1 %************************************************************************
3 \section[Driver-asm-fiddling]{Fiddling with assembler files}
5 %************************************************************************
10 Utterly stomp out C functions' prologues and epilogues; i.e., the
11 stuff to do with the C stack.
13 Any other required tidying up.
16 General note [chak]: Many regexps are very fragile because they rely on white
17 space being in the right place. This caused trouble with gcc 2.95 (at least
18 on Linux), where the use of white space in .s files generated by gcc suddenly
19 changed. To guarantee compatibility across different versions of gcc, make
20 sure (at least on i386-.*-linux) that regexps tolerate varying amounts of white
21 space between an assembler statement and its arguments as well as after a the
22 comma separating multiple arguments.
24 \emph{For the time being, I have corrected the regexps for i386-.*-linux. I
25 didn't touch all the regexps for other i386 platforms, as I don't have
26 a box to test these changes.}
31 The HP linker is very picky about symbols being in the appropriate
32 space (code vs. data). When we mangle the threaded code to put the
33 info tables just prior to the code, they wind up in code space
34 rather than data space. This means that references to *_info from
35 un-mangled parts of the RTS (e.g. unthreaded GC code) get
36 unresolved symbols. Solution: mini-mangler for .c files on HP. I
37 think this should really be triggered in the driver by a new -rts
38 option, so that user code doesn't get mangled inappropriately.
40 With reversed tables, jumps are to the _info label rather than to
41 the _entry label. The _info label is just an address in code
42 space, rather than an entry point with the descriptive blob we
43 talked about yesterday. As a result, you can't use the call-style
44 JMP_ macro. However, some JMP_ macros take _info labels as targets
45 and some take code entry points within the RTS. The latter won't
46 work with the goto-style JMP_ macro. Sigh. Solution: Use the goto
47 style JMP_ macro, and mangle some more assembly, changing all
48 "RP'literal" and "LP'literal" references to "R'literal" and
49 "L'literal," so that you get the real address of the code, rather
50 than the descriptive blob. Also change all ".word P%literal"
51 entries in info tables and vector tables to just ".word literal,"
52 for the same reason. Advantage: No more ridiculous call sequences.
55 %************************************************************************
57 \subsection{Top-level code}
59 %************************************************************************
62 $TargetPlatform = $TARGETPLATFORM;
64 ($Pgm = $0) =~ s|.*/||;
68 if ( $TargetPlatform =~ /^i386-/ ) {
72 $StolenX86Regs = $ARGV[2];
76 &mangle_asm($ifile,$ofile);
81 %************************************************************************
83 \subsection{Constants for various architectures}
85 %************************************************************************
88 sub init_TARGET_STUFF {
90 #--------------------------------------------------------#
91 if ( $TargetPlatform =~ /^alpha-.*-.*/ ) {
93 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
94 $T_US = ''; # _ if symbols have an underscore on the front
95 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
96 $T_CONST_LBL = '^\$L?C(\d+):$'; # regexp for what such a lbl looks like
99 $T_MOVE_DIRVS = '^(\s*(\$.*\.\.ng:|\.align\s+\d+|\.(globl|ent)\s+\S+|\#.*|\.(file|loc)\s+\S+\s+\S+|\.text|\.r?data)\n)';
100 $T_COPY_DIRVS = '^\s*(\$.*\.\.ng:|\#|\.(file|globl|ent|loc))';
102 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
103 $T_DOT_WORD = '\.(long|quad|byte|word)';
104 $T_DOT_GLOBAL = '^\t\.globl';
105 $T_HDR_literal = "\.rdata\n\t\.align 3\n";
106 $T_HDR_misc = "\.text\n\t\.align 3\n";
107 $T_HDR_data = "\.data\n\t\.align 3\n";
108 $T_HDR_consist = "\.text\n";
109 $T_HDR_closure = "\.data\n\t\.align 3\n";
110 $T_HDR_srt = "\.text\n\t\.align 3\n";
111 $T_HDR_info = "\.text\n\t\.align 3\n";
112 $T_HDR_entry = "\.text\n\t\.align 3\n";
113 $T_HDR_vector = "\.text\n\t\.align 3\n";
114 $T_HDR_direct = "\.text\n\t\.align 3\n";
116 #--------------------------------------------------------#
117 } elsif ( $TargetPlatform =~ /^hppa/ ) {
119 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
120 $T_US = ''; # _ if symbols have an underscore on the front
121 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
122 $T_CONST_LBL = '^L\$C(\d+)$'; # regexp for what such a lbl looks like
125 $T_MOVE_DIRVS = '^((\s+\.(IMPORT|EXPORT|PARAM).*|\s+\.align\s+\d+|\s+\.(SPACE|SUBSPA)\s+\S+|\s*)\n)';
126 $T_COPY_DIRVS = '^\s+\.(IMPORT|EXPORT)';
128 $T_hsc_cc_PAT = '\.STRING.*\)(hsc|cc) (.*)\\\\x09(.*)\\\\x00"';
129 $T_DOT_WORD = '\.(blockz|word|half|byte)';
130 $T_DOT_GLOBAL = '^\s+\.EXPORT';
131 $T_HDR_literal = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
132 $T_HDR_misc = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
133 $T_HDR_data = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
134 $T_HDR_consist = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
135 $T_HDR_closure = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
136 $T_HDR_srt = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
137 $T_HDR_info = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
138 $T_HDR_entry = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
139 $T_HDR_vector = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
140 $T_HDR_direct = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
142 #--------------------------------------------------------#
143 } elsif ( $TargetPlatform =~ /^i386-.*-(linuxaout|freebsd2|nextstep3|cygwin32|mingw32)$/ ) {
144 # NeXT added but not tested. CaS
146 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
147 $T_US = '_'; # _ if symbols have an underscore on the front
148 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
149 $T_CONST_LBL = '^LC(\d+):$';
151 $T_X86_PRE_LLBL_PAT = 'L';
152 $T_X86_PRE_LLBL = 'L';
153 $T_X86_BADJMP = '^\tjmp [^L\*]';
155 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s+\d+(,0x90)?|\.globl\s+\S+|\.text|\.data|\.stab[^n].*|\.type\s+.*|\.size\s+.*)\n)';
156 $T_COPY_DIRVS = '\.(globl|stab)';
157 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
158 $T_DOT_WORD = '\.(long|word|value|byte|space)';
159 $T_DOT_GLOBAL = '\.globl';
160 $T_HDR_literal = "\.text\n\t\.align 2\n";
161 $T_HDR_misc = "\.text\n\t\.align 2,0x90\n";
162 $T_HDR_data = "\.data\n\t\.align 2\n";
163 $T_HDR_consist = "\.text\n";
164 $T_HDR_closure = "\.data\n\t\.align 2\n";
165 $T_HDR_srt = "\.text\n\t\.align 2\n";
166 $T_HDR_info = "\.text\n\t\.align 2\n"; # NB: requires padding
167 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
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 ( $TargetPlatform =~ /^i386-.*-(solaris2|linux|freebsd|netbsd|openbsd)$/ ) {
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)$/) ? '#' : '/' ;
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+.*|\.size\s+\S+\s*,\s*\d+|\.ident.*|\.local.*)\n)';
185 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
187 if ( $TargetPlatform =~ /freebsd|netbsd/ ) {
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|word|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_vector = "\.text\n\t\.align 4\n"; # NB: requires padding
204 $T_HDR_direct = "\.text\n\t\.align 4\n";
206 #--------------------------------------------------------#
207 } elsif ( $TargetPlatform =~ /^ia64-.*-linux$/ ) {
209 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
210 $T_US = ''; # _ if symbols have an underscore on the front
212 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
215 $T_MOVE_DIRVS = '^(\s*\.(global|proc|pred\.safe_across_calls|text|data|section|subsection|align|size|type|ident)\s+.*\n)';
216 $T_COPY_DIRVS = '\.(global|proc)';
218 $T_hsc_cc_PAT = '\.string.*\)(hsc|cc) (.*)\\\\t(.*)"';
219 $T_DOT_WORD = '\.(long|value|byte|zero)';
220 $T_DOT_GLOBAL = '\.global';
221 $T_HDR_literal = "\.section\t\.rodata\n";
222 $T_HDR_misc = "\.text\n\t\.align 8\n";
223 $T_HDR_data = "\.data\n\t\.align 8\n";
224 $T_HDR_consist = "\.text\n";
225 $T_HDR_closure = "\.data\n\t\.align 8\n";
226 $T_HDR_srt = "\.text\n\t\.align 8\n";
227 $T_HDR_info = "\.text\n\t\.align 8\n";
228 $T_HDR_entry = "\.text\n\t\.align 16\n";
229 $T_HDR_vector = "\.text\n\t\.align 8\n";
230 $T_HDR_direct = "\.text\n\t\.align 8\n";
232 #--------------------------------------------------------#
233 } elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) {
235 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
236 $T_US = '_'; # _ if symbols have an underscore on the front
237 $T_PRE_APP = '^# MAY NOT APPLY'; # regexp that says what comes before APP/NO_APP
238 $T_CONST_LBL = '^LC(\d+):$';
241 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.proc\s+\d+|\.const|\.cstring|\.globl\s+\S+|\.text|\.data|\.even|\.stab[^n].*)\n)';
242 $T_COPY_DIRVS = '\.(globl|proc|stab)';
243 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
245 $T_DOT_WORD = '\.long';
246 $T_DOT_GLOBAL = '\.globl';
247 $T_HDR_literal = "\.text\n\t\.even\n";
248 $T_HDR_misc = "\.text\n\t\.even\n";
249 $T_HDR_data = "\.data\n\t\.even\n";
250 $T_HDR_consist = "\.text\n";
251 $T_HDR_closure = "\.data\n\t\.even\n";
252 $T_HDR_srt = "\.text\n\t\.even\n";
253 $T_HDR_info = "\.text\n\t\.even\n";
254 $T_HDR_entry = "\.text\n\t\.even\n";
255 $T_HDR_vector = "\.text\n\t\.even\n";
256 $T_HDR_direct = "\.text\n\t\.even\n";
258 #--------------------------------------------------------#
259 } elsif ( $TargetPlatform =~ /^mips-.*/ ) {
261 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
262 $T_US = ''; # _ if symbols have an underscore on the front
263 $T_PRE_APP = '^\s*#'; # regexp that says what comes before APP/NO_APP
264 $T_CONST_LBL = '^\$LC(\d+):$'; # regexp for what such a lbl looks like
267 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.(globl|ent)\s+\S+|\.text|\.r?data)\n)';
268 $T_COPY_DIRVS = '\.(globl|ent)';
270 $T_hsc_cc_PAT = 'I WAS TOO LAZY TO DO THIS BIT (WDP 95/05)';
271 $T_DOT_WORD = '\.word';
272 $T_DOT_GLOBAL = '^\t\.globl';
273 $T_HDR_literal = "\t\.rdata\n\t\.align 2\n";
274 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
275 $T_HDR_data = "\t\.data\n\t\.align 2\n";
276 $T_HDR_consist = 'TOO LAZY TO DO THIS TOO';
277 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
278 $T_HDR_srt = "\t\.text\n\t\.align 2\n";
279 $T_HDR_info = "\t\.text\n\t\.align 2\n";
280 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
281 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
282 $T_HDR_direct = "\t\.text\n\t\.align 2\n";
284 #--------------------------------------------------------#
285 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ ) {
286 # Apple PowerPC Darwin/MacOS X.
287 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
288 $T_US = '_'; # _ if symbols have an underscore on the front
289 $T_PRE_APP = 'WHAT IS THIS'; # regexp that says what comes before APP/NO_APP
290 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
293 $T_MOVE_DIRVS = '^(\s*(\.align \d+|\.text|\.data|\.const_data|\.cstring|\.non_lazy_symbol_pointer|\.const|\.static_const|\.literal4|\.literal8|\.static_data|\.globl \S+|\.section .*)\n)';
294 $T_COPY_DIRVS = '\.(globl)';
296 $T_hsc_cc_PAT = '\.byte.*\)(hsc|cc) (.*)"\n\t\.byte \d+\n\t\.byte "(.*)"\n\t\.byte \d+';
297 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
298 $T_DOT_GLOBAL = '\.globl';
299 $T_HDR_toc = "\.toc\n";
300 $T_HDR_literal = "\t\.const_data\n\t\.align 2\n";
301 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
302 $T_HDR_data = "\t\.data\n\t\.align 2\n";
303 $T_HDR_consist = "\t\.text\n\t\.align 2\n";
304 $T_HDR_closure = "\t\.const_data\n\t\.align 2\n";
305 $T_HDR_srt = "\t\.text\n\t\.align 2\n";
306 $T_HDR_info = "\t\.text\n\t\.align 2\n";
307 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
308 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
309 $T_HDR_direct = "\t\.text\n\t\.align 2\n";
311 #--------------------------------------------------------#
312 } elsif ( $TargetPlatform =~ /^sparc-.*-(solaris2|openbsd)/ ) {
314 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
315 $T_US = ''; # _ if symbols have an underscore on the front
316 $T_PRE_APP = 'DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
317 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
320 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*|\.section.*|\s+\.type.*|\s+\.size.*)\n)';
321 $T_COPY_DIRVS = '\.(global|proc|stab)';
323 $T_hsc_cc_PAT = '\.asciz.*\)(hsc|cc) (.*)\\\\t(.*)"';
324 $T_DOT_WORD = '\.(long|word|byte|half|skip|uahalf|uaword)';
325 $T_DOT_GLOBAL = '^\t\.global';
326 $T_HDR_literal = "\.text\n\t\.align 8\n";
327 $T_HDR_misc = "\.text\n\t\.align 4\n";
328 $T_HDR_data = "\.data\n\t\.align 8\n";
329 $T_HDR_consist = "\.text\n";
330 $T_HDR_closure = "\.data\n\t\.align 4\n";
331 $T_HDR_srt = "\.data\n\t\.align 4\n";
332 $T_HDR_info = "\.text\n\t\.align 4\n";
333 $T_HDR_entry = "\.text\n\t\.align 4\n";
334 $T_HDR_vector = "\.text\n\t\.align 4\n";
335 $T_HDR_direct = "\.text\n\t\.align 4\n";
337 #--------------------------------------------------------#
338 } elsif ( $TargetPlatform =~ /^sparc-.*-sunos4/ ) {
340 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
341 $T_US = '_'; # _ if symbols have an underscore on the front
342 $T_PRE_APP = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
343 $T_CONST_LBL = '^LC(\d+):$';
346 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*)\n)';
347 $T_COPY_DIRVS = '\.(global|proc|stab)';
348 $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"';
350 $T_DOT_WORD = '\.word';
351 $T_DOT_GLOBAL = '^\t\.global';
352 $T_HDR_literal = "\.text\n\t\.align 8\n";
353 $T_HDR_misc = "\.text\n\t\.align 4\n";
354 $T_HDR_data = "\.data\n\t\.align 8\n";
355 $T_HDR_consist = "\.text\n";
356 $T_HDR_closure = "\.data\n\t\.align 4\n";
357 $T_HDR_srt = "\.data\n\t\.align 4\n";
358 $T_HDR_info = "\.text\n\t\.align 4\n";
359 $T_HDR_entry = "\.text\n\t\.align 4\n";
360 $T_HDR_vector = "\.text\n\t\.align 4\n";
361 $T_HDR_direct = "\.text\n\t\.align 4\n";
363 #--------------------------------------------------------#
364 } elsif ( $TargetPlatform =~ /^sparc-.*-linux/ ) {
365 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
366 $T_US = ''; # _ if symbols have an underscore on the front
367 $T_PRE_APP = '#'; # regexp that says what comes before APP/NO_APP
368 # Probably doesn't apply anyway
369 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
372 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.seg|\.stab.*|\.section.*|\s+\.type.*|\s+\.size.*)\n)';
373 $T_COPY_DIRVS = '\.(global|globl|proc|stab)';
375 $T_hsc_cc_PAT = '\.asciz.*\)(hsc|cc) (.*)\\\\t(.*)"';
376 $T_DOT_WORD = '\.(long|word|nword|xword|byte|half|short|skip|uahalf|uaword)';
377 $T_DOT_GLOBAL = '^\t\.global';
378 $T_HDR_literal = "\.text\n\t\.align 8\n";
379 $T_HDR_misc = "\.text\n\t\.align 4\n";
380 $T_HDR_data = "\.data\n\t\.align 8\n";
381 $T_HDR_consist = "\.text\n";
382 $T_HDR_closure = "\.data\n\t\.align 4\n";
383 $T_HDR_srt = "\.data\n\t\.align 4\n";
384 $T_HDR_info = "\.text\n\t\.align 4\n";
385 $T_HDR_entry = "\.text\n\t\.align 4\n";
386 $T_HDR_vector = "\.text\n\t\.align 4\n";
387 $T_HDR_direct = "\.text\n\t\.align 4\n";
389 #--------------------------------------------------------#
391 print STDERR "$Pgm: don't know how to mangle assembly language for: $TargetPlatform\n";
396 print STDERR "T_STABBY: $T_STABBY\n";
397 print STDERR "T_US: $T_US\n";
398 print STDERR "T_PRE_APP: $T_PRE_APP\n";
399 print STDERR "T_CONST_LBL: $T_CONST_LBL\n";
400 print STDERR "T_POST_LBL: $T_POST_LBL\n";
401 if ( $TargetPlatform =~ /^i386-/ ) {
402 print STDERR "T_X86_PRE_LLBL_PAT: $T_X86_PRE_LLBL_PAT\n";
403 print STDERR "T_X86_PRE_LLBL: $T_X86_PRE_LLBL\n";
404 print STDERR "T_X86_BADJMP: $T_X86_BADJMP\n";
406 print STDERR "T_MOVE_DIRVS: $T_MOVE_DIRVS\n";
407 print STDERR "T_COPY_DIRVS: $T_COPY_DIRVS\n";
408 print STDERR "T_hsc_cc_PAT: $T_hsc_cc_PAT\n";
409 print STDERR "T_DOT_WORD: $T_DOT_WORD\n";
410 print STDERR "T_HDR_literal: $T_HDR_literal\n";
411 print STDERR "T_HDR_misc: $T_HDR_misc\n";
412 print STDERR "T_HDR_data: $T_HDR_data\n";
413 print STDERR "T_HDR_consist: $T_HDR_consist\n";
414 print STDERR "T_HDR_closure: $T_HDR_closure\n";
415 print STDERR "T_HDR_info: $T_HDR_info\n";
416 print STDERR "T_HDR_entry: $T_HDR_entry\n";
417 print STDERR "T_HDR_vector: $T_HDR_vector\n";
418 print STDERR "T_HDR_direct: $T_HDR_direct\n";
424 %************************************************************************
426 \subsection{Mangle away}
428 %************************************************************************
432 local($in_asmf, $out_asmf) = @_;
434 # multi-line regexp matching:
439 &init_TARGET_STUFF();
440 &init_FUNNY_THINGS();
442 open(INASM, "< $in_asmf")
443 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
444 open(OUTASM,"> $out_asmf")
445 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
447 # read whole file, divide into "chunks":
448 # record some info about what we've found...
450 @chk = (); # contents of the chunk
451 $numchks = 0; # number of them
452 @chkcat = (); # what category of thing in each chunk
453 @chksymb = (); # what symbol(base) is defined in this chunk
454 %entrychk = (); # ditto, its entry code
455 %closurechk = (); # ditto, the (static) closure
456 %srtchk = (); # ditto, its SRT (for top-level things)
457 %infochk = (); # given a symbol base, say what chunk its info tbl is in
458 %vectorchk = (); # ditto, return vector table
459 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
461 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
464 tr/\r//d if $TargetPlatform =~ /-mingw32$/; # In case Perl doesn't convert line endings
465 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
466 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
467 next if /^\t\.def.*endef$/;
468 next if /${T_PRE_APP}(NO_)?APP/o;
469 next if /^;/ && $TargetPlatform =~ /^hppa/;
471 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|ia64)-/;
473 if ( $TargetPlatform =~ /^mips-/
474 && /^\t\.(globl\S+\.text|comm\t)/ ) {
475 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
476 # Treat .comm variables as data. These show up in two (known) places:
478 # - the module_registered variable used in the __stginit fragment.
479 # even though these are declared static and initialised, gcc 3.3
480 # likes to make them .comm, presumably to save space in the
483 # - global variables used to pass arguments from C to STG in
484 # a foreign export. (is this still true? --SDM)
486 } elsif ( /^\t\.comm.*$/ ) {
488 $chkcat[$i] = 'data';
491 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
492 # duplicated from the bottom
496 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
497 # Alphas: Local labels not to be confused with new chunks
500 # NB: all the rest start with a non-space
502 } elsif ( $TargetPlatform =~ /^mips-/
503 && /^\d+:/ ) { # a funny-looking very-local label
506 } elsif ( /$T_CONST_LBL/o ) {
508 $chkcat[$i] = 'literal';
511 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
513 $chkcat[$i] = 'splitmarker';
516 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
519 $chkcat[$i] = 'infotbl';
520 $chksymb[$i] = $symb;
522 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
524 $infochk{$symb} = $i;
526 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
528 $chkcat[$i] = 'entry';
533 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
535 $chkcat[$i] = 'closure';
538 $closurechk{$1} = $i;
540 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
547 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
549 $chkcat[$i] = 'data';
552 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
555 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
556 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
557 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
558 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
561 $chkcat[$i] = 'data';
564 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
569 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
570 # all CC_ symbols go in the data section...
572 $chkcat[$i] = 'data';
575 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
577 $chkcat[$i] = 'misc';
579 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
581 $chkcat[$i] = 'vector';
586 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
587 && /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
588 # Some Solaris system headers contain function definitions (as
589 # opposed to mere prototypes), which end up in the .hc file when
590 # a Haskell module foreign imports the corresponding system
591 # functions (most notably stat()). We put them into the text
592 # segment. Note that this currently does not extend to function
593 # names starting with an underscore.
596 $chkcat[$i] = 'misc';
599 } elsif ( /^${T_US}[A-Za-z0-9_]/o
600 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
601 || ! /^L\$\d+$/ ) ) {
605 print STDERR "Warning: retaining unknown function \`$thing' in output from C compiler\n"
606 unless # $KNOWN_FUNNY_THING{$thing}
607 /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
608 || /^${T_US}__stg_.*${T_POST_LBL}$/o # more RTS internals
609 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
610 || /^${T_US}.*_slow${T_POST_LBL}$/o # slow entry
611 || /^${T_US}__stginit.*${T_POST_LBL}$/o # __stginit<module>
612 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
613 || /^${T_US}.*_srtd${T_POST_LBL}$/o # large bitmaps
614 || /^${T_US}.*_fast${T_POST_LBL}$/o # primops
615 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
616 || /^_uname:/o; # x86/Solaris2
618 $chkcat[$i] = 'misc';
621 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && (
623 || /^\.section __TEXT,__picsymbol_stub1,.*/
625 || /^\.section __TEXT,__symbol_stub1,.*/
626 || /^\.lazy_symbol_pointer/
627 || /^\.non_lazy_symbol_pointer/ ))
630 $chkcat[$i] = 'dyld';
632 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && /^\.data/ && $chkcat[$i] eq 'dyld')
633 { # non_lazy_symbol_ptrs that point to local symbols
635 $chkcat[$i] = 'dyld';
637 } else { # simple line (duplicated at the top)
642 $numchks = $#chk + 1;
644 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
645 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
648 # the division into chunks is imperfect;
649 # we throw some things over the fence into the next
652 # also, there are things we would like to know
653 # about the whole module before we start spitting
656 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
657 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
659 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
661 # Alphas: NB: we start meddling at chunk 1, not chunk 0
662 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
663 # spits a ".quad 0" in after the very first ".rdata"; we
664 # detect this special case (tossing the ".quad 0")!
665 local($magic_rdata_seen) = 0;
667 # HPPAs, MIPSen: also start medding at chunk 1
669 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
670 $c = $chk[$i]; # convenience copy
672 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
674 # toss all prologue stuff; HPPA is pretty weird
676 $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
678 # be slightly paranoid to make sure there's
679 # nothing surprising in there
680 if ( $c =~ /--- BEGIN ---/ ) {
681 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
683 # remove junk whitespace around the split point
687 if ($TargetPlatform =~ /^i386-/) {
688 $p =~ s/^\tpushl\s+\%edi\n//;
689 $p =~ s/^\tpushl\s+\%esi\n//;
690 $p =~ s/^\tpushl\s+\%ebx\n//;
691 $p =~ s/^\tmovl\s+\%esi,\s*\d*\(\%esp\)\n//;
692 $p =~ s/^\tmovl\s+\%edi,\s*\d*\(\%esp\)\n//;
693 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
694 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-cygwin32/);
696 # GCC 3.1 is in the habit of adding spurious writes to the
697 # stack in the prologue. Just to be on the safe side,
698 # chuck these over the fence into the main code.
699 while ($p =~ /^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n/) {
700 # print "Spurious instruction: $&";
705 } elsif ($TargetPlatform =~ /^ia64-/) {
706 $p =~ s/^\t\.prologue .*\n//;
707 $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, 3[12], \d+, 0\n//;
708 $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
709 $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
710 $p =~ s/^\t\.(mii|mmi)\n//g; # bundling is no longer sensible
711 $p =~ s/^\t;;\n//g; # discard stops
712 $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
714 # GCC 3.3 saves r1 in the prologue, move this to the body
715 if ($p =~ /^\tmov r\d+ = r1\n/) {
719 } elsif ($TargetPlatform =~ /^m68k-/) {
720 $p =~ s/^\tlink a6,#-?\d.*\n//;
721 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
722 # The above showed up in the asm code,
723 # so I added it here.
724 # I hope it's correct.
726 $p =~ s/^\tmovel d2,sp\@-\n//;
727 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
728 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
729 } elsif ($TargetPlatform =~ /^mips-/) {
730 # the .frame/.mask/.fmask that we use is the same
731 # as that produced by GCC for miniInterpret; this
732 # gives GDB some chance of figuring out what happened
733 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
734 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
735 $p =~ s/^\t\.(mask|fmask).*\n//g;
736 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
737 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
738 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
739 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
740 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
741 $p =~ s/__FRAME__/$FRAME/;
742 } elsif ($TargetPlatform =~ /^powerpc-apple-.*/) {
744 $pcrel_label =~ s/(.|\n)*^(L\d+\$pb):\n(.|\n)*/$2/ or $pcrel_label = "";
746 $p =~ s/^\tmflr r0\n//;
747 $p =~ s/^\tbl saveFP # f\d+\n//;
748 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
749 $p =~ s/^L\d+\$pb:\n//;
750 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
751 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
752 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
753 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
754 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
755 $p =~ s/^\tbcl 20,31,L\d+\$pb\n//;
756 $p =~ s/^L\d+\$pb:\n//;
757 $p =~ s/^\tmflr r31\n//;
759 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
760 # under some circumstances, only when generating position dependent code.
761 # I have no idea why, and I don't think it is necessary, so let's toss it.
762 $p =~ s/^\tli r\d+,0\n//g;
763 $p =~ s/^\tstw r\d+,\d+\(r1\)\n//g;
765 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
768 # HWL HACK: dont die, just print a warning
769 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
770 die "Prologue junk?: $p\n" if $p =~ /^\s+[^\s\.]/;
772 if ($TargetPlatform =~ /^powerpc-apple-.*/ && $pcrel_label ne "") {
773 # on PowerPC, we have to keep a part of the prologue
774 # (which loads the current instruction pointer into register r31)
775 $p .= "bcl 20,31,$pcrel_label\n";
776 $p .= "$pcrel_label:\n";
777 $p .= "\tmflr r31\n";
780 # glue together what's left
785 if ( $TargetPlatform =~ /^mips-/ ) {
786 # MIPS: first, this basic sequence may occur "--- END ---" or not
787 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
790 # toss all epilogue stuff; again, paranoidly
791 if ( $c =~ /--- END ---/ ) {
792 if (($r, $e) = split(/--- END ---/, $c)) {
793 if ($TargetPlatform =~ /^i386-/) {
795 $e =~ s/^\tpopl\s+\%edi\n//;
796 $e =~ s/^\tpopl\s+\%esi\n//;
797 $e =~ s/^\tpopl\s+\%edx\n//;
798 $e =~ s/^\tpopl\s+\%ecx\n//;
799 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
800 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
801 } elsif ($TargetPlatform =~ /^ia64-/) {
802 $e =~ s/^\tmov ar\.pfs = r\d+\n//;
803 $e =~ s/^\tmov b0 = r\d+\n//;
804 $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
805 $e =~ s/^\tbr\.ret\.sptk\.many b0\n//;
806 $e =~ s/^\t\.(mii|mmi|mib)\n//g; # bundling is no longer sensible
807 $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
808 $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
809 } elsif ($TargetPlatform =~ /^m68k-/) {
810 $e =~ s/^\tunlk a6\n//;
812 } elsif ($TargetPlatform =~ /^mips-/) {
813 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
814 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
815 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
816 $e =~ s/^\tj\t\$31\n//;
817 } elsif ($TargetPlatform =~ /^powerpc-apple-.*/) {
818 $e =~ s/^\taddi r1,r1,\d+\n//;
819 $e =~ s/^\tcal r1,\d+\(r1\)\n//;
820 $e =~ s/^\tlw?z? r\d+,\d+\(r1\)\n//;
821 $e =~ s/^\tmtlr r0\n//;
824 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
827 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
829 # glue together what's left
831 $c =~ s/\n\t\n/\n/; # junk blank line
835 # On SPARCs, we don't do --- BEGIN/END ---, we just
836 # toss the register-windowing save/restore/ret* instructions
838 if ( $TargetPlatform =~ /^sparc-/ ) {
839 $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
840 # throw away PROLOGUE comments
841 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
844 # On Alphas, the prologue mangling is done a little later (below)
846 # toss all calls to __DISCARD__
847 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
848 $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
849 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-.*/;
851 # IA64: mangle tailcalls into jumps here
852 if ($TargetPlatform =~ /^ia64-/) {
853 while ($c =~ s/^\tbr\.call\.sptk\.many b0 = (.*)\n(?:^\.L([0-9]*):\n)?(?:\t;;\n)?(?:\tmov r1 = r\d+\n)?(?:\t;;\n)?\t--- TAILCALL ---\n(?:\t;;\n\tbr \.L\d+\n)?/\tbr\.few $1\n/) {
854 # Eek, the gcc optimiser is getting smarter... if we see a jump to the --- TAILCALL ---
855 # marker then we reapply the substitution at the source sites
856 $c =~ s/^\tbr \.L$2\n/\t--- TAILCALL ---\n/g if ($2);
860 # MIPS: that may leave some gratuitous asm macros around
861 # (no harm done; but we get rid of them to be tidier)
862 $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/
863 if $TargetPlatform =~ /^mips-/;
865 # toss stack adjustment after DoSparks
866 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
867 if $TargetPlatform =~ /^m68k-/; # this looks old...
869 if ( $TargetPlatform =~ /^alpha-/ &&
870 ! $magic_rdata_seen &&
871 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
872 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
873 $magic_rdata_seen = 1;
876 # pick some end-things and move them to the next chunk
878 # pin a funny end-thing on (for easier matching):
879 $c .= 'FUNNY#END#THING';
881 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
885 # on x86 we try not to copy any directives into a literal
886 # chunk, rather we keep looking for the next real chunk. This
887 # is because we get things like
889 # .globl blah_closure
895 if ( $TargetPlatform =~ /^(i386|sparc|powerpc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
897 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
900 if ( $j < $numchks ) {
901 $chk[$j] = $to_move . $chk[$j];
905 elsif ( $i < ($numchks - 1)
906 && ( $to_move =~ /${T_COPY_DIRVS}/
907 || ($TargetPlatform =~ /^hppa/ && $to_move =~ /align/ && $chkcat[$i+1] eq 'literal') )) {
908 $chk[$i + 1] = $to_move . $chk[$i + 1];
909 # otherwise they're tossed
912 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
915 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
917 # toss all prologue stuff, except for loading gp, and the ..ng address
918 unless ($c =~ /\.ent.*\n\$.*\.\.ng:/) {
919 if (($p, $r) = split(/^\t\.prologue/, $c)) {
920 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
921 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
922 $keep =~ s/^\t\.(mask|fmask).*\n//g;
923 $c = $keep . "..ng:\n";
925 print STDERR "malformed code block ($ent)?\n"
928 $c .= "\t.prologue" . $r;
932 $c =~ s/FUNNY#END#THING//;
934 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
936 $chk[$i] = $c; # update w/ convenience copy
939 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
940 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
943 if ( $TargetPlatform =~ /^alpha-/ ) {
944 # print out the header stuff first
945 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
946 print OUTASM $chk[0];
948 } elsif ( $TargetPlatform =~ /^hppa/ ) {
949 print OUTASM $chk[0];
951 } elsif ( $TargetPlatform =~ /^mips-/ ) {
952 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
954 # get rid of horrible "<dollar>Revision: .*$" strings
955 local(@lines0) = split(/\n/, $chk[0]);
957 while ( $z <= $#lines0 ) {
958 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
961 while ( $z <= $#lines0 ) {
963 last if $lines0[$z] =~ /[,\t]0x0$/;
969 $chk[0] = join("\n", @lines0);
970 $chk[0] =~ s/\n\n+/\n/;
971 print OUTASM $chk[0];
974 # print out all the literal strings next
975 for ($i = 0; $i < $numchks; $i++) {
976 if ( $chkcat[$i] eq 'literal' ) {
977 print OUTASM $T_HDR_literal, $chk[$i];
978 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
980 $chkcat[$i] = 'DONE ALREADY';
984 # on the HPPA, print out all the bss next
985 if ( $TargetPlatform =~ /^hppa/ ) {
986 for ($i = 1; $i < $numchks; $i++) {
987 if ( $chkcat[$i] eq 'bss' ) {
988 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
989 print OUTASM $chk[$i];
991 $chkcat[$i] = 'DONE ALREADY';
996 for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) {
997 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
999 next if $chkcat[$i] eq 'DONE ALREADY';
1001 if ( $chkcat[$i] eq 'misc' ) {
1002 if ($chk[$i] ne '') {
1003 print OUTASM $T_HDR_misc;
1004 &print_doctored($chk[$i], 0);
1007 } elsif ( $chkcat[$i] eq 'toss' ) {
1008 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
1010 } elsif ( $chkcat[$i] eq 'data' ) {
1011 if ($chk[$i] ne '') {
1012 print OUTASM $T_HDR_data;
1013 print OUTASM $chk[$i];
1016 } elsif ( $chkcat[$i] eq 'consist' ) {
1017 if ( $chk[$i] =~ /$T_hsc_cc_PAT/o ) {
1018 local($consist) = "$1.$2.$3";
1019 $consist =~ s/,/./g;
1020 $consist =~ s/\//./g;
1021 $consist =~ s/-/_/g;
1022 $consist =~ s/[^A-Za-z0-9_.]/ZZ/g; # ToDo: properly?
1024 # Using a cygnus-2.7-96q4 gcc build on hppas, the
1025 # consistency chunk for ghc_cc_ID often (but not always!)
1026 # gets lumped with a bunch of .IMPORT directives containing info on
1027 # the code or data space nature of external symbols. We can't
1028 # toss these, so once the consistency ID has been turned into
1029 # a representable symbol, we substitute it for the symbol
1030 # that the string was attached to in the first place (ghc_cc_ID.)
1031 # (The original string is also substituted away.)
1033 # This change may affect the code output on other platforms in
1034 # adverse ways, hence we restrict this hack hppa targets only.
1037 if ( $TargetPlatform =~ /^hppa/ ) {
1038 $chk[$i] =~ s/^${T_US}ghc.*c_ID$TPOSTLBL/$consist/o;
1039 $chk[$i] =~ s/\t$T_hsc_cc_PAT/$T_HDR_misc/o;
1040 $consist = $chk[$i]; #clumsily
1042 print OUTASM $T_HDR_consist, "${consist}${T_POST_LBL}\n";
1044 } elsif ( $TargetPlatform !~ /^(mips)-/ ) { # we just don't try in those case (ToDo)
1045 # on mips: consistency string is just a v
1046 # horrible bunch of .bytes,
1047 # which I am too lazy to sort out (WDP 95/05)
1049 print STDERR "Couldn't grok consistency: ", $chk[$i];
1052 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1053 # we can just re-constitute this one...
1054 # NB: we emit _three_ underscores no matter what,
1055 # so ghc-split doesn't have to care.
1056 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1058 } elsif ( $chkcat[$i] eq 'closure'
1059 || $chkcat[$i] eq 'srt'
1060 || $chkcat[$i] eq 'infotbl'
1061 || $chkcat[$i] eq 'entry') { # do them in that order
1062 $symb = $chksymb[$i];
1065 if ( defined($closurechk{$symb}) ) {
1066 print OUTASM $T_HDR_closure;
1067 print OUTASM $chk[$closurechk{$symb}];
1068 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1072 if ( defined($srtchk{$symb}) ) {
1073 print OUTASM $T_HDR_srt;
1074 print OUTASM $chk[$srtchk{$symb}];
1075 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1079 if ( defined($infochk{$symb}) ) {
1081 print OUTASM $T_HDR_info;
1082 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1084 # entry code will be put here!
1086 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1090 if ( defined($entrychk{$symb}) ) {
1092 $c = $chk[$entrychk{$symb}];
1094 print OUTASM $T_HDR_entry;
1096 &print_doctored($c, 1); # NB: the 1!!!
1098 $chkcat[$entrychk{$symb}] = 'DONE ALREADY';
1101 } elsif ( $chkcat[$i] eq 'vector'
1102 || $chkcat[$i] eq 'direct' ) { # do them in that order
1103 $symb = $chksymb[$i];
1106 if ( defined($vectorchk{$symb}) ) {
1107 print OUTASM $T_HDR_vector;
1108 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1110 # direct return code will be put here!
1111 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1113 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1114 # Alphas: the commented nop is for the splitter, to ensure
1115 # that no module ends with a label as the very last
1116 # thing. (The linker will adjust the label to point
1117 # to the first code word of the next module linked in,
1118 # even if alignment constraints cause the label to move!)
1120 print OUTASM "\t# nop\n";
1123 } elsif ( $chkcat[$i] eq 'toc' ) {
1124 # silly optimisation to print tocs, since they come in groups...
1125 print OUTASM $T_HDR_toc;
1127 while ($chkcat[$j] eq 'toc')
1128 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1131 print OUTASM $chk[$j];
1133 $chkcat[$j] = 'DONE ALREADY';
1137 } elsif ( $TargetPlatform =~ /^powerpc-apple-.*/ && $chkcat[$i] eq 'dyld' ) {
1138 # powerpc-apple: dynamic linker stubs
1139 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1140 { # print them out unchanged, but remove the stubs for __DISCARD__
1141 print OUTASM $chk[$i];
1144 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1148 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1151 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1152 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1157 sub hppa_mash_prologue { # OK, epilogue, too
1160 # toss all prologue stuff
1161 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1163 # Lie about our .CALLINFO
1164 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1171 # toss all epilogue stuff
1172 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1174 # Sorry; we moved the _info stuff to the code segment.
1175 s/_info,DATA/_info,CODE/g;
1182 sub print_doctored {
1183 local($_, $need_fallthru_patch) = @_;
1185 if ( $TargetPlatform !~ /^i386-/
1186 || ! /^\t[a-z]/ # no instructions in here, apparently
1187 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1191 # OK, must do some x86 **HACKING**
1193 local($entry_patch) = '';
1194 local($exit_patch) = '';
1196 # gotta watch out for weird instructions that
1197 # invisibly smash various regs:
1198 # rep* %ecx used for counting
1199 # scas* %edi used for destination index
1200 # cmps* %e[sd]i used for indices
1201 # loop* %ecx used for counting
1206 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1208 # * GCC used an "STG reg" for its own purposes
1210 # * some secret uses of machine reg, requiring STG reg
1211 # to be saved/restored
1213 # The most dangerous "GCC uses" of an "STG reg" are when
1214 # the reg holds the target of a jmp -- it's tricky to
1215 # insert the patch-up code before we get to the target!
1216 # So here we change the jmps:
1218 # --------------------------------------------------------
1219 # it can happen that we have jumps of the form...
1220 # jmp *<something involving %esp>
1222 # jmp <something involving another naughty register...>
1224 # a reasonably-common case is:
1226 # movl $_blah,<bad-reg>
1230 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[abcd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1232 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1233 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1234 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1235 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1236 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1237 if /(jmp|call)\s+.*\%esi/;
1239 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1240 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1241 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1242 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1243 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1244 if /(jmp|call)\s+.*\%edi/;
1247 # OK, now we can decide what our patch-up code is going to
1250 # Offsets into register table - you'd better update these magic
1251 # numbers should you change its contents!
1252 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1253 local($OFFSET_Hp)=88;
1255 # Note funky ".=" stuff; we're *adding* to these _patch guys
1256 if ( $StolenX86Regs <= 2
1257 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1258 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1259 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1261 # nothing for call_{entry,exit} because %esi is callee-save
1263 if ( $StolenX86Regs <= 3
1264 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1265 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1266 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1268 # nothing for call_{entry,exit} because %edi is callee-save
1271 # --------------------------------------------------------
1272 # next, here we go with non-%esp patching!
1274 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1276 # Before calling GC we must set up the exit condition before the call
1277 # and entry condition when we come back
1279 # fix _all_ non-local jumps:
1281 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1282 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1284 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1286 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1287 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1289 if ($StolenX86Regs == 2 ) {
1290 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1291 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1292 } elsif ($StolenX86Regs == 3 ) {
1293 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1294 if /^\t(jmp|call)\s+.*\%edi/;
1297 # --------------------------------------------------------
1298 # that's it -- print it
1300 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1304 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1305 print OUTASM $exit_patch;
1306 # ToDo: make it not print if there is a "jmp" at the end
1312 sub init_FUNNY_THINGS {
1313 %KNOWN_FUNNY_THING = (
1315 # "${T_US}stg_.*{T_POST_LBL}", 1,
1320 The following table reversal is used for both info tables and return
1321 vectors. In both cases, we remove the first entry from the table,
1322 reverse the table, put the label at the end, and paste some code
1323 (that which is normally referred to by the first entry in the table)
1324 right after the table itself. (The code pasting is done elsewhere.)
1328 local($symb, $tbl, $discard1) = @_;
1330 return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1332 local($before) = '';
1334 local(@imports) = (); # hppa only
1337 local(@lines) = split(/\n/, $tbl);
1340 # Deal with the header...
1341 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1342 $label .= $lines[$i] . "\n",
1343 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1344 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1345 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1347 $before .= $lines[$i] . "\n"; # otherwise...
1350 # Grab the table data...
1351 if ( $TargetPlatform !~ /^hppa/ ) {
1352 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1353 push(@words, $lines[$i]);
1355 } else { # hppa weirdness
1356 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1357 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1358 push(@imports, $lines[$i]);
1360 # We don't use HP's ``function pointers''
1361 # We just use labels in code space, like normal people
1362 $lines[$i] =~ s/P%//;
1363 push(@words, $lines[$i]);
1368 # Now throw away any initial zero word from the table. This is a hack
1369 # that lets us reduce the size of info tables when the SRT field is not
1370 # needed: see comments StgFunInfoTable in InfoTables.h.
1372 # The .zero business is for Linux/ELF.
1373 # The .skip business is for Sparc/Solaris/ELF.
1374 # The .blockz business is for HPPA.
1376 # if ($words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1381 for (; $i <= $#lines; $i++) {
1382 $after .= $lines[$i] . "\n";
1385 # Alphas: If we have anonymous text (not part of a procedure), the
1386 # linker may complain about missing exception information. Bleh.
1387 # To suppress this, we place a .ent/.end pair around the code.
1388 # At the same time, we have to be careful and not enclose any leading
1389 # .file/.loc directives.
1390 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1391 local ($ident) = $1;
1392 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1393 $after .= "\t.end $ident\n";
1396 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1397 # assembler (!) wherein .quad constants inside .text sections are
1398 # first narrowed to 32 bits then sign-extended back to 64 bits.
1399 # This obviously screws up our 64-bit bitmaps, so we work around
1400 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1401 if ( $TargetPlatform =~ /^alpha-/ ) {
1403 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1404 local ($number) = $1;
1405 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1406 local ($sign, $base, $digits) = ($1, $2, $3);
1407 $base = (10, 8, 16)[length $base];
1408 local ($hi, $lo) = (0, 0);
1409 foreach $i (split(//, $digits)) {
1410 $j = $lo * $base + $i;
1411 $lo = $j % 4294967296;
1412 $hi = $hi * $base + ($j - $lo) / 4294967296;
1414 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1416 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1417 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1419 print STDERR "Cannot handle \".quad $number\" in info table\n";
1427 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
1428 . join("\n", @words) . "\n"
1431 # print STDERR "before=$before\n";
1432 # print STDERR "label=$label\n";
1433 # print STDERR "words=",(reverse @words),"\n";
1434 # print STDERR "after=$after\n";
1440 The HP is a major nuisance. The threaded code mangler moved info
1441 tables from data space to code space, but unthreaded code in the RTS
1442 still has references to info tables in data space. Since the HP
1443 linker is very precise about where symbols live, we need to patch the
1444 references in the unthreaded RTS as well.
1447 sub mini_mangle_asm_hppa {
1448 local($in_asmf, $out_asmf) = @_;
1450 open(INASM, "< $in_asmf")
1451 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1452 open(OUTASM,"> $out_asmf")
1453 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1456 s/_info,DATA/_info,CODE/; # Move _info references to code space
1462 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1463 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1469 sub tidy_up_and_die {
1470 local($return_val, $msg) = @_;
1472 exit (($return_val == 0) ? 0 : 1);