used for destination index
1229 # cmps* %e[sd]i used for indices
1230 # loop* %ecx used for counting
1235 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1237 # * GCC used an "STG reg" for its own purposes
1239 # * some secret uses of machine reg, requiring STG reg
1240 # to be saved/restored
1242 # The most dangerous "GCC uses" of an "STG reg" are when
1243 # the reg holds the target of a jmp -- it's tricky to
1244 # insert the patch-up code before we get to the target!
1245 # So here we change the jmps:
1247 # --------------------------------------------------------
1248 # it can happen that we have jumps of the form...
1249 # jmp *<something involving %esp>
1251 # jmp <something involving another naughty register...>
1253 # a reasonably-common case is:
1255 # movl $_blah,<bad-reg>
1259 s/^\tmovl\s+\$${T_US}(.*),(\%e[abcd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1261 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1262 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1263 s/^\tjmp\s+\*(-?\d*)\((.*\%esi.*)\)\n/\tmovl $2,\%eax\n\tjmp \*$1\(\%eax\)\n/g;
1264 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1265 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1266 if /(jmp|call)\s+.*\%esi/;
1268 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1269 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1270 s/^\tjmp\s+\*(-?\d*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1271 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1272 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1273 if /(jmp|call)\s+.*\%edi/;
1276 # OK, now we can decide what our patch-up code is going to
1279 # Offsets into register table - you'd better update these magic
1280 # numbers should you change its contents!
1281 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1282 local($OFFSET_Hp)=92;
1284 # Note funky ".=" stuff; we're *adding* to these _patch guys
1285 if ( $StolenX86Regs <= 2
1286 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1287 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1288 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1290 # nothing for call_{entry,exit} because %esi is callee-save
1292 if ( $StolenX86Regs <= 3
1293 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1294 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1295 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1297 # nothing for call_{entry,exit} because %edi is callee-save
1300 # --------------------------------------------------------
1301 # next, here we go with non-%esp patching!
1303 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1305 # Before calling GC we must set up the exit condition before the call
1306 # and entry condition when we come back
1308 # fix _all_ non-local jumps:
1310 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1311 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1313 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1315 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1316 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1318 if ($StolenX86Regs == 2 ) {
1319 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1320 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1321 } elsif ($StolenX86Regs == 3 ) {
1322 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1323 if /^\t(jmp|call)\s+.*\%edi/;
1326 # --------------------------------------------------------
1327 # that's it -- print it
1329 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1333 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1334 print OUTASM $exit_patch;
1335 # ToDo: make it not print if there is a "jmp" at the end
1341 sub init_FUNNY_THINGS {
1342 %KNOWN_FUNNY_THING = (
1344 # "${T_US}stg_.*{T_POST_LBL}", 1,
1349 The following table reversal is used for both info tables and return
1350 vectors. In both cases, we remove the first entry from the table,
1351 reverse the table, put the label at the end, and paste some code
1352 (that which is normally referred to by the first entry in the table)
1353 right after the table itself. (The code pasting is done elsewhere.)
1357 local($symb, $tbl, $discard1) = @_;
1359 local($before) = '';
1361 local(@imports) = (); # hppa only
1364 local(@lines) = split(/\n/, $tbl);
1365 local($i, $j); #local ($i, $extra, $words_to_pad, $j);
1367 # Deal with the header...
1368 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1369 $label .= $lines[$i] . "\n",
1370 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1371 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1372 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1374 $before .= $lines[$i] . "\n"; # otherwise...
1377 # Grab the table data...
1378 if ( $TargetPlatform !~ /^hppa/ ) {
1379 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1380 push(@words, $lines[$i]);
1382 } else { # hppa weirdness
1383 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1384 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1385 push(@imports, $lines[$i]);
1387 # We don't use HP's ``function pointers''
1388 # We just use labels in code space, like normal people
1389 $lines[$i] =~ s/P%//;
1390 push(@words, $lines[$i]);
1395 # now throw away the first word (SRT) iff it is empty.
1396 # The .zero business is for Linux/ELF.
1397 # The .skip business is for Sparc/Solaris/ELF.
1398 # The .blockz business is for HPPA.
1399 if ($discard1 && $words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1403 # Padding removed to reduce code size and improve performance on Pentiums.
1405 # for 486-cache-friendliness, we want our tables aligned
1406 # on 16-byte boundaries (.align 4). Let's pad:
1407 # $extra = ($#words + 1) % 4;
1408 # $words_to_pad = ($extra == 0) ? 0 : 4 - $extra;
1409 # for ($j = 0; $j < $words_to_pad; $j++) { push(@words, "\t${T_DOT_WORD} 0"); }
1411 for (; $i <= $#lines; $i++) {
1412 $after .= $lines[$i] . "\n";
1415 # Alphas:If we have anonymous text (not part of a procedure), the
1416 # linker may complain about missing exception information. Bleh.
1417 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1418 $before = "\t.ent $1\n" . $before;
1419 $after .= "\t.end $1\n";
1423 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
1424 . join("\n", @words) . "\n"
1427 # print STDERR "before=$before\n";
1428 # print STDERR "label=$label\n";
1429 # print STDERR "words=",(reverse @words),"\n";
1430 # print STDERR "after=$after\n";
1436 The HP is a major nuisance. The threaded code mangler moved info
1437 tables from data space to code space, but unthreaded code in the RTS
1438 still has references to info tables in data space. Since the HP
1439 linker is very precise about where symbols live, we need to patch the
1440 references in the unthreaded RTS as well.
1443 sub mini_mangle_asm_hppa {
1444 local($in_asmf, $out_asmf) = @_;
1446 open(INASM, "< $in_asmf")
1447 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1448 open(OUTASM,"> $out_asmf")
1449 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1452 s/_info,DATA/_info,CODE/; # Move _info references to code space
1458 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1459 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1465 sub mini_mangle_asm_powerpc {
1466 local($in_asmf, $out_asmf) = @_;
1468 open(INASM, "< $in_asmf")
1469 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1470 open(OUTASM,"> $out_asmf")
1471 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1474 s/long _PRStart/long ._PRStart/;
1475 s/long _PRIn_/long ._PRIn_/;
1476 s/long _Dummy_(\S+)_entry/long ._Dummy_\1_entry/;
1477 s/long _PRMarking_MarkNextRoot\[DS\]/long ._PRMarking_MarkNextRoot/;
1478 s/long _PRMarking_MarkNextCAF\[DS\]/long ._PRMarking_MarkNextCAF/;
1479 s/long _PRMarking_MarkNextAStack\[DS\]/long ._PRMarking_MarkNextAStack/;
1480 s/long _PRMarking_MarkNextBStack\[DS\]/long ._PRMarking_MarkNextBStack/;
1481 s/\.tc EnterNodeCode\[TC]\,EnterNodeCode\[DS\]/\.tc EnterNodeCode\[TC]\,.EnterNodeCode/; # CONC
1482 s/\.tc CheckHeapCode\[TC]\,CheckHeapCode\[DS\]/\.tc CheckHeapCode\[TC]\,.CheckHeapCode/; # CONC
1487 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1488 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1491 sub mangle_powerpc_tailjump {
1493 local($maybe_more) = 1;
1494 while (($c =~ /\tlw?z? \d+,LC\.\.\d+\(2\)\n\tmtctr \d+\n\tbctr\n/) && $maybe_more)
1497 $lcsymb =~ s/(.*\n)*\tlw?z? \d+,LC\.\.(\d+)\(2\)\n\tmtctr \d+\n\tbctr\n(.*\n)*/\2/;
1498 # the checks for r1 and r2 are mostly paranoia...
1500 $r1 =~ s/(.*\n)*\tlw?z? (\d+),LC\.\.\d+\(2\)\n\tmtctr \d+\n\tbctr\n(.*\n)*/\2/;
1502 $r2 =~ s/(.*\n)*\tlw?z? \d+,LC\.\.(\d+)\(2\)\n\tmtctr (\d+)\n\tbctr\n(.*\n)*/\3/;
1505 $c =~ s/((.*\n)*)\tlw?z? \d+,LC\.\.\d+\(2\)\n\tmtctr \d+\n\tbctr\n((.*\n)*)/\1\tb $tocequiv{$lcsymb}\n\3/;
1513 sub tidy_up_and_die {
1514 local($return_val, $msg) = @_;
1516 exit (($return_val == 0) ? 0 : 1);