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 if ( $TargetPlatform =~ /solaris2/ ) {
177 # newer Solaris linkers are picky about .size information, so
178 # omit it (see #1421)
179 $T_COPY_DIRVS = '^\s*\.(globl|local)';
181 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
184 $T_DOT_WORD = '\.(long|value|word|byte|zero)';
185 $T_DOT_GLOBAL = '\.globl';
186 $T_HDR_literal = "\.section\t\.rodata\n"; # or just use .text??? (WDP 95/11)
187 $T_HDR_misc = "\.text\n\t\.align 4\n";
188 $T_HDR_data = "\.data\n\t\.align 4\n";
189 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 4\n";
190 $T_HDR_closure = "\.data\n\t\.align 4\n";
191 $T_HDR_info = "\.text\n\t\.align 4\n";
192 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
193 $T_HDR_vector = "\.text\n\t\.align 4\n"; # NB: requires padding
195 #--------------------------------------------------------#
196 } elsif ( $TargetPlatform =~ /^ia64-.*-linux$/ ) {
198 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
199 $T_US = ''; # _ if symbols have an underscore on the front
201 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
204 $T_MOVE_DIRVS = '^(\s*\.(global|proc|pred\.safe_across_calls|text|data|section|subsection|align|size|type|ident)\s+.*\n)';
205 $T_COPY_DIRVS = '\.(global|proc)';
207 $T_DOT_WORD = '\.(long|value|byte|zero)';
208 $T_DOT_GLOBAL = '\.global';
209 $T_HDR_literal = "\.section\t\.rodata\n";
210 $T_HDR_misc = "\.text\n\t\.align 16\n"; # May contain code; align like 'entry'
211 $T_HDR_data = "\.data\n\t\.align 8\n";
212 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 8\n";
213 $T_HDR_closure = "\.data\n\t\.align 8\n";
214 $T_HDR_info = "\.text\n\t\.align 8\n";
215 $T_HDR_entry = "\.text\n\t\.align 16\n";
216 $T_HDR_vector = "\.text\n\t\.align 8\n";
218 #--------------------------------------------------------#
219 } elsif ( $TargetPlatform =~ /^x86_64-.*-(linux|openbsd)$/ ) {
221 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
222 $T_US = ''; # _ if symbols have an underscore on the front
224 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
227 $T_MOVE_DIRVS = '^(\s*\.(globl|text|data|section|align|size|type|ident|local)([ \t].*)?\n)';
228 $T_COPY_DIRVS = '\.(globl|type|size|local)';
230 $T_DOT_WORD = '\.(quad|long|value|byte|zero)';
231 $T_DOT_GLOBAL = '\.global';
233 $T_HDR_literal16 = "\.section\t\.rodata.cst16\n\t.align 16\n";
234 $T_HDR_literal = "\.section\t\.rodata\n";
236 $T_HDR_misc = "\.text\n\t\.align 8\n";
237 $T_HDR_data = "\.data\n\t\.align 8\n";
238 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 8\n";
240 # the assembler on x86_64/Linux refuses to generate code for
242 # where x is in the text section and y in the rodata section.
243 # It works if y is in the text section, though. This is probably
244 # going to cause difficulties for PIC, I imagine.
246 # See Note [x86-64-relative] in includes/InfoTables.h
247 $T_HDR_relrodata= "\.text\n\t\.align 8\n";
249 $T_HDR_closure = "\.data\n\t\.align 8\n";
250 $T_HDR_info = "\.text\n\t\.align 8\n";
251 $T_HDR_entry = "\.text\n\t\.align 8\n";
252 $T_HDR_vector = "\.text\n\t\.align 8\n";
254 #--------------------------------------------------------#
255 } elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) {
257 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
258 $T_US = '_'; # _ if symbols have an underscore on the front
259 $T_PRE_APP = '^# MAY NOT APPLY'; # regexp that says what comes before APP/NO_APP
260 $T_CONST_LBL = '^LC(\d+):$';
263 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.proc\s+\d+|\.const|\.cstring|\.globl\s+\S+|\.text|\.data|\.even|\.stab[^n].*)\n)';
264 $T_COPY_DIRVS = '\.(globl|proc|stab)';
266 $T_DOT_WORD = '\.long';
267 $T_DOT_GLOBAL = '\.globl';
268 $T_HDR_literal = "\.text\n\t\.even\n";
269 $T_HDR_misc = "\.text\n\t\.even\n";
270 $T_HDR_data = "\.data\n\t\.even\n";
271 $T_HDR_rodata = "\.text\n\t\.even\n";
272 $T_HDR_closure = "\.data\n\t\.even\n";
273 $T_HDR_info = "\.text\n\t\.even\n";
274 $T_HDR_entry = "\.text\n\t\.even\n";
275 $T_HDR_vector = "\.text\n\t\.even\n";
277 #--------------------------------------------------------#
278 } elsif ( $TargetPlatform =~ /^mips-.*/ ) {
280 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
281 $T_US = ''; # _ if symbols have an underscore on the front
282 $T_PRE_APP = '^\s*#'; # regexp that says what comes before APP/NO_APP
283 $T_CONST_LBL = '^\$LC(\d+):$'; # regexp for what such a lbl looks like
286 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.(globl|ent)\s+\S+|\.text|\.r?data)\n)';
287 $T_COPY_DIRVS = '\.(globl|ent)';
289 $T_DOT_WORD = '\.word';
290 $T_DOT_GLOBAL = '^\t\.globl';
291 $T_HDR_literal = "\t\.rdata\n\t\.align 2\n";
292 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
293 $T_HDR_data = "\t\.data\n\t\.align 2\n";
294 $T_HDR_rodata = "\t\.rdata\n\t\.align 2\n";
295 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
296 $T_HDR_info = "\t\.text\n\t\.align 2\n";
297 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
298 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
300 #--------------------------------------------------------#
301 } elsif ( $TargetPlatform =~ /^powerpc-apple-darwin.*/ ) {
302 # Apple PowerPC Darwin/MacOS X.
303 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
304 $T_US = '_'; # _ if symbols have an underscore on the front
305 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
306 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
309 $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)';
310 $T_COPY_DIRVS = '\.(globl|lcomm)';
312 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
313 $T_DOT_GLOBAL = '\.globl';
314 $T_HDR_toc = "\.toc\n";
315 $T_HDR_literal = "\t\.const\n\t\.align 2\n";
316 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
317 $T_HDR_data = "\t\.data\n\t\.align 2\n";
318 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
319 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
320 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
321 $T_HDR_info = "\t\.text\n\t\.align 2\n";
322 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
323 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
325 #--------------------------------------------------------#
326 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ ) {
327 # Apple PowerPC Darwin/MacOS X.
328 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
329 $T_US = '_'; # _ if symbols have an underscore on the front
330 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
331 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
333 $T_X86_PRE_LLBL_PAT = 'L';
334 $T_X86_PRE_LLBL = 'L';
335 $T_X86_BADJMP = '^\tjmp [^L\*]';
337 $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)';
338 $T_COPY_DIRVS = '\.(globl|lcomm)';
340 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
341 $T_DOT_GLOBAL = '\.globl';
342 $T_HDR_toc = "\.toc\n";
343 $T_HDR_literal16= "\t\.literal8\n\t\.align 4\n";
344 $T_HDR_literal = "\t\.const\n\t\.align 4\n";
345 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
346 $T_HDR_data = "\t\.data\n\t\.align 2\n";
347 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
348 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
349 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
350 $T_HDR_info = "\t\.text\n\t\.align 2\n";
351 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
352 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
354 #--------------------------------------------------------#
355 } elsif ( $TargetPlatform =~ /^x86_64-apple-darwin.*/ ) {
356 # Apple PowerPC Darwin/MacOS X.
357 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
358 $T_US = '_'; # _ if symbols have an underscore on the front
359 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
360 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
363 $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)';
364 $T_COPY_DIRVS = '\.(globl|lcomm)';
366 $T_DOT_WORD = '\.(quad|long|short|byte|fill|space)';
367 $T_DOT_GLOBAL = '\.globl';
368 $T_HDR_toc = "\.toc\n";
369 $T_HDR_literal16= "\t\.literal8\n\t\.align 4\n";
370 $T_HDR_literal = "\t\.const\n\t\.align 4\n";
371 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
372 $T_HDR_data = "\t\.data\n\t\.align 2\n";
373 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
374 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
375 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
376 $T_HDR_info = "\t\.text\n\t\.align 2\n";
377 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
378 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
380 #--------------------------------------------------------#
381 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ ) {
383 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
384 $T_US = ''; # _ if symbols have an underscore on the front
385 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
386 $T_CONST_LBL = '^\.LC\d+:'; # regexp for what such a lbl looks like
389 $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)';
390 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
392 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
393 $T_DOT_GLOBAL = '\.globl';
394 $T_HDR_toc = "\.toc\n";
395 $T_HDR_literal = "\t\.section\t.rodata\n\t\.align 2\n";
396 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
397 $T_HDR_data = "\t\.data\n\t\.align 2\n";
398 $T_HDR_rodata = "\t\.section\t.rodata\n\t\.align 2\n";
399 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
400 $T_HDR_info = "\t\.text\n\t\.align 2\n";
401 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
402 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
404 #--------------------------------------------------------#
405 } elsif ( $TargetPlatform =~ /^powerpc64-.*-linux/ ) {
407 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
408 $T_US = '\.'; # _ if symbols have an underscore on the front
409 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
410 $T_CONST_LBL = '^\.LC\d+:'; # regexp for what such a lbl looks like
413 $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)';
414 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
416 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
417 $T_DOT_GLOBAL = '\.globl';
418 $T_HDR_toc = "\.toc\n";
419 $T_HDR_literal = "\t\.section\t\".toc\",\"aw\"\n";
420 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
421 $T_HDR_data = "\t\.data\n\t\.align 2\n";
422 $T_HDR_rodata = "\t\.section\t.rodata\n\t\.align 2\n";
423 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
424 $T_HDR_info = "\t\.text\n\t\.align 2\n";
425 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
426 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
428 #--------------------------------------------------------#
429 } elsif ( $TargetPlatform =~ /^sparc-.*-(solaris2|openbsd)/ ) {
431 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
432 $T_US = ''; # _ if symbols have an underscore on the front
433 $T_PRE_APP = 'DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
434 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
437 $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)';
438 $T_COPY_DIRVS = '\.(global|local|proc|stab)';
440 $T_DOT_WORD = '\.(long|word|byte|half|skip|uahalf|uaword)';
441 $T_DOT_GLOBAL = '^\t\.global';
442 $T_HDR_literal = "\.text\n\t\.align 8\n";
443 $T_HDR_misc = "\.text\n\t\.align 4\n";
444 $T_HDR_data = "\.data\n\t\.align 8\n";
445 $T_HDR_rodata = "\.text\n\t\.align 4\n";
446 $T_HDR_closure = "\.data\n\t\.align 4\n";
447 $T_HDR_info = "\.text\n\t\.align 4\n";
448 $T_HDR_entry = "\.text\n\t\.align 4\n";
449 $T_HDR_vector = "\.text\n\t\.align 4\n";
451 #--------------------------------------------------------#
452 } elsif ( $TargetPlatform =~ /^sparc-.*-sunos4/ ) {
454 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
455 $T_US = '_'; # _ if symbols have an underscore on the front
456 $T_PRE_APP = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
457 $T_CONST_LBL = '^LC(\d+):$';
460 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*)\n)';
461 $T_COPY_DIRVS = '\.(global|proc|stab)';
463 $T_DOT_WORD = '\.word';
464 $T_DOT_GLOBAL = '^\t\.global';
465 $T_HDR_literal = "\.text\n\t\.align 8\n";
466 $T_HDR_misc = "\.text\n\t\.align 4\n";
467 $T_HDR_data = "\.data\n\t\.align 8\n";
468 $T_HDR_rodata = "\.text\n\t\.align 4\n";
469 $T_HDR_closure = "\.data\n\t\.align 4\n";
470 $T_HDR_info = "\.text\n\t\.align 4\n";
471 $T_HDR_entry = "\.text\n\t\.align 4\n";
472 $T_HDR_vector = "\.text\n\t\.align 4\n";
474 #--------------------------------------------------------#
475 } elsif ( $TargetPlatform =~ /^sparc-.*-linux/ ) {
476 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
477 $T_US = ''; # _ if symbols have an underscore on the front
478 $T_PRE_APP = '#'; # regexp that says what comes before APP/NO_APP
479 # Probably doesn't apply anyway
480 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
483 $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)';
484 $T_COPY_DIRVS = '\.(global|local|globl|proc|stab)';
486 $T_DOT_WORD = '\.(long|word|nword|xword|byte|half|short|skip|uahalf|uaword)';
487 $T_DOT_GLOBAL = '^\t\.global';
488 $T_HDR_literal = "\.text\n\t\.align 8\n";
489 $T_HDR_misc = "\.text\n\t\.align 4\n";
490 $T_HDR_data = "\.data\n\t\.align 8\n";
491 $T_HDR_rodata = "\.text\n\t\.align 4\n";
492 $T_HDR_closure = "\.data\n\t\.align 4\n";
493 $T_HDR_info = "\.text\n\t\.align 4\n";
494 $T_HDR_entry = "\.text\n\t\.align 4\n";
495 $T_HDR_vector = "\.text\n\t\.align 4\n";
497 #--------------------------------------------------------#
499 print STDERR "$Pgm: don't know how to mangle assembly language for: $TargetPlatform\n";
503 if($T_HDR_relrodata eq "") {
505 # relrodata defaults to rodata.
506 $T_HDR_relrodata = $T_HDR_rodata;
510 print STDERR "T_STABBY: $T_STABBY\n";
511 print STDERR "T_US: $T_US\n";
512 print STDERR "T_PRE_APP: $T_PRE_APP\n";
513 print STDERR "T_CONST_LBL: $T_CONST_LBL\n";
514 print STDERR "T_POST_LBL: $T_POST_LBL\n";
515 if ( $TargetPlatform =~ /^i386-/ ) {
516 print STDERR "T_X86_PRE_LLBL_PAT: $T_X86_PRE_LLBL_PAT\n";
517 print STDERR "T_X86_PRE_LLBL: $T_X86_PRE_LLBL\n";
518 print STDERR "T_X86_BADJMP: $T_X86_BADJMP\n";
520 print STDERR "T_MOVE_DIRVS: $T_MOVE_DIRVS\n";
521 print STDERR "T_COPY_DIRVS: $T_COPY_DIRVS\n";
522 print STDERR "T_DOT_WORD: $T_DOT_WORD\n";
523 print STDERR "T_HDR_literal: $T_HDR_literal\n";
524 print STDERR "T_HDR_misc: $T_HDR_misc\n";
525 print STDERR "T_HDR_data: $T_HDR_data\n";
526 print STDERR "T_HDR_rodata: $T_HDR_rodata\n";
527 print STDERR "T_HDR_closure: $T_HDR_closure\n";
528 print STDERR "T_HDR_info: $T_HDR_info\n";
529 print STDERR "T_HDR_entry: $T_HDR_entry\n";
530 print STDERR "T_HDR_vector: $T_HDR_vector\n";
536 %************************************************************************
538 \subsection{Mangle away}
540 %************************************************************************
544 local($in_asmf, $out_asmf) = @_;
546 # multi-line regexp matching:
550 # ia64-specific information for code chunks
554 &init_TARGET_STUFF();
555 &init_FUNNY_THINGS();
557 open(INASM, "< $in_asmf")
558 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
559 open(OUTASM,"> $out_asmf")
560 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
562 # read whole file, divide into "chunks":
563 # record some info about what we've found...
565 @chk = (); # contents of the chunk
566 $numchks = 0; # number of them
567 @chkcat = (); # what category of thing in each chunk
568 @chksymb = (); # what symbol(base) is defined in this chunk
569 %entrychk = (); # ditto, its entry code
570 %closurechk = (); # ditto, the (static) closure
571 %srtchk = (); # ditto, its SRT (for top-level things)
572 %infochk = (); # given a symbol base, say what chunk its info tbl is in
573 %vectorchk = (); # ditto, return vector table
574 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
576 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
579 tr/\r//d if $TargetPlatform =~ /-mingw32$/; # In case Perl doesn't convert line endings
580 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
581 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
582 next if /^\t\.def.*endef$/;
583 next if /${T_PRE_APP}(NO_)?APP/o;
584 next if /^;/ && $TargetPlatform =~ /^hppa/;
586 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|ia64)-/;
588 if ( $TargetPlatform =~ /^mips-/
589 && /^\t\.(globl\S+\.text|comm\t)/ ) {
590 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
591 # Treat .comm variables as data. These show up in two (known) places:
593 # - the module_registered variable used in the __stginit fragment.
594 # even though these are declared static and initialised, gcc 3.3
595 # likes to make them .comm, presumably to save space in the
598 # - global variables used to pass arguments from C to STG in
599 # a foreign export. (is this still true? --SDM)
601 } elsif ( /^\t\.comm.*$/ ) {
603 $chkcat[$i] = 'data';
606 # Labels ending "_str": these are literal strings.
607 } elsif ( /^${T_US}([A-Za-z0-9_]+)_str${T_POST_LBL}$/ ) {
609 $chkcat[$i] = 'relrodata';
611 } elsif ( $TargetPlatform =~ /-darwin/
612 && (/^\s*\.subsections_via_symbols/
613 ||/^\s*\.no_dead_strip.*/)) {
614 # Don't allow Apple's linker to do any dead-stripping of symbols
615 # in this file, because it will mess up info-tables in mangled
617 # The .no_dead_strip directives are actually put there by
618 # the gcc3 "used" attribute on entry points.
620 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && (
621 /^\s*\.picsymbol_stub/
622 || /^\s*\.section __TEXT,__picsymbol_stub\d,.*/
623 || /^\s*\.section __TEXT,__picsymbolstub\d,.*/
624 || /^\s*\.symbol_stub/
625 || /^\s*\.section __TEXT,__symbol_stub\d,.*/
626 || /^\s*\.section __TEXT,__symbolstub\d,.*/
627 || /^\s*\.lazy_symbol_pointer/
628 || /^\s*\.non_lazy_symbol_pointer/
629 || /^\s*\.section __IMPORT.*/))
632 $chkcat[$i] = 'dyld';
636 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.data/)
637 { # non_lazy_symbol_ptrs that point to local symbols
639 $chkcat[$i] = 'dyld';
642 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.align/)
643 { # non_lazy_symbol_ptrs that point to local symbols
645 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^L_.*:$/)
646 { # non_lazy_symbol_ptrs that point to local symbols
647 $chk[++$i] = $dyld_section . $_;
648 $chkcat[$i] = 'dyld';
651 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
652 # duplicated from the bottom
656 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
657 # Alphas: Local labels not to be confused with new chunks
659 # NB: all the rest start with a non-space
661 } elsif ( $TargetPlatform =~ /^mips-/
662 && /^\d+:/ ) { # a funny-looking very-local label
665 } elsif ( /$T_CONST_LBL/o ) {
667 $chkcat[$i] = 'literal';
670 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
672 $chkcat[$i] = 'splitmarker';
675 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
678 $chkcat[$i] = 'infotbl';
679 $chksymb[$i] = $symb;
681 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
683 $infochk{$symb} = $i;
685 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
687 $chkcat[$i] = 'entry';
692 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
694 $chkcat[$i] = 'closure';
697 $closurechk{$1} = $i;
699 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
706 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
708 $chkcat[$i] = 'data';
711 } elsif ( /^${T_US}(stg_ap_stack_entries|stg_stack_save_entries|stg_arg_bitmaps)${T_POST_LBL}$/o ) {
713 $chkcat[$i] = 'data';
716 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
719 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
720 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
721 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
722 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
725 $chkcat[$i] = 'data';
728 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
733 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
734 # all CC_ symbols go in the data section...
736 $chkcat[$i] = 'data';
739 } elsif ( /^${T_US}([A-Za-z0-9_]+)_hpc${T_POST_LBL}$/o ) {
740 # hpc shares tick boxes across modules
742 $chkcat[$i] = 'data';
745 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
747 $chkcat[$i] = 'misc';
749 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
751 $chkcat[$i] = 'vector';
756 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
757 && /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
758 # Some Solaris system headers contain function definitions (as
759 # opposed to mere prototypes), which end up in the .hc file when
760 # a Haskell module foreign imports the corresponding system
761 # functions (most notably stat()). We put them into the text
762 # segment. Note that this currently does not extend to function
763 # names starting with an underscore.
766 $chkcat[$i] = 'misc';
769 } elsif ( $TargetPlatform =~ /^i386-apple-darwin/ && /^(___i686\.get_pc_thunk\.[abcd]x):/o) {
770 # To handle PIC on Darwin/x86, we need to appropriately pass through
771 # the get_pc_thunk functions. The need to be put into a special section
772 # marked as coalesced (otherwise the .weak_definition doesn't work
775 $chkcat[$i] = 'get_pc_thunk';
778 } elsif ( /^${T_US}[A-Za-z0-9_]/o
779 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
781 && ( $TargetPlatform !~ /^powerpc64/ # we need to avoid local labels in this case
782 || ! /^\.L\d+:$/ ) ) {
789 /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
790 || /^${T_US}__stg_.*${T_POST_LBL}$/o # more RTS internals
791 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
792 || /^${T_US}.*_slow${T_POST_LBL}$/o # slow entry
793 || /^${T_US}__stginit.*${T_POST_LBL}$/o # __stginit<module>
794 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
795 || /^${T_US}.*_fast${T_POST_LBL}$/o # primops
796 || /^_uname:/o # x86/Solaris2
799 $chkcat[$i] = 'misc';
801 /^${T_US}.*_srtd${T_POST_LBL}$/o # large bitmaps
802 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
805 $chkcat[$i] = 'relrodata';
808 print STDERR "Warning: retaining unknown function \`$thing' in output from C compiler\n";
809 $chkcat[$i] = 'unknown';
812 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ && /^\.LCTOC1 = /o ) {
813 # PowerPC Linux's large-model PIC (-fPIC) generates a gobal offset
814 # table "by hand". Be sure to copy it over.
815 # Note that this label and all entries in the table should actually
816 # go into the .got2 section, but it isn't easy to distinguish them
817 # from other constant literals (.LC\d+), so we just put everything
820 $chkcat[$i] = 'literal';
821 $chksymb[$i] = 'LCTOC1';
822 } else { # simple line (duplicated at the top)
827 $numchks = $#chk + 1;
828 $chk[$numchks] = ''; # We might push .note.GNU-stack into this
829 $chkcat[$numchks] = 'verbatim'; # If we do, write it straight back out
831 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
832 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
835 # the division into chunks is imperfect;
836 # we throw some things over the fence into the next
839 # also, there are things we would like to know
840 # about the whole module before we start spitting
843 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
844 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
846 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
848 # Alphas: NB: we start meddling at chunk 1, not chunk 0
849 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
850 # spits a ".quad 0" in after the very first ".rdata"; we
851 # detect this special case (tossing the ".quad 0")!
852 local($magic_rdata_seen) = 0;
854 # HPPAs, MIPSen: also start medding at chunk 1
856 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
857 $c = $chk[$i]; # convenience copy
859 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
861 # toss all prologue stuff; HPPA is pretty weird
863 $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
868 # be slightly paranoid to make sure there's
869 # nothing surprising in there
870 if ( $c =~ /--- BEGIN ---/ ) {
871 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
873 # remove junk whitespace around the split point
877 if ($TargetPlatform =~ /^i386-/) {
878 if ($p =~ /^\tsubl\s+\$(\d+),\s*\%esp\n/) {
880 die "Error: reserved stack space exceeded!\n Possible workarounds: compile with -fasm, or try another version of gcc.\n"
884 # gcc 3.4.3 puts this kind of stuff in the prologue, eg.
885 # when compiling PrimOps.cmm with -optc-O2:
888 # movl %ecx, 16(%esp)
889 # movl %edx, 20(%esp)
890 # but then the code of the function doesn't assume
891 # anything about the contnets of these stack locations.
892 # I think it's to do with the use of inline functions for
893 # PK_Word64() and friends, where gcc is initialising the
894 # contents of the struct to zero, and failing to optimise
895 # away the initialisation. Let's live dangerously and
896 # discard these initalisations.
898 $p =~ s/^\tpushl\s+\%e(di|si|bx)\n//g;
899 $p =~ s/^\txorl\s+\%e(ax|cx|dx),\s*\%e(ax|cx|dx)\n//g;
900 $p =~ s/^\tmovl\s+\%e(ax|cx|dx|si|di),\s*\d*\(\%esp\)\n//g;
901 $p =~ s/^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n//g;
902 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
903 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-(cygwin32|mingw32)/);
905 if ($TargetPlatform =~ /^i386-apple-darwin/) {
907 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
909 $pcrel_reg =~ s/(.|\n)*.*___i686\.get_pc_thunk\.([abcd]x)\n(.|\n)*/$2/ or $pcrel_reg = "";
910 $p =~ s/^\s+call\s+___i686\.get_pc_thunk\..x//;
911 $p =~ s/^\"?L\d+\$pb\"?:\n//;
913 if ($pcrel_reg eq "bx") {
914 # Bad gcc. Goes and uses %ebx, our BaseReg, for PIC. Bad gcc.
915 die "Darwin/x86: -fPIC -via-C doesn't work yet, use -fasm. Aborting."
919 } elsif ($TargetPlatform =~ /^x86_64-/) {
920 $p =~ s/^\tpushq\s+\%r(bx|bp|12|13|14)\n//g;
921 $p =~ s/^\tmovq\s+\%r(bx|bp|12|13|14),\s*\d*\(\%rsp\)\n//g;
922 $p =~ s/^\tsubq\s+\$\d+,\s*\%rsp\n//;
924 } elsif ($TargetPlatform =~ /^ia64-/) {
925 $p =~ s/^\t\.prologue .*\n//;
927 # Record the number of local and out registers for register relocation later
928 $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, (\d+), (\d+), 0\n//;
932 $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
933 $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
935 # Ignore save/restore of these registers; they're taken
936 # care of in StgRun()
937 $p =~ s/^\t\.save ar\.lc, r\d+\n//;
938 $p =~ s/^\t\.save pr, r\d+\n//;
939 $p =~ s/^\tmov r\d+ = ar\.lc\n//;
940 $p =~ s/^\tmov r\d+ = pr\n//;
942 # Remove .proc and .body directives
943 $p =~ s/^\t\.proc [a-zA-Z0-9_.]+#\n//;
944 $p =~ s/^\t\.body\n//;
946 # If there's a label, move it to the body
947 if ($p =~ /^[a-zA-Z0-9.]+:\n/) {
952 # Remove floating-point spill instructions.
953 # Only fp registers 2-5 and 16-23 are saved by the runtime.
954 if ($p =~ s/^\tstf\.spill \[r1[4-9]\] = f([2-5]|1[6-9]|2[0-3])(, [0-9]+)?\n//g) {
955 # Being paranoid, only try to remove these if we saw a
957 $p =~ s/^\tmov r1[4-9] = r12\n//;
958 $p =~ s/^\tadds r1[4-9] = -[0-9]+, r12\n//g;
959 $p =~ s/^\t\.save\.f 0x[0-9a-fA-F]\n//g;
960 $p =~ s/^\t\.save\.gf 0x0, 0x[0-9a-fA-F]+\n//g;
963 $p =~ s/^\tnop(?:\.[mifb])?\s+\d+\n//g; # remove nop instructions
964 $p =~ s/^\t\.(mii|mmi|mfi)\n//g; # bundling is no longer sensible
965 $p =~ s/^\t;;\n//g; # discard stops
966 $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
968 # GCC 3.3 saves r1 in the prologue, move this to the body
969 # (Does this register get restored anywhere?)
970 if ($p =~ /^\tmov r\d+ = r1\n/) {
974 } elsif ($TargetPlatform =~ /^m68k-/) {
975 $p =~ s/^\tlink a6,#-?\d.*\n//;
976 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
977 # The above showed up in the asm code,
978 # so I added it here.
979 # I hope it's correct.
981 $p =~ s/^\tmovel d2,sp\@-\n//;
982 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
983 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
984 } elsif ($TargetPlatform =~ /^mips-/) {
985 # the .frame/.mask/.fmask that we use is the same
986 # as that produced by GCC for miniInterpret; this
987 # gives GDB some chance of figuring out what happened
988 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
989 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
990 $p =~ s/^\t\.(mask|fmask).*\n//g;
991 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
992 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
993 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
994 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
995 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
996 $p =~ s/__FRAME__/$FRAME/;
997 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
999 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
1001 $p =~ s/^\tmflr r0\n//;
1002 $p =~ s/^\tbl saveFP # f\d+\n//;
1003 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
1004 $p =~ s/^\"?L\d+\$pb\"?:\n//;
1005 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
1006 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
1007 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
1008 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
1009 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
1010 $p =~ s/^\tbcl 20,31,\"?L\d+\$pb\"?\n//;
1011 $p =~ s/^\"?L\d+\$pb\"?:\n//;
1012 $p =~ s/^\tmflr r31\n//;
1014 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1015 # under some circumstances, only when generating position dependent code.
1016 # I have no idea why, and I don't think it is necessary, so let's toss it.
1017 $p =~ s/^\tli r\d+,0\n//g;
1018 $p =~ s/^\tstw r\d+,\d+\(r1\)\n//g;
1019 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/) {
1020 $p =~ s/^\tmflr 0\n//;
1021 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
1022 $p =~ s/^\tstfd \d+,\d+\(1\)\n//g;
1023 $p =~ s/^\tstw r0,8\(1\)\n//;
1024 $p =~ s/^\tstwu 1,-\d+\(1\)\n//;
1025 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1027 # GCC's "large-model" PIC (-fPIC)
1029 $pcrel_label =~ s/(.|\n)*^.LCF(\d+):\n(.|\n)*/$2/ or $pcrel_label = "";
1031 $p =~ s/^\tbcl 20,31,.LCF\d+\n//;
1032 $p =~ s/^.LCF\d+:\n//;
1033 $p =~ s/^\tmflr 30\n//;
1034 $p =~ s/^\tlwz 0,\.LCL\d+-\.LCF\d+\(30\)\n//;
1035 $p =~ s/^\tadd 30,0,30\n//;
1037 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1038 # under some circumstances, only when generating position dependent code.
1039 # I have no idea why, and I don't think it is necessary, so let's toss it.
1040 $p =~ s/^\tli \d+,0\n//g;
1041 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1042 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1043 $p =~ s/^\tmr 31,1\n//;
1044 $p =~ s/^\tmflr 0\n//;
1045 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
1046 $p =~ s/^\tstfd \d+,-?\d+\(1\)\n//g;
1047 $p =~ s/^\tstd r0,8\(1\)\n//;
1048 $p =~ s/^\tstdu 1,-\d+\(1\)\n//;
1049 $p =~ s/^\tstd \d+,-?\d+\(1\)\n//g;
1051 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1052 # under some circumstances, only when generating position dependent code.
1053 # I have no idea why, and I don't think it is necessary, so let's toss it.
1054 $p =~ s/^\tli \d+,0\n//g;
1055 $p =~ s/^\tstd \d+,\d+\(1\)\n//g;
1057 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
1060 # HWL HACK: dont die, just print a warning
1061 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
1062 die "Prologue junk?: $p\n" if $p =~ /^\s+[^\s\.]/;
1064 # For PIC, we want to keep part of the prologue
1065 if ($TargetPlatform =~ /^powerpc-apple-darwin.*/ && $pcrel_label ne "") {
1066 # Darwin: load the current instruction pointer into register r31
1067 $p .= "bcl 20,31,$pcrel_label\n";
1068 $p .= "$pcrel_label:\n";
1069 $p .= "\tmflr r31\n";
1070 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/ && $pcrel_label ne "") {
1071 # Linux: load the GOT pointer into register 30
1072 $p .= "\tbcl 20,31,.LCF$pcrel_label\n";
1073 $p .= ".LCF$pcrel_label:\n";
1074 $p .= "\tmflr 30\n";
1075 $p .= "\tlwz 0,.LCL$pcrel_label-.LCF$pcrel_label(30)\n";
1076 $p .= "\tadd 30,0,30\n";
1077 } elsif ($TargetPlatform =~ /^i386-apple-darwin.*/ && $pcrel_label ne "") {
1078 $p .= "\tcall ___i686.get_pc_thunk.$pcrel_reg\n";
1079 $p .= "$pcrel_label:\n";
1082 # glue together what's left
1087 if ( $TargetPlatform =~ /^mips-/ ) {
1088 # MIPS: first, this basic sequence may occur "--- END ---" or not
1089 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
1092 # toss all epilogue stuff; again, paranoidly
1093 if ( $c =~ /--- END ---/ ) {
1094 # Gcc may decide to replicate the function epilogue. We want
1095 # to process all epilogues, so we split the function and then
1097 @fragments = split(/--- END ---/, $c);
1098 $r = shift(@fragments);
1100 # Rebuild `c'; processed fragments will be appended to `c'
1103 foreach $e (@fragments) {
1104 # etail holds code that is after the epilogue in the assembly-code
1105 # layout and should not be filtered as part of the epilogue.
1107 if ($TargetPlatform =~ /^i386-/) {
1109 $e =~ s/^\tpopl\s+\%edi\n//;
1110 $e =~ s/^\tpopl\s+\%esi\n//;
1111 $e =~ s/^\tpopl\s+\%edx\n//;
1112 $e =~ s/^\tpopl\s+\%ecx\n//;
1113 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
1114 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
1115 } elsif ($TargetPlatform =~ /^ia64-/) {
1116 # The epilogue is first split into:
1117 # $e, the epilogue code (up to the return instruction)
1118 # $etail, non-epilogue code (after the return instruction)
1119 # The return instruction is stripped in the process.
1120 if (!(($e, $etail) = split(/^\tbr\.ret\.sptk\.many b0\n/, $e))) {
1121 die "Epilogue doesn't seem to have one return instruction: $e\n";
1123 # Remove 'endp' directive from the tail
1124 $etail =~ s/^\t\.endp [a-zA-Z0-9_.]+#\n//;
1126 # If a return value is saved here, discard it
1127 $e =~ s/^\tmov r8 = r14\n//;
1129 # Remove floating-point fill instructions.
1130 # Only fp registers 2-5 and 16-23 are saved by the runtime.
1131 if ($e =~ s/^\tldf\.fill f([2-5]|1[6-9]|2[0-3]) = \[r1[4-9]\](, [0-9]+)?\n//g) {
1132 # Being paranoid, only try to remove this if we saw a fill
1134 $e =~ s/^\tadds r1[4-9] = [0-9]+, r12//g;
1137 $e =~ s/^\tnop(?:\.[mifb])?\s+\d+\n//g; # remove nop instructions
1138 $e =~ s/^\tmov ar\.pfs = r\d+\n//;
1139 $e =~ s/^\tmov ar\.lc = r\d+\n//;
1140 $e =~ s/^\tmov pr = r\d+, -1\n//;
1141 $e =~ s/^\tmov b0 = r\d+\n//;
1142 $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
1143 #$e =~ s/^\tbr\.ret\.sptk\.many b0\n//; # already removed
1144 $e =~ s/^\t\.(mii|mmi|mfi|mib)\n//g; # bundling is no longer sensible
1145 $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
1146 $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
1147 } elsif ($TargetPlatform =~ /^m68k-/) {
1148 $e =~ s/^\tunlk a6\n//;
1150 } elsif ($TargetPlatform =~ /^mips-/) {
1151 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
1152 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
1153 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
1154 $e =~ s/^\tj\t\$31\n//;
1155 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
1156 $e =~ s/^\taddi r1,r1,\d+\n//;
1157 $e =~ s/^\tlwz r\d+,\d+\(r1\)\n//;
1158 $e =~ s/^\tlmw r\d+,-\d+\(r1\)\n//;
1159 $e =~ s/^\tmtlr r0\n//;
1161 $e =~ s/^\tb restFP ;.*\n//;
1162 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1163 $e =~ s/^\tmr 3,0\n//;
1164 $e =~ s/^\taddi 1,1,\d+\n//;
1165 $e =~ s/^\tld 0,16\(1\)\n//;
1166 $e =~ s/^\tmtlr 0\n//;
1168 # callee-save registers
1169 $e =~ s/^\tld \d+,-?\d+\(1\)\n//g;
1170 $e =~ s/^\tlfd \d+,-?\d+\(1\)\n//g;
1172 # get rid of the debug junk along with the blr
1173 $e =~ s/^\tblr\n\t.long .*\n\t.byte .*\n//;
1175 # incase we missed it with the last one get the blr alone
1178 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
1181 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
1183 # glue together what's left
1186 $c =~ s/\n\t\n/\n/; # junk blank line
1189 if ($TargetPlatform =~ /^ia64-/) {
1190 # On IA64, remove an .endp directive even if no epilogue was found.
1191 # Code optimizations may have removed the "--- END ---" token.
1192 $c =~ s/^\t\.endp [a-zA-Z0-9_.]+#\n//;
1196 # On SPARCs, we don't do --- BEGIN/END ---, we just
1197 # toss the register-windowing save/restore/ret* instructions
1198 # directly unless they've been generated by function definitions in header
1200 if ( $TargetPlatform =~ /^sparc-/ ) {
1201 if ( ! ( $TargetPlatform =~ /solaris2$/ && $chkcat[$i] eq 'unknown' )) {
1202 $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
1204 # throw away PROLOGUE comments
1205 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
1208 # On Alphas, the prologue mangling is done a little later (below)
1210 # toss all calls to __DISCARD__
1211 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
1212 $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
1213 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-darwin.*/;
1214 $c =~ s/^\tbl\s+__DISCARD__(\@plt)?\n//go if $TargetPlatform =~ /^powerpc-.*-linux/;
1215 $c =~ s/^\tbl\s+\.__DISCARD__\n\s+nop\n//go if $TargetPlatform =~ /^powerpc64-.*-linux/;
1216 $c =~ s/^\tcall\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /i386-apple-darwin.*/;
1218 # IA64: fix register allocation; mangle tailcalls into jumps
1219 if ($TargetPlatform =~ /^ia64-/) {
1220 ia64_rename_registers($ia64_locnum, $ia64_outnum) if (defined($ia64_locnum));
1221 ia64_mangle_tailcalls();
1224 # MIPS: that may leave some gratuitous asm macros around
1225 # (no harm done; but we get rid of them to be tidier)
1226 $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/
1227 if $TargetPlatform =~ /^mips-/;
1229 # toss stack adjustment after DoSparks
1230 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
1231 if $TargetPlatform =~ /^m68k-/; # this looks old...
1233 if ( $TargetPlatform =~ /^alpha-/ &&
1234 ! $magic_rdata_seen &&
1235 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
1236 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
1237 $magic_rdata_seen = 1;
1240 # pick some end-things and move them to the next chunk
1242 # pin a funny end-thing on (for easier matching):
1243 $c .= 'FUNNY#END#THING';
1245 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
1249 # on x86 we try not to copy any directives into a literal
1250 # chunk, rather we keep looking for the next real chunk. This
1251 # is because we get things like
1253 # .globl blah_closure
1259 if ( $TargetPlatform =~ /^(i386|sparc|powerpc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
1261 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
1264 if ( $j < $numchks ) {
1265 $chk[$j] = $to_move . $chk[$j];
1269 elsif ( ( $i < ($numchks - 1)
1270 && ( $to_move =~ /${T_COPY_DIRVS}/
1271 || ( $TargetPlatform =~ /^hppa/
1272 && $to_move =~ /align/
1273 && $chkcat[$i+1] eq 'literal')
1276 || ($to_move =~ /^[ \t]*\.section[ \t]+\.note\.GNU-stack,/)
1278 $chk[$i + 1] = $to_move . $chk[$i + 1];
1279 # otherwise they're tossed
1282 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
1285 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
1287 # toss all prologue stuff, except for loading gp, and the ..ng address
1288 unless ($c =~ /\.ent.*\n\$.*\.\.ng:/) {
1289 if (($p, $r) = split(/^\t\.prologue/, $c)) {
1290 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
1291 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
1292 $keep =~ s/^\t\.(mask|fmask).*\n//g;
1293 $c = $keep . "..ng:\n";
1295 print STDERR "malformed code block ($ent)?\n"
1298 $c .= "\t.prologue" . $r;
1302 $c =~ s/FUNNY#END#THING//;
1304 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
1306 $chk[$i] = $c; # update w/ convenience copy
1309 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
1310 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
1313 if ( $TargetPlatform =~ /^alpha-/ ) {
1314 # print out the header stuff first
1315 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
1316 print OUTASM $chk[0];
1318 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1319 print OUTASM $chk[0];
1321 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1322 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
1324 # get rid of horrible "<dollar>Revision: .*$" strings
1325 local(@lines0) = split(/\n/, $chk[0]);
1327 while ( $z <= $#lines0 ) {
1328 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
1331 while ( $z <= $#lines0 ) {
1333 last if $lines0[$z] =~ /[,\t]0x0$/;
1339 $chk[0] = join("\n", @lines0);
1340 $chk[0] =~ s/\n\n+/\n/;
1341 print OUTASM $chk[0];
1344 # print out all the literal strings next
1345 for ($i = 0; $i < $numchks; $i++) {
1346 if ( $chkcat[$i] eq 'literal' ) {
1348 # HACK: try to detect 16-byte constants and align them
1349 # on a 16-byte boundary. x86_64 sometimes needs 128-bit
1350 # aligned constants, and so does Darwin/x86.
1351 if ( $TargetPlatform =~ /^x86_64/
1352 || $TargetPlatform =~ /^i386-apple-darwin/ ) {
1354 if ($z =~ /(\.long.*\n.*\.long.*\n.*\.long.*\n.*\.long|\.quad.*\n.*\.quad)/) {
1355 print OUTASM $T_HDR_literal16;
1357 print OUTASM $T_HDR_literal;
1360 print OUTASM $T_HDR_literal;
1363 print OUTASM $chk[$i];
1364 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
1366 $chkcat[$i] = 'DONE ALREADY';
1370 # on the HPPA, print out all the bss next
1371 if ( $TargetPlatform =~ /^hppa/ ) {
1372 for ($i = 1; $i < $numchks; $i++) {
1373 if ( $chkcat[$i] eq 'bss' ) {
1374 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
1375 print OUTASM $chk[$i];
1377 $chkcat[$i] = 'DONE ALREADY';
1382 # $numchks + 1 as we have the extra one for .note.GNU-stack
1383 for ($i = $FIRST_MANGLABLE; $i < $numchks + 1; $i++) {
1384 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
1386 next if $chkcat[$i] eq 'DONE ALREADY';
1388 if ( $chkcat[$i] eq 'misc' || $chkcat[$i] eq 'unknown' ) {
1389 if ($chk[$i] ne '') {
1390 print OUTASM $T_HDR_misc;
1391 &print_doctored($chk[$i], 0);
1394 } elsif ( $chkcat[$i] eq 'verbatim' ) {
1395 print OUTASM $chk[$i];
1397 } elsif ( $chkcat[$i] eq 'toss' ) {
1398 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
1400 } elsif ( $chkcat[$i] eq 'data' ) {
1401 if ($chk[$i] ne '') {
1402 print OUTASM $T_HDR_data;
1403 print OUTASM $chk[$i];
1406 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1407 # we can just re-constitute this one...
1408 # NB: we emit _three_ underscores no matter what,
1409 # so ghc-split doesn't have to care.
1410 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1412 } elsif ( $chkcat[$i] eq 'closure'
1413 || $chkcat[$i] eq 'srt'
1414 || $chkcat[$i] eq 'infotbl'
1415 || $chkcat[$i] eq 'entry') { # do them in that order
1416 $symb = $chksymb[$i];
1419 if ( defined($closurechk{$symb}) ) {
1420 print OUTASM $T_HDR_closure;
1421 print OUTASM $chk[$closurechk{$symb}];
1422 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1426 if ( defined($srtchk{$symb}) ) {
1427 print OUTASM $T_HDR_relrodata;
1428 print OUTASM $chk[$srtchk{$symb}];
1429 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1433 if ( defined($infochk{$symb}) ) {
1435 print OUTASM $T_HDR_info;
1436 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1438 # entry code will be put here!
1440 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1444 if ( defined($entrychk{$symb}) ) {
1446 $c = $chk[$entrychk{$symb}];
1448 # If this is an entry point with an info table,
1449 # eliminate the entry symbol and all directives involving it.
1450 if (defined($infochk{$symb}) && $TargetPlatform !~ /^ia64-/) {
1452 foreach $l (split(/\n/,$c)) {
1453 next if $l =~ /^.*$symb_(entry|ret)${T_POST_LBL}/;
1455 # If we have .type/.size direrctives involving foo_entry,
1456 # then make them refer to foo_info instead. The information
1457 # in these directives is used by the cachegrind annotator,
1458 # so it is worthwhile keeping.
1459 if ($l =~ /^\s*\.(type|size).*$symb_(entry|ret)/) {
1460 $l =~ s/$symb(_entry|_ret)/${symb}_info/g;
1464 next if $l =~ /^\s*\..*$symb.*\n?/;
1467 $c = join("\n",@o) . "\n";
1470 print OUTASM $T_HDR_entry;
1472 &print_doctored($c, 1); # NB: the 1!!!
1474 $chkcat[$entrychk{$symb}] = 'DONE ALREADY';
1477 } elsif ( $chkcat[$i] eq 'vector' ) {
1478 $symb = $chksymb[$i];
1481 if ( defined($vectorchk{$symb}) ) {
1482 print OUTASM $T_HDR_vector;
1483 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1485 # direct return code will be put here!
1486 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1488 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1489 # Alphas: the commented nop is for the splitter, to ensure
1490 # that no module ends with a label as the very last
1491 # thing. (The linker will adjust the label to point
1492 # to the first code word of the next module linked in,
1493 # even if alignment constraints cause the label to move!)
1495 print OUTASM "\t# nop\n";
1498 } elsif ( $chkcat[$i] eq 'rodata' ) {
1499 print OUTASM $T_HDR_rodata;
1500 print OUTASM $chk[$i];
1501 $chkcat[$i] = 'DONE ALREADY';
1502 } elsif ( $chkcat[$i] eq 'relrodata' ) {
1503 print OUTASM $T_HDR_relrodata;
1504 print OUTASM $chk[$i];
1505 $chkcat[$i] = 'DONE ALREADY';
1506 } elsif ( $chkcat[$i] eq 'toc' ) {
1507 # silly optimisation to print tocs, since they come in groups...
1508 print OUTASM $T_HDR_toc;
1510 while ($chkcat[$j] eq 'toc')
1511 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1514 print OUTASM $chk[$j];
1516 $chkcat[$j] = 'DONE ALREADY';
1520 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' ) {
1521 # apple-darwin: dynamic linker stubs
1522 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1523 { # print them out unchanged, but remove the stubs for __DISCARD__
1524 print OUTASM $chk[$i];
1526 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ && $chkcat[$i] eq 'get_pc_thunk' ) {
1527 # i386-apple-darwin: __i686.get_pc_thunk.[abcd]x
1528 print OUTASM ".section __TEXT,__textcoal_nt,coalesced,no_toc\n";
1529 print OUTASM $chk[$i];
1531 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1535 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1538 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1539 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1543 On IA64, tail calls are converted to branches at this point. The mangler
1544 searches for function calls immediately followed by a '--- TAILCALL ---'
1545 token. Since the compiler can put various combinations of labels, bundling
1546 directives, nop instructions, stops, and a move of the return value
1547 between the branch and the tail call, proper matching of the tail call
1548 gets a little hairy. This subroutine does the mangling.
1550 Here is an example of a tail call before mangling:
1553 br.call.sptk.many b0 = b6
1568 sub ia64_mangle_tailcalls {
1569 # Function input and output are in $c
1571 # Construct the tailcall-mangling expression the first time this function
1573 if (!defined($IA64_MATCH_TAILCALL)) {
1574 # One-line pattern matching constructs. None of these
1575 # should bind references; all parenthesized terms
1576 # should be (?:) terms.
1577 my $stop = q/(?:\t;;\n)/;
1578 my $bundle = q/(?:\t\.(?:mii|mib|mmi|mmb|mfi|mfb|mbb|bbb)\n)/;
1579 my $nop = q/(?:\tnop(?:\.[mifb])?\s+\d+\n)/;
1580 my $movgp = q/(?:\tmov r1 = r\d+\n)/;
1581 my $postbr = q/(?:\tbr \.L\d+\n)/;
1583 my $noeffect = "(?:$stop$bundle?|$nop)*";
1584 my $postbundle = "(?:$bundle?$nop?$nop?$postbr)?";
1586 # Important parts of the pattern match. The branch target
1587 # and subsequent jump label are bound to $1 and $2
1588 # respectively. Sometimes there is no label.
1589 my $callbr = q/^\tbr\.call\.sptk\.many b0 = (.*)\n/;
1590 my $label = q/(?:^\.L([0-9]*):\n)/;
1591 my $tailcall = q/\t--- TAILCALL ---\n/;
1593 $IA64_MATCH_TAILCALL =
1594 $callbr . $label . '?' . $noeffect . $movgp . '?' . $noeffect .
1595 $tailcall . $stop . '?' . '(?:' . $postbundle . ')?';
1598 # Find and mangle tailcalls
1599 while ($c =~ s/$IA64_MATCH_TAILCALL/\tbr\.few $1\n/o) {
1600 # Eek, the gcc optimiser is getting smarter... if we see a jump to the
1601 # --- TAILCALL --- marker then we reapply the substitution at the source sites
1602 $c =~ s/^\tbr \.L$2\n/\t--- TAILCALL ---\n/g if ($2);
1605 # Verify that all instances of TAILCALL were processed
1606 if ($c =~ /^\t--- TAILCALL ---\n/) {
1607 die "Unmangled TAILCALL tokens remain after mangling"
1612 The number of registers allocated on the IA64 register stack is set
1613 upon entry to the runtime with an `alloc' instruction at the entry
1614 point of \verb+StgRun()+. Gcc uses its own `alloc' to allocate
1615 however many registers it likes in each function. When we discard
1616 gcc's alloc, we have to reconcile its register assignment with what
1619 There are three stack areas: fixed registers, input/local registers,
1620 and output registers. We move the output registers to the output
1621 register space and leave the other registers where they are.
1624 sub ia64_rename_registers() {
1625 # The text to be mangled is in $c
1626 # Find number of registers in each stack area
1627 my ($loc, $out) = @_;
1633 # These are the register numbers used in the STG runtime
1634 my $STG_FIRST_OUT_REG = 32 + 34;
1635 my $STG_LAST_OUT_REG = $STG_FIRST_OUT_REG + 7;
1637 $first_out_reg = 32 + $loc;
1639 if ($first_out_reg > $STG_FIRST_OUT_REG) {
1640 die "Too many local registers allocated by gcc";
1643 # Split the string into fragments containing one register name each.
1644 # Rename the register in each fragment and concatenate.
1646 foreach $fragment (split(/(?=r\d+[^a-zA-Z0-9_.])/s, $c)) {
1647 if ($fragment =~ /^r(\d+)((?:[^a-zA-Z0-9_.].*)?)$/s) {
1650 if ($regnum < $first_out_reg) {
1651 # This is a local or fixed register
1653 # Local registers 32 and 33 (r64 and r65) are
1654 # used to hold saved state; they shouldn't be touched
1655 if ($regnum == 64 || $regnum == 65) {
1656 die "Reserved register $regnum is in use";
1660 # This is an output register
1661 $regnum = $regnum - $first_out_reg + $STG_FIRST_OUT_REG;
1662 if ($regnum > $STG_LAST_OUT_REG) {
1663 die "Register number ($regnum) is out of expected range";
1667 # Update this fragment
1668 $fragment = "r" . $regnum . $2;
1679 sub hppa_mash_prologue { # OK, epilogue, too
1682 # toss all prologue stuff
1683 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1685 # Lie about our .CALLINFO
1686 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1693 # toss all epilogue stuff
1694 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1696 # Sorry; we moved the _info stuff to the code segment.
1697 s/_info,DATA/_info,CODE/g;
1704 sub print_doctored {
1705 local($_, $need_fallthru_patch) = @_;
1707 if ( $TargetPlatform =~ /^x86_64-/ ) {
1710 # movq -4(%ebp), %rax
1715 s/^\tmovq\s+(-?\d*\(\%r(bx|bp|13)\)),\s*(\%r(ax|cx|dx|10|11))\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1716 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e(ax|cx|si|di))\n\tjmp\s+\*\%r\3/\tjmp\t$T_US$1/g;
1719 if ( $TargetPlatform !~ /^i386-/
1720 || ! /^\t[a-z]/ # no instructions in here, apparently
1721 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1726 # OK, must do some x86 **HACKING**
1728 local($entry_patch) = '';
1729 local($exit_patch) = '';
1731 # gotta watch out for weird instructions that
1732 # invisibly smash various regs:
1733 # rep* %ecx used for counting
1734 # scas* %edi used for destination index
1735 # cmps* %e[sd]i used for indices
1736 # loop* %ecx used for counting
1741 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1743 # * GCC used an "STG reg" for its own purposes
1745 # * some secret uses of machine reg, requiring STG reg
1746 # to be saved/restored
1748 # The most dangerous "GCC uses" of an "STG reg" are when
1749 # the reg holds the target of a jmp -- it's tricky to
1750 # insert the patch-up code before we get to the target!
1751 # So here we change the jmps:
1753 # --------------------------------------------------------
1754 # it can happen that we have jumps of the form...
1755 # jmp *<something involving %esp>
1757 # jmp <something involving another naughty register...>
1759 # a reasonably-common case is:
1761 # movl $_blah,<bad-reg>
1764 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[acd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1768 # movl -4(%ebx), %eax
1773 s/^\tmovl\s+(-?\d*\(\%e(bx|si)\)),\s*(\%e[acd]x)\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1775 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1776 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1777 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1778 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1779 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1780 if /(jmp|call)\s+.*\%esi/;
1782 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1783 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1784 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1785 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1786 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1787 if /(jmp|call)\s+.*\%edi/;
1790 # OK, now we can decide what our patch-up code is going to
1793 # Offsets into register table - you'd better update these magic
1794 # numbers should you change its contents!
1795 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1796 local($OFFSET_Hp)=88;
1798 # Note funky ".=" stuff; we're *adding* to these _patch guys
1799 if ( $StolenX86Regs <= 2
1800 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1801 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1802 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1804 # nothing for call_{entry,exit} because %esi is callee-save
1806 if ( $StolenX86Regs <= 3
1807 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1808 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1809 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1811 # nothing for call_{entry,exit} because %edi is callee-save
1814 # --------------------------------------------------------
1815 # next, here we go with non-%esp patching!
1817 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1819 # Before calling GC we must set up the exit condition before the call
1820 # and entry condition when we come back
1822 # fix _all_ non-local jumps:
1824 if ( $TargetPlatform =~ /^.*-apple-darwin.*/ ) {
1825 # On Darwin, we've got local-looking jumps that are
1826 # actually global (i.e. jumps to Lfoo$stub or via
1827 # Lfoo$non_lazy_ptr), so we fix those first.
1828 # In fact, we just fix everything that contains a dollar
1829 # because false positives don't hurt here.
1831 s/^(\tjmp\s+\*?L.*\$.*\n)/$exit_patch$1/g;
1834 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1835 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1837 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1839 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1840 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1842 if ($StolenX86Regs == 2 ) {
1843 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1844 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1845 } elsif ($StolenX86Regs == 3 ) {
1846 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1847 if /^\t(jmp|call)\s+.*\%edi/;
1850 # --------------------------------------------------------
1851 # that's it -- print it
1853 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1857 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1858 print OUTASM $exit_patch;
1859 # ToDo: make it not print if there is a "jmp" at the end
1865 sub init_FUNNY_THINGS {
1866 %KNOWN_FUNNY_THING = (
1868 # "${T_US}stg_.*{T_POST_LBL}", 1,
1873 The following table reversal is used for both info tables and return
1874 vectors. In both cases, we remove the first entry from the table,
1875 reverse the table, put the label at the end, and paste some code
1876 (that which is normally referred to by the first entry in the table)
1877 right after the table itself. (The code pasting is done elsewhere.)
1881 local($symb, $tbl, $discard1) = @_;
1883 return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1885 local($before) = '';
1887 local(@imports) = (); # hppa only
1890 local(@lines) = split(/\n/, $tbl);
1893 # Deal with the header...
1894 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1895 $label .= $lines[$i] . "\n",
1896 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1897 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1898 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1900 $before .= $lines[$i] . "\n"; # otherwise...
1904 $infoname =~ s/(.|\n)*^([A-Za-z0-9_]+_info)${T_POST_LBL}$(.|\n)*/\2/;
1906 # Grab the table data...
1907 if ( $TargetPlatform !~ /^hppa/ ) {
1908 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1910 # Convert addresses of SRTs, slow entrypoints and large bitmaps
1911 # to offsets (relative to the info label),
1912 # in order to support position independent code.
1913 $line =~ s/$infoname/0/
1914 || $line =~ s/([A-Za-z0-9_]+_srtd)$/\1 - $infoname/
1915 || $line =~ s/([A-Za-z0-9_]+_srt(\+\d+)?)$/\1 - $infoname/
1916 || $line =~ s/([A-Za-z0-9_]+_str)$/\1 - $infoname/
1917 || $line =~ s/([A-Za-z0-9_]+_slow)$/\1 - $infoname/
1918 || $line =~ s/([A-Za-z0-9_]+_btm)$/\1 - $infoname/
1919 || $line =~ s/([A-Za-z0-9_]+_alt)$/\1 - $infoname/
1920 || $line =~ s/([A-Za-z0-9_]+_dflt)$/\1 - $infoname/
1921 || $line =~ s/([A-Za-z0-9_]+_ret)$/\1 - $infoname/;
1922 push(@words, $line);
1924 } else { # hppa weirdness
1925 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1926 # FIXME: the RTS now expects offsets instead of addresses
1927 # for all labels in info tables.
1928 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1929 push(@imports, $lines[$i]);
1931 # We don't use HP's ``function pointers''
1932 # We just use labels in code space, like normal people
1933 $lines[$i] =~ s/P%//;
1934 push(@words, $lines[$i]);
1939 # Now throw away any initial zero word from the table. This is a hack
1940 # that lets us reduce the size of info tables when the SRT field is not
1941 # needed: see comments StgFunInfoTable in InfoTables.h.
1943 # The .zero business is for Linux/ELF.
1944 # The .skip business is for Sparc/Solaris/ELF.
1945 # The .blockz business is for HPPA.
1947 # if ($words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1952 for (; $i <= $#lines; $i++) {
1953 $after .= $lines[$i] . "\n";
1956 # Alphas: If we have anonymous text (not part of a procedure), the
1957 # linker may complain about missing exception information. Bleh.
1958 # To suppress this, we place a .ent/.end pair around the code.
1959 # At the same time, we have to be careful and not enclose any leading
1960 # .file/.loc directives.
1961 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1962 local ($ident) = $1;
1963 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1964 $after .= "\t.end $ident\n";
1967 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1968 # assembler (!) wherein .quad constants inside .text sections are
1969 # first narrowed to 32 bits then sign-extended back to 64 bits.
1970 # This obviously screws up our 64-bit bitmaps, so we work around
1971 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1972 if ( $TargetPlatform =~ /^alpha-/ ) {
1974 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1975 local ($number) = $1;
1976 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1977 local ($sign, $base, $digits) = ($1, $2, $3);
1978 $base = (10, 8, 16)[length $base];
1979 local ($hi, $lo) = (0, 0);
1980 foreach $i (split(//, $digits)) {
1981 $j = $lo * $base + $i;
1982 $lo = $j % 4294967296;
1983 $hi = $hi * $base + ($j - $lo) / 4294967296;
1985 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1987 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1988 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1990 print STDERR "Cannot handle \".quad $number\" in info table\n";
1997 if ( $TargetPlatform =~ /x86_64-apple-darwin/ ) {
1998 # Tack a label to the front of the info table, too.
1999 # For now, this just serves to work around a crash in Apple's new
2000 # 64-bit linker (it seems to assume that there is no data before the
2001 # first label in a section).
2003 # The plan for the future is to do this on all Darwin platforms, and
2004 # to add a reference to this label after the entry code, just as the
2005 # NCG does, so we can enable dead-code-stripping in the linker without
2006 # losing our info tables. (Hence the name _dsp, for dead-strip preventer)
2008 $before .= "\n${infoname}_dsp:\n";
2012 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
2013 . join("\n", @words) . "\n"
2016 # print STDERR "before=$before\n";
2017 # print STDERR "label=$label\n";
2018 # print STDERR "words=",(reverse @words),"\n";
2019 # print STDERR "after=$after\n";
2025 The HP is a major nuisance. The threaded code mangler moved info
2026 tables from data space to code space, but unthreaded code in the RTS
2027 still has references to info tables in data space. Since the HP
2028 linker is very precise about where symbols live, we need to patch the
2029 references in the unthreaded RTS as well.
2032 sub mini_mangle_asm_hppa {
2033 local($in_asmf, $out_asmf) = @_;
2035 open(INASM, "< $in_asmf")
2036 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
2037 open(OUTASM,"> $out_asmf")
2038 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
2041 s/_info,DATA/_info,CODE/; # Move _info references to code space
2047 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
2048 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
2054 sub tidy_up_and_die {
2055 local($return_val, $msg) = @_;
2057 exit (($return_val == 0) ? 0 : 1);