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 tr/\r//d if $TargetPlatform =~ /-mingw32$/;
421 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
422 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
423 next if /^\t\.def.*endef$/;
424 next if /${T_PRE_APP}(NO_)?APP/o;
425 next if /^;/ && $TargetPlatform =~ /^hppa/;
427 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|powerpc|rs6000)-/;
429 last if /^_section_\.text:$/ && $TargetPlatform =~ /^powerpc-|^rs6000-/;
431 if ( $TargetPlatform =~ /^mips-/
432 && /^\t\.(globl \S+ \.text|comm\t)/ ) {
433 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
435 # As a temporary solution for compiling "foreign export" declarations,
436 # we use global variables to pass arguments from C to STG land.
437 # These declarations live in the .hc file and not in the generated C
438 # stub file, so we let them pass through here.
439 } elsif ( /^\t\.comm\t__fexp_.*$/ ) {
441 $chkcat[$i] = 'data';
444 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
445 # duplicated from the bottom
449 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
450 # Alphas: Local labels not to be confused with new chunks
453 # NB: all the rest start with a non-space
455 } elsif ( $TargetPlatform =~ /^mips-/
456 && /^\d+:/ ) { # a funny-looking very-local label
459 } elsif ( /$T_CONST_LBL/o ) {
461 $chkcat[$i] = 'literal';
464 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
466 $chkcat[$i] = 'splitmarker';
469 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
472 $chkcat[$i] = 'infotbl';
473 $chksymb[$i] = $symb;
475 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
477 $infochk{$symb} = $i;
479 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
481 $chkcat[$i] = 'slow';
486 } elsif ( /^${T_US}([A-Za-z0-9_]+)_fast\d*${T_POST_LBL}$/o ) {
488 $chkcat[$i] = 'fast';
493 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
495 $chkcat[$i] = 'closure';
498 $closurechk{$1} = $i;
500 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
507 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
509 $chkcat[$i] = 'data';
512 } elsif ( /^${T_US}ghc.*c_ID${T_POST_LBL}/o ) {
514 $chkcat[$i] = 'consist';
516 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
519 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
520 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
521 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
522 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
525 $chkcat[$i] = 'data';
528 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
533 } elsif ( $TargetPlatform =~ /^powerpc-|^rs6000-/ && /^LC\.\.([0-9]+)/ ) {
538 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
539 # all CC_ symbols go in the data section...
541 $chkcat[$i] = 'data';
544 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
546 $chkcat[$i] = 'misc';
549 #$chksymb[$i] = $symbtmp if ($TargetPlatform =~ /^powerpc-|^rs6000-/) ; #rm andre
551 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
553 $chkcat[$i] = 'vector';
558 # As a temporary solution for compiling "foreign export" declarations,
559 # we use global variables to pass arguments from C to STG land.
560 # These declarations live in the .hc file and not in the generated C
561 # stub file, so we let them pass through here.
562 } elsif ( /^[\t ]+\.comm[\t ]+__fexp_.*$/ ) {
564 $chkcat[$i] = 'data';
567 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
568 && /^(_uname|uname|stat|fstat):/ ) {
569 # for some utterly bizarre reason, this platform
570 # likes to drop little local C routines with these names
571 # into each and every .o file that #includes the
572 # relevant system .h file. Yuck. We just don't
573 # tolerate them in .hc files (which we are processing
574 # here). If you need to call one of these things from
575 # Haskell, make a call to your own C wrapper, then
576 # put that C wrapper (which calls one of these) in a
577 # plain .c file. WDP 95/12
579 $chkcat[$i] = 'toss';
582 } elsif ( /^${T_US}[A-Za-z0-9_]/o
583 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
585 && ( $TargetPlatform !~ /^powerpc|^rs6000/ # ditto
586 || ! /^(L\.\.\d+|LT\.\..*):$/ ) ) {
589 print "Funny global thing?: $_"
590 unless $KNOWN_FUNNY_THING{$thing}
591 || /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
592 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
593 || /^${T_US}__init.*${T_POST_LBL}$/o # __init<module>
594 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
595 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o; # closure tables
597 $chkcat[$i] = 'misc';
598 if ($TargetPlatform =~ /^powerpc-|^rs6000-/)
599 { $chksymb[$i] = $thing; }
600 else { $chksymb[$i] = ''; };
602 } else { # simple line (duplicated at the top)
607 $numchks = $#chk + 1;
609 # the division into chunks is imperfect;
610 # we throw some things over the fence into the next
613 # also, there are things we would like to know
614 # about the whole module before we start spitting
617 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
619 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
621 # Alphas: NB: we start meddling at chunk 1, not chunk 0
622 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
623 # spits a ".quad 0" in after the v first ".rdata"; we
624 # detect this special case (tossing the ".quad 0")!
625 local($magic_rdata_seen) = 0;
627 # HPPAs, MIPSen: also start medding at chunk 1
629 if ($TargetPlatform =~ /^powerpc|^rs6000/) {
630 print OUTASM $T_HDR_toc; # yes, we have to put a .toc
631 # in the beginning of every file!
632 %tocequiv = (); # maps toc symbol number to toc symbol
633 %revtocequiv = (); # maps toc symbol to toc symbol number
634 for ($i = 1; $i < $numchks; $i++) {
635 $chk[$i] =~ s/\[RW\]//g;
636 $chk[$i] =~ s/\[DS\]//g;
637 $chk[$i] =~ s/^\.csect .*\[DS\]$//g;
639 if ( $chkcat[$i] eq 'toc' && $chk[$i] !~ /\.byte/ )
640 #ToDo: instead of all these changes, just fix mangle_powerpc_tailjump and delete/ignore these tocs?
641 { $chk[$i] =~ s/$T_MOVE_DIRVS//g;
642 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(\S+_fast\d+)/\t\.tc \1\[TC\],\.\2/;
643 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(\S+_entry)\n/\t\.tc \1\[TC\],\.\2\n/;
644 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(ret_\S+)/\t\.tc \1\[TC\],\.\2/;
645 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(alt_\S+)/\t\.tc \1\[TC\],\.\2/;
646 $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(vtbl_\S+)/\t\.tc \1\[TC\],\.\2/;
648 $tocnumber = $chksymb[$i];
650 $tocsymb =~ s/^LC\.\.\d+:\n//;
651 $tocsymb =~ s/^\t\.tc \S+,(\S+)\n/\1/;
652 $tocequiv{$tocnumber} = $tocsymb;
654 } elsif ( $chkcat[$i] eq 'toc' && $chk[$i] =~ /\.byte/ ) {
655 $chkcat[$i] = 'literal';
660 for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) {
661 $c = $chk[$i]; # convenience copy
663 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
665 # toss all prologue stuff; HPPA is pretty weird
667 $c = &mash_hppa_prologue($c) if $TargetPlatform =~ /^hppa/;
669 # be slightly paranoid to make sure there's
670 # nothing surprising in there
671 if ( $c =~ /--- BEGIN ---/ ) {
672 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
674 if ($TargetPlatform =~ /^i386-/) {
675 $p =~ s/^\tpushl\s+\%edi\n//;
676 $p =~ s/^\tpushl\s+\%esi\n//;
677 $p =~ s/^\tpushl\s+\%ebx\n//;
678 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
679 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-cygwin32/);
680 } elsif ($TargetPlatform =~ /^m68k-/) {
681 $p =~ s/^\tlink a6,#-?\d.*\n//;
682 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
683 # The above showed up in the asm code,
684 # so I added it here.
685 # I hope it's correct.
687 $p =~ s/^\tmovel d2,sp\@-\n//;
688 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
689 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
690 } elsif ($TargetPlatform =~ /^mips-/) {
691 # the .frame/.mask/.fmask that we use is the same
692 # as that produced by GCC for miniInterpret; this
693 # gives GDB some chance of figuring out what happened
694 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
695 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
696 $p =~ s/^\t\.(mask|fmask).*\n//g;
697 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
698 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
699 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
700 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
701 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
702 $p =~ s/__FRAME__/$FRAME/;
703 } elsif ($TargetPlatform =~ /^powerpc-|^rs6000/) {
704 $p =~ s/^\tmflr 0\n//;
705 $p =~ s/^\tstm \d+,-\d+\(1\)\n//;
706 $p =~ s/^\tstw? 0,\d+\(1\)\n//g;
707 $p =~ s/^\tstw? 1,\d+\(1\)\n//g; #mc
708 $p =~ s/^\tlw?z 0,0\(1\)\n//g; #mc
709 $p =~ s/^\tstw?u 1,-\d+\(1\)\n//;
710 $p =~ s/^\tstw? \d+,-\d+\(1\)\n//g;
711 $p =~ s/^\tstfd \d+,-\d+\(1\)\n//g;
713 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
716 # HWL HACK: dont die, just print a warning
717 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/
718 # && $TargetPlatform !~ /^powerpc-/; #ToDo: remove test
719 die "Prologue junk?: $p\n" if $p =~ /^\t[^\.]/
720 && $TargetPlatform !~ /^powerpc-/; #ToDo: remove test
722 # glue together what's left
724 $c =~ s/\n\t\n/\n/; # junk blank line
728 if ( $TargetPlatform =~ /^mips-/ ) {
729 # MIPS: first, this basic sequence may occur "--- END ---" or not
730 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
733 # toss all epilogue stuff; again, paranoidly
734 if ( $c =~ /--- END ---/ ) {
735 if (($r, $e) = split(/--- END ---/, $c)) {
736 if ($TargetPlatform =~ /^i386-/) {
738 $e =~ s/^\tpopl\s+\%edi\n//;
739 $e =~ s/^\tpopl\s+\%esi\n//;
740 $e =~ s/^\tpopl\s+\%edx\n//;
741 $e =~ s/^\tpopl\s+\%ecx\n//;
742 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
743 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
744 } elsif ($TargetPlatform =~ /^m68k-/) {
745 $e =~ s/^\tunlk a6\n//;
747 } elsif ($TargetPlatform =~ /^mips-/) {
748 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
749 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
750 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
751 $e =~ s/^\tj\t\$31\n//;
752 } elsif ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
753 $e =~ s/^\taddi 1,1,\d+\n//;
754 $e =~ s/^\tcal 1,\d+\(1\)\n//;
755 $e =~ s/^\tlw?z? \d+,\d+\(1\)\n//;
756 $e =~ s/^\tmtlr 0\n//;
759 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
762 print "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/
763 && $TargetPlatform !~ /^powerpc-/; #ToDo: remove test
765 # glue together what's left
767 $c =~ s/\n\t\n/\n/; # junk blank line
771 # On SPARCs, we don't do --- BEGIN/END ---, we just
772 # toss the register-windowing save/restore/ret* instructions
774 if ( $TargetPlatform =~ /^sparc-/ ) {
775 $c =~ s/^\t(save.*|restore|ret|retl)\n//g;
776 # throw away PROLOGUE comments
777 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
780 # On Alphas, the prologue mangling is done a little later (below)
782 # toss all calls to __DISCARD__
783 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
785 # MIPS: that may leave some gratuitous asm macros around
786 # (no harm done; but we get rid of them to be tidier)
787 $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/
788 if $TargetPlatform =~ /^mips-/;
790 # toss stack adjustment after DoSparks
791 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
792 if $TargetPlatform =~ /^m68k-/; # this looks old...
794 if ( $TargetPlatform =~ /^alpha-/ &&
795 ! $magic_rdata_seen &&
796 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
797 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
798 $magic_rdata_seen = 1;
801 # pick some end-things and move them to the next chunk
803 # pin a funny end-thing on (for easier matching):
804 $c .= 'FUNNY#END#THING';
806 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
810 # on x86 we try not to copy any directives into a literal
811 # chunk, rather we keep looking for the next real chunk. This
812 # is because we get things like
814 # .globl blah_closure
820 if ( $TargetPlatform =~ /^(i386|sparc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
822 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
825 if ( $j < $numchks ) {
826 $chk[$j] = $to_move . $chk[$j];
830 elsif ( $i < ($numchks - 1)
831 && ( $to_move =~ /${T_COPY_DIRVS}/
832 || ($TargetPlatform =~ /^hppa/ && $to_move =~ /align/ && $chkcat[$i+1] eq 'literal') )) {
833 $chk[$i + 1] = $to_move . $chk[$i + 1];
834 # otherwise they're tossed
837 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
840 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
842 # toss all prologue stuff, except for loading gp, and the ..ng address
843 if (($p, $r) = split(/^\t\.prologue/, $c)) {
844 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
845 $c = $keep . "..ng:\n";
847 print STDERR "malformed code block ($ent)?\n"
850 $c .= "\t.frame \$30,0,\$26,0\n\t.prologue" . $r;
853 $c =~ s/FUNNY#END#THING//;
855 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
857 $chk[$i] = $c; # update w/ convenience copy
860 if ( $TargetPlatform =~ /^alpha-/ ) {
861 # print out the header stuff first
862 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
863 print OUTASM $chk[0];
865 } elsif ( $TargetPlatform =~ /^hppa/ ) {
866 print OUTASM $chk[0];
868 } elsif ( $TargetPlatform =~ /^mips-/ ) {
869 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
871 # get rid of horrible "<dollar>Revision: .*$" strings
872 local(@lines0) = split(/\n/, $chk[0]);
874 while ( $z <= $#lines0 ) {
875 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
878 while ( $z <= $#lines0 ) {
880 last if $lines0[$z] =~ /[,\t]0x0$/;
886 $chk[0] = join("\n", @lines0);
887 $chk[0] =~ s/\n\n+/\n/;
888 print OUTASM $chk[0];
891 # print out all the literal strings next
892 for ($i = 0; $i < $numchks; $i++) {
893 if ( $chkcat[$i] eq 'literal' ) {
894 print OUTASM $T_HDR_literal, $chk[$i];
895 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
897 $chkcat[$i] = 'DONE ALREADY';
901 # on the HPPA, print out all the bss next
902 if ( $TargetPlatform =~ /^hppa/ ) {
903 for ($i = 1; $i < $numchks; $i++) {
904 if ( $chkcat[$i] eq 'bss' ) {
905 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
906 print OUTASM $chk[$i];
908 $chkcat[$i] = 'DONE ALREADY';
913 for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) {
914 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
916 next if $chkcat[$i] eq 'DONE ALREADY';
918 if ( $chkcat[$i] eq 'misc' ) {
919 if ($chk[$i] ne '') {
920 print OUTASM $T_HDR_misc;
921 if ($TargetPlatform =~ /^powerpc-|^rs6000/) {
922 $chksymb[$i] =~ s/://;
923 #new if ($chksymb[$i] =~ /ret.*upd/ || $KNOWN_FUNNY_THING{$chksymb[$i]}
924 #new || $chksymb[$i] =~ /^$.{T_US}_(PRIn|PRStart).*${T_POST_LBL}$/o )
925 #new { print OUTASM "\t\.globl $chksymb[$i]\n"; }
926 # if ($chksymb[$i] ne '' && $chksymb[$i] !~ /ret_[a-z]/ && $chksymb[$i] !~ /djn_[a-z]/)
927 if ($chksymb[$i] ne '')
928 { print OUTASM "\t\.globl \.$chksymb[$i]\n"; };
929 if ($chk[$i] =~ /TOC\[tc0\], 0\n/)
930 { ($p, $r) = split(/TOC\[tc0\], 0\n/, $chk[$i]); $printDS = 1;}
931 else { $r = $chk[$i]; $printDS = 0; };
932 $chk[$i] = &mangle_powerpc_tailjump($r);
934 &print_doctored($chk[$i], 0);
935 if ($TargetPlatform =~ /^powerpc-|^rs6000-/ && $printDS) {
936 #ok if ($chksymb[$i] !~ /\__init_Main/) {
937 print OUTASM "\.csect ${chksymb[$i]}[DS]\n";
938 print OUTASM "${p}TOC[tc0], 0\n";
943 } elsif ( $chkcat[$i] eq 'toss' ) {
944 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
946 } elsif ( $chkcat[$i] eq 'data' ) {
947 if ($chk[$i] ne '') {
948 print OUTASM $T_HDR_data;
949 print OUTASM $chk[$i];
952 } elsif ( $chkcat[$i] eq 'consist' ) {
953 if ( $chk[$i] =~ /$T_hsc_cc_PAT/o ) {
954 local($consist) = "$1.$2.$3";
956 $consist =~ s/\//./g;
958 $consist =~ s/[^A-Za-z0-9_.]/ZZ/g; # ToDo: properly?
960 # Using a cygnus-2.7-96q4 gcc build on hppas, the
961 # consistency chunk for ghc_cc_ID often (but not always!)
962 # gets lumped with a bunch of .IMPORT directives containing info on
963 # the code or data space nature of external symbols. We can't
964 # toss these, so once the consistency ID has been turned into
965 # a representable symbol, we substitute it for the symbol
966 # that the string was attached to in the first place (ghc_cc_ID.)
967 # (The original string is also substituted away.)
969 # This change may affect the code output on other platforms in
970 # adverse ways, hence we restrict this hack hppa targets only.
973 if ( $TargetPlatform =~ /^hppa/ ) {
974 $chk[$i] =~ s/^${T_US}ghc.*c_ID$TPOSTLBL/$consist/o;
975 $chk[$i] =~ s/\t$T_hsc_cc_PAT/$T_HDR_misc/o;
976 $consist = $chk[$i]; #clumsily
978 print OUTASM $T_HDR_consist, "${consist}${T_POST_LBL}\n";
980 } elsif ( $TargetPlatform !~ /^(mips)-/ ) { # we just don't try in those case (ToDo)
981 # on mips: consistency string is just a v
982 # horrible bunch of .bytes,
983 # which I am too lazy to sort out (WDP 95/05)
985 print STDERR "Couldn't grok consistency: ", $chk[$i];
988 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
989 # we can just re-constitute this one...
990 # NB: we emit _three_ underscores no matter what,
991 # so ghc-split doesn't have to care.
992 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
994 } elsif ( $chkcat[$i] eq 'closure'
995 || $chkcat[$i] eq 'srt'
996 || $chkcat[$i] eq 'infotbl'
997 || $chkcat[$i] eq 'slow'
998 || $chkcat[$i] eq 'fast' ) { # do them in that order
999 $symb = $chksymb[$i];
1002 if ( defined($closurechk{$symb}) ) {
1003 print OUTASM $T_HDR_closure;
1004 print OUTASM $chk[$closurechk{$symb}];
1005 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1009 if ( defined($srtchk{$symb}) ) {
1010 print OUTASM $T_HDR_srt;
1011 print OUTASM $chk[$srtchk{$symb}];
1012 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1016 if ( defined($infochk{$symb}) ) {
1018 print OUTASM $T_HDR_info;
1019 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1020 if ( !defined($slowchk{$symb}) && defined($fastchk{$symb}) ) {
1021 $fastname = $chk[$fastchk{$symb}];
1022 $fastname =~ s/([_A-Za-z]*_fast\d+):.*(.*\n)*/\1/;
1023 $chk[$infochk{$symb}] =~ s/\.long StdErrorCode/\.long $fastname/;
1025 $chk[$infochk{$symb}] =~ s/\.long ([_A-Za-z]\S+_entry)/\.long \.\1/;
1026 $chk[$infochk{$symb}] =~ s/\.long ([A-Za-z]\S+_upd)/\.long \.\1/;
1027 print OUTASM $chk[$infochk{$symb}];
1029 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1031 # entry code will be put here!
1033 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1037 if ( defined($slowchk{$symb}) ) {
1039 # teach it to drop through to the fast entry point:
1040 $c = $chk[$slowchk{$symb}];
1042 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1043 ($p, $r) = split(/TOC\[tc0\], 0\n/, $c);
1044 if ($symb =~ /^[_A-Z]/)
1046 print OUTASM "\t\.globl \.${chksymb[$i]}_entry\n";
1047 print OUTASM "\.csect ${symb}_entry[DS]\n";
1048 print OUTASM "${p}TOC[tc0], 0\n";
1050 $r =~ s/\.csect \.text\[PR\]\n//; # todo: properly - andre
1051 $c = &mangle_powerpc_tailjump($r);
1054 if ( defined($fastchk{$symb}) ) {
1055 if ( $TargetPlatform =~ /^alpha-/ ) {
1056 $c =~ s/^\tjmp \$31,\(\$27\),0\n\t\.align 4\n\t\.end/\t.align 4\n\t.end/;
1057 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1058 $c =~ s/^\s+ldil.*\n\s+ldo.*\n\s+bv.*\n(.*\n)?\s+\.EXIT/$1\t.EXIT/;
1059 } elsif ( $TargetPlatform =~ /^i386-/ ) {
1060 # Reg alloc depending, gcc generated code may jump to the fast entry point via
1061 # a number of registers.
1062 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%edx\n\tjmp\s+\*\%edx\n//;
1063 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%ecx\n\tjmp\s+\*\%ecx\n//;
1064 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%eax\n\tjmp\s+\*\%eax\n//;
1065 # The next two only apply if we're not stealing %esi or %edi.
1066 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%esi\n\tjmp\s+\*\%esi\n// if ($StolenX86Regs < 3);
1067 $c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%edi\n\tjmp\s+\*\%edi\n// if ($StolenX86Regs < 4);
1068 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1069 $c =~ s/^\tjmp \$31,\(\$27\),0\n\t\.align 4\n\t\.end/\t.align 4\n\t.end/;
1070 } elsif ( $TargetPlatform =~ /^m68k-/ ) {
1071 $c =~ s/^\tjmp ${T_US}${symb}_fast\d+.*\n\tnop\n//;
1072 $c =~ s/^\tjmp ${T_US}${symb}_fast\d+.*\n//;
1073 } elsif ( $TargetPlatform =~ /^powerpc-|^rs6000-/ ) {
1074 $c =~ s/^\tb \.${T_US}${symb}_fast\d+\n//;
1075 } elsif ( $TargetPlatform =~ /^sparc-/ ) {
1076 $c =~ s/^\tcall\s+${T_US}${symb}_fast\d+,.*\n\t\s*nop\n//;
1077 $c =~ s/^\tcall\s+${T_US}${symb}_fast\d+,.*\n(\t\s*[a-z].*\n)/$1/;
1079 print STDERR "$Pgm: mystery slow-fast dropthrough: $TargetPlatform\n";
1083 if ( $TargetPlatform !~ /^(alpha-|hppa|mips-)/ ) {
1084 # On alphas, hppa: no very good way to look for "dangling"
1085 # references to fast-entry point.
1086 # (questionable re hppa and mips...)
1087 print STDERR "still has jump to fast entry point:\n$c"
1088 if $c =~ /\b${T_US}${symb}_fast/;
1091 print OUTASM $T_HDR_entry;
1093 &print_doctored($c, 1); # NB: the 1!!!
1095 $chkcat[$slowchk{$symb}] = 'DONE ALREADY';
1099 if ( defined($fastchk{$symb}) ) {
1100 $c = $chk[$fastchk{$symb}];
1101 if ( ! defined($slowchk{$symb})
1102 # ToDo: the || clause can go once we're no longer
1103 # concerned about producing exactly the same output as before
1104 #OLD: || $TargetPlatform =~ /^(m68k|sparc|i386)-/
1106 print OUTASM $T_HDR_fast;
1108 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1109 local(@lbls) = split(/:/, $c);
1110 $fullname = $lbls[0];
1111 $fullname =~ s/$T_MOVE_DIRVS//g;
1112 if ( $fullname =~ /^[A-Z]/)
1113 { print OUTASM "\t\.globl \.${fullname}\n";
1115 # print OUTASM "\t\.lglobl \.${fullname}\n"; #todo: rm - andre
1117 $c =~ s/((.*\n)*)\t.long \S+, TOC\[tc0\], 0\n\.csect \.text\[PR\]\n((.*\n)*)/\1\3/;
1118 $c = &mangle_powerpc_tailjump($c);
1120 &print_doctored($c, 0);
1121 $chkcat[$fastchk{$symb}] = 'DONE ALREADY';
1124 } elsif ( $chkcat[$i] eq 'vector'
1125 || $chkcat[$i] eq 'direct' ) { # do them in that order
1126 $symb = $chksymb[$i];
1129 if ( defined($vectorchk{$symb}) ) {
1130 print OUTASM $T_HDR_vector;
1131 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1132 if ( $symb =~ /^[A-Z]/) {
1133 print OUTASM "\t\.globl \.${symb}_vtbl\n";
1134 print OUTASM "\t\.globl ${symb}_vtbl\n";
1136 $chk[$vectorchk{$symb}] =~ s/\.long (\S+)/\.long \.\1/g;
1137 print OUTASM ".${symb}_vtbl:\n";
1138 print OUTASM $chk[$vectorchk{$symb}];
1140 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1142 # direct return code will be put here!
1143 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1145 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1146 # Alphas: the commented nop is for the splitter, to ensure
1147 # that no module ends with a label as the very last
1148 # thing. (The linker will adjust the label to point
1149 # to the first code word of the next module linked in,
1150 # even if alignment constraints cause the label to move!)
1152 print OUTASM "\t# nop\n";
1155 } elsif ( $chkcat[$i] eq 'toc' ) {
1156 # silly optimisation to print tocs, since they come in groups...
1157 print OUTASM $T_HDR_toc;
1159 while ($chkcat[$j] eq 'toc')
1160 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1163 print OUTASM $chk[$j];
1165 $chkcat[$j] = 'DONE ALREADY';
1170 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1174 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1176 if ($TargetPlatform =~ /^powerpc-|^rs6000-/) {
1177 print OUTASM ".csect .text[PR]\n_section_.text:\n.csect .data[RW]\n\t.long _section_.text\n"
1181 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1182 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1187 sub mash_hppa_prologue { # OK, epilogue, too
1190 # toss all prologue stuff
1191 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1193 # Lie about our .CALLINFO
1194 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1201 # toss all epilogue stuff
1202 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1204 # Sorry; we moved the _info stuff to the code segment.
1205 s/_info,DATA/_info,CODE/g;
1212 sub print_doctored {
1213 local($_, $need_fallthru_patch) = @_;
1215 if ( $TargetPlatform !~ /^i386-/
1216 || ! /^\t[a-z]/ # no instructions in here, apparently
1217 || /^${T_US}__init_[A-Za-z0-9_]+${T_POST_LBL}/) {
1221 # OK, must do some x86 **HACKING**
1223 local($entry_patch) = '';
1224 local($exit_patch) = '';
1226 # gotta watch out for weird instructions that
1227 # invisibly smash various regs:
1228 # rep* %ecx used for counting
1229 # scas* %edi used for destination index
1230 # cmps* %e[sd]i used for indices
1231 # loop* %ecx used for counting
1236 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1238 # * GCC used an "STG reg" for its own purposes
1240 # * some secret uses of machine reg, requiring STG reg
1241 # to be saved/restored
1243 # The most dangerous "GCC uses" of an "STG reg" are when
1244 # the reg holds the target of a jmp -- it's tricky to
1245 # insert the patch-up code before we get to the target!
1246 # So here we change the jmps:
1248 # --------------------------------------------------------
1249 # it can happen that we have jumps of the form...
1250 # jmp *<something involving %esp>
1252 # jmp <something involving another naughty register...>
1254 # a reasonably-common case is:
1256 # movl $_blah,<bad-reg>
1260 s/^\tmovl\s+\$${T_US}(.*),(\%e[abcd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1262 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1263 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1264 s/^\tjmp\s+\*(-?\d*)\((.*\%esi.*)\)\n/\tmovl $2,\%eax\n\tjmp \*$1\(\%eax\)\n/g;
1265 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1266 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1267 if /(jmp|call)\s+.*\%esi/;
1269 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1270 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1271 s/^\tjmp\s+\*(-?\d*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1272 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1273 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1274 if /(jmp|call)\s+.*\%edi/;
1277 # OK, now we can decide what our patch-up code is going to
1280 # Offsets into register table - you'd better update these magic
1281 # numbers should you change its contents!
1282 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1283 local($OFFSET_Hp)=92;
1285 # Note funky ".=" stuff; we're *adding* to these _patch guys
1286 if ( $StolenX86Regs <= 2
1287 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1288 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1289 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1291 # nothing for call_{entry,exit} because %esi is callee-save
1293 if ( $StolenX86Regs <= 3
1294 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1295 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1296 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1298 # nothing for call_{entry,exit} because %edi is callee-save
1301 # --------------------------------------------------------
1302 # next, here we go with non-%esp patching!
1304 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1306 # Before calling GC we must set up the exit condition before the call
1307 # and entry condition when we come back
1309 # fix _all_ non-local jumps:
1311 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1312 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1314 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1316 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1317 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1319 if ($StolenX86Regs == 2 ) {
1320 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1321 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1322 } elsif ($StolenX86Regs == 3 ) {
1323 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1324 if /^\t(jmp|call)\s+.*\%edi/;
1327 # --------------------------------------------------------
1328 # that's it -- print it
1330 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1334 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1335 print OUTASM $exit_patch;
1336 # ToDo: make it not print if there is a "jmp" at the end
1342 sub init_FUNNY_THINGS {
1343 %KNOWN_FUNNY_THING = (
1345 # "${T_US}stg_.*{T_POST_LBL}", 1,
1350 The following table reversal is used for both info tables and return
1351 vectors. In both cases, we remove the first entry from the table,
1352 reverse the table, put the label at the end, and paste some code
1353 (that which is normally referred to by the first entry in the table)
1354 right after the table itself. (The code pasting is done elsewhere.)
1358 local($symb, $tbl, $discard1) = @_;
1360 local($before) = '';
1362 local(@imports) = (); # hppa only
1365 local(@lines) = split(/\n/, $tbl);
1366 local($i, $j); #local ($i, $extra, $words_to_pad, $j);
1368 # Deal with the header...
1369 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1370 $label .= $lines[$i] . "\n",
1371 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1372 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1373 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1375 $before .= $lines[$i] . "\n"; # otherwise...
1378 # Grab the table data...
1379 if ( $TargetPlatform !~ /^hppa/ ) {
1380 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1381 push(@words, $lines[$i]);
1383 } else { # hppa weirdness
1384 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1385 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1386 push(@imports, $lines[$i]);
1388 # We don't use HP's ``function pointers''
1389 # We just use labels in code space, like normal people
1390 $lines[$i] =~ s/P%//;
1391 push(@words, $lines[$i]);
1396 # now throw away the first word (SRT) iff it is empty.
1397 # The .zero business is for Linux/ELF.
1398 # The .skip business is for Sparc/Solaris/ELF.
1399 # The .blockz business is for HPPA.
1400 if ($discard1 && $words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1404 # Padding removed to reduce code size and improve performance on Pentiums.
1406 # for 486-cache-friendliness, we want our tables aligned
1407 # on 16-byte boundaries (.align 4). Let's pad:
1408 # $extra = ($#words + 1) % 4;
1409 # $words_to_pad = ($extra == 0) ? 0 : 4 - $extra;
1410 # for ($j = 0; $j < $words_to_pad; $j++) { push(@words, "\t${T_DOT_WORD} 0"); }
1412 for (; $i <= $#lines; $i++) {
1413 $after .= $lines[$i] . "\n";
1416 # Alphas:If we have anonymous text (not part of a procedure), the
1417 # linker may complain about missing exception information. Bleh.
1418 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1419 $before = "\t.ent $1\n" . $before;
1420 $after .= "\t.end $1\n";
1424 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
1425 . join("\n", @words) . "\n"
1428 # print STDERR "before=$before\n";
1429 # print STDERR "label=$label\n";
1430 # print STDERR "words=",(reverse @words),"\n";
1431 # print STDERR "after=$after\n";
1437 The HP is a major nuisance. The threaded code mangler moved info
1438 tables from data space to code space, but unthreaded code in the RTS
1439 still has references to info tables in data space. Since the HP
1440 linker is very precise about where symbols live, we need to patch the
1441 references in the unthreaded RTS as well.
1444 sub mini_mangle_asm_hppa {
1445 local($in_asmf, $out_asmf) = @_;
1447 open(INASM, "< $in_asmf")
1448 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1449 open(OUTASM,"> $out_asmf")
1450 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1453 s/_info,DATA/_info,CODE/; # Move _info references to code space
1459 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1460 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1466 sub mini_mangle_asm_powerpc {
1467 local($in_asmf, $out_asmf) = @_;
1469 open(INASM, "< $in_asmf")
1470 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1471 open(OUTASM,"> $out_asmf")
1472 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1475 s/long _PRStart/long ._PRStart/;
1476 s/long _PRIn_/long ._PRIn_/;
1477 s/long _Dummy_(\S+)_entry/long ._Dummy_\1_entry/;
1478 s/long _PRMarking_MarkNextRoot\[DS\]/long ._PRMarking_MarkNextRoot/;
1479 s/long _PRMarking_MarkNextCAF\[DS\]/long ._PRMarking_MarkNextCAF/;
1480 s/long _PRMarking_MarkNextAStack\[DS\]/long ._PRMarking_MarkNextAStack/;
1481 s/long _PRMarking_MarkNextBStack\[DS\]/long ._PRMarking_MarkNextBStack/;
1482 s/\.tc EnterNodeCode\[TC]\,EnterNodeCode\[DS\]/\.tc EnterNodeCode\[TC]\,.EnterNodeCode/; # CONC
1483 s/\.tc CheckHeapCode\[TC]\,CheckHeapCode\[DS\]/\.tc CheckHeapCode\[TC]\,.CheckHeapCode/; # CONC
1488 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1489 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1492 sub mangle_powerpc_tailjump {
1494 local($maybe_more) = 1;
1495 while (($c =~ /\tlw?z? \d+,LC\.\.\d+\(2\)\n\tmtctr \d+\n\tbctr\n/) && $maybe_more)
1498 $lcsymb =~ s/(.*\n)*\tlw?z? \d+,LC\.\.(\d+)\(2\)\n\tmtctr \d+\n\tbctr\n(.*\n)*/\2/;
1499 # the checks for r1 and r2 are mostly paranoia...
1501 $r1 =~ s/(.*\n)*\tlw?z? (\d+),LC\.\.\d+\(2\)\n\tmtctr \d+\n\tbctr\n(.*\n)*/\2/;
1503 $r2 =~ s/(.*\n)*\tlw?z? \d+,LC\.\.(\d+)\(2\)\n\tmtctr (\d+)\n\tbctr\n(.*\n)*/\3/;
1506 $c =~ s/((.*\n)*)\tlw?z? \d+,LC\.\.\d+\(2\)\n\tmtctr \d+\n\tbctr\n((.*\n)*)/\1\tb $tocequiv{$lcsymb}\n\3/;
1514 sub tidy_up_and_die {
1515 local($return_val, $msg) = @_;
1517 exit (($return_val == 0) ? 0 : 1);