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_DOT_WORD = '\.(long|quad|byte|word)';
103 $T_DOT_GLOBAL = '^\t\.globl';
104 $T_HDR_literal = "\.rdata\n\t\.align 3\n";
105 $T_HDR_misc = "\.text\n\t\.align 3\n";
106 $T_HDR_data = "\.data\n\t\.align 3\n";
107 $T_HDR_rodata = "\.rdata\n\t\.align 3\n";
108 $T_HDR_closure = "\.data\n\t\.align 3\n";
109 $T_HDR_info = "\.text\n\t\.align 3\n";
110 $T_HDR_entry = "\.text\n\t\.align 3\n";
111 $T_HDR_vector = "\.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_DOT_WORD = '\.(blockz|word|half|byte)';
126 $T_DOT_GLOBAL = '^\s+\.EXPORT';
127 $T_HDR_literal = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
128 $T_HDR_misc = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
129 $T_HDR_data = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
130 $T_HDR_rodata = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
131 $T_HDR_closure = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
132 $T_HDR_info = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
133 $T_HDR_entry = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
134 $T_HDR_vector = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
136 #--------------------------------------------------------#
137 } elsif ( $TargetPlatform =~ /^i386-.*-(linuxaout|freebsd2|nextstep3|cygwin32|mingw32)$/ ) {
138 # NeXT added but not tested. CaS
140 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
141 $T_US = '_'; # _ if symbols have an underscore on the front
142 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
143 $T_CONST_LBL = '^LC(\d+):$';
145 $T_X86_PRE_LLBL_PAT = 'L';
146 $T_X86_PRE_LLBL = 'L';
147 $T_X86_BADJMP = '^\tjmp [^L\*]';
149 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s.*|\.globl\s+\S+|\.text|\.data|\.stab[^n].*|\.type\s+.*|\.size\s+.*|\.lcomm.*)\n)';
150 $T_COPY_DIRVS = '\.(globl|stab|lcomm)';
151 $T_DOT_WORD = '\.(long|word|value|byte|space)';
152 $T_DOT_GLOBAL = '\.globl';
153 $T_HDR_literal = "\.text\n\t\.align 2\n";
154 $T_HDR_misc = "\.text\n\t\.align 2,0x90\n";
155 $T_HDR_data = "\.data\n\t\.align 2\n";
156 $T_HDR_rodata = "\.text\n\t\.align 2\n";
157 $T_HDR_closure = "\.data\n\t\.align 2\n";
158 $T_HDR_info = "\.text\n\t\.align 2\n"; # NB: requires padding
159 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
160 $T_HDR_vector = "\.text\n\t\.align 2\n"; # NB: requires padding
162 #--------------------------------------------------------#
163 } elsif ( $TargetPlatform =~ /^i386-.*-(solaris2|linux|gnu|freebsd|netbsd|openbsd|kfreebsdgnu)$/ ) {
165 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
166 $T_US = ''; # _ if symbols have an underscore on the front
167 $T_PRE_APP = # regexp that says what comes before APP/NO_APP
168 ($TargetPlatform =~ /-(linux|gnu|freebsd|netbsd|openbsd)$/) ? '#' : '/' ;
169 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
171 $T_X86_PRE_LLBL_PAT = '\.L';
172 $T_X86_PRE_LLBL = '.L';
173 $T_X86_BADJMP = '^\tjmp\s+[^\.\*]';
175 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s.*|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.size\s+\S+\s*,\s*\d+|\.ident.*|\.local.*)\n)';
176 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
178 $T_DOT_WORD = '\.(long|value|word|byte|zero)';
179 $T_DOT_GLOBAL = '\.globl';
180 $T_HDR_literal = "\.section\t\.rodata\n"; # or just use .text??? (WDP 95/11)
181 $T_HDR_misc = "\.text\n\t\.align 4\n";
182 $T_HDR_data = "\.data\n\t\.align 4\n";
183 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 4\n";
184 $T_HDR_closure = "\.data\n\t\.align 4\n";
185 $T_HDR_info = "\.text\n\t\.align 4\n";
186 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
187 $T_HDR_vector = "\.text\n\t\.align 4\n"; # NB: requires padding
189 #--------------------------------------------------------#
190 } elsif ( $TargetPlatform =~ /^ia64-.*-linux$/ ) {
192 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
193 $T_US = ''; # _ if symbols have an underscore on the front
195 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
198 $T_MOVE_DIRVS = '^(\s*\.(global|proc|pred\.safe_across_calls|text|data|section|subsection|align|size|type|ident)\s+.*\n)';
199 $T_COPY_DIRVS = '\.(global|proc)';
201 $T_DOT_WORD = '\.(long|value|byte|zero)';
202 $T_DOT_GLOBAL = '\.global';
203 $T_HDR_literal = "\.section\t\.rodata\n";
204 $T_HDR_misc = "\.text\n\t\.align 16\n"; # May contain code; align like 'entry'
205 $T_HDR_data = "\.data\n\t\.align 8\n";
206 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 8\n";
207 $T_HDR_closure = "\.data\n\t\.align 8\n";
208 $T_HDR_info = "\.text\n\t\.align 8\n";
209 $T_HDR_entry = "\.text\n\t\.align 16\n";
210 $T_HDR_vector = "\.text\n\t\.align 8\n";
212 #--------------------------------------------------------#
213 } elsif ( $TargetPlatform =~ /^x86_64-.*-(linux|openbsd)$/ ) {
215 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
216 $T_US = ''; # _ if symbols have an underscore on the front
218 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
221 $T_MOVE_DIRVS = '^(\s*\.(globl|text|data|section|align|size|type|ident|local)([ \t].*)?\n)';
222 $T_COPY_DIRVS = '\.(globl|type|size|local)';
224 $T_DOT_WORD = '\.(quad|long|value|byte|zero)';
225 $T_DOT_GLOBAL = '\.global';
227 $T_HDR_literal16 = "\.section\t\.rodata.cst16\n\t.align 16\n";
228 $T_HDR_literal = "\.section\t\.rodata\n";
230 $T_HDR_misc = "\.text\n\t\.align 8\n";
231 $T_HDR_data = "\.data\n\t\.align 8\n";
232 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 8\n";
234 # the assembler on x86_64/Linux refuses to generate code for
236 # where x is in the text section and y in the rodata section.
237 # It works if y is in the text section, though. This is probably
238 # going to cause difficulties for PIC, I imagine.
240 # See Note [x86-64-relative] in includes/InfoTables.h
241 $T_HDR_relrodata= "\.text\n\t\.align 8\n";
243 $T_HDR_closure = "\.data\n\t\.align 8\n";
244 $T_HDR_info = "\.text\n\t\.align 8\n";
245 $T_HDR_entry = "\.text\n\t\.align 8\n";
246 $T_HDR_vector = "\.text\n\t\.align 8\n";
248 #--------------------------------------------------------#
249 } elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) {
251 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
252 $T_US = '_'; # _ if symbols have an underscore on the front
253 $T_PRE_APP = '^# MAY NOT APPLY'; # regexp that says what comes before APP/NO_APP
254 $T_CONST_LBL = '^LC(\d+):$';
257 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.proc\s+\d+|\.const|\.cstring|\.globl\s+\S+|\.text|\.data|\.even|\.stab[^n].*)\n)';
258 $T_COPY_DIRVS = '\.(globl|proc|stab)';
260 $T_DOT_WORD = '\.long';
261 $T_DOT_GLOBAL = '\.globl';
262 $T_HDR_literal = "\.text\n\t\.even\n";
263 $T_HDR_misc = "\.text\n\t\.even\n";
264 $T_HDR_data = "\.data\n\t\.even\n";
265 $T_HDR_rodata = "\.text\n\t\.even\n";
266 $T_HDR_closure = "\.data\n\t\.even\n";
267 $T_HDR_info = "\.text\n\t\.even\n";
268 $T_HDR_entry = "\.text\n\t\.even\n";
269 $T_HDR_vector = "\.text\n\t\.even\n";
271 #--------------------------------------------------------#
272 } elsif ( $TargetPlatform =~ /^mips-.*/ ) {
274 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
275 $T_US = ''; # _ if symbols have an underscore on the front
276 $T_PRE_APP = '^\s*#'; # regexp that says what comes before APP/NO_APP
277 $T_CONST_LBL = '^\$LC(\d+):$'; # regexp for what such a lbl looks like
280 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.(globl|ent)\s+\S+|\.text|\.r?data)\n)';
281 $T_COPY_DIRVS = '\.(globl|ent)';
283 $T_DOT_WORD = '\.word';
284 $T_DOT_GLOBAL = '^\t\.globl';
285 $T_HDR_literal = "\t\.rdata\n\t\.align 2\n";
286 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
287 $T_HDR_data = "\t\.data\n\t\.align 2\n";
288 $T_HDR_rodata = "\t\.rdata\n\t\.align 2\n";
289 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
290 $T_HDR_info = "\t\.text\n\t\.align 2\n";
291 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
292 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
294 #--------------------------------------------------------#
295 } elsif ( $TargetPlatform =~ /^powerpc-apple-darwin.*/ ) {
296 # Apple PowerPC Darwin/MacOS X.
297 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
298 $T_US = '_'; # _ if symbols have an underscore on the front
299 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
300 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
303 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s.*|\.text|\.data|\.const_data|\.cstring|\.non_lazy_symbol_pointer|\.const|\.static_const|\.literal4|\.literal8|\.static_data|\.globl \S+|\.section .*|\.lcomm.*)\n)';
304 $T_COPY_DIRVS = '\.(globl|lcomm)';
306 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
307 $T_DOT_GLOBAL = '\.globl';
308 $T_HDR_toc = "\.toc\n";
309 $T_HDR_literal = "\t\.const\n\t\.align 2\n";
310 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
311 $T_HDR_data = "\t\.data\n\t\.align 2\n";
312 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
313 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
314 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
315 $T_HDR_info = "\t\.text\n\t\.align 2\n";
316 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
317 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
319 #--------------------------------------------------------#
320 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ ) {
321 # Apple PowerPC Darwin/MacOS X.
322 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
323 $T_US = '_'; # _ if symbols have an underscore on the front
324 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
325 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
327 $T_X86_PRE_LLBL_PAT = 'L';
328 $T_X86_PRE_LLBL = 'L';
329 $T_X86_BADJMP = '^\tjmp [^L\*]';
331 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s.*|\.text|\.data|\.const_data|\.cstring|\.non_lazy_symbol_pointer|\.const|\.static_const|\.literal4|\.literal8|\.static_data|\.globl \S+|\.section .*|\.lcomm.*)\n)';
332 $T_COPY_DIRVS = '\.(globl|lcomm)';
334 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
335 $T_DOT_GLOBAL = '\.globl';
336 $T_HDR_toc = "\.toc\n";
337 $T_HDR_literal16= "\t\.literal8\n\t\.align 4\n";
338 $T_HDR_literal = "\t\.const\n\t\.align 4\n";
339 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
340 $T_HDR_data = "\t\.data\n\t\.align 2\n";
341 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
342 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
343 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
344 $T_HDR_info = "\t\.text\n\t\.align 2\n";
345 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
346 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
348 #--------------------------------------------------------#
349 } elsif ( $TargetPlatform =~ /^x86_64-apple-darwin.*/ ) {
350 # Apple PowerPC Darwin/MacOS X.
351 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
352 $T_US = '_'; # _ if symbols have an underscore on the front
353 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
354 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
357 $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 .*|\.lcomm.*)\n)';
358 $T_COPY_DIRVS = '\.(globl|lcomm)';
360 $T_DOT_WORD = '\.(quad|long|short|byte|fill|space)';
361 $T_DOT_GLOBAL = '\.globl';
362 $T_HDR_toc = "\.toc\n";
363 $T_HDR_literal16= "\t\.literal8\n\t\.align 4\n";
364 $T_HDR_literal = "\t\.const\n\t\.align 4\n";
365 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
366 $T_HDR_data = "\t\.data\n\t\.align 2\n";
367 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
368 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
369 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
370 $T_HDR_info = "\t\.text\n\t\.align 2\n";
371 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
372 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
374 #--------------------------------------------------------#
375 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ ) {
377 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
378 $T_US = ''; # _ if symbols have an underscore on the front
379 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
380 $T_CONST_LBL = '^\.LC\d+:'; # regexp for what such a lbl looks like
383 $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)';
384 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
386 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
387 $T_DOT_GLOBAL = '\.globl';
388 $T_HDR_toc = "\.toc\n";
389 $T_HDR_literal = "\t\.section\t.rodata\n\t\.align 2\n";
390 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
391 $T_HDR_data = "\t\.data\n\t\.align 2\n";
392 $T_HDR_rodata = "\t\.section\t.rodata\n\t\.align 2\n";
393 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
394 $T_HDR_info = "\t\.text\n\t\.align 2\n";
395 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
396 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
398 #--------------------------------------------------------#
399 } elsif ( $TargetPlatform =~ /^powerpc64-.*-linux/ ) {
401 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
402 $T_US = '\.'; # _ if symbols have an underscore on the front
403 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
404 $T_CONST_LBL = '^\.LC\d+:'; # regexp for what such a lbl looks like
407 $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)';
408 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
410 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
411 $T_DOT_GLOBAL = '\.globl';
412 $T_HDR_toc = "\.toc\n";
413 $T_HDR_literal = "\t\.section\t\".toc\",\"aw\"\n";
414 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
415 $T_HDR_data = "\t\.data\n\t\.align 2\n";
416 $T_HDR_rodata = "\t\.section\t.rodata\n\t\.align 2\n";
417 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
418 $T_HDR_info = "\t\.text\n\t\.align 2\n";
419 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
420 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
422 #--------------------------------------------------------#
423 } elsif ( $TargetPlatform =~ /^sparc-.*-(solaris2|openbsd)/ ) {
425 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
426 $T_US = ''; # _ if symbols have an underscore on the front
427 $T_PRE_APP = 'DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
428 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
431 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\s+\.local\s+\S+|\.text|\.data|\.stab.*|\s*\.section.*|\s+\.type.*|\s+\.size.*)\n)';
432 $T_COPY_DIRVS = '\.(global|local|proc|stab)';
434 $T_DOT_WORD = '\.(long|word|byte|half|skip|uahalf|uaword)';
435 $T_DOT_GLOBAL = '^\t\.global';
436 $T_HDR_literal = "\.text\n\t\.align 8\n";
437 $T_HDR_misc = "\.text\n\t\.align 4\n";
438 $T_HDR_data = "\.data\n\t\.align 8\n";
439 $T_HDR_rodata = "\.text\n\t\.align 4\n";
440 $T_HDR_closure = "\.data\n\t\.align 4\n";
441 $T_HDR_info = "\.text\n\t\.align 4\n";
442 $T_HDR_entry = "\.text\n\t\.align 4\n";
443 $T_HDR_vector = "\.text\n\t\.align 4\n";
445 #--------------------------------------------------------#
446 } elsif ( $TargetPlatform =~ /^sparc-.*-sunos4/ ) {
448 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
449 $T_US = '_'; # _ if symbols have an underscore on the front
450 $T_PRE_APP = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
451 $T_CONST_LBL = '^LC(\d+):$';
454 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*)\n)';
455 $T_COPY_DIRVS = '\.(global|proc|stab)';
457 $T_DOT_WORD = '\.word';
458 $T_DOT_GLOBAL = '^\t\.global';
459 $T_HDR_literal = "\.text\n\t\.align 8\n";
460 $T_HDR_misc = "\.text\n\t\.align 4\n";
461 $T_HDR_data = "\.data\n\t\.align 8\n";
462 $T_HDR_rodata = "\.text\n\t\.align 4\n";
463 $T_HDR_closure = "\.data\n\t\.align 4\n";
464 $T_HDR_info = "\.text\n\t\.align 4\n";
465 $T_HDR_entry = "\.text\n\t\.align 4\n";
466 $T_HDR_vector = "\.text\n\t\.align 4\n";
468 #--------------------------------------------------------#
469 } elsif ( $TargetPlatform =~ /^sparc-.*-linux/ ) {
470 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
471 $T_US = ''; # _ if symbols have an underscore on the front
472 $T_PRE_APP = '#'; # regexp that says what comes before APP/NO_APP
473 # Probably doesn't apply anyway
474 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
477 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\s+\.local\s+\S+|\.text|\.data|\.seg|\.stab.*|\s+?\.section.*|\s+\.type.*|\s+\.size.*)\n)';
478 $T_COPY_DIRVS = '\.(global|local|globl|proc|stab)';
480 $T_DOT_WORD = '\.(long|word|nword|xword|byte|half|short|skip|uahalf|uaword)';
481 $T_DOT_GLOBAL = '^\t\.global';
482 $T_HDR_literal = "\.text\n\t\.align 8\n";
483 $T_HDR_misc = "\.text\n\t\.align 4\n";
484 $T_HDR_data = "\.data\n\t\.align 8\n";
485 $T_HDR_rodata = "\.text\n\t\.align 4\n";
486 $T_HDR_closure = "\.data\n\t\.align 4\n";
487 $T_HDR_info = "\.text\n\t\.align 4\n";
488 $T_HDR_entry = "\.text\n\t\.align 4\n";
489 $T_HDR_vector = "\.text\n\t\.align 4\n";
491 #--------------------------------------------------------#
493 print STDERR "$Pgm: don't know how to mangle assembly language for: $TargetPlatform\n";
497 if($T_HDR_relrodata eq "") {
499 # relrodata defaults to rodata.
500 $T_HDR_relrodata = $T_HDR_rodata;
504 print STDERR "T_STABBY: $T_STABBY\n";
505 print STDERR "T_US: $T_US\n";
506 print STDERR "T_PRE_APP: $T_PRE_APP\n";
507 print STDERR "T_CONST_LBL: $T_CONST_LBL\n";
508 print STDERR "T_POST_LBL: $T_POST_LBL\n";
509 if ( $TargetPlatform =~ /^i386-/ ) {
510 print STDERR "T_X86_PRE_LLBL_PAT: $T_X86_PRE_LLBL_PAT\n";
511 print STDERR "T_X86_PRE_LLBL: $T_X86_PRE_LLBL\n";
512 print STDERR "T_X86_BADJMP: $T_X86_BADJMP\n";
514 print STDERR "T_MOVE_DIRVS: $T_MOVE_DIRVS\n";
515 print STDERR "T_COPY_DIRVS: $T_COPY_DIRVS\n";
516 print STDERR "T_DOT_WORD: $T_DOT_WORD\n";
517 print STDERR "T_HDR_literal: $T_HDR_literal\n";
518 print STDERR "T_HDR_misc: $T_HDR_misc\n";
519 print STDERR "T_HDR_data: $T_HDR_data\n";
520 print STDERR "T_HDR_rodata: $T_HDR_rodata\n";
521 print STDERR "T_HDR_closure: $T_HDR_closure\n";
522 print STDERR "T_HDR_info: $T_HDR_info\n";
523 print STDERR "T_HDR_entry: $T_HDR_entry\n";
524 print STDERR "T_HDR_vector: $T_HDR_vector\n";
530 %************************************************************************
532 \subsection{Mangle away}
534 %************************************************************************
538 local($in_asmf, $out_asmf) = @_;
540 # multi-line regexp matching:
544 # ia64-specific information for code chunks
548 &init_TARGET_STUFF();
549 &init_FUNNY_THINGS();
551 open(INASM, "< $in_asmf")
552 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
553 open(OUTASM,"> $out_asmf")
554 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
556 # read whole file, divide into "chunks":
557 # record some info about what we've found...
559 @chk = (); # contents of the chunk
560 $numchks = 0; # number of them
561 @chkcat = (); # what category of thing in each chunk
562 @chksymb = (); # what symbol(base) is defined in this chunk
563 %entrychk = (); # ditto, its entry code
564 %closurechk = (); # ditto, the (static) closure
565 %srtchk = (); # ditto, its SRT (for top-level things)
566 %infochk = (); # given a symbol base, say what chunk its info tbl is in
567 %vectorchk = (); # ditto, return vector table
568 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
570 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
573 tr/\r//d if $TargetPlatform =~ /-mingw32$/; # In case Perl doesn't convert line endings
574 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
575 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
576 next if /^\t\.def.*endef$/;
577 next if /${T_PRE_APP}(NO_)?APP/o;
578 next if /^;/ && $TargetPlatform =~ /^hppa/;
580 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|ia64)-/;
582 if ( $TargetPlatform =~ /^mips-/
583 && /^\t\.(globl\S+\.text|comm\t)/ ) {
584 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
585 # Treat .comm variables as data. These show up in two (known) places:
587 # - the module_registered variable used in the __stginit fragment.
588 # even though these are declared static and initialised, gcc 3.3
589 # likes to make them .comm, presumably to save space in the
592 # - global variables used to pass arguments from C to STG in
593 # a foreign export. (is this still true? --SDM)
595 } elsif ( /^\t\.comm.*$/ ) {
597 $chkcat[$i] = 'data';
600 # Labels ending "_str": these are literal strings.
601 } elsif ( /^${T_US}([A-Za-z0-9_]+)_str${T_POST_LBL}$/ ) {
603 $chkcat[$i] = 'relrodata';
605 } elsif ( $TargetPlatform =~ /-darwin/
606 && (/^\s*\.subsections_via_symbols/
607 ||/^\s*\.no_dead_strip.*/)) {
608 # Don't allow Apple's linker to do any dead-stripping of symbols
609 # in this file, because it will mess up info-tables in mangled
611 # The .no_dead_strip directives are actually put there by
612 # the gcc3 "used" attribute on entry points.
614 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && (
615 /^\s*\.picsymbol_stub/
616 || /^\s*\.section __TEXT,__picsymbol_stub\d,.*/
617 || /^\s*\.section __TEXT,__picsymbolstub\d,.*/
618 || /^\s*\.symbol_stub/
619 || /^\s*\.section __TEXT,__symbol_stub\d,.*/
620 || /^\s*\.section __TEXT,__symbolstub\d,.*/
621 || /^\s*\.lazy_symbol_pointer/
622 || /^\s*\.non_lazy_symbol_pointer/
623 || /^\s*\.section __IMPORT.*/))
626 $chkcat[$i] = 'dyld';
630 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.data/)
631 { # non_lazy_symbol_ptrs that point to local symbols
633 $chkcat[$i] = 'dyld';
636 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.align/)
637 { # non_lazy_symbol_ptrs that point to local symbols
639 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^L_.*:$/)
640 { # non_lazy_symbol_ptrs that point to local symbols
641 $chk[++$i] = $dyld_section . $_;
642 $chkcat[$i] = 'dyld';
645 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
646 # duplicated from the bottom
650 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
651 # Alphas: Local labels not to be confused with new chunks
653 # NB: all the rest start with a non-space
655 } elsif ( $TargetPlatform =~ /^mips-/
656 && /^\d+:/ ) { # a funny-looking very-local label
659 } elsif ( /$T_CONST_LBL/o ) {
661 $chkcat[$i] = 'literal';
664 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
666 $chkcat[$i] = 'splitmarker';
669 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
672 $chkcat[$i] = 'infotbl';
673 $chksymb[$i] = $symb;
675 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
677 $infochk{$symb} = $i;
679 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
681 $chkcat[$i] = 'entry';
686 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
688 $chkcat[$i] = 'closure';
691 $closurechk{$1} = $i;
693 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
700 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
702 $chkcat[$i] = 'data';
705 } elsif ( /^${T_US}(stg_ap_stack_entries|stg_stack_save_entries|stg_arg_bitmaps)${T_POST_LBL}$/o ) {
707 $chkcat[$i] = 'data';
710 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
713 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
714 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
715 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
716 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
719 $chkcat[$i] = 'data';
722 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
727 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
728 # all CC_ symbols go in the data section...
730 $chkcat[$i] = 'data';
733 } elsif ( /^${T_US}([A-Za-z0-9_]+)_hpc${T_POST_LBL}$/o ) {
734 # hpc shares tick boxes across modules
736 $chkcat[$i] = 'data';
739 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
741 $chkcat[$i] = 'misc';
743 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
745 $chkcat[$i] = 'vector';
750 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
751 && /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
752 # Some Solaris system headers contain function definitions (as
753 # opposed to mere prototypes), which end up in the .hc file when
754 # a Haskell module foreign imports the corresponding system
755 # functions (most notably stat()). We put them into the text
756 # segment. Note that this currently does not extend to function
757 # names starting with an underscore.
760 $chkcat[$i] = 'misc';
763 } elsif ( $TargetPlatform =~ /^i386-apple-darwin/ && /^(___i686\.get_pc_thunk\.[abcd]x):/o) {
764 # To handle PIC on Darwin/x86, we need to appropriately pass through
765 # the get_pc_thunk functions. The need to be put into a special section
766 # marked as coalesced (otherwise the .weak_definition doesn't work
769 $chkcat[$i] = 'get_pc_thunk';
772 } elsif ( /^${T_US}[A-Za-z0-9_]/o
773 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
775 && ( $TargetPlatform !~ /^powerpc64/ # we need to avoid local labels in this case
776 || ! /^\.L\d+:$/ ) ) {
783 /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
784 || /^${T_US}__stg_.*${T_POST_LBL}$/o # more RTS internals
785 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
786 || /^${T_US}.*_slow${T_POST_LBL}$/o # slow entry
787 || /^${T_US}__stginit.*${T_POST_LBL}$/o # __stginit<module>
788 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
789 || /^${T_US}.*_fast${T_POST_LBL}$/o # primops
790 || /^_uname:/o # x86/Solaris2
793 $chkcat[$i] = 'misc';
795 /^${T_US}.*_srtd${T_POST_LBL}$/o # large bitmaps
796 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
799 $chkcat[$i] = 'relrodata';
802 print STDERR "Warning: retaining unknown function \`$thing' in output from C compiler\n";
803 $chkcat[$i] = 'unknown';
806 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ && /^\.LCTOC1 = /o ) {
807 # PowerPC Linux's large-model PIC (-fPIC) generates a gobal offset
808 # table "by hand". Be sure to copy it over.
809 # Note that this label and all entries in the table should actually
810 # go into the .got2 section, but it isn't easy to distinguish them
811 # from other constant literals (.LC\d+), so we just put everything
814 $chkcat[$i] = 'literal';
815 $chksymb[$i] = 'LCTOC1';
816 } else { # simple line (duplicated at the top)
821 $numchks = $#chk + 1;
822 $chk[$numchks] = ''; # We might push .note.GNU-stack into this
823 $chkcat[$numchks] = 'verbatim'; # If we do, write it straight back out
825 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
826 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
829 # the division into chunks is imperfect;
830 # we throw some things over the fence into the next
833 # also, there are things we would like to know
834 # about the whole module before we start spitting
837 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
838 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
840 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
842 # Alphas: NB: we start meddling at chunk 1, not chunk 0
843 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
844 # spits a ".quad 0" in after the very first ".rdata"; we
845 # detect this special case (tossing the ".quad 0")!
846 local($magic_rdata_seen) = 0;
848 # HPPAs, MIPSen: also start medding at chunk 1
850 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
851 $c = $chk[$i]; # convenience copy
853 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
855 # toss all prologue stuff; HPPA is pretty weird
857 $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
862 # be slightly paranoid to make sure there's
863 # nothing surprising in there
864 if ( $c =~ /--- BEGIN ---/ ) {
865 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
867 # remove junk whitespace around the split point
871 if ($TargetPlatform =~ /^i386-/) {
872 if ($p =~ /^\tsubl\s+\$(\d+),\s*\%esp\n/) {
874 die "Error: reserved stack space exceeded!\n Possible workarounds: compile with -fasm, or try another version of gcc.\n"
878 # gcc 3.4.3 puts this kind of stuff in the prologue, eg.
879 # when compiling PrimOps.cmm with -optc-O2:
882 # movl %ecx, 16(%esp)
883 # movl %edx, 20(%esp)
884 # but then the code of the function doesn't assume
885 # anything about the contnets of these stack locations.
886 # I think it's to do with the use of inline functions for
887 # PK_Word64() and friends, where gcc is initialising the
888 # contents of the struct to zero, and failing to optimise
889 # away the initialisation. Let's live dangerously and
890 # discard these initalisations.
892 $p =~ s/^\tpushl\s+\%e(di|si|bx)\n//g;
893 $p =~ s/^\txorl\s+\%e(ax|cx|dx),\s*\%e(ax|cx|dx)\n//g;
894 $p =~ s/^\tmovl\s+\%e(ax|cx|dx|si|di),\s*\d*\(\%esp\)\n//g;
895 $p =~ s/^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n//g;
896 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
897 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-(cygwin32|mingw32)/);
899 if ($TargetPlatform =~ /^i386-apple-darwin/) {
901 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
903 $pcrel_reg =~ s/(.|\n)*.*___i686\.get_pc_thunk\.([abcd]x)\n(.|\n)*/$2/ or $pcrel_reg = "";
904 $p =~ s/^\s+call\s+___i686\.get_pc_thunk\..x//;
905 $p =~ s/^\"?L\d+\$pb\"?:\n//;
907 if ($pcrel_reg eq "bx") {
908 # Bad gcc. Goes and uses %ebx, our BaseReg, for PIC. Bad gcc.
909 die "Darwin/x86: -fPIC -via-C doesn't work yet, use -fasm. Aborting."
913 } elsif ($TargetPlatform =~ /^x86_64-/) {
914 $p =~ s/^\tpushq\s+\%r(bx|bp|12|13|14)\n//g;
915 $p =~ s/^\tmovq\s+\%r(bx|bp|12|13|14),\s*\d*\(\%rsp\)\n//g;
916 $p =~ s/^\tsubq\s+\$\d+,\s*\%rsp\n//;
918 } elsif ($TargetPlatform =~ /^ia64-/) {
919 $p =~ s/^\t\.prologue .*\n//;
921 # Record the number of local and out registers for register relocation later
922 $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, (\d+), (\d+), 0\n//;
926 $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
927 $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
929 # Ignore save/restore of these registers; they're taken
930 # care of in StgRun()
931 $p =~ s/^\t\.save ar\.lc, r\d+\n//;
932 $p =~ s/^\t\.save pr, r\d+\n//;
933 $p =~ s/^\tmov r\d+ = ar\.lc\n//;
934 $p =~ s/^\tmov r\d+ = pr\n//;
936 # Remove .proc and .body directives
937 $p =~ s/^\t\.proc [a-zA-Z0-9_.]+#\n//;
938 $p =~ s/^\t\.body\n//;
940 # If there's a label, move it to the body
941 if ($p =~ /^[a-zA-Z0-9.]+:\n/) {
946 # Remove floating-point spill instructions.
947 # Only fp registers 2-5 and 16-23 are saved by the runtime.
948 if ($p =~ s/^\tstf\.spill \[r1[4-9]\] = f([2-5]|1[6-9]|2[0-3])(, [0-9]+)?\n//g) {
949 # Being paranoid, only try to remove these if we saw a
951 $p =~ s/^\tmov r1[4-9] = r12\n//;
952 $p =~ s/^\tadds r1[4-9] = -[0-9]+, r12\n//g;
953 $p =~ s/^\t\.save\.f 0x[0-9a-fA-F]\n//g;
954 $p =~ s/^\t\.save\.gf 0x0, 0x[0-9a-fA-F]+\n//g;
957 $p =~ s/^\tnop(?:\.[mifb])?\s+\d+\n//g; # remove nop instructions
958 $p =~ s/^\t\.(mii|mmi|mfi)\n//g; # bundling is no longer sensible
959 $p =~ s/^\t;;\n//g; # discard stops
960 $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
962 # GCC 3.3 saves r1 in the prologue, move this to the body
963 # (Does this register get restored anywhere?)
964 if ($p =~ /^\tmov r\d+ = r1\n/) {
968 } elsif ($TargetPlatform =~ /^m68k-/) {
969 $p =~ s/^\tlink a6,#-?\d.*\n//;
970 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
971 # The above showed up in the asm code,
972 # so I added it here.
973 # I hope it's correct.
975 $p =~ s/^\tmovel d2,sp\@-\n//;
976 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
977 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
978 } elsif ($TargetPlatform =~ /^mips-/) {
979 # the .frame/.mask/.fmask that we use is the same
980 # as that produced by GCC for miniInterpret; this
981 # gives GDB some chance of figuring out what happened
982 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
983 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
984 $p =~ s/^\t\.(mask|fmask).*\n//g;
985 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
986 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
987 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
988 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
989 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
990 $p =~ s/__FRAME__/$FRAME/;
991 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
993 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
995 $p =~ s/^\tmflr r0\n//;
996 $p =~ s/^\tbl saveFP # f\d+\n//;
997 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
998 $p =~ s/^\"?L\d+\$pb\"?:\n//;
999 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
1000 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
1001 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
1002 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
1003 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
1004 $p =~ s/^\tbcl 20,31,\"?L\d+\$pb\"?\n//;
1005 $p =~ s/^\"?L\d+\$pb\"?:\n//;
1006 $p =~ s/^\tmflr r31\n//;
1008 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1009 # under some circumstances, only when generating position dependent code.
1010 # I have no idea why, and I don't think it is necessary, so let's toss it.
1011 $p =~ s/^\tli r\d+,0\n//g;
1012 $p =~ s/^\tstw r\d+,\d+\(r1\)\n//g;
1013 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/) {
1014 $p =~ s/^\tmflr 0\n//;
1015 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
1016 $p =~ s/^\tstfd \d+,\d+\(1\)\n//g;
1017 $p =~ s/^\tstw r0,8\(1\)\n//;
1018 $p =~ s/^\tstwu 1,-\d+\(1\)\n//;
1019 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1021 # GCC's "large-model" PIC (-fPIC)
1023 $pcrel_label =~ s/(.|\n)*^.LCF(\d+):\n(.|\n)*/$2/ or $pcrel_label = "";
1025 $p =~ s/^\tbcl 20,31,.LCF\d+\n//;
1026 $p =~ s/^.LCF\d+:\n//;
1027 $p =~ s/^\tmflr 30\n//;
1028 $p =~ s/^\tlwz 0,\.LCL\d+-\.LCF\d+\(30\)\n//;
1029 $p =~ s/^\tadd 30,0,30\n//;
1031 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1032 # under some circumstances, only when generating position dependent code.
1033 # I have no idea why, and I don't think it is necessary, so let's toss it.
1034 $p =~ s/^\tli \d+,0\n//g;
1035 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1036 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1037 $p =~ s/^\tmr 31,1\n//;
1038 $p =~ s/^\tmflr 0\n//;
1039 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
1040 $p =~ s/^\tstfd \d+,-?\d+\(1\)\n//g;
1041 $p =~ s/^\tstd r0,8\(1\)\n//;
1042 $p =~ s/^\tstdu 1,-\d+\(1\)\n//;
1043 $p =~ s/^\tstd \d+,-?\d+\(1\)\n//g;
1045 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1046 # under some circumstances, only when generating position dependent code.
1047 # I have no idea why, and I don't think it is necessary, so let's toss it.
1048 $p =~ s/^\tli \d+,0\n//g;
1049 $p =~ s/^\tstd \d+,\d+\(1\)\n//g;
1051 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
1054 # HWL HACK: dont die, just print a warning
1055 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
1056 die "Prologue junk?: $p\n" if $p =~ /^\s+[^\s\.]/;
1058 # For PIC, we want to keep part of the prologue
1059 if ($TargetPlatform =~ /^powerpc-apple-darwin.*/ && $pcrel_label ne "") {
1060 # Darwin: load the current instruction pointer into register r31
1061 $p .= "bcl 20,31,$pcrel_label\n";
1062 $p .= "$pcrel_label:\n";
1063 $p .= "\tmflr r31\n";
1064 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/ && $pcrel_label ne "") {
1065 # Linux: load the GOT pointer into register 30
1066 $p .= "\tbcl 20,31,.LCF$pcrel_label\n";
1067 $p .= ".LCF$pcrel_label:\n";
1068 $p .= "\tmflr 30\n";
1069 $p .= "\tlwz 0,.LCL$pcrel_label-.LCF$pcrel_label(30)\n";
1070 $p .= "\tadd 30,0,30\n";
1071 } elsif ($TargetPlatform =~ /^i386-apple-darwin.*/ && $pcrel_label ne "") {
1072 $p .= "\tcall ___i686.get_pc_thunk.$pcrel_reg\n";
1073 $p .= "$pcrel_label:\n";
1076 # glue together what's left
1081 if ( $TargetPlatform =~ /^mips-/ ) {
1082 # MIPS: first, this basic sequence may occur "--- END ---" or not
1083 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
1086 # toss all epilogue stuff; again, paranoidly
1087 if ( $c =~ /--- END ---/ ) {
1088 # Gcc may decide to replicate the function epilogue. We want
1089 # to process all epilogues, so we split the function and then
1091 @fragments = split(/--- END ---/, $c);
1092 $r = shift(@fragments);
1094 # Rebuild `c'; processed fragments will be appended to `c'
1097 foreach $e (@fragments) {
1098 # etail holds code that is after the epilogue in the assembly-code
1099 # layout and should not be filtered as part of the epilogue.
1101 if ($TargetPlatform =~ /^i386-/) {
1103 $e =~ s/^\tpopl\s+\%edi\n//;
1104 $e =~ s/^\tpopl\s+\%esi\n//;
1105 $e =~ s/^\tpopl\s+\%edx\n//;
1106 $e =~ s/^\tpopl\s+\%ecx\n//;
1107 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
1108 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
1109 } elsif ($TargetPlatform =~ /^ia64-/) {
1110 # The epilogue is first split into:
1111 # $e, the epilogue code (up to the return instruction)
1112 # $etail, non-epilogue code (after the return instruction)
1113 # The return instruction is stripped in the process.
1114 if (!(($e, $etail) = split(/^\tbr\.ret\.sptk\.many b0\n/, $e))) {
1115 die "Epilogue doesn't seem to have one return instruction: $e\n";
1117 # Remove 'endp' directive from the tail
1118 $etail =~ s/^\t\.endp [a-zA-Z0-9_.]+#\n//;
1120 # If a return value is saved here, discard it
1121 $e =~ s/^\tmov r8 = r14\n//;
1123 # Remove floating-point fill instructions.
1124 # Only fp registers 2-5 and 16-23 are saved by the runtime.
1125 if ($e =~ s/^\tldf\.fill f([2-5]|1[6-9]|2[0-3]) = \[r1[4-9]\](, [0-9]+)?\n//g) {
1126 # Being paranoid, only try to remove this if we saw a fill
1128 $e =~ s/^\tadds r1[4-9] = [0-9]+, r12//g;
1131 $e =~ s/^\tnop(?:\.[mifb])?\s+\d+\n//g; # remove nop instructions
1132 $e =~ s/^\tmov ar\.pfs = r\d+\n//;
1133 $e =~ s/^\tmov ar\.lc = r\d+\n//;
1134 $e =~ s/^\tmov pr = r\d+, -1\n//;
1135 $e =~ s/^\tmov b0 = r\d+\n//;
1136 $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
1137 #$e =~ s/^\tbr\.ret\.sptk\.many b0\n//; # already removed
1138 $e =~ s/^\t\.(mii|mmi|mfi|mib)\n//g; # bundling is no longer sensible
1139 $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
1140 $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
1141 } elsif ($TargetPlatform =~ /^m68k-/) {
1142 $e =~ s/^\tunlk a6\n//;
1144 } elsif ($TargetPlatform =~ /^mips-/) {
1145 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
1146 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
1147 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
1148 $e =~ s/^\tj\t\$31\n//;
1149 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
1150 $e =~ s/^\taddi r1,r1,\d+\n//;
1151 $e =~ s/^\tlwz r\d+,\d+\(r1\)\n//;
1152 $e =~ s/^\tlmw r\d+,-\d+\(r1\)\n//;
1153 $e =~ s/^\tmtlr r0\n//;
1155 $e =~ s/^\tb restFP ;.*\n//;
1156 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1157 $e =~ s/^\tmr 3,0\n//;
1158 $e =~ s/^\taddi 1,1,\d+\n//;
1159 $e =~ s/^\tld 0,16\(1\)\n//;
1160 $e =~ s/^\tmtlr 0\n//;
1162 # callee-save registers
1163 $e =~ s/^\tld \d+,-?\d+\(1\)\n//g;
1164 $e =~ s/^\tlfd \d+,-?\d+\(1\)\n//g;
1166 # get rid of the debug junk along with the blr
1167 $e =~ s/^\tblr\n\t.long .*\n\t.byte .*\n//;
1169 # incase we missed it with the last one get the blr alone
1172 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
1175 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
1177 # glue together what's left
1180 $c =~ s/\n\t\n/\n/; # junk blank line
1183 if ($TargetPlatform =~ /^ia64-/) {
1184 # On IA64, remove an .endp directive even if no epilogue was found.
1185 # Code optimizations may have removed the "--- END ---" token.
1186 $c =~ s/^\t\.endp [a-zA-Z0-9_.]+#\n//;
1190 # On SPARCs, we don't do --- BEGIN/END ---, we just
1191 # toss the register-windowing save/restore/ret* instructions
1192 # directly unless they've been generated by function definitions in header
1194 if ( $TargetPlatform =~ /^sparc-/ ) {
1195 if ( ! ( $TargetPlatform =~ /solaris2$/ && $chkcat[$i] eq 'unknown' )) {
1196 $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
1198 # throw away PROLOGUE comments
1199 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
1202 # On Alphas, the prologue mangling is done a little later (below)
1204 # toss all calls to __DISCARD__
1205 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
1206 $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
1207 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-darwin.*/;
1208 $c =~ s/^\tbl\s+__DISCARD__(\@plt)?\n//go if $TargetPlatform =~ /^powerpc-.*-linux/;
1209 $c =~ s/^\tbl\s+\.__DISCARD__\n\s+nop\n//go if $TargetPlatform =~ /^powerpc64-.*-linux/;
1210 $c =~ s/^\tcall\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /i386-apple-darwin.*/;
1212 # IA64: fix register allocation; mangle tailcalls into jumps
1213 if ($TargetPlatform =~ /^ia64-/) {
1214 ia64_rename_registers($ia64_locnum, $ia64_outnum) if (defined($ia64_locnum));
1215 ia64_mangle_tailcalls();
1218 # MIPS: that may leave some gratuitous asm macros around
1219 # (no harm done; but we get rid of them to be tidier)
1220 $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/
1221 if $TargetPlatform =~ /^mips-/;
1223 # toss stack adjustment after DoSparks
1224 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
1225 if $TargetPlatform =~ /^m68k-/; # this looks old...
1227 if ( $TargetPlatform =~ /^alpha-/ &&
1228 ! $magic_rdata_seen &&
1229 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
1230 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
1231 $magic_rdata_seen = 1;
1234 # pick some end-things and move them to the next chunk
1236 # pin a funny end-thing on (for easier matching):
1237 $c .= 'FUNNY#END#THING';
1239 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
1243 # on x86 we try not to copy any directives into a literal
1244 # chunk, rather we keep looking for the next real chunk. This
1245 # is because we get things like
1247 # .globl blah_closure
1253 if ( $TargetPlatform =~ /^(i386|sparc|powerpc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
1255 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
1258 if ( $j < $numchks ) {
1259 $chk[$j] = $to_move . $chk[$j];
1263 elsif ( ( $i < ($numchks - 1)
1264 && ( $to_move =~ /${T_COPY_DIRVS}/
1265 || ( $TargetPlatform =~ /^hppa/
1266 && $to_move =~ /align/
1267 && $chkcat[$i+1] eq 'literal')
1270 || ($to_move =~ /^[ \t]*\.section[ \t]+\.note\.GNU-stack,/)
1272 $chk[$i + 1] = $to_move . $chk[$i + 1];
1273 # otherwise they're tossed
1276 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
1279 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
1281 # toss all prologue stuff, except for loading gp, and the ..ng address
1282 unless ($c =~ /\.ent.*\n\$.*\.\.ng:/) {
1283 if (($p, $r) = split(/^\t\.prologue/, $c)) {
1284 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
1285 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
1286 $keep =~ s/^\t\.(mask|fmask).*\n//g;
1287 $c = $keep . "..ng:\n";
1289 print STDERR "malformed code block ($ent)?\n"
1292 $c .= "\t.prologue" . $r;
1296 $c =~ s/FUNNY#END#THING//;
1298 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
1300 $chk[$i] = $c; # update w/ convenience copy
1303 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
1304 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
1307 if ( $TargetPlatform =~ /^alpha-/ ) {
1308 # print out the header stuff first
1309 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
1310 print OUTASM $chk[0];
1312 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1313 print OUTASM $chk[0];
1315 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1316 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
1318 # get rid of horrible "<dollar>Revision: .*$" strings
1319 local(@lines0) = split(/\n/, $chk[0]);
1321 while ( $z <= $#lines0 ) {
1322 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
1325 while ( $z <= $#lines0 ) {
1327 last if $lines0[$z] =~ /[,\t]0x0$/;
1333 $chk[0] = join("\n", @lines0);
1334 $chk[0] =~ s/\n\n+/\n/;
1335 print OUTASM $chk[0];
1338 # print out all the literal strings next
1339 for ($i = 0; $i < $numchks; $i++) {
1340 if ( $chkcat[$i] eq 'literal' ) {
1342 # HACK: try to detect 16-byte constants and align them
1343 # on a 16-byte boundary. x86_64 sometimes needs 128-bit
1344 # aligned constants, and so does Darwin/x86.
1345 if ( $TargetPlatform =~ /^x86_64/
1346 || $TargetPlatform =~ /^i386-apple-darwin/ ) {
1348 if ($z =~ /(\.long.*\n.*\.long.*\n.*\.long.*\n.*\.long|\.quad.*\n.*\.quad)/) {
1349 print OUTASM $T_HDR_literal16;
1351 print OUTASM $T_HDR_literal;
1354 print OUTASM $T_HDR_literal;
1357 print OUTASM $chk[$i];
1358 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
1360 $chkcat[$i] = 'DONE ALREADY';
1364 # on the HPPA, print out all the bss next
1365 if ( $TargetPlatform =~ /^hppa/ ) {
1366 for ($i = 1; $i < $numchks; $i++) {
1367 if ( $chkcat[$i] eq 'bss' ) {
1368 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
1369 print OUTASM $chk[$i];
1371 $chkcat[$i] = 'DONE ALREADY';
1376 # $numchks + 1 as we have the extra one for .note.GNU-stack
1377 for ($i = $FIRST_MANGLABLE; $i < $numchks + 1; $i++) {
1378 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
1380 next if $chkcat[$i] eq 'DONE ALREADY';
1382 if ( $chkcat[$i] eq 'misc' || $chkcat[$i] eq 'unknown' ) {
1383 if ($chk[$i] ne '') {
1384 print OUTASM $T_HDR_misc;
1385 &print_doctored($chk[$i], 0);
1388 } elsif ( $chkcat[$i] eq 'verbatim' ) {
1389 print OUTASM $chk[$i];
1391 } elsif ( $chkcat[$i] eq 'toss' ) {
1392 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
1394 } elsif ( $chkcat[$i] eq 'data' ) {
1395 if ($chk[$i] ne '') {
1396 print OUTASM $T_HDR_data;
1397 print OUTASM $chk[$i];
1400 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1401 # we can just re-constitute this one...
1402 # NB: we emit _three_ underscores no matter what,
1403 # so ghc-split doesn't have to care.
1404 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1406 } elsif ( $chkcat[$i] eq 'closure'
1407 || $chkcat[$i] eq 'srt'
1408 || $chkcat[$i] eq 'infotbl'
1409 || $chkcat[$i] eq 'entry') { # do them in that order
1410 $symb = $chksymb[$i];
1413 if ( defined($closurechk{$symb}) ) {
1414 print OUTASM $T_HDR_closure;
1415 print OUTASM $chk[$closurechk{$symb}];
1416 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1420 if ( defined($srtchk{$symb}) ) {
1421 print OUTASM $T_HDR_relrodata;
1422 print OUTASM $chk[$srtchk{$symb}];
1423 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1427 if ( defined($infochk{$symb}) ) {
1429 print OUTASM $T_HDR_info;
1430 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1432 # entry code will be put here!
1434 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1438 if ( defined($entrychk{$symb}) ) {
1440 $c = $chk[$entrychk{$symb}];
1442 # If this is an entry point with an info table,
1443 # eliminate the entry symbol and all directives involving it.
1444 if (defined($infochk{$symb}) && $TargetPlatform !~ /^ia64-/) {
1446 foreach $l (split(/\n/,$c)) {
1447 next if $l =~ /^.*$symb_(entry|ret)${T_POST_LBL}/;
1449 # If we have .type/.size direrctives involving foo_entry,
1450 # then make them refer to foo_info instead. The information
1451 # in these directives is used by the cachegrind annotator,
1452 # so it is worthwhile keeping.
1453 if ($l =~ /^\s*\.(type|size).*$symb_(entry|ret)/) {
1454 $l =~ s/$symb(_entry|_ret)/${symb}_info/g;
1458 next if $l =~ /^\s*\..*$symb.*\n?/;
1461 $c = join("\n",@o) . "\n";
1464 print OUTASM $T_HDR_entry;
1466 &print_doctored($c, 1); # NB: the 1!!!
1468 $chkcat[$entrychk{$symb}] = 'DONE ALREADY';
1471 } elsif ( $chkcat[$i] eq 'vector' ) {
1472 $symb = $chksymb[$i];
1475 if ( defined($vectorchk{$symb}) ) {
1476 print OUTASM $T_HDR_vector;
1477 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1479 # direct return code will be put here!
1480 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1482 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1483 # Alphas: the commented nop is for the splitter, to ensure
1484 # that no module ends with a label as the very last
1485 # thing. (The linker will adjust the label to point
1486 # to the first code word of the next module linked in,
1487 # even if alignment constraints cause the label to move!)
1489 print OUTASM "\t# nop\n";
1492 } elsif ( $chkcat[$i] eq 'rodata' ) {
1493 print OUTASM $T_HDR_rodata;
1494 print OUTASM $chk[$i];
1495 $chkcat[$i] = 'DONE ALREADY';
1496 } elsif ( $chkcat[$i] eq 'relrodata' ) {
1497 print OUTASM $T_HDR_relrodata;
1498 print OUTASM $chk[$i];
1499 $chkcat[$i] = 'DONE ALREADY';
1500 } elsif ( $chkcat[$i] eq 'toc' ) {
1501 # silly optimisation to print tocs, since they come in groups...
1502 print OUTASM $T_HDR_toc;
1504 while ($chkcat[$j] eq 'toc')
1505 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1508 print OUTASM $chk[$j];
1510 $chkcat[$j] = 'DONE ALREADY';
1514 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' ) {
1515 # apple-darwin: dynamic linker stubs
1516 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1517 { # print them out unchanged, but remove the stubs for __DISCARD__
1518 print OUTASM $chk[$i];
1520 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ && $chkcat[$i] eq 'get_pc_thunk' ) {
1521 # i386-apple-darwin: __i686.get_pc_thunk.[abcd]x
1522 print OUTASM ".section __TEXT,__textcoal_nt,coalesced,no_toc\n";
1523 print OUTASM $chk[$i];
1525 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1529 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1532 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1533 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1537 On IA64, tail calls are converted to branches at this point. The mangler
1538 searches for function calls immediately followed by a '--- TAILCALL ---'
1539 token. Since the compiler can put various combinations of labels, bundling
1540 directives, nop instructions, stops, and a move of the return value
1541 between the branch and the tail call, proper matching of the tail call
1542 gets a little hairy. This subroutine does the mangling.
1544 Here is an example of a tail call before mangling:
1547 br.call.sptk.many b0 = b6
1562 sub ia64_mangle_tailcalls {
1563 # Function input and output are in $c
1565 # Construct the tailcall-mangling expression the first time this function
1567 if (!defined($IA64_MATCH_TAILCALL)) {
1568 # One-line pattern matching constructs. None of these
1569 # should bind references; all parenthesized terms
1570 # should be (?:) terms.
1571 my $stop = q/(?:\t;;\n)/;
1572 my $bundle = q/(?:\t\.(?:mii|mib|mmi|mmb|mfi|mfb|mbb|bbb)\n)/;
1573 my $nop = q/(?:\tnop(?:\.[mifb])?\s+\d+\n)/;
1574 my $movgp = q/(?:\tmov r1 = r\d+\n)/;
1575 my $postbr = q/(?:\tbr \.L\d+\n)/;
1577 my $noeffect = "(?:$stop$bundle?|$nop)*";
1578 my $postbundle = "(?:$bundle?$nop?$nop?$postbr)?";
1580 # Important parts of the pattern match. The branch target
1581 # and subsequent jump label are bound to $1 and $2
1582 # respectively. Sometimes there is no label.
1583 my $callbr = q/^\tbr\.call\.sptk\.many b0 = (.*)\n/;
1584 my $label = q/(?:^\.L([0-9]*):\n)/;
1585 my $tailcall = q/\t--- TAILCALL ---\n/;
1587 $IA64_MATCH_TAILCALL =
1588 $callbr . $label . '?' . $noeffect . $movgp . '?' . $noeffect .
1589 $tailcall . $stop . '?' . '(?:' . $postbundle . ')?';
1592 # Find and mangle tailcalls
1593 while ($c =~ s/$IA64_MATCH_TAILCALL/\tbr\.few $1\n/o) {
1594 # Eek, the gcc optimiser is getting smarter... if we see a jump to the
1595 # --- TAILCALL --- marker then we reapply the substitution at the source sites
1596 $c =~ s/^\tbr \.L$2\n/\t--- TAILCALL ---\n/g if ($2);
1599 # Verify that all instances of TAILCALL were processed
1600 if ($c =~ /^\t--- TAILCALL ---\n/) {
1601 die "Unmangled TAILCALL tokens remain after mangling"
1606 The number of registers allocated on the IA64 register stack is set
1607 upon entry to the runtime with an `alloc' instruction at the entry
1608 point of \verb+StgRun()+. Gcc uses its own `alloc' to allocate
1609 however many registers it likes in each function. When we discard
1610 gcc's alloc, we have to reconcile its register assignment with what
1613 There are three stack areas: fixed registers, input/local registers,
1614 and output registers. We move the output registers to the output
1615 register space and leave the other registers where they are.
1618 sub ia64_rename_registers() {
1619 # The text to be mangled is in $c
1620 # Find number of registers in each stack area
1621 my ($loc, $out) = @_;
1627 # These are the register numbers used in the STG runtime
1628 my $STG_FIRST_OUT_REG = 32 + 34;
1629 my $STG_LAST_OUT_REG = $STG_FIRST_OUT_REG + 7;
1631 $first_out_reg = 32 + $loc;
1633 if ($first_out_reg > $STG_FIRST_OUT_REG) {
1634 die "Too many local registers allocated by gcc";
1637 # Split the string into fragments containing one register name each.
1638 # Rename the register in each fragment and concatenate.
1640 foreach $fragment (split(/(?=r\d+[^a-zA-Z0-9_.])/s, $c)) {
1641 if ($fragment =~ /^r(\d+)((?:[^a-zA-Z0-9_.].*)?)$/s) {
1644 if ($regnum < $first_out_reg) {
1645 # This is a local or fixed register
1647 # Local registers 32 and 33 (r64 and r65) are
1648 # used to hold saved state; they shouldn't be touched
1649 if ($regnum == 64 || $regnum == 65) {
1650 die "Reserved register $regnum is in use";
1654 # This is an output register
1655 $regnum = $regnum - $first_out_reg + $STG_FIRST_OUT_REG;
1656 if ($regnum > $STG_LAST_OUT_REG) {
1657 die "Register number ($regnum) is out of expected range";
1661 # Update this fragment
1662 $fragment = "r" . $regnum . $2;
1673 sub hppa_mash_prologue { # OK, epilogue, too
1676 # toss all prologue stuff
1677 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1679 # Lie about our .CALLINFO
1680 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1687 # toss all epilogue stuff
1688 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1690 # Sorry; we moved the _info stuff to the code segment.
1691 s/_info,DATA/_info,CODE/g;
1698 sub print_doctored {
1699 local($_, $need_fallthru_patch) = @_;
1701 if ( $TargetPlatform =~ /^x86_64-/ ) {
1704 # movq -4(%ebp), %rax
1709 s/^\tmovq\s+(-?\d*\(\%r(bx|bp|13)\)),\s*(\%r(ax|cx|dx|10|11))\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1710 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e(ax|cx|si|di))\n\tjmp\s+\*\%r\3/\tjmp\t$T_US$1/g;
1713 if ( $TargetPlatform !~ /^i386-/
1714 || ! /^\t[a-z]/ # no instructions in here, apparently
1715 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1720 # OK, must do some x86 **HACKING**
1722 local($entry_patch) = '';
1723 local($exit_patch) = '';
1725 # gotta watch out for weird instructions that
1726 # invisibly smash various regs:
1727 # rep* %ecx used for counting
1728 # scas* %edi used for destination index
1729 # cmps* %e[sd]i used for indices
1730 # loop* %ecx used for counting
1735 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1737 # * GCC used an "STG reg" for its own purposes
1739 # * some secret uses of machine reg, requiring STG reg
1740 # to be saved/restored
1742 # The most dangerous "GCC uses" of an "STG reg" are when
1743 # the reg holds the target of a jmp -- it's tricky to
1744 # insert the patch-up code before we get to the target!
1745 # So here we change the jmps:
1747 # --------------------------------------------------------
1748 # it can happen that we have jumps of the form...
1749 # jmp *<something involving %esp>
1751 # jmp <something involving another naughty register...>
1753 # a reasonably-common case is:
1755 # movl $_blah,<bad-reg>
1758 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[acd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1762 # movl -4(%ebx), %eax
1767 s/^\tmovl\s+(-?\d*\(\%e(bx|si)\)),\s*(\%e[acd]x)\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1769 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1770 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1771 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1772 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1773 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1774 if /(jmp|call)\s+.*\%esi/;
1776 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1777 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1778 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1779 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1780 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1781 if /(jmp|call)\s+.*\%edi/;
1784 # OK, now we can decide what our patch-up code is going to
1787 # Offsets into register table - you'd better update these magic
1788 # numbers should you change its contents!
1789 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1790 local($OFFSET_Hp)=88;
1792 # Note funky ".=" stuff; we're *adding* to these _patch guys
1793 if ( $StolenX86Regs <= 2
1794 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1795 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1796 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1798 # nothing for call_{entry,exit} because %esi is callee-save
1800 if ( $StolenX86Regs <= 3
1801 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1802 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1803 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1805 # nothing for call_{entry,exit} because %edi is callee-save
1808 # --------------------------------------------------------
1809 # next, here we go with non-%esp patching!
1811 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1813 # Before calling GC we must set up the exit condition before the call
1814 # and entry condition when we come back
1816 # fix _all_ non-local jumps:
1818 if ( $TargetPlatform =~ /^.*-apple-darwin.*/ ) {
1819 # On Darwin, we've got local-looking jumps that are
1820 # actually global (i.e. jumps to Lfoo$stub or via
1821 # Lfoo$non_lazy_ptr), so we fix those first.
1822 # In fact, we just fix everything that contains a dollar
1823 # because false positives don't hurt here.
1825 s/^(\tjmp\s+\*?L.*\$.*\n)/$exit_patch$1/g;
1828 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1829 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1831 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1833 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1834 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1836 if ($StolenX86Regs == 2 ) {
1837 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1838 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1839 } elsif ($StolenX86Regs == 3 ) {
1840 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1841 if /^\t(jmp|call)\s+.*\%edi/;
1844 # --------------------------------------------------------
1845 # that's it -- print it
1847 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1851 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1852 print OUTASM $exit_patch;
1853 # ToDo: make it not print if there is a "jmp" at the end
1859 sub init_FUNNY_THINGS {
1860 %KNOWN_FUNNY_THING = (
1862 # "${T_US}stg_.*{T_POST_LBL}", 1,
1867 The following table reversal is used for both info tables and return
1868 vectors. In both cases, we remove the first entry from the table,
1869 reverse the table, put the label at the end, and paste some code
1870 (that which is normally referred to by the first entry in the table)
1871 right after the table itself. (The code pasting is done elsewhere.)
1875 local($symb, $tbl, $discard1) = @_;
1877 return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1879 local($before) = '';
1881 local(@imports) = (); # hppa only
1884 local(@lines) = split(/\n/, $tbl);
1887 # Deal with the header...
1888 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1889 $label .= $lines[$i] . "\n",
1890 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1891 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1892 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1894 $before .= $lines[$i] . "\n"; # otherwise...
1898 $infoname =~ s/(.|\n)*^([A-Za-z0-9_]+_info)${T_POST_LBL}$(.|\n)*/\2/;
1900 # Grab the table data...
1901 if ( $TargetPlatform !~ /^hppa/ ) {
1902 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1904 # Convert addresses of SRTs, slow entrypoints and large bitmaps
1905 # to offsets (relative to the info label),
1906 # in order to support position independent code.
1907 $line =~ s/$infoname/0/
1908 || $line =~ s/([A-Za-z0-9_]+_srtd)$/\1 - $infoname/
1909 || $line =~ s/([A-Za-z0-9_]+_srt(\+\d+)?)$/\1 - $infoname/
1910 || $line =~ s/([A-Za-z0-9_]+_str)$/\1 - $infoname/
1911 || $line =~ s/([A-Za-z0-9_]+_slow)$/\1 - $infoname/
1912 || $line =~ s/([A-Za-z0-9_]+_btm)$/\1 - $infoname/
1913 || $line =~ s/([A-Za-z0-9_]+_alt)$/\1 - $infoname/
1914 || $line =~ s/([A-Za-z0-9_]+_dflt)$/\1 - $infoname/
1915 || $line =~ s/([A-Za-z0-9_]+_ret)$/\1 - $infoname/;
1916 push(@words, $line);
1918 } else { # hppa weirdness
1919 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1920 # FIXME: the RTS now expects offsets instead of addresses
1921 # for all labels in info tables.
1922 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1923 push(@imports, $lines[$i]);
1925 # We don't use HP's ``function pointers''
1926 # We just use labels in code space, like normal people
1927 $lines[$i] =~ s/P%//;
1928 push(@words, $lines[$i]);
1933 # Now throw away any initial zero word from the table. This is a hack
1934 # that lets us reduce the size of info tables when the SRT field is not
1935 # needed: see comments StgFunInfoTable in InfoTables.h.
1937 # The .zero business is for Linux/ELF.
1938 # The .skip business is for Sparc/Solaris/ELF.
1939 # The .blockz business is for HPPA.
1941 # if ($words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1946 for (; $i <= $#lines; $i++) {
1947 $after .= $lines[$i] . "\n";
1950 # Alphas: If we have anonymous text (not part of a procedure), the
1951 # linker may complain about missing exception information. Bleh.
1952 # To suppress this, we place a .ent/.end pair around the code.
1953 # At the same time, we have to be careful and not enclose any leading
1954 # .file/.loc directives.
1955 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1956 local ($ident) = $1;
1957 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1958 $after .= "\t.end $ident\n";
1961 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1962 # assembler (!) wherein .quad constants inside .text sections are
1963 # first narrowed to 32 bits then sign-extended back to 64 bits.
1964 # This obviously screws up our 64-bit bitmaps, so we work around
1965 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1966 if ( $TargetPlatform =~ /^alpha-/ ) {
1968 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1969 local ($number) = $1;
1970 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1971 local ($sign, $base, $digits) = ($1, $2, $3);
1972 $base = (10, 8, 16)[length $base];
1973 local ($hi, $lo) = (0, 0);
1974 foreach $i (split(//, $digits)) {
1975 $j = $lo * $base + $i;
1976 $lo = $j % 4294967296;
1977 $hi = $hi * $base + ($j - $lo) / 4294967296;
1979 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1981 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1982 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1984 print STDERR "Cannot handle \".quad $number\" in info table\n";
1991 if ( $TargetPlatform =~ /x86_64-apple-darwin/ ) {
1992 # Tack a label to the front of the info table, too.
1993 # For now, this just serves to work around a crash in Apple's new
1994 # 64-bit linker (it seems to assume that there is no data before the
1995 # first label in a section).
1997 # The plan for the future is to do this on all Darwin platforms, and
1998 # to add a reference to this label after the entry code, just as the
1999 # NCG does, so we can enable dead-code-stripping in the linker without
2000 # losing our info tables. (Hence the name _dsp, for dead-strip preventer)
2002 $before .= "\n${infoname}_dsp:\n";
2006 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
2007 . join("\n", @words) . "\n"
2010 # print STDERR "before=$before\n";
2011 # print STDERR "label=$label\n";
2012 # print STDERR "words=",(reverse @words),"\n";
2013 # print STDERR "after=$after\n";
2019 The HP is a major nuisance. The threaded code mangler moved info
2020 tables from data space to code space, but unthreaded code in the RTS
2021 still has references to info tables in data space. Since the HP
2022 linker is very precise about where symbols live, we need to patch the
2023 references in the unthreaded RTS as well.
2026 sub mini_mangle_asm_hppa {
2027 local($in_asmf, $out_asmf) = @_;
2029 open(INASM, "< $in_asmf")
2030 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
2031 open(OUTASM,"> $out_asmf")
2032 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
2035 s/_info,DATA/_info,CODE/; # Move _info references to code space
2041 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
2042 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
2048 sub tidy_up_and_die {
2049 local($return_val, $msg) = @_;
2051 exit (($return_val == 0) ? 0 : 1);