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) = @_;
547 # multi-line regexp matching:
549 # the above line used to work, but Perl 5.10 removes $*, so we uses an
550 # equivalent construct that works in Perl 5.6 and later.
551 BEGIN { require overload; overload::constant( qr => sub { "(?m:$_[1])" } ) }
553 # ia64-specific information for code chunks
557 &init_TARGET_STUFF();
558 &init_FUNNY_THINGS();
560 open(INASM, "< $in_asmf")
561 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
562 open(OUTASM,"> $out_asmf")
563 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
565 # read whole file, divide into "chunks":
566 # record some info about what we've found...
568 @chk = (); # contents of the chunk
569 $numchks = 0; # number of them
570 @chkcat = (); # what category of thing in each chunk
571 @chksymb = (); # what symbol(base) is defined in this chunk
572 %entrychk = (); # ditto, its entry code
573 %closurechk = (); # ditto, the (static) closure
574 %srtchk = (); # ditto, its SRT (for top-level things)
575 %infochk = (); # given a symbol base, say what chunk its info tbl is in
576 %vectorchk = (); # ditto, return vector table
577 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
579 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
582 tr/\r//d if $TargetPlatform =~ /-mingw32$/; # In case Perl doesn't convert line endings
583 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
584 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
585 next if /^\t\.def.*endef$/;
586 next if /${T_PRE_APP}(NO_)?APP/o;
587 next if /^;/ && $TargetPlatform =~ /^hppa/;
589 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|ia64)-/;
591 if ( $TargetPlatform =~ /^mips-/
592 && /^\t\.(globl\S+\.text|comm\t)/ ) {
593 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
594 # Treat .comm variables as data. These show up in two (known) places:
596 # - the module_registered variable used in the __stginit fragment.
597 # even though these are declared static and initialised, gcc 3.3
598 # likes to make them .comm, presumably to save space in the
601 # - global variables used to pass arguments from C to STG in
602 # a foreign export. (is this still true? --SDM)
604 } elsif ( /^\t\.comm.*$/ ) {
606 $chkcat[$i] = 'data';
609 # Labels ending "_str": these are literal strings.
610 } elsif ( /^${T_US}([A-Za-z0-9_]+)_str${T_POST_LBL}$/ ) {
612 $chkcat[$i] = 'relrodata';
614 } elsif ( $TargetPlatform =~ /-darwin/
615 && (/^\s*\.subsections_via_symbols/
616 ||/^\s*\.no_dead_strip.*/)) {
617 # Don't allow Apple's linker to do any dead-stripping of symbols
618 # in this file, because it will mess up info-tables in mangled
620 # The .no_dead_strip directives are actually put there by
621 # the gcc3 "used" attribute on entry points.
623 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && (
624 /^\s*\.picsymbol_stub/
625 || /^\s*\.section __TEXT,__picsymbol_stub\d,.*/
626 || /^\s*\.section __TEXT,__picsymbolstub\d,.*/
627 || /^\s*\.symbol_stub/
628 || /^\s*\.section __TEXT,__symbol_stub\d,.*/
629 || /^\s*\.section __TEXT,__symbolstub\d,.*/
630 || /^\s*\.lazy_symbol_pointer/
631 || /^\s*\.non_lazy_symbol_pointer/
632 || /^\s*\.section __IMPORT.*/))
635 $chkcat[$i] = 'dyld';
639 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.data/)
640 { # non_lazy_symbol_ptrs that point to local symbols
642 $chkcat[$i] = 'dyld';
645 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.align/)
646 { # non_lazy_symbol_ptrs that point to local symbols
648 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^L_.*:$/)
649 { # non_lazy_symbol_ptrs that point to local symbols
650 $chk[++$i] = $dyld_section . $_;
651 $chkcat[$i] = 'dyld';
654 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
655 # duplicated from the bottom
659 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
660 # Alphas: Local labels not to be confused with new chunks
662 # NB: all the rest start with a non-space
664 } elsif ( $TargetPlatform =~ /^mips-/
665 && /^\d+:/ ) { # a funny-looking very-local label
668 } elsif ( /$T_CONST_LBL/o ) {
670 $chkcat[$i] = 'literal';
673 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
675 $chkcat[$i] = 'splitmarker';
678 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
681 $chkcat[$i] = 'infotbl';
682 $chksymb[$i] = $symb;
684 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
686 $infochk{$symb} = $i;
688 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
690 $chkcat[$i] = 'entry';
695 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
697 $chkcat[$i] = 'closure';
700 $closurechk{$1} = $i;
702 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
709 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
711 $chkcat[$i] = 'data';
714 } elsif ( /^${T_US}(stg_ap_stack_entries|stg_stack_save_entries|stg_arg_bitmaps)${T_POST_LBL}$/o ) {
716 $chkcat[$i] = 'data';
719 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
722 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
723 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
724 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
725 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
728 $chkcat[$i] = 'data';
731 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
736 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
737 # all CC_ symbols go in the data section...
739 $chkcat[$i] = 'data';
742 } elsif ( /^${T_US}([A-Za-z0-9_]+)_hpc${T_POST_LBL}$/o ) {
743 # hpc shares tick boxes across modules
745 $chkcat[$i] = 'data';
748 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
750 $chkcat[$i] = 'misc';
752 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
754 $chkcat[$i] = 'vector';
759 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
760 && /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
761 # Some Solaris system headers contain function definitions (as
762 # opposed to mere prototypes), which end up in the .hc file when
763 # a Haskell module foreign imports the corresponding system
764 # functions (most notably stat()). We put them into the text
765 # segment. Note that this currently does not extend to function
766 # names starting with an underscore.
769 $chkcat[$i] = 'misc';
772 } elsif ( $TargetPlatform =~ /^i386-apple-darwin/ && /^(___i686\.get_pc_thunk\.[abcd]x):/o) {
773 # To handle PIC on Darwin/x86, we need to appropriately pass through
774 # the get_pc_thunk functions. The need to be put into a special section
775 # marked as coalesced (otherwise the .weak_definition doesn't work
778 $chkcat[$i] = 'get_pc_thunk';
781 } elsif ( /^${T_US}[A-Za-z0-9_]/o
782 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
784 && ( $TargetPlatform !~ /^powerpc64/ # we need to avoid local labels in this case
785 || ! /^\.L\d+:$/ ) ) {
792 /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
793 || /^${T_US}__stg_.*${T_POST_LBL}$/o # more RTS internals
794 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
795 || /^${T_US}.*_slow${T_POST_LBL}$/o # slow entry
796 || /^${T_US}__stginit.*${T_POST_LBL}$/o # __stginit<module>
797 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
798 || /^${T_US}.*_fast${T_POST_LBL}$/o # primops
799 || /^_uname:/o # x86/Solaris2
802 $chkcat[$i] = 'misc';
804 /^${T_US}.*_srtd${T_POST_LBL}$/o # large bitmaps
805 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
808 $chkcat[$i] = 'relrodata';
811 print STDERR "Warning: retaining unknown function \`$thing' in output from C compiler\n";
812 $chkcat[$i] = 'unknown';
815 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ && /^\.LCTOC1 = /o ) {
816 # PowerPC Linux's large-model PIC (-fPIC) generates a gobal offset
817 # table "by hand". Be sure to copy it over.
818 # Note that this label and all entries in the table should actually
819 # go into the .got2 section, but it isn't easy to distinguish them
820 # from other constant literals (.LC\d+), so we just put everything
823 $chkcat[$i] = 'literal';
824 $chksymb[$i] = 'LCTOC1';
825 } else { # simple line (duplicated at the top)
830 $numchks = $#chk + 1;
831 $chk[$numchks] = ''; # We might push .note.GNU-stack into this
832 $chkcat[$numchks] = 'verbatim'; # If we do, write it straight back out
834 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
835 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
838 # the division into chunks is imperfect;
839 # we throw some things over the fence into the next
842 # also, there are things we would like to know
843 # about the whole module before we start spitting
846 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
847 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
849 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
851 # Alphas: NB: we start meddling at chunk 1, not chunk 0
852 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
853 # spits a ".quad 0" in after the very first ".rdata"; we
854 # detect this special case (tossing the ".quad 0")!
855 local($magic_rdata_seen) = 0;
857 # HPPAs, MIPSen: also start medding at chunk 1
859 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
860 $c = $chk[$i]; # convenience copy
862 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
864 # toss all prologue stuff; HPPA is pretty weird
866 $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
871 # be slightly paranoid to make sure there's
872 # nothing surprising in there
873 if ( $c =~ /--- BEGIN ---/ ) {
874 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
876 # remove junk whitespace around the split point
880 if ($TargetPlatform =~ /^i386-/) {
881 if ($p =~ /^\tsubl\s+\$(\d+),\s*\%esp\n/) {
883 die "Error: reserved stack space exceeded!\n Possible workarounds: compile with -fasm, or try another version of gcc.\n"
887 # gcc 3.4.3 puts this kind of stuff in the prologue, eg.
888 # when compiling PrimOps.cmm with -optc-O2:
891 # movl %ecx, 16(%esp)
892 # movl %edx, 20(%esp)
893 # but then the code of the function doesn't assume
894 # anything about the contnets of these stack locations.
895 # I think it's to do with the use of inline functions for
896 # PK_Word64() and friends, where gcc is initialising the
897 # contents of the struct to zero, and failing to optimise
898 # away the initialisation. Let's live dangerously and
899 # discard these initalisations.
901 $p =~ s/^\tpushl\s+\%e(di|si|bx)\n//g;
902 $p =~ s/^\txorl\s+\%e(ax|cx|dx),\s*\%e(ax|cx|dx)\n//g;
903 $p =~ s/^\tmovl\s+\%e(ax|cx|dx|si|di),\s*\d*\(\%esp\)\n//g;
904 $p =~ s/^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n//g;
905 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
906 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-(cygwin32|mingw32)/);
908 if ($TargetPlatform =~ /^i386-apple-darwin/) {
910 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
912 $pcrel_reg =~ s/(.|\n)*.*___i686\.get_pc_thunk\.([abcd]x)\n(.|\n)*/$2/ or $pcrel_reg = "";
913 $p =~ s/^\s+call\s+___i686\.get_pc_thunk\..x//;
914 $p =~ s/^\"?L\d+\$pb\"?:\n//;
916 if ($pcrel_reg eq "bx") {
917 # Bad gcc. Goes and uses %ebx, our BaseReg, for PIC. Bad gcc.
918 die "Darwin/x86: -fPIC -via-C doesn't work yet, use -fasm. Aborting."
922 } elsif ($TargetPlatform =~ /^x86_64-/) {
923 $p =~ s/^\tpushq\s+\%r(bx|bp|12|13|14)\n//g;
924 $p =~ s/^\tmovq\s+\%r(bx|bp|12|13|14),\s*\d*\(\%rsp\)\n//g;
925 $p =~ s/^\tsubq\s+\$\d+,\s*\%rsp\n//;
927 } elsif ($TargetPlatform =~ /^ia64-/) {
928 $p =~ s/^\t\.prologue .*\n//;
930 # Record the number of local and out registers for register relocation later
931 $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, (\d+), (\d+), 0\n//;
935 $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
936 $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
938 # Ignore save/restore of these registers; they're taken
939 # care of in StgRun()
940 $p =~ s/^\t\.save ar\.lc, r\d+\n//;
941 $p =~ s/^\t\.save pr, r\d+\n//;
942 $p =~ s/^\tmov r\d+ = ar\.lc\n//;
943 $p =~ s/^\tmov r\d+ = pr\n//;
945 # Remove .proc and .body directives
946 $p =~ s/^\t\.proc [a-zA-Z0-9_.]+#\n//;
947 $p =~ s/^\t\.body\n//;
949 # If there's a label, move it to the body
950 if ($p =~ /^[a-zA-Z0-9.]+:\n/) {
955 # Remove floating-point spill instructions.
956 # Only fp registers 2-5 and 16-23 are saved by the runtime.
957 if ($p =~ s/^\tstf\.spill \[r1[4-9]\] = f([2-5]|1[6-9]|2[0-3])(, [0-9]+)?\n//g) {
958 # Being paranoid, only try to remove these if we saw a
960 $p =~ s/^\tmov r1[4-9] = r12\n//;
961 $p =~ s/^\tadds r1[4-9] = -[0-9]+, r12\n//g;
962 $p =~ s/^\t\.save\.f 0x[0-9a-fA-F]\n//g;
963 $p =~ s/^\t\.save\.gf 0x0, 0x[0-9a-fA-F]+\n//g;
966 $p =~ s/^\tnop(?:\.[mifb])?\s+\d+\n//g; # remove nop instructions
967 $p =~ s/^\t\.(mii|mmi|mfi)\n//g; # bundling is no longer sensible
968 $p =~ s/^\t;;\n//g; # discard stops
969 $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
971 # GCC 3.3 saves r1 in the prologue, move this to the body
972 # (Does this register get restored anywhere?)
973 if ($p =~ /^\tmov r\d+ = r1\n/) {
977 } elsif ($TargetPlatform =~ /^m68k-/) {
978 $p =~ s/^\tlink a6,#-?\d.*\n//;
979 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
980 # The above showed up in the asm code,
981 # so I added it here.
982 # I hope it's correct.
984 $p =~ s/^\tmovel d2,sp\@-\n//;
985 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
986 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
987 } elsif ($TargetPlatform =~ /^mips-/) {
988 # the .frame/.mask/.fmask that we use is the same
989 # as that produced by GCC for miniInterpret; this
990 # gives GDB some chance of figuring out what happened
991 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
992 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
993 $p =~ s/^\t\.(mask|fmask).*\n//g;
994 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
995 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
996 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
997 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
998 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
999 $p =~ s/__FRAME__/$FRAME/;
1000 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
1002 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
1004 $p =~ s/^\tmflr r0\n//;
1005 $p =~ s/^\tbl saveFP # f\d+\n//;
1006 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
1007 $p =~ s/^\"?L\d+\$pb\"?:\n//;
1008 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
1009 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
1010 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
1011 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
1012 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
1013 $p =~ s/^\tbcl 20,31,\"?L\d+\$pb\"?\n//;
1014 $p =~ s/^\"?L\d+\$pb\"?:\n//;
1015 $p =~ s/^\tmflr r31\n//;
1017 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1018 # under some circumstances, only when generating position dependent code.
1019 # I have no idea why, and I don't think it is necessary, so let's toss it.
1020 $p =~ s/^\tli r\d+,0\n//g;
1021 $p =~ s/^\tstw r\d+,\d+\(r1\)\n//g;
1022 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/) {
1023 $p =~ s/^\tmflr 0\n//;
1024 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
1025 $p =~ s/^\tstfd \d+,\d+\(1\)\n//g;
1026 $p =~ s/^\tstw r0,8\(1\)\n//;
1027 $p =~ s/^\tstwu 1,-\d+\(1\)\n//;
1028 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1030 # GCC's "large-model" PIC (-fPIC)
1032 $pcrel_label =~ s/(.|\n)*^.LCF(\d+):\n(.|\n)*/$2/ or $pcrel_label = "";
1034 $p =~ s/^\tbcl 20,31,.LCF\d+\n//;
1035 $p =~ s/^.LCF\d+:\n//;
1036 $p =~ s/^\tmflr 30\n//;
1037 $p =~ s/^\tlwz 0,\.LCL\d+-\.LCF\d+\(30\)\n//;
1038 $p =~ s/^\tadd 30,0,30\n//;
1040 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1041 # under some circumstances, only when generating position dependent code.
1042 # I have no idea why, and I don't think it is necessary, so let's toss it.
1043 $p =~ s/^\tli \d+,0\n//g;
1044 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1045 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1046 $p =~ s/^\tmr 31,1\n//;
1047 $p =~ s/^\tmflr 0\n//;
1048 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
1049 $p =~ s/^\tstfd \d+,-?\d+\(1\)\n//g;
1050 $p =~ s/^\tstd r0,8\(1\)\n//;
1051 $p =~ s/^\tstdu 1,-\d+\(1\)\n//;
1052 $p =~ s/^\tstd \d+,-?\d+\(1\)\n//g;
1054 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1055 # under some circumstances, only when generating position dependent code.
1056 # I have no idea why, and I don't think it is necessary, so let's toss it.
1057 $p =~ s/^\tli \d+,0\n//g;
1058 $p =~ s/^\tstd \d+,\d+\(1\)\n//g;
1060 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
1063 # HWL HACK: dont die, just print a warning
1064 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
1065 die "Prologue junk?: $p\n" if $p =~ /^\s+[^\s\.]/;
1067 # For PIC, we want to keep part of the prologue
1068 if ($TargetPlatform =~ /^powerpc-apple-darwin.*/ && $pcrel_label ne "") {
1069 # Darwin: load the current instruction pointer into register r31
1070 $p .= "bcl 20,31,$pcrel_label\n";
1071 $p .= "$pcrel_label:\n";
1072 $p .= "\tmflr r31\n";
1073 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/ && $pcrel_label ne "") {
1074 # Linux: load the GOT pointer into register 30
1075 $p .= "\tbcl 20,31,.LCF$pcrel_label\n";
1076 $p .= ".LCF$pcrel_label:\n";
1077 $p .= "\tmflr 30\n";
1078 $p .= "\tlwz 0,.LCL$pcrel_label-.LCF$pcrel_label(30)\n";
1079 $p .= "\tadd 30,0,30\n";
1080 } elsif ($TargetPlatform =~ /^i386-apple-darwin.*/ && $pcrel_label ne "") {
1081 $p .= "\tcall ___i686.get_pc_thunk.$pcrel_reg\n";
1082 $p .= "$pcrel_label:\n";
1085 # glue together what's left
1090 if ( $TargetPlatform =~ /^mips-/ ) {
1091 # MIPS: first, this basic sequence may occur "--- END ---" or not
1092 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
1095 # toss all epilogue stuff; again, paranoidly
1096 if ( $c =~ /--- END ---/ ) {
1097 # Gcc may decide to replicate the function epilogue. We want
1098 # to process all epilogues, so we split the function and then
1100 @fragments = split(/--- END ---/, $c);
1101 $r = shift(@fragments);
1103 # Rebuild `c'; processed fragments will be appended to `c'
1106 foreach $e (@fragments) {
1107 # etail holds code that is after the epilogue in the assembly-code
1108 # layout and should not be filtered as part of the epilogue.
1110 if ($TargetPlatform =~ /^i386-/) {
1112 $e =~ s/^\tpopl\s+\%edi\n//;
1113 $e =~ s/^\tpopl\s+\%esi\n//;
1114 $e =~ s/^\tpopl\s+\%edx\n//;
1115 $e =~ s/^\tpopl\s+\%ecx\n//;
1116 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
1117 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
1118 } elsif ($TargetPlatform =~ /^ia64-/) {
1119 # The epilogue is first split into:
1120 # $e, the epilogue code (up to the return instruction)
1121 # $etail, non-epilogue code (after the return instruction)
1122 # The return instruction is stripped in the process.
1123 if (!(($e, $etail) = split(/^\tbr\.ret\.sptk\.many b0\n/, $e))) {
1124 die "Epilogue doesn't seem to have one return instruction: $e\n";
1126 # Remove 'endp' directive from the tail
1127 $etail =~ s/^\t\.endp [a-zA-Z0-9_.]+#\n//;
1129 # If a return value is saved here, discard it
1130 $e =~ s/^\tmov r8 = r14\n//;
1132 # Remove floating-point fill instructions.
1133 # Only fp registers 2-5 and 16-23 are saved by the runtime.
1134 if ($e =~ s/^\tldf\.fill f([2-5]|1[6-9]|2[0-3]) = \[r1[4-9]\](, [0-9]+)?\n//g) {
1135 # Being paranoid, only try to remove this if we saw a fill
1137 $e =~ s/^\tadds r1[4-9] = [0-9]+, r12//g;
1140 $e =~ s/^\tnop(?:\.[mifb])?\s+\d+\n//g; # remove nop instructions
1141 $e =~ s/^\tmov ar\.pfs = r\d+\n//;
1142 $e =~ s/^\tmov ar\.lc = r\d+\n//;
1143 $e =~ s/^\tmov pr = r\d+, -1\n//;
1144 $e =~ s/^\tmov b0 = r\d+\n//;
1145 $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
1146 #$e =~ s/^\tbr\.ret\.sptk\.many b0\n//; # already removed
1147 $e =~ s/^\t\.(mii|mmi|mfi|mib)\n//g; # bundling is no longer sensible
1148 $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
1149 $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
1150 } elsif ($TargetPlatform =~ /^m68k-/) {
1151 $e =~ s/^\tunlk a6\n//;
1153 } elsif ($TargetPlatform =~ /^mips-/) {
1154 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
1155 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
1156 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
1157 $e =~ s/^\tj\t\$31\n//;
1158 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
1159 $e =~ s/^\taddi r1,r1,\d+\n//;
1160 $e =~ s/^\tlwz r\d+,\d+\(r1\)\n//;
1161 $e =~ s/^\tlmw r\d+,-\d+\(r1\)\n//;
1162 $e =~ s/^\tmtlr r0\n//;
1164 $e =~ s/^\tb restFP ;.*\n//;
1165 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1166 $e =~ s/^\tmr 3,0\n//;
1167 $e =~ s/^\taddi 1,1,\d+\n//;
1168 $e =~ s/^\tld 0,16\(1\)\n//;
1169 $e =~ s/^\tmtlr 0\n//;
1171 # callee-save registers
1172 $e =~ s/^\tld \d+,-?\d+\(1\)\n//g;
1173 $e =~ s/^\tlfd \d+,-?\d+\(1\)\n//g;
1175 # get rid of the debug junk along with the blr
1176 $e =~ s/^\tblr\n\t.long .*\n\t.byte .*\n//;
1178 # incase we missed it with the last one get the blr alone
1181 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
1184 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
1186 # glue together what's left
1189 $c =~ s/\n\t\n/\n/; # junk blank line
1192 if ($TargetPlatform =~ /^ia64-/) {
1193 # On IA64, remove an .endp directive even if no epilogue was found.
1194 # Code optimizations may have removed the "--- END ---" token.
1195 $c =~ s/^\t\.endp [a-zA-Z0-9_.]+#\n//;
1199 # On SPARCs, we don't do --- BEGIN/END ---, we just
1200 # toss the register-windowing save/restore/ret* instructions
1201 # directly unless they've been generated by function definitions in header
1203 if ( $TargetPlatform =~ /^sparc-/ ) {
1204 if ( ! ( $TargetPlatform =~ /solaris2$/ && $chkcat[$i] eq 'unknown' )) {
1205 $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
1207 # throw away PROLOGUE comments
1208 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
1211 # On Alphas, the prologue mangling is done a little later (below)
1213 # toss all calls to __DISCARD__
1214 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
1215 $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
1216 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-darwin.*/;
1217 $c =~ s/^\tbl\s+__DISCARD__(\@plt)?\n//go if $TargetPlatform =~ /^powerpc-.*-linux/;
1218 $c =~ s/^\tbl\s+\.__DISCARD__\n\s+nop\n//go if $TargetPlatform =~ /^powerpc64-.*-linux/;
1219 $c =~ s/^\tcall\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /i386-apple-darwin.*/;
1221 # IA64: fix register allocation; mangle tailcalls into jumps
1222 if ($TargetPlatform =~ /^ia64-/) {
1223 ia64_rename_registers($ia64_locnum, $ia64_outnum) if (defined($ia64_locnum));
1224 ia64_mangle_tailcalls();
1227 # MIPS: that may leave some gratuitous asm macros around
1228 # (no harm done; but we get rid of them to be tidier)
1229 $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/
1230 if $TargetPlatform =~ /^mips-/;
1232 # toss stack adjustment after DoSparks
1233 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
1234 if $TargetPlatform =~ /^m68k-/; # this looks old...
1236 if ( $TargetPlatform =~ /^alpha-/ &&
1237 ! $magic_rdata_seen &&
1238 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
1239 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
1240 $magic_rdata_seen = 1;
1243 # pick some end-things and move them to the next chunk
1245 # pin a funny end-thing on (for easier matching):
1246 $c .= 'FUNNY#END#THING';
1248 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
1252 # on x86 we try not to copy any directives into a literal
1253 # chunk, rather we keep looking for the next real chunk. This
1254 # is because we get things like
1256 # .globl blah_closure
1262 if ( $TargetPlatform =~ /^(i386|sparc|powerpc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
1264 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
1267 if ( $j < $numchks ) {
1268 $chk[$j] = $to_move . $chk[$j];
1272 elsif ( ( $i < ($numchks - 1)
1273 && ( $to_move =~ /${T_COPY_DIRVS}/
1274 || ( $TargetPlatform =~ /^hppa/
1275 && $to_move =~ /align/
1276 && $chkcat[$i+1] eq 'literal')
1279 || ($to_move =~ /^[ \t]*\.section[ \t]+\.note\.GNU-stack,/)
1281 $chk[$i + 1] = $to_move . $chk[$i + 1];
1282 # otherwise they're tossed
1285 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
1288 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
1290 # toss all prologue stuff, except for loading gp, and the ..ng address
1291 unless ($c =~ /\.ent.*\n\$.*\.\.ng:/) {
1292 if (($p, $r) = split(/^\t\.prologue/, $c)) {
1293 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
1294 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
1295 $keep =~ s/^\t\.(mask|fmask).*\n//g;
1296 $c = $keep . "..ng:\n";
1298 print STDERR "malformed code block ($ent)?\n"
1301 $c .= "\t.prologue" . $r;
1305 $c =~ s/FUNNY#END#THING//;
1307 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
1309 $chk[$i] = $c; # update w/ convenience copy
1312 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
1313 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
1316 if ( $TargetPlatform =~ /^alpha-/ ) {
1317 # print out the header stuff first
1318 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
1319 print OUTASM $chk[0];
1321 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1322 print OUTASM $chk[0];
1324 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1325 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
1327 # get rid of horrible "<dollar>Revision: .*$" strings
1328 local(@lines0) = split(/\n/, $chk[0]);
1330 while ( $z <= $#lines0 ) {
1331 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
1334 while ( $z <= $#lines0 ) {
1336 last if $lines0[$z] =~ /[,\t]0x0$/;
1342 $chk[0] = join("\n", @lines0);
1343 $chk[0] =~ s/\n\n+/\n/;
1344 print OUTASM $chk[0];
1347 # print out all the literal strings next
1348 for ($i = 0; $i < $numchks; $i++) {
1349 if ( $chkcat[$i] eq 'literal' ) {
1351 # HACK: try to detect 16-byte constants and align them
1352 # on a 16-byte boundary. x86_64 sometimes needs 128-bit
1353 # aligned constants, and so does Darwin/x86.
1354 if ( $TargetPlatform =~ /^x86_64/
1355 || $TargetPlatform =~ /^i386-apple-darwin/ ) {
1357 if ($z =~ /(\.long.*\n.*\.long.*\n.*\.long.*\n.*\.long|\.quad.*\n.*\.quad)/) {
1358 print OUTASM $T_HDR_literal16;
1360 print OUTASM $T_HDR_literal;
1363 print OUTASM $T_HDR_literal;
1366 print OUTASM $chk[$i];
1367 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
1369 $chkcat[$i] = 'DONE ALREADY';
1373 # on the HPPA, print out all the bss next
1374 if ( $TargetPlatform =~ /^hppa/ ) {
1375 for ($i = 1; $i < $numchks; $i++) {
1376 if ( $chkcat[$i] eq 'bss' ) {
1377 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
1378 print OUTASM $chk[$i];
1380 $chkcat[$i] = 'DONE ALREADY';
1385 # $numchks + 1 as we have the extra one for .note.GNU-stack
1386 for ($i = $FIRST_MANGLABLE; $i < $numchks + 1; $i++) {
1387 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
1389 next if $chkcat[$i] eq 'DONE ALREADY';
1391 if ( $chkcat[$i] eq 'misc' || $chkcat[$i] eq 'unknown' ) {
1392 if ($chk[$i] ne '') {
1393 print OUTASM $T_HDR_misc;
1394 &print_doctored($chk[$i], 0);
1397 } elsif ( $chkcat[$i] eq 'verbatim' ) {
1398 print OUTASM $chk[$i];
1400 } elsif ( $chkcat[$i] eq 'toss' ) {
1401 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
1403 } elsif ( $chkcat[$i] eq 'data' ) {
1404 if ($chk[$i] ne '') {
1405 print OUTASM $T_HDR_data;
1406 print OUTASM $chk[$i];
1409 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1410 # we can just re-constitute this one...
1411 # NB: we emit _three_ underscores no matter what,
1412 # so ghc-split doesn't have to care.
1413 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1415 } elsif ( $chkcat[$i] eq 'closure'
1416 || $chkcat[$i] eq 'srt'
1417 || $chkcat[$i] eq 'infotbl'
1418 || $chkcat[$i] eq 'entry') { # do them in that order
1419 $symb = $chksymb[$i];
1422 if ( defined($closurechk{$symb}) ) {
1423 print OUTASM $T_HDR_closure;
1424 print OUTASM $chk[$closurechk{$symb}];
1425 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1429 if ( defined($srtchk{$symb}) ) {
1430 print OUTASM $T_HDR_relrodata;
1431 print OUTASM $chk[$srtchk{$symb}];
1432 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1436 if ( defined($infochk{$symb}) ) {
1438 print OUTASM $T_HDR_info;
1439 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1441 # entry code will be put here!
1443 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1447 if ( defined($entrychk{$symb}) ) {
1449 $c = $chk[$entrychk{$symb}];
1451 # If this is an entry point with an info table,
1452 # eliminate the entry symbol and all directives involving it.
1453 if (defined($infochk{$symb}) && $TargetPlatform !~ /^ia64-/) {
1455 foreach $l (split(/\n/,$c)) {
1456 next if $l =~ /^.*$symb_(entry|ret)${T_POST_LBL}/;
1458 # If we have .type/.size direrctives involving foo_entry,
1459 # then make them refer to foo_info instead. The information
1460 # in these directives is used by the cachegrind annotator,
1461 # so it is worthwhile keeping.
1462 if ($l =~ /^\s*\.(type|size).*$symb_(entry|ret)/) {
1463 $l =~ s/$symb(_entry|_ret)/${symb}_info/g;
1467 next if $l =~ /^\s*\..*$symb.*\n?/;
1470 $c = join("\n",@o) . "\n";
1473 print OUTASM $T_HDR_entry;
1475 &print_doctored($c, 1); # NB: the 1!!!
1477 $chkcat[$entrychk{$symb}] = 'DONE ALREADY';
1480 } elsif ( $chkcat[$i] eq 'vector' ) {
1481 $symb = $chksymb[$i];
1484 if ( defined($vectorchk{$symb}) ) {
1485 print OUTASM $T_HDR_vector;
1486 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1488 # direct return code will be put here!
1489 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1491 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1492 # Alphas: the commented nop is for the splitter, to ensure
1493 # that no module ends with a label as the very last
1494 # thing. (The linker will adjust the label to point
1495 # to the first code word of the next module linked in,
1496 # even if alignment constraints cause the label to move!)
1498 print OUTASM "\t# nop\n";
1501 } elsif ( $chkcat[$i] eq 'rodata' ) {
1502 print OUTASM $T_HDR_rodata;
1503 print OUTASM $chk[$i];
1504 $chkcat[$i] = 'DONE ALREADY';
1505 } elsif ( $chkcat[$i] eq 'relrodata' ) {
1506 print OUTASM $T_HDR_relrodata;
1507 print OUTASM $chk[$i];
1508 $chkcat[$i] = 'DONE ALREADY';
1509 } elsif ( $chkcat[$i] eq 'toc' ) {
1510 # silly optimisation to print tocs, since they come in groups...
1511 print OUTASM $T_HDR_toc;
1513 while ($chkcat[$j] eq 'toc')
1514 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1517 print OUTASM $chk[$j];
1519 $chkcat[$j] = 'DONE ALREADY';
1523 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' ) {
1524 # apple-darwin: dynamic linker stubs
1525 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1526 { # print them out unchanged, but remove the stubs for __DISCARD__
1527 print OUTASM $chk[$i];
1529 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ && $chkcat[$i] eq 'get_pc_thunk' ) {
1530 # i386-apple-darwin: __i686.get_pc_thunk.[abcd]x
1531 print OUTASM ".section __TEXT,__textcoal_nt,coalesced,no_toc\n";
1532 print OUTASM $chk[$i];
1534 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1538 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1541 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1542 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1546 On IA64, tail calls are converted to branches at this point. The mangler
1547 searches for function calls immediately followed by a '--- TAILCALL ---'
1548 token. Since the compiler can put various combinations of labels, bundling
1549 directives, nop instructions, stops, and a move of the return value
1550 between the branch and the tail call, proper matching of the tail call
1551 gets a little hairy. This subroutine does the mangling.
1553 Here is an example of a tail call before mangling:
1556 br.call.sptk.many b0 = b6
1571 sub ia64_mangle_tailcalls {
1572 # Function input and output are in $c
1574 # Construct the tailcall-mangling expression the first time this function
1576 if (!defined($IA64_MATCH_TAILCALL)) {
1577 # One-line pattern matching constructs. None of these
1578 # should bind references; all parenthesized terms
1579 # should be (?:) terms.
1580 my $stop = q/(?:\t;;\n)/;
1581 my $bundle = q/(?:\t\.(?:mii|mib|mmi|mmb|mfi|mfb|mbb|bbb)\n)/;
1582 my $nop = q/(?:\tnop(?:\.[mifb])?\s+\d+\n)/;
1583 my $movgp = q/(?:\tmov r1 = r\d+\n)/;
1584 my $postbr = q/(?:\tbr \.L\d+\n)/;
1586 my $noeffect = "(?:$stop$bundle?|$nop)*";
1587 my $postbundle = "(?:$bundle?$nop?$nop?$postbr)?";
1589 # Important parts of the pattern match. The branch target
1590 # and subsequent jump label are bound to $1 and $2
1591 # respectively. Sometimes there is no label.
1592 my $callbr = q/^\tbr\.call\.sptk\.many b0 = (.*)\n/;
1593 my $label = q/(?:^\.L([0-9]*):\n)/;
1594 my $tailcall = q/\t--- TAILCALL ---\n/;
1596 $IA64_MATCH_TAILCALL =
1597 $callbr . $label . '?' . $noeffect . $movgp . '?' . $noeffect .
1598 $tailcall . $stop . '?' . '(?:' . $postbundle . ')?';
1601 # Find and mangle tailcalls
1602 while ($c =~ s/$IA64_MATCH_TAILCALL/\tbr\.few $1\n/o) {
1603 # Eek, the gcc optimiser is getting smarter... if we see a jump to the
1604 # --- TAILCALL --- marker then we reapply the substitution at the source sites
1605 $c =~ s/^\tbr \.L$2\n/\t--- TAILCALL ---\n/g if ($2);
1608 # Verify that all instances of TAILCALL were processed
1609 if ($c =~ /^\t--- TAILCALL ---\n/) {
1610 die "Unmangled TAILCALL tokens remain after mangling"
1615 The number of registers allocated on the IA64 register stack is set
1616 upon entry to the runtime with an `alloc' instruction at the entry
1617 point of \verb+StgRun()+. Gcc uses its own `alloc' to allocate
1618 however many registers it likes in each function. When we discard
1619 gcc's alloc, we have to reconcile its register assignment with what
1622 There are three stack areas: fixed registers, input/local registers,
1623 and output registers. We move the output registers to the output
1624 register space and leave the other registers where they are.
1627 sub ia64_rename_registers() {
1628 # The text to be mangled is in $c
1629 # Find number of registers in each stack area
1630 my ($loc, $out) = @_;
1636 # These are the register numbers used in the STG runtime
1637 my $STG_FIRST_OUT_REG = 32 + 34;
1638 my $STG_LAST_OUT_REG = $STG_FIRST_OUT_REG + 7;
1640 $first_out_reg = 32 + $loc;
1642 if ($first_out_reg > $STG_FIRST_OUT_REG) {
1643 die "Too many local registers allocated by gcc";
1646 # Split the string into fragments containing one register name each.
1647 # Rename the register in each fragment and concatenate.
1649 foreach $fragment (split(/(?=r\d+[^a-zA-Z0-9_.])/s, $c)) {
1650 if ($fragment =~ /^r(\d+)((?:[^a-zA-Z0-9_.].*)?)$/s) {
1653 if ($regnum < $first_out_reg) {
1654 # This is a local or fixed register
1656 # Local registers 32 and 33 (r64 and r65) are
1657 # used to hold saved state; they shouldn't be touched
1658 if ($regnum == 64 || $regnum == 65) {
1659 die "Reserved register $regnum is in use";
1663 # This is an output register
1664 $regnum = $regnum - $first_out_reg + $STG_FIRST_OUT_REG;
1665 if ($regnum > $STG_LAST_OUT_REG) {
1666 die "Register number ($regnum) is out of expected range";
1670 # Update this fragment
1671 $fragment = "r" . $regnum . $2;
1682 sub hppa_mash_prologue { # OK, epilogue, too
1685 # toss all prologue stuff
1686 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1688 # Lie about our .CALLINFO
1689 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1696 # toss all epilogue stuff
1697 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1699 # Sorry; we moved the _info stuff to the code segment.
1700 s/_info,DATA/_info,CODE/g;
1707 sub print_doctored {
1708 local($_, $need_fallthru_patch) = @_;
1710 if ( $TargetPlatform =~ /^x86_64-/ ) {
1713 # movq -4(%ebp), %rax
1718 s/^\tmovq\s+(-?\d*\(\%r(bx|bp|13)\)),\s*(\%r(ax|cx|dx|10|11))\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1719 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e(ax|cx|si|di))\n\tjmp\s+\*\%r\3/\tjmp\t$T_US$1/g;
1722 if ( $TargetPlatform !~ /^i386-/
1723 || ! /^\t[a-z]/ # no instructions in here, apparently
1724 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1729 # OK, must do some x86 **HACKING**
1731 local($entry_patch) = '';
1732 local($exit_patch) = '';
1734 # gotta watch out for weird instructions that
1735 # invisibly smash various regs:
1736 # rep* %ecx used for counting
1737 # scas* %edi used for destination index
1738 # cmps* %e[sd]i used for indices
1739 # loop* %ecx used for counting
1744 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1746 # * GCC used an "STG reg" for its own purposes
1748 # * some secret uses of machine reg, requiring STG reg
1749 # to be saved/restored
1751 # The most dangerous "GCC uses" of an "STG reg" are when
1752 # the reg holds the target of a jmp -- it's tricky to
1753 # insert the patch-up code before we get to the target!
1754 # So here we change the jmps:
1756 # --------------------------------------------------------
1757 # it can happen that we have jumps of the form...
1758 # jmp *<something involving %esp>
1760 # jmp <something involving another naughty register...>
1762 # a reasonably-common case is:
1764 # movl $_blah,<bad-reg>
1767 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[acd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1771 # movl -4(%ebx), %eax
1776 s/^\tmovl\s+(-?\d*\(\%e(bx|si)\)),\s*(\%e[acd]x)\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1778 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1779 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1780 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1781 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1782 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1783 if /(jmp|call)\s+.*\%esi/;
1785 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1786 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1787 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1788 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1789 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1790 if /(jmp|call)\s+.*\%edi/;
1793 # OK, now we can decide what our patch-up code is going to
1796 # Offsets into register table - you'd better update these magic
1797 # numbers should you change its contents!
1798 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1799 local($OFFSET_Hp)=88;
1801 # Note funky ".=" stuff; we're *adding* to these _patch guys
1802 if ( $StolenX86Regs <= 2
1803 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1804 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1805 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1807 # nothing for call_{entry,exit} because %esi is callee-save
1809 if ( $StolenX86Regs <= 3
1810 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1811 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1812 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1814 # nothing for call_{entry,exit} because %edi is callee-save
1817 # --------------------------------------------------------
1818 # next, here we go with non-%esp patching!
1820 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1822 # Before calling GC we must set up the exit condition before the call
1823 # and entry condition when we come back
1825 # fix _all_ non-local jumps:
1827 if ( $TargetPlatform =~ /^.*-apple-darwin.*/ ) {
1828 # On Darwin, we've got local-looking jumps that are
1829 # actually global (i.e. jumps to Lfoo$stub or via
1830 # Lfoo$non_lazy_ptr), so we fix those first.
1831 # In fact, we just fix everything that contains a dollar
1832 # because false positives don't hurt here.
1834 s/^(\tjmp\s+\*?L.*\$.*\n)/$exit_patch$1/g;
1837 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1838 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1840 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1842 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1843 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1845 if ($StolenX86Regs == 2 ) {
1846 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1847 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1848 } elsif ($StolenX86Regs == 3 ) {
1849 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1850 if /^\t(jmp|call)\s+.*\%edi/;
1853 # --------------------------------------------------------
1854 # that's it -- print it
1856 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1860 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1861 print OUTASM $exit_patch;
1862 # ToDo: make it not print if there is a "jmp" at the end
1868 sub init_FUNNY_THINGS {
1869 %KNOWN_FUNNY_THING = (
1871 # "${T_US}stg_.*{T_POST_LBL}", 1,
1876 The following table reversal is used for both info tables and return
1877 vectors. In both cases, we remove the first entry from the table,
1878 reverse the table, put the label at the end, and paste some code
1879 (that which is normally referred to by the first entry in the table)
1880 right after the table itself. (The code pasting is done elsewhere.)
1884 local($symb, $tbl, $discard1) = @_;
1886 return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1888 local($before) = '';
1890 local(@imports) = (); # hppa only
1893 local(@lines) = split(/\n/, $tbl);
1896 # Deal with the header...
1897 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1898 $label .= $lines[$i] . "\n",
1899 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1900 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1901 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1903 $before .= $lines[$i] . "\n"; # otherwise...
1907 $infoname =~ s/(.|\n)*^([A-Za-z0-9_]+_info)${T_POST_LBL}$(.|\n)*/\2/;
1909 # Grab the table data...
1910 if ( $TargetPlatform !~ /^hppa/ ) {
1911 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1913 # Convert addresses of SRTs, slow entrypoints and large bitmaps
1914 # to offsets (relative to the info label),
1915 # in order to support position independent code.
1916 $line =~ s/$infoname/0/
1917 || $line =~ s/([A-Za-z0-9_]+_srtd)$/\1 - $infoname/
1918 || $line =~ s/([A-Za-z0-9_]+_srt(\+\d+)?)$/\1 - $infoname/
1919 || $line =~ s/([A-Za-z0-9_]+_str)$/\1 - $infoname/
1920 || $line =~ s/([A-Za-z0-9_]+_slow)$/\1 - $infoname/
1921 || $line =~ s/([A-Za-z0-9_]+_btm)$/\1 - $infoname/
1922 || $line =~ s/([A-Za-z0-9_]+_alt)$/\1 - $infoname/
1923 || $line =~ s/([A-Za-z0-9_]+_dflt)$/\1 - $infoname/
1924 || $line =~ s/([A-Za-z0-9_]+_ret)$/\1 - $infoname/;
1925 push(@words, $line);
1927 } else { # hppa weirdness
1928 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1929 # FIXME: the RTS now expects offsets instead of addresses
1930 # for all labels in info tables.
1931 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1932 push(@imports, $lines[$i]);
1934 # We don't use HP's ``function pointers''
1935 # We just use labels in code space, like normal people
1936 $lines[$i] =~ s/P%//;
1937 push(@words, $lines[$i]);
1942 # Now throw away any initial zero word from the table. This is a hack
1943 # that lets us reduce the size of info tables when the SRT field is not
1944 # needed: see comments StgFunInfoTable in InfoTables.h.
1946 # The .zero business is for Linux/ELF.
1947 # The .skip business is for Sparc/Solaris/ELF.
1948 # The .blockz business is for HPPA.
1950 # if ($words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1955 for (; $i <= $#lines; $i++) {
1956 $after .= $lines[$i] . "\n";
1959 # Alphas: If we have anonymous text (not part of a procedure), the
1960 # linker may complain about missing exception information. Bleh.
1961 # To suppress this, we place a .ent/.end pair around the code.
1962 # At the same time, we have to be careful and not enclose any leading
1963 # .file/.loc directives.
1964 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1965 local ($ident) = $1;
1966 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1967 $after .= "\t.end $ident\n";
1970 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1971 # assembler (!) wherein .quad constants inside .text sections are
1972 # first narrowed to 32 bits then sign-extended back to 64 bits.
1973 # This obviously screws up our 64-bit bitmaps, so we work around
1974 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1975 if ( $TargetPlatform =~ /^alpha-/ ) {
1977 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1978 local ($number) = $1;
1979 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1980 local ($sign, $base, $digits) = ($1, $2, $3);
1981 $base = (10, 8, 16)[length $base];
1982 local ($hi, $lo) = (0, 0);
1983 foreach $i (split(//, $digits)) {
1984 $j = $lo * $base + $i;
1985 $lo = $j % 4294967296;
1986 $hi = $hi * $base + ($j - $lo) / 4294967296;
1988 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1990 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1991 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1993 print STDERR "Cannot handle \".quad $number\" in info table\n";
2000 if ( $TargetPlatform =~ /x86_64-apple-darwin/ ) {
2001 # Tack a label to the front of the info table, too.
2002 # For now, this just serves to work around a crash in Apple's new
2003 # 64-bit linker (it seems to assume that there is no data before the
2004 # first label in a section).
2006 # The plan for the future is to do this on all Darwin platforms, and
2007 # to add a reference to this label after the entry code, just as the
2008 # NCG does, so we can enable dead-code-stripping in the linker without
2009 # losing our info tables. (Hence the name _dsp, for dead-strip preventer)
2011 $before .= "\n${infoname}_dsp:\n";
2015 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
2016 . join("\n", @words) . "\n"
2019 # print STDERR "before=$before\n";
2020 # print STDERR "label=$label\n";
2021 # print STDERR "words=",(reverse @words),"\n";
2022 # print STDERR "after=$after\n";
2028 The HP is a major nuisance. The threaded code mangler moved info
2029 tables from data space to code space, but unthreaded code in the RTS
2030 still has references to info tables in data space. Since the HP
2031 linker is very precise about where symbols live, we need to patch the
2032 references in the unthreaded RTS as well.
2035 sub mini_mangle_asm_hppa {
2036 local($in_asmf, $out_asmf) = @_;
2038 open(INASM, "< $in_asmf")
2039 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
2040 open(OUTASM,"> $out_asmf")
2041 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
2044 s/_info,DATA/_info,CODE/; # Move _info references to code space
2050 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
2051 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
2057 sub tidy_up_and_die {
2058 local($return_val, $msg) = @_;
2060 exit (($return_val == 0) ? 0 : 1);