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 ############################################################################
63 # Make all regexp matching multi-line aware. This replaces the line below
64 # originally found in "sub mangle_asm":
68 # This used to work, but Perl 5.10 removes support for $*, so we uses an
69 # equivalent construct that works in Perl 5.6 and later.
71 BEGIN { if ($] >= 5.010) {
72 require overload; overload::constant( qr => sub { "(?m:$_[1])" } );
74 ############################################################################
76 $TargetPlatform = $TARGETPLATFORM;
78 ($Pgm = $0) =~ s|.*/||;
82 if ( $TargetPlatform =~ /^i386-/ ) {
86 $StolenX86Regs = $ARGV[2];
90 &mangle_asm($ifile,$ofile);
95 %************************************************************************
97 \subsection{Constants for various architectures}
99 %************************************************************************
102 sub init_TARGET_STUFF {
104 #--------------------------------------------------------#
105 if ( $TargetPlatform =~ /^alpha-.*-.*/ ) {
107 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
108 $T_US = ''; # _ if symbols have an underscore on the front
109 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
110 $T_CONST_LBL = '^\$L?C(\d+):$'; # regexp for what such a lbl looks like
113 $T_MOVE_DIRVS = '^(\s*(\$.*\.\.ng:|\.align\s+\d+|\.(globl|ent)\s+\S+|\#.*|\.(file|loc)\s+\S+\s+\S+|\.text|\.r?data)\n)';
114 $T_COPY_DIRVS = '^\s*(\$.*\.\.ng:|\#|\.(file|globl|ent|loc))';
116 $T_DOT_WORD = '\.(long|quad|byte|word)';
117 $T_DOT_GLOBAL = '^\t\.globl';
118 $T_HDR_literal = "\.rdata\n\t\.align 3\n";
119 $T_HDR_misc = "\.text\n\t\.align 3\n";
120 $T_HDR_data = "\.data\n\t\.align 3\n";
121 $T_HDR_rodata = "\.rdata\n\t\.align 3\n";
122 $T_HDR_closure = "\.data\n\t\.align 3\n";
123 $T_HDR_info = "\.text\n\t\.align 3\n";
124 $T_HDR_entry = "\.text\n\t\.align 3\n";
125 $T_HDR_vector = "\.text\n\t\.align 3\n";
127 #--------------------------------------------------------#
128 } elsif ( $TargetPlatform =~ /^hppa/ ) {
130 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
131 $T_US = ''; # _ if symbols have an underscore on the front
132 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
133 $T_CONST_LBL = '^L\$C(\d+)$'; # regexp for what such a lbl looks like
136 $T_MOVE_DIRVS = '^((\s+\.(IMPORT|EXPORT|PARAM).*|\s+\.align\s+\d+|\s+\.(SPACE|SUBSPA)\s+\S+|\s*)\n)';
137 $T_COPY_DIRVS = '^\s+\.(IMPORT|EXPORT)';
139 $T_DOT_WORD = '\.(blockz|word|half|byte)';
140 $T_DOT_GLOBAL = '^\s+\.EXPORT';
141 $T_HDR_literal = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
142 $T_HDR_misc = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
143 $T_HDR_data = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
144 $T_HDR_rodata = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
145 $T_HDR_closure = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
146 $T_HDR_info = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
147 $T_HDR_entry = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
148 $T_HDR_vector = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
150 #--------------------------------------------------------#
151 } elsif ( $TargetPlatform =~ /^i386-.*-(linuxaout|freebsd2|nextstep3|cygwin32|mingw32)$/ ) {
152 # NeXT added but not tested. CaS
154 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
155 $T_US = '_'; # _ if symbols have an underscore on the front
156 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
157 $T_CONST_LBL = '^LC(\d+):$';
159 $T_X86_PRE_LLBL_PAT = 'L';
160 $T_X86_PRE_LLBL = 'L';
161 $T_X86_BADJMP = '^\tjmp [^L\*]';
163 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s.*|\.globl\s+\S+|\.text|\.data|\.stab[^n].*|\.type\s+.*|\.size\s+.*|\.lcomm.*)\n)';
164 $T_COPY_DIRVS = '\.(globl|stab|lcomm)';
165 $T_DOT_WORD = '\.(long|word|value|byte|space)';
166 $T_DOT_GLOBAL = '\.globl';
167 $T_HDR_literal = "\.text\n\t\.align 2\n";
168 $T_HDR_misc = "\.text\n\t\.align 2,0x90\n";
169 $T_HDR_data = "\.data\n\t\.align 2\n";
170 $T_HDR_rodata = "\.text\n\t\.align 2\n";
171 $T_HDR_closure = "\.data\n\t\.align 2\n";
172 $T_HDR_info = "\.text\n\t\.align 2\n"; # NB: requires padding
173 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
174 $T_HDR_vector = "\.text\n\t\.align 2\n"; # NB: requires padding
176 #--------------------------------------------------------#
177 } elsif ( $TargetPlatform =~ /^i386-.*-(solaris2|linux|gnu|freebsd|netbsd|openbsd|kfreebsdgnu)$/ ) {
179 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
180 $T_US = ''; # _ if symbols have an underscore on the front
181 $T_PRE_APP = # regexp that says what comes before APP/NO_APP
182 ($TargetPlatform =~ /-(linux|gnu|freebsd|netbsd|openbsd)$/) ? '#' : '/' ;
183 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
185 $T_X86_PRE_LLBL_PAT = '\.L';
186 $T_X86_PRE_LLBL = '.L';
187 $T_X86_BADJMP = '^\tjmp\s+[^\.\*]';
189 $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)';
190 if ( $TargetPlatform =~ /solaris2/ ) {
191 # newer Solaris linkers are picky about .size information, so
192 # omit it (see #1421)
193 $T_COPY_DIRVS = '^\s*\.(globl|local)';
195 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
198 $T_DOT_WORD = '\.(long|value|word|byte|zero)';
199 $T_DOT_GLOBAL = '\.globl';
200 $T_HDR_literal = "\.section\t\.rodata\n"; # or just use .text??? (WDP 95/11)
201 $T_HDR_misc = "\.text\n\t\.align 4\n";
202 $T_HDR_data = "\.data\n\t\.align 4\n";
203 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 4\n";
204 $T_HDR_closure = "\.data\n\t\.align 4\n";
205 $T_HDR_info = "\.text\n\t\.align 4\n";
206 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
207 $T_HDR_vector = "\.text\n\t\.align 4\n"; # NB: requires padding
209 #--------------------------------------------------------#
210 } elsif ( $TargetPlatform =~ /^ia64-.*-linux$/ ) {
212 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
213 $T_US = ''; # _ if symbols have an underscore on the front
215 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
218 $T_MOVE_DIRVS = '^(\s*\.(global|proc|pred\.safe_across_calls|text|data|section|subsection|align|size|type|ident)\s+.*\n)';
219 $T_COPY_DIRVS = '\.(global|proc)';
221 $T_DOT_WORD = '\.(long|value|byte|zero)';
222 $T_DOT_GLOBAL = '\.global';
223 $T_HDR_literal = "\.section\t\.rodata\n";
224 $T_HDR_misc = "\.text\n\t\.align 16\n"; # May contain code; align like 'entry'
225 $T_HDR_data = "\.data\n\t\.align 8\n";
226 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 8\n";
227 $T_HDR_closure = "\.data\n\t\.align 8\n";
228 $T_HDR_info = "\.text\n\t\.align 8\n";
229 $T_HDR_entry = "\.text\n\t\.align 16\n";
230 $T_HDR_vector = "\.text\n\t\.align 8\n";
232 #--------------------------------------------------------#
233 } elsif ( $TargetPlatform =~ /^x86_64-.*-(linux|openbsd)$/ ) {
235 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
236 $T_US = ''; # _ if symbols have an underscore on the front
238 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
241 $T_MOVE_DIRVS = '^(\s*\.(globl|text|data|section|align|size|type|ident|local)([ \t].*)?\n)';
242 $T_COPY_DIRVS = '\.(globl|type|size|local)';
244 $T_DOT_WORD = '\.(quad|long|value|byte|zero)';
245 $T_DOT_GLOBAL = '\.global';
247 $T_HDR_literal16 = "\.section\t\.rodata.cst16\n\t.align 16\n";
248 $T_HDR_literal = "\.section\t\.rodata\n";
250 $T_HDR_misc = "\.text\n\t\.align 8\n";
251 $T_HDR_data = "\.data\n\t\.align 8\n";
252 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 8\n";
254 # the assembler on x86_64/Linux refuses to generate code for
256 # where x is in the text section and y in the rodata section.
257 # It works if y is in the text section, though. This is probably
258 # going to cause difficulties for PIC, I imagine.
260 # See Note [x86-64-relative] in includes/InfoTables.h
261 $T_HDR_relrodata= "\.text\n\t\.align 8\n";
263 $T_HDR_closure = "\.data\n\t\.align 8\n";
264 $T_HDR_info = "\.text\n\t\.align 8\n";
265 $T_HDR_entry = "\.text\n\t\.align 8\n";
266 $T_HDR_vector = "\.text\n\t\.align 8\n";
268 #--------------------------------------------------------#
269 } elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) {
271 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
272 $T_US = '_'; # _ if symbols have an underscore on the front
273 $T_PRE_APP = '^# MAY NOT APPLY'; # regexp that says what comes before APP/NO_APP
274 $T_CONST_LBL = '^LC(\d+):$';
277 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.proc\s+\d+|\.const|\.cstring|\.globl\s+\S+|\.text|\.data|\.even|\.stab[^n].*)\n)';
278 $T_COPY_DIRVS = '\.(globl|proc|stab)';
280 $T_DOT_WORD = '\.long';
281 $T_DOT_GLOBAL = '\.globl';
282 $T_HDR_literal = "\.text\n\t\.even\n";
283 $T_HDR_misc = "\.text\n\t\.even\n";
284 $T_HDR_data = "\.data\n\t\.even\n";
285 $T_HDR_rodata = "\.text\n\t\.even\n";
286 $T_HDR_closure = "\.data\n\t\.even\n";
287 $T_HDR_info = "\.text\n\t\.even\n";
288 $T_HDR_entry = "\.text\n\t\.even\n";
289 $T_HDR_vector = "\.text\n\t\.even\n";
291 #--------------------------------------------------------#
292 } elsif ( $TargetPlatform =~ /^mips-.*/ ) {
294 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
295 $T_US = ''; # _ if symbols have an underscore on the front
296 $T_PRE_APP = '^\s*#'; # regexp that says what comes before APP/NO_APP
297 $T_CONST_LBL = '^\$LC(\d+):$'; # regexp for what such a lbl looks like
300 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.(globl|ent)\s+\S+|\.text|\.r?data)\n)';
301 $T_COPY_DIRVS = '\.(globl|ent)';
303 $T_DOT_WORD = '\.word';
304 $T_DOT_GLOBAL = '^\t\.globl';
305 $T_HDR_literal = "\t\.rdata\n\t\.align 2\n";
306 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
307 $T_HDR_data = "\t\.data\n\t\.align 2\n";
308 $T_HDR_rodata = "\t\.rdata\n\t\.align 2\n";
309 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
310 $T_HDR_info = "\t\.text\n\t\.align 2\n";
311 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
312 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
314 #--------------------------------------------------------#
315 } elsif ( $TargetPlatform =~ /^powerpc-apple-darwin.*/ ) {
316 # Apple PowerPC Darwin/MacOS X.
317 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
318 $T_US = '_'; # _ if symbols have an underscore on the front
319 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
320 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
323 $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)';
324 $T_COPY_DIRVS = '\.(globl|lcomm)';
326 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
327 $T_DOT_GLOBAL = '\.globl';
328 $T_HDR_toc = "\.toc\n";
329 $T_HDR_literal = "\t\.const\n\t\.align 2\n";
330 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
331 $T_HDR_data = "\t\.data\n\t\.align 2\n";
332 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
333 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
334 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
335 $T_HDR_info = "\t\.text\n\t\.align 2\n";
336 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
337 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
339 #--------------------------------------------------------#
340 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ ) {
341 # Apple PowerPC Darwin/MacOS X.
342 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
343 $T_US = '_'; # _ if symbols have an underscore on the front
344 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
345 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
347 $T_X86_PRE_LLBL_PAT = 'L';
348 $T_X86_PRE_LLBL = 'L';
349 $T_X86_BADJMP = '^\tjmp [^L\*]';
351 $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)';
352 $T_COPY_DIRVS = '\.(globl|lcomm)';
354 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
355 $T_DOT_GLOBAL = '\.globl';
356 $T_HDR_toc = "\.toc\n";
357 $T_HDR_literal16= "\t\.literal8\n\t\.align 4\n";
358 $T_HDR_literal = "\t\.const\n\t\.align 4\n";
359 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
360 $T_HDR_data = "\t\.data\n\t\.align 2\n";
361 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
362 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
363 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
364 $T_HDR_info = "\t\.text\n\t\.align 2\n";
365 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
366 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
368 #--------------------------------------------------------#
369 } elsif ( $TargetPlatform =~ /^x86_64-apple-darwin.*/ ) {
370 # Apple PowerPC Darwin/MacOS X.
371 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
372 $T_US = '_'; # _ if symbols have an underscore on the front
373 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
374 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
377 $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)';
378 $T_COPY_DIRVS = '\.(globl|lcomm)';
380 $T_DOT_WORD = '\.(quad|long|short|byte|fill|space)';
381 $T_DOT_GLOBAL = '\.globl';
382 $T_HDR_toc = "\.toc\n";
383 $T_HDR_literal16= "\t\.literal8\n\t\.align 4\n";
384 $T_HDR_literal = "\t\.const\n\t\.align 4\n";
385 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
386 $T_HDR_data = "\t\.data\n\t\.align 2\n";
387 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
388 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
389 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
390 $T_HDR_info = "\t\.text\n\t\.align 2\n";
391 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
392 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
394 #--------------------------------------------------------#
395 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ ) {
397 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
398 $T_US = ''; # _ if symbols have an underscore on the front
399 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
400 $T_CONST_LBL = '^\.LC\d+:'; # regexp for what such a lbl looks like
403 $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)';
404 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
406 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
407 $T_DOT_GLOBAL = '\.globl';
408 $T_HDR_toc = "\.toc\n";
409 $T_HDR_literal = "\t\.section\t.rodata\n\t\.align 2\n";
410 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
411 $T_HDR_data = "\t\.data\n\t\.align 2\n";
412 $T_HDR_rodata = "\t\.section\t.rodata\n\t\.align 2\n";
413 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
414 $T_HDR_info = "\t\.text\n\t\.align 2\n";
415 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
416 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
418 #--------------------------------------------------------#
419 } elsif ( $TargetPlatform =~ /^powerpc64-.*-linux/ ) {
421 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
422 $T_US = '\.'; # _ if symbols have an underscore on the front
423 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
424 $T_CONST_LBL = '^\.LC\d+:'; # regexp for what such a lbl looks like
427 $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)';
428 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
430 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
431 $T_DOT_GLOBAL = '\.globl';
432 $T_HDR_toc = "\.toc\n";
433 $T_HDR_literal = "\t\.section\t\".toc\",\"aw\"\n";
434 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
435 $T_HDR_data = "\t\.data\n\t\.align 2\n";
436 $T_HDR_rodata = "\t\.section\t.rodata\n\t\.align 2\n";
437 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
438 $T_HDR_info = "\t\.text\n\t\.align 2\n";
439 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
440 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
442 #--------------------------------------------------------#
443 } elsif ( $TargetPlatform =~ /^sparc-.*-(solaris2|openbsd)/ ) {
445 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
446 $T_US = ''; # _ if symbols have an underscore on the front
447 $T_PRE_APP = 'DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
448 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
451 $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)';
452 $T_COPY_DIRVS = '\.(global|local|proc|stab)';
454 $T_DOT_WORD = '\.(long|word|byte|half|skip|uahalf|uaword)';
455 $T_DOT_GLOBAL = '^\t\.global';
456 $T_HDR_literal = "\.text\n\t\.align 8\n";
457 $T_HDR_misc = "\.text\n\t\.align 4\n";
458 $T_HDR_data = "\.data\n\t\.align 8\n";
459 $T_HDR_rodata = "\.text\n\t\.align 4\n";
460 $T_HDR_closure = "\.data\n\t\.align 4\n";
461 $T_HDR_info = "\.text\n\t\.align 4\n";
462 $T_HDR_entry = "\.text\n\t\.align 4\n";
463 $T_HDR_vector = "\.text\n\t\.align 4\n";
465 #--------------------------------------------------------#
466 } elsif ( $TargetPlatform =~ /^sparc-.*-sunos4/ ) {
468 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
469 $T_US = '_'; # _ if symbols have an underscore on the front
470 $T_PRE_APP = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
471 $T_CONST_LBL = '^LC(\d+):$';
474 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*)\n)';
475 $T_COPY_DIRVS = '\.(global|proc|stab)';
477 $T_DOT_WORD = '\.word';
478 $T_DOT_GLOBAL = '^\t\.global';
479 $T_HDR_literal = "\.text\n\t\.align 8\n";
480 $T_HDR_misc = "\.text\n\t\.align 4\n";
481 $T_HDR_data = "\.data\n\t\.align 8\n";
482 $T_HDR_rodata = "\.text\n\t\.align 4\n";
483 $T_HDR_closure = "\.data\n\t\.align 4\n";
484 $T_HDR_info = "\.text\n\t\.align 4\n";
485 $T_HDR_entry = "\.text\n\t\.align 4\n";
486 $T_HDR_vector = "\.text\n\t\.align 4\n";
488 #--------------------------------------------------------#
489 } elsif ( $TargetPlatform =~ /^sparc-.*-linux/ ) {
490 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
491 $T_US = ''; # _ if symbols have an underscore on the front
492 $T_PRE_APP = '#'; # regexp that says what comes before APP/NO_APP
493 # Probably doesn't apply anyway
494 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
497 $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)';
498 $T_COPY_DIRVS = '\.(global|local|globl|proc|stab)';
500 $T_DOT_WORD = '\.(long|word|nword|xword|byte|half|short|skip|uahalf|uaword)';
501 $T_DOT_GLOBAL = '^\t\.global';
502 $T_HDR_literal = "\.text\n\t\.align 8\n";
503 $T_HDR_misc = "\.text\n\t\.align 4\n";
504 $T_HDR_data = "\.data\n\t\.align 8\n";
505 $T_HDR_rodata = "\.text\n\t\.align 4\n";
506 $T_HDR_closure = "\.data\n\t\.align 4\n";
507 $T_HDR_info = "\.text\n\t\.align 4\n";
508 $T_HDR_entry = "\.text\n\t\.align 4\n";
509 $T_HDR_vector = "\.text\n\t\.align 4\n";
511 #--------------------------------------------------------#
513 print STDERR "$Pgm: don't know how to mangle assembly language for: $TargetPlatform\n";
517 if($T_HDR_relrodata eq "") {
519 # relrodata defaults to rodata.
520 $T_HDR_relrodata = $T_HDR_rodata;
524 print STDERR "T_STABBY: $T_STABBY\n";
525 print STDERR "T_US: $T_US\n";
526 print STDERR "T_PRE_APP: $T_PRE_APP\n";
527 print STDERR "T_CONST_LBL: $T_CONST_LBL\n";
528 print STDERR "T_POST_LBL: $T_POST_LBL\n";
529 if ( $TargetPlatform =~ /^i386-/ ) {
530 print STDERR "T_X86_PRE_LLBL_PAT: $T_X86_PRE_LLBL_PAT\n";
531 print STDERR "T_X86_PRE_LLBL: $T_X86_PRE_LLBL\n";
532 print STDERR "T_X86_BADJMP: $T_X86_BADJMP\n";
534 print STDERR "T_MOVE_DIRVS: $T_MOVE_DIRVS\n";
535 print STDERR "T_COPY_DIRVS: $T_COPY_DIRVS\n";
536 print STDERR "T_DOT_WORD: $T_DOT_WORD\n";
537 print STDERR "T_HDR_literal: $T_HDR_literal\n";
538 print STDERR "T_HDR_misc: $T_HDR_misc\n";
539 print STDERR "T_HDR_data: $T_HDR_data\n";
540 print STDERR "T_HDR_rodata: $T_HDR_rodata\n";
541 print STDERR "T_HDR_closure: $T_HDR_closure\n";
542 print STDERR "T_HDR_info: $T_HDR_info\n";
543 print STDERR "T_HDR_entry: $T_HDR_entry\n";
544 print STDERR "T_HDR_vector: $T_HDR_vector\n";
550 %************************************************************************
552 \subsection{Mangle away}
554 %************************************************************************
558 local($in_asmf, $out_asmf) = @_;
562 # ia64-specific information for code chunks
566 &init_TARGET_STUFF();
567 &init_FUNNY_THINGS();
569 open(INASM, "< $in_asmf")
570 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
571 open(OUTASM,"> $out_asmf")
572 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
574 # read whole file, divide into "chunks":
575 # record some info about what we've found...
577 @chk = (); # contents of the chunk
578 $numchks = 0; # number of them
579 @chkcat = (); # what category of thing in each chunk
580 @chksymb = (); # what symbol(base) is defined in this chunk
581 %entrychk = (); # ditto, its entry code
582 %closurechk = (); # ditto, the (static) closure
583 %srtchk = (); # ditto, its SRT (for top-level things)
584 %infochk = (); # given a symbol base, say what chunk its info tbl is in
585 %vectorchk = (); # ditto, return vector table
586 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
588 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
591 tr/\r//d if $TargetPlatform =~ /-mingw32$/; # In case Perl doesn't convert line endings
592 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
593 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
594 next if /^\t\.def.*endef$/;
595 next if /${T_PRE_APP}(NO_)?APP/o;
596 next if /^;/ && $TargetPlatform =~ /^hppa/;
598 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|ia64)-/;
600 if ( $TargetPlatform =~ /^mips-/
601 && /^\t\.(globl\S+\.text|comm\t)/ ) {
602 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
603 # Treat .comm variables as data. These show up in two (known) places:
605 # - the module_registered variable used in the __stginit fragment.
606 # even though these are declared static and initialised, gcc 3.3
607 # likes to make them .comm, presumably to save space in the
610 # - global variables used to pass arguments from C to STG in
611 # a foreign export. (is this still true? --SDM)
613 } elsif ( /^\t\.comm.*$/ ) {
615 $chkcat[$i] = 'data';
618 # Labels ending "_str": these are literal strings.
619 } elsif ( /^${T_US}([A-Za-z0-9_]+)_str${T_POST_LBL}$/ ) {
621 $chkcat[$i] = 'relrodata';
623 } elsif ( $TargetPlatform =~ /-darwin/
624 && (/^\s*\.subsections_via_symbols/
625 ||/^\s*\.no_dead_strip.*/)) {
626 # Don't allow Apple's linker to do any dead-stripping of symbols
627 # in this file, because it will mess up info-tables in mangled
629 # The .no_dead_strip directives are actually put there by
630 # the gcc3 "used" attribute on entry points.
632 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && (
633 /^\s*\.picsymbol_stub/
634 || /^\s*\.section __TEXT,__picsymbol_stub\d,.*/
635 || /^\s*\.section __TEXT,__picsymbolstub\d,.*/
636 || /^\s*\.symbol_stub/
637 || /^\s*\.section __TEXT,__symbol_stub\d,.*/
638 || /^\s*\.section __TEXT,__symbolstub\d,.*/
639 || /^\s*\.lazy_symbol_pointer/
640 || /^\s*\.non_lazy_symbol_pointer/
641 || /^\s*\.section __IMPORT.*/))
644 $chkcat[$i] = 'dyld';
648 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.data/)
649 { # non_lazy_symbol_ptrs that point to local symbols
651 $chkcat[$i] = 'dyld';
654 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.align/)
655 { # non_lazy_symbol_ptrs that point to local symbols
657 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^L_.*:$/)
658 { # non_lazy_symbol_ptrs that point to local symbols
659 $chk[++$i] = $dyld_section . $_;
660 $chkcat[$i] = 'dyld';
663 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
664 # duplicated from the bottom
668 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
669 # Alphas: Local labels not to be confused with new chunks
671 # NB: all the rest start with a non-space
673 } elsif ( $TargetPlatform =~ /^mips-/
674 && /^\d+:/ ) { # a funny-looking very-local label
677 } elsif ( /$T_CONST_LBL/o ) {
679 $chkcat[$i] = 'literal';
682 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
684 $chkcat[$i] = 'splitmarker';
687 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
690 $chkcat[$i] = 'infotbl';
691 $chksymb[$i] = $symb;
693 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
695 $infochk{$symb} = $i;
697 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
699 $chkcat[$i] = 'entry';
704 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
706 $chkcat[$i] = 'closure';
709 $closurechk{$1} = $i;
711 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
718 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
720 $chkcat[$i] = 'data';
723 } elsif ( /^${T_US}(stg_ap_stack_entries|stg_stack_save_entries|stg_arg_bitmaps)${T_POST_LBL}$/o ) {
725 $chkcat[$i] = 'data';
728 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
731 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
732 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
733 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
734 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
737 $chkcat[$i] = 'data';
740 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
745 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
746 # all CC_ symbols go in the data section...
748 $chkcat[$i] = 'data';
751 } elsif ( /^${T_US}([A-Za-z0-9_]+)_hpc${T_POST_LBL}$/o ) {
752 # hpc shares tick boxes across modules
754 $chkcat[$i] = 'data';
757 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
759 $chkcat[$i] = 'misc';
761 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
763 $chkcat[$i] = 'vector';
768 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
769 && /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
770 # Some Solaris system headers contain function definitions (as
771 # opposed to mere prototypes), which end up in the .hc file when
772 # a Haskell module foreign imports the corresponding system
773 # functions (most notably stat()). We put them into the text
774 # segment. Note that this currently does not extend to function
775 # names starting with an underscore.
778 $chkcat[$i] = 'misc';
781 } elsif ( $TargetPlatform =~ /^i386-apple-darwin/ && /^(___i686\.get_pc_thunk\.[abcd]x):/o) {
782 # To handle PIC on Darwin/x86, we need to appropriately pass through
783 # the get_pc_thunk functions. The need to be put into a special section
784 # marked as coalesced (otherwise the .weak_definition doesn't work
787 $chkcat[$i] = 'get_pc_thunk';
790 } elsif ( /^${T_US}[A-Za-z0-9_]/o
791 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
793 && ( $TargetPlatform !~ /^powerpc64/ # we need to avoid local labels in this case
794 || ! /^\.L\d+:$/ ) ) {
801 /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
802 || /^${T_US}__stg_.*${T_POST_LBL}$/o # more RTS internals
803 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
804 || /^${T_US}.*_slow${T_POST_LBL}$/o # slow entry
805 || /^${T_US}__stginit.*${T_POST_LBL}$/o # __stginit<module>
806 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
807 || /^${T_US}.*_fast${T_POST_LBL}$/o # primops
808 || /^_uname:/o # x86/Solaris2
811 $chkcat[$i] = 'misc';
813 /^${T_US}.*_srtd${T_POST_LBL}$/o # large bitmaps
814 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
817 $chkcat[$i] = 'relrodata';
820 print STDERR "Warning: retaining unknown function \`$thing' in output from C compiler\n";
821 $chkcat[$i] = 'unknown';
824 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ && /^\.LCTOC1 = /o ) {
825 # PowerPC Linux's large-model PIC (-fPIC) generates a gobal offset
826 # table "by hand". Be sure to copy it over.
827 # Note that this label and all entries in the table should actually
828 # go into the .got2 section, but it isn't easy to distinguish them
829 # from other constant literals (.LC\d+), so we just put everything
832 $chkcat[$i] = 'literal';
833 $chksymb[$i] = 'LCTOC1';
834 } else { # simple line (duplicated at the top)
839 $numchks = $#chk + 1;
840 $chk[$numchks] = ''; # We might push .note.GNU-stack into this
841 $chkcat[$numchks] = 'verbatim'; # If we do, write it straight back out
843 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
844 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
847 # the division into chunks is imperfect;
848 # we throw some things over the fence into the next
851 # also, there are things we would like to know
852 # about the whole module before we start spitting
855 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
856 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
858 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
860 # Alphas: NB: we start meddling at chunk 1, not chunk 0
861 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
862 # spits a ".quad 0" in after the very first ".rdata"; we
863 # detect this special case (tossing the ".quad 0")!
864 local($magic_rdata_seen) = 0;
866 # HPPAs, MIPSen: also start medding at chunk 1
868 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
869 $c = $chk[$i]; # convenience copy
871 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
873 # toss all prologue stuff; HPPA is pretty weird
875 $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
880 # be slightly paranoid to make sure there's
881 # nothing surprising in there
882 if ( $c =~ /--- BEGIN ---/ ) {
883 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
885 # remove junk whitespace around the split point
889 if ($TargetPlatform =~ /^i386-/) {
890 if ($p =~ /^\tsubl\s+\$(\d+),\s*\%esp\n/) {
892 die "Error: reserved stack space exceeded!\n Possible workarounds: compile with -fasm, or try another version of gcc.\n"
896 # gcc 3.4.3 puts this kind of stuff in the prologue, eg.
897 # when compiling PrimOps.cmm with -optc-O2:
900 # movl %ecx, 16(%esp)
901 # movl %edx, 20(%esp)
902 # but then the code of the function doesn't assume
903 # anything about the contnets of these stack locations.
904 # I think it's to do with the use of inline functions for
905 # PK_Word64() and friends, where gcc is initialising the
906 # contents of the struct to zero, and failing to optimise
907 # away the initialisation. Let's live dangerously and
908 # discard these initalisations.
910 $p =~ s/^\tpushl\s+\%e(di|si|bx)\n//g;
911 $p =~ s/^\txorl\s+\%e(ax|cx|dx),\s*\%e(ax|cx|dx)\n//g;
912 $p =~ s/^\tmovl\s+\%e(ax|cx|dx|si|di),\s*\d*\(\%esp\)\n//g;
913 $p =~ s/^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n//g;
914 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
915 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-(cygwin32|mingw32)/);
917 if ($TargetPlatform =~ /^i386-apple-darwin/) {
919 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
921 $pcrel_reg =~ s/(.|\n)*.*___i686\.get_pc_thunk\.([abcd]x)\n(.|\n)*/$2/ or $pcrel_reg = "";
922 $p =~ s/^\s+call\s+___i686\.get_pc_thunk\..x//;
923 $p =~ s/^\"?L\d+\$pb\"?:\n//;
925 if ($pcrel_reg eq "bx") {
926 # Bad gcc. Goes and uses %ebx, our BaseReg, for PIC. Bad gcc.
927 die "Darwin/x86: -fPIC -via-C doesn't work yet, use -fasm. Aborting."
931 } elsif ($TargetPlatform =~ /^x86_64-/) {
932 $p =~ s/^\tpushq\s+\%r(bx|bp|12|13|14)\n//g;
933 $p =~ s/^\tmovq\s+\%r(bx|bp|12|13|14),\s*\d*\(\%rsp\)\n//g;
934 $p =~ s/^\tsubq\s+\$\d+,\s*\%rsp\n//;
936 } elsif ($TargetPlatform =~ /^ia64-/) {
937 $p =~ s/^\t\.prologue .*\n//;
939 # Record the number of local and out registers for register relocation later
940 $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, (\d+), (\d+), 0\n//;
944 $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
945 $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
947 # Ignore save/restore of these registers; they're taken
948 # care of in StgRun()
949 $p =~ s/^\t\.save ar\.lc, r\d+\n//;
950 $p =~ s/^\t\.save pr, r\d+\n//;
951 $p =~ s/^\tmov r\d+ = ar\.lc\n//;
952 $p =~ s/^\tmov r\d+ = pr\n//;
954 # Remove .proc and .body directives
955 $p =~ s/^\t\.proc [a-zA-Z0-9_.]+#\n//;
956 $p =~ s/^\t\.body\n//;
958 # If there's a label, move it to the body
959 if ($p =~ /^[a-zA-Z0-9.]+:\n/) {
964 # Remove floating-point spill instructions.
965 # Only fp registers 2-5 and 16-23 are saved by the runtime.
966 if ($p =~ s/^\tstf\.spill \[r1[4-9]\] = f([2-5]|1[6-9]|2[0-3])(, [0-9]+)?\n//g) {
967 # Being paranoid, only try to remove these if we saw a
969 $p =~ s/^\tmov r1[4-9] = r12\n//;
970 $p =~ s/^\tadds r1[4-9] = -[0-9]+, r12\n//g;
971 $p =~ s/^\t\.save\.f 0x[0-9a-fA-F]\n//g;
972 $p =~ s/^\t\.save\.gf 0x0, 0x[0-9a-fA-F]+\n//g;
975 $p =~ s/^\tnop(?:\.[mifb])?\s+\d+\n//g; # remove nop instructions
976 $p =~ s/^\t\.(mii|mmi|mfi)\n//g; # bundling is no longer sensible
977 $p =~ s/^\t;;\n//g; # discard stops
978 $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
980 # GCC 3.3 saves r1 in the prologue, move this to the body
981 # (Does this register get restored anywhere?)
982 if ($p =~ /^\tmov r\d+ = r1\n/) {
986 } elsif ($TargetPlatform =~ /^m68k-/) {
987 $p =~ s/^\tlink a6,#-?\d.*\n//;
988 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
989 # The above showed up in the asm code,
990 # so I added it here.
991 # I hope it's correct.
993 $p =~ s/^\tmovel d2,sp\@-\n//;
994 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
995 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
996 } elsif ($TargetPlatform =~ /^mips-/) {
997 # the .frame/.mask/.fmask that we use is the same
998 # as that produced by GCC for miniInterpret; this
999 # gives GDB some chance of figuring out what happened
1000 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
1001 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
1002 $p =~ s/^\t\.(mask|fmask).*\n//g;
1003 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
1004 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
1005 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
1006 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
1007 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
1008 $p =~ s/__FRAME__/$FRAME/;
1009 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
1011 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
1013 $p =~ s/^\tmflr r0\n//;
1014 $p =~ s/^\tbl saveFP # f\d+\n//;
1015 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
1016 $p =~ s/^\"?L\d+\$pb\"?:\n//;
1017 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
1018 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
1019 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
1020 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
1021 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
1022 $p =~ s/^\tbcl 20,31,\"?L\d+\$pb\"?\n//;
1023 $p =~ s/^\"?L\d+\$pb\"?:\n//;
1024 $p =~ s/^\tmflr r31\n//;
1026 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1027 # under some circumstances, only when generating position dependent code.
1028 # I have no idea why, and I don't think it is necessary, so let's toss it.
1029 $p =~ s/^\tli r\d+,0\n//g;
1030 $p =~ s/^\tstw r\d+,\d+\(r1\)\n//g;
1031 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/) {
1032 $p =~ s/^\tmflr 0\n//;
1033 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
1034 $p =~ s/^\tstfd \d+,\d+\(1\)\n//g;
1035 $p =~ s/^\tstw r0,8\(1\)\n//;
1036 $p =~ s/^\tstwu 1,-\d+\(1\)\n//;
1037 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1039 # GCC's "large-model" PIC (-fPIC)
1041 $pcrel_label =~ s/(.|\n)*^.LCF(\d+):\n(.|\n)*/$2/ or $pcrel_label = "";
1043 $p =~ s/^\tbcl 20,31,.LCF\d+\n//;
1044 $p =~ s/^.LCF\d+:\n//;
1045 $p =~ s/^\tmflr 30\n//;
1046 $p =~ s/^\tlwz 0,\.LCL\d+-\.LCF\d+\(30\)\n//;
1047 $p =~ s/^\tadd 30,0,30\n//;
1049 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1050 # under some circumstances, only when generating position dependent code.
1051 # I have no idea why, and I don't think it is necessary, so let's toss it.
1052 $p =~ s/^\tli \d+,0\n//g;
1053 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1054 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1055 $p =~ s/^\tmr 31,1\n//;
1056 $p =~ s/^\tmflr 0\n//;
1057 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
1058 $p =~ s/^\tstfd \d+,-?\d+\(1\)\n//g;
1059 $p =~ s/^\tstd r0,8\(1\)\n//;
1060 $p =~ s/^\tstdu 1,-\d+\(1\)\n//;
1061 $p =~ s/^\tstd \d+,-?\d+\(1\)\n//g;
1063 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1064 # under some circumstances, only when generating position dependent code.
1065 # I have no idea why, and I don't think it is necessary, so let's toss it.
1066 $p =~ s/^\tli \d+,0\n//g;
1067 $p =~ s/^\tstd \d+,\d+\(1\)\n//g;
1069 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
1072 # HWL HACK: dont die, just print a warning
1073 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
1074 die "Prologue junk?: $p\n" if $p =~ /^\s+[^\s\.]/;
1076 # For PIC, we want to keep part of the prologue
1077 if ($TargetPlatform =~ /^powerpc-apple-darwin.*/ && $pcrel_label ne "") {
1078 # Darwin: load the current instruction pointer into register r31
1079 $p .= "bcl 20,31,$pcrel_label\n";
1080 $p .= "$pcrel_label:\n";
1081 $p .= "\tmflr r31\n";
1082 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/ && $pcrel_label ne "") {
1083 # Linux: load the GOT pointer into register 30
1084 $p .= "\tbcl 20,31,.LCF$pcrel_label\n";
1085 $p .= ".LCF$pcrel_label:\n";
1086 $p .= "\tmflr 30\n";
1087 $p .= "\tlwz 0,.LCL$pcrel_label-.LCF$pcrel_label(30)\n";
1088 $p .= "\tadd 30,0,30\n";
1089 } elsif ($TargetPlatform =~ /^i386-apple-darwin.*/ && $pcrel_label ne "") {
1090 $p .= "\tcall ___i686.get_pc_thunk.$pcrel_reg\n";
1091 $p .= "$pcrel_label:\n";
1094 # glue together what's left
1099 if ( $TargetPlatform =~ /^mips-/ ) {
1100 # MIPS: first, this basic sequence may occur "--- END ---" or not
1101 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
1104 # toss all epilogue stuff; again, paranoidly
1105 if ( $c =~ /--- END ---/ ) {
1106 # Gcc may decide to replicate the function epilogue. We want
1107 # to process all epilogues, so we split the function and then
1109 @fragments = split(/--- END ---/, $c);
1110 $r = shift(@fragments);
1112 # Rebuild `c'; processed fragments will be appended to `c'
1115 foreach $e (@fragments) {
1116 # etail holds code that is after the epilogue in the assembly-code
1117 # layout and should not be filtered as part of the epilogue.
1119 if ($TargetPlatform =~ /^i386-/) {
1121 $e =~ s/^\tpopl\s+\%edi\n//;
1122 $e =~ s/^\tpopl\s+\%esi\n//;
1123 $e =~ s/^\tpopl\s+\%edx\n//;
1124 $e =~ s/^\tpopl\s+\%ecx\n//;
1125 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
1126 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
1127 } elsif ($TargetPlatform =~ /^ia64-/) {
1128 # The epilogue is first split into:
1129 # $e, the epilogue code (up to the return instruction)
1130 # $etail, non-epilogue code (after the return instruction)
1131 # The return instruction is stripped in the process.
1132 if (!(($e, $etail) = split(/^\tbr\.ret\.sptk\.many b0\n/, $e))) {
1133 die "Epilogue doesn't seem to have one return instruction: $e\n";
1135 # Remove 'endp' directive from the tail
1136 $etail =~ s/^\t\.endp [a-zA-Z0-9_.]+#\n//;
1138 # If a return value is saved here, discard it
1139 $e =~ s/^\tmov r8 = r14\n//;
1141 # Remove floating-point fill instructions.
1142 # Only fp registers 2-5 and 16-23 are saved by the runtime.
1143 if ($e =~ s/^\tldf\.fill f([2-5]|1[6-9]|2[0-3]) = \[r1[4-9]\](, [0-9]+)?\n//g) {
1144 # Being paranoid, only try to remove this if we saw a fill
1146 $e =~ s/^\tadds r1[4-9] = [0-9]+, r12//g;
1149 $e =~ s/^\tnop(?:\.[mifb])?\s+\d+\n//g; # remove nop instructions
1150 $e =~ s/^\tmov ar\.pfs = r\d+\n//;
1151 $e =~ s/^\tmov ar\.lc = r\d+\n//;
1152 $e =~ s/^\tmov pr = r\d+, -1\n//;
1153 $e =~ s/^\tmov b0 = r\d+\n//;
1154 $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
1155 #$e =~ s/^\tbr\.ret\.sptk\.many b0\n//; # already removed
1156 $e =~ s/^\t\.(mii|mmi|mfi|mib)\n//g; # bundling is no longer sensible
1157 $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
1158 $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
1159 } elsif ($TargetPlatform =~ /^m68k-/) {
1160 $e =~ s/^\tunlk a6\n//;
1162 } elsif ($TargetPlatform =~ /^mips-/) {
1163 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
1164 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
1165 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
1166 $e =~ s/^\tj\t\$31\n//;
1167 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
1168 $e =~ s/^\taddi r1,r1,\d+\n//;
1169 $e =~ s/^\tlwz r\d+,\d+\(r1\)\n//;
1170 $e =~ s/^\tlmw r\d+,-\d+\(r1\)\n//;
1171 $e =~ s/^\tmtlr r0\n//;
1173 $e =~ s/^\tb restFP ;.*\n//;
1174 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1175 $e =~ s/^\tmr 3,0\n//;
1176 $e =~ s/^\taddi 1,1,\d+\n//;
1177 $e =~ s/^\tld 0,16\(1\)\n//;
1178 $e =~ s/^\tmtlr 0\n//;
1180 # callee-save registers
1181 $e =~ s/^\tld \d+,-?\d+\(1\)\n//g;
1182 $e =~ s/^\tlfd \d+,-?\d+\(1\)\n//g;
1184 # get rid of the debug junk along with the blr
1185 $e =~ s/^\tblr\n\t.long .*\n\t.byte .*\n//;
1187 # incase we missed it with the last one get the blr alone
1190 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
1193 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
1195 # glue together what's left
1198 $c =~ s/\n\t\n/\n/; # junk blank line
1201 if ($TargetPlatform =~ /^ia64-/) {
1202 # On IA64, remove an .endp directive even if no epilogue was found.
1203 # Code optimizations may have removed the "--- END ---" token.
1204 $c =~ s/^\t\.endp [a-zA-Z0-9_.]+#\n//;
1208 # On SPARCs, we don't do --- BEGIN/END ---, we just
1209 # toss the register-windowing save/restore/ret* instructions
1210 # directly unless they've been generated by function definitions in header
1212 if ( $TargetPlatform =~ /^sparc-/ ) {
1213 if ( ! ( $TargetPlatform =~ /solaris2$/ && $chkcat[$i] eq 'unknown' )) {
1214 $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
1216 # throw away PROLOGUE comments
1217 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
1220 # On Alphas, the prologue mangling is done a little later (below)
1222 # toss all calls to __DISCARD__
1223 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
1224 $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
1225 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-darwin.*/;
1226 $c =~ s/^\tbl\s+__DISCARD__(\@plt)?\n//go if $TargetPlatform =~ /^powerpc-.*-linux/;
1227 $c =~ s/^\tbl\s+\.__DISCARD__\n\s+nop\n//go if $TargetPlatform =~ /^powerpc64-.*-linux/;
1228 $c =~ s/^\tcall\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /i386-apple-darwin.*/;
1230 # IA64: fix register allocation; mangle tailcalls into jumps
1231 if ($TargetPlatform =~ /^ia64-/) {
1232 ia64_rename_registers($ia64_locnum, $ia64_outnum) if (defined($ia64_locnum));
1233 ia64_mangle_tailcalls();
1236 # MIPS: that may leave some gratuitous asm macros around
1237 # (no harm done; but we get rid of them to be tidier)
1238 $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/
1239 if $TargetPlatform =~ /^mips-/;
1241 # toss stack adjustment after DoSparks
1242 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
1243 if $TargetPlatform =~ /^m68k-/; # this looks old...
1245 if ( $TargetPlatform =~ /^alpha-/ &&
1246 ! $magic_rdata_seen &&
1247 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
1248 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
1249 $magic_rdata_seen = 1;
1252 # pick some end-things and move them to the next chunk
1254 # pin a funny end-thing on (for easier matching):
1255 $c .= 'FUNNY#END#THING';
1257 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
1261 # on x86 we try not to copy any directives into a literal
1262 # chunk, rather we keep looking for the next real chunk. This
1263 # is because we get things like
1265 # .globl blah_closure
1271 if ( $TargetPlatform =~ /^(i386|sparc|powerpc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
1273 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
1276 if ( $j < $numchks ) {
1277 $chk[$j] = $to_move . $chk[$j];
1281 elsif ( ( $i < ($numchks - 1)
1282 && ( $to_move =~ /${T_COPY_DIRVS}/
1283 || ( $TargetPlatform =~ /^hppa/
1284 && $to_move =~ /align/
1285 && $chkcat[$i+1] eq 'literal')
1288 || ($to_move =~ /^[ \t]*\.section[ \t]+\.note\.GNU-stack,/)
1290 $chk[$i + 1] = $to_move . $chk[$i + 1];
1291 # otherwise they're tossed
1294 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
1297 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
1299 # toss all prologue stuff, except for loading gp, and the ..ng address
1300 unless ($c =~ /\.ent.*\n\$.*\.\.ng:/) {
1301 if (($p, $r) = split(/^\t\.prologue/, $c)) {
1302 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
1303 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
1304 $keep =~ s/^\t\.(mask|fmask).*\n//g;
1305 $c = $keep . "..ng:\n";
1307 print STDERR "malformed code block ($ent)?\n"
1310 $c .= "\t.prologue" . $r;
1314 $c =~ s/FUNNY#END#THING//;
1316 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
1318 $chk[$i] = $c; # update w/ convenience copy
1321 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
1322 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
1325 if ( $TargetPlatform =~ /^alpha-/ ) {
1326 # print out the header stuff first
1327 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
1328 print OUTASM $chk[0];
1330 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1331 print OUTASM $chk[0];
1333 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1334 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
1336 # get rid of horrible "<dollar>Revision: .*$" strings
1337 local(@lines0) = split(/\n/, $chk[0]);
1339 while ( $z <= $#lines0 ) {
1340 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
1343 while ( $z <= $#lines0 ) {
1345 last if $lines0[$z] =~ /[,\t]0x0$/;
1351 $chk[0] = join("\n", @lines0);
1352 $chk[0] =~ s/\n\n+/\n/;
1353 print OUTASM $chk[0];
1356 # print out all the literal strings next
1357 for ($i = 0; $i < $numchks; $i++) {
1358 if ( $chkcat[$i] eq 'literal' ) {
1360 # HACK: try to detect 16-byte constants and align them
1361 # on a 16-byte boundary. x86_64 sometimes needs 128-bit
1362 # aligned constants, and so does Darwin/x86.
1363 if ( $TargetPlatform =~ /^x86_64/
1364 || $TargetPlatform =~ /^i386-apple-darwin/ ) {
1366 if ($z =~ /(\.long.*\n.*\.long.*\n.*\.long.*\n.*\.long|\.quad.*\n.*\.quad)/) {
1367 print OUTASM $T_HDR_literal16;
1369 print OUTASM $T_HDR_literal;
1372 print OUTASM $T_HDR_literal;
1375 print OUTASM $chk[$i];
1376 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
1378 $chkcat[$i] = 'DONE ALREADY';
1382 # on the HPPA, print out all the bss next
1383 if ( $TargetPlatform =~ /^hppa/ ) {
1384 for ($i = 1; $i < $numchks; $i++) {
1385 if ( $chkcat[$i] eq 'bss' ) {
1386 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
1387 print OUTASM $chk[$i];
1389 $chkcat[$i] = 'DONE ALREADY';
1394 # $numchks + 1 as we have the extra one for .note.GNU-stack
1395 for ($i = $FIRST_MANGLABLE; $i < $numchks + 1; $i++) {
1396 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
1398 next if $chkcat[$i] eq 'DONE ALREADY';
1400 if ( $chkcat[$i] eq 'misc' || $chkcat[$i] eq 'unknown' ) {
1401 if ($chk[$i] ne '') {
1402 print OUTASM $T_HDR_misc;
1403 &print_doctored($chk[$i], 0);
1406 } elsif ( $chkcat[$i] eq 'verbatim' ) {
1407 print OUTASM $chk[$i];
1409 } elsif ( $chkcat[$i] eq 'toss' ) {
1410 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
1412 } elsif ( $chkcat[$i] eq 'data' ) {
1413 if ($chk[$i] ne '') {
1414 print OUTASM $T_HDR_data;
1415 print OUTASM $chk[$i];
1418 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1419 # we can just re-constitute this one...
1420 # NB: we emit _three_ underscores no matter what,
1421 # so ghc-split doesn't have to care.
1422 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1424 } elsif ( $chkcat[$i] eq 'closure'
1425 || $chkcat[$i] eq 'srt'
1426 || $chkcat[$i] eq 'infotbl'
1427 || $chkcat[$i] eq 'entry') { # do them in that order
1428 $symb = $chksymb[$i];
1431 if ( defined($closurechk{$symb}) ) {
1432 print OUTASM $T_HDR_closure;
1433 print OUTASM $chk[$closurechk{$symb}];
1434 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1438 if ( defined($srtchk{$symb}) ) {
1439 print OUTASM $T_HDR_relrodata;
1440 print OUTASM $chk[$srtchk{$symb}];
1441 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1445 if ( defined($infochk{$symb}) ) {
1447 print OUTASM $T_HDR_info;
1448 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1450 # entry code will be put here!
1452 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1456 if ( defined($entrychk{$symb}) ) {
1458 $c = $chk[$entrychk{$symb}];
1460 # If this is an entry point with an info table,
1461 # eliminate the entry symbol and all directives involving it.
1462 if (defined($infochk{$symb}) && $TargetPlatform !~ /^ia64-/) {
1464 foreach $l (split(/\n/,$c)) {
1465 next if $l =~ /^.*$symb_(entry|ret)${T_POST_LBL}/;
1467 # If we have .type/.size direrctives involving foo_entry,
1468 # then make them refer to foo_info instead. The information
1469 # in these directives is used by the cachegrind annotator,
1470 # so it is worthwhile keeping.
1471 if ($l =~ /^\s*\.(type|size).*$symb_(entry|ret)/) {
1472 $l =~ s/$symb(_entry|_ret)/${symb}_info/g;
1476 next if $l =~ /^\s*\..*$symb.*\n?/;
1479 $c = join("\n",@o) . "\n";
1482 print OUTASM $T_HDR_entry;
1484 &print_doctored($c, 1); # NB: the 1!!!
1486 $chkcat[$entrychk{$symb}] = 'DONE ALREADY';
1489 } elsif ( $chkcat[$i] eq 'vector' ) {
1490 $symb = $chksymb[$i];
1493 if ( defined($vectorchk{$symb}) ) {
1494 print OUTASM $T_HDR_vector;
1495 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1497 # direct return code will be put here!
1498 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1500 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1501 # Alphas: the commented nop is for the splitter, to ensure
1502 # that no module ends with a label as the very last
1503 # thing. (The linker will adjust the label to point
1504 # to the first code word of the next module linked in,
1505 # even if alignment constraints cause the label to move!)
1507 print OUTASM "\t# nop\n";
1510 } elsif ( $chkcat[$i] eq 'rodata' ) {
1511 print OUTASM $T_HDR_rodata;
1512 print OUTASM $chk[$i];
1513 $chkcat[$i] = 'DONE ALREADY';
1514 } elsif ( $chkcat[$i] eq 'relrodata' ) {
1515 print OUTASM $T_HDR_relrodata;
1516 print OUTASM $chk[$i];
1517 $chkcat[$i] = 'DONE ALREADY';
1518 } elsif ( $chkcat[$i] eq 'toc' ) {
1519 # silly optimisation to print tocs, since they come in groups...
1520 print OUTASM $T_HDR_toc;
1522 while ($chkcat[$j] eq 'toc')
1523 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1526 print OUTASM $chk[$j];
1528 $chkcat[$j] = 'DONE ALREADY';
1532 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' ) {
1533 # apple-darwin: dynamic linker stubs
1534 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1535 { # print them out unchanged, but remove the stubs for __DISCARD__
1536 print OUTASM $chk[$i];
1538 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ && $chkcat[$i] eq 'get_pc_thunk' ) {
1539 # i386-apple-darwin: __i686.get_pc_thunk.[abcd]x
1540 print OUTASM ".section __TEXT,__textcoal_nt,coalesced,no_toc\n";
1541 print OUTASM $chk[$i];
1543 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1547 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1550 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1551 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1555 On IA64, tail calls are converted to branches at this point. The mangler
1556 searches for function calls immediately followed by a '--- TAILCALL ---'
1557 token. Since the compiler can put various combinations of labels, bundling
1558 directives, nop instructions, stops, and a move of the return value
1559 between the branch and the tail call, proper matching of the tail call
1560 gets a little hairy. This subroutine does the mangling.
1562 Here is an example of a tail call before mangling:
1565 br.call.sptk.many b0 = b6
1580 sub ia64_mangle_tailcalls {
1581 # Function input and output are in $c
1583 # Construct the tailcall-mangling expression the first time this function
1585 if (!defined($IA64_MATCH_TAILCALL)) {
1586 # One-line pattern matching constructs. None of these
1587 # should bind references; all parenthesized terms
1588 # should be (?:) terms.
1589 my $stop = q/(?:\t;;\n)/;
1590 my $bundle = q/(?:\t\.(?:mii|mib|mmi|mmb|mfi|mfb|mbb|bbb)\n)/;
1591 my $nop = q/(?:\tnop(?:\.[mifb])?\s+\d+\n)/;
1592 my $movgp = q/(?:\tmov r1 = r\d+\n)/;
1593 my $postbr = q/(?:\tbr \.L\d+\n)/;
1595 my $noeffect = "(?:$stop$bundle?|$nop)*";
1596 my $postbundle = "(?:$bundle?$nop?$nop?$postbr)?";
1598 # Important parts of the pattern match. The branch target
1599 # and subsequent jump label are bound to $1 and $2
1600 # respectively. Sometimes there is no label.
1601 my $callbr = q/^\tbr\.call\.sptk\.many b0 = (.*)\n/;
1602 my $label = q/(?:^\.L([0-9]*):\n)/;
1603 my $tailcall = q/\t--- TAILCALL ---\n/;
1605 $IA64_MATCH_TAILCALL =
1606 $callbr . $label . '?' . $noeffect . $movgp . '?' . $noeffect .
1607 $tailcall . $stop . '?' . '(?:' . $postbundle . ')?';
1610 # Find and mangle tailcalls
1611 while ($c =~ s/$IA64_MATCH_TAILCALL/\tbr\.few $1\n/o) {
1612 # Eek, the gcc optimiser is getting smarter... if we see a jump to the
1613 # --- TAILCALL --- marker then we reapply the substitution at the source sites
1614 $c =~ s/^\tbr \.L$2\n/\t--- TAILCALL ---\n/g if ($2);
1617 # Verify that all instances of TAILCALL were processed
1618 if ($c =~ /^\t--- TAILCALL ---\n/) {
1619 die "Unmangled TAILCALL tokens remain after mangling"
1624 The number of registers allocated on the IA64 register stack is set
1625 upon entry to the runtime with an `alloc' instruction at the entry
1626 point of \verb+StgRun()+. Gcc uses its own `alloc' to allocate
1627 however many registers it likes in each function. When we discard
1628 gcc's alloc, we have to reconcile its register assignment with what
1631 There are three stack areas: fixed registers, input/local registers,
1632 and output registers. We move the output registers to the output
1633 register space and leave the other registers where they are.
1636 sub ia64_rename_registers() {
1637 # The text to be mangled is in $c
1638 # Find number of registers in each stack area
1639 my ($loc, $out) = @_;
1645 # These are the register numbers used in the STG runtime
1646 my $STG_FIRST_OUT_REG = 32 + 34;
1647 my $STG_LAST_OUT_REG = $STG_FIRST_OUT_REG + 7;
1649 $first_out_reg = 32 + $loc;
1651 if ($first_out_reg > $STG_FIRST_OUT_REG) {
1652 die "Too many local registers allocated by gcc";
1655 # Split the string into fragments containing one register name each.
1656 # Rename the register in each fragment and concatenate.
1658 foreach $fragment (split(/(?=r\d+[^a-zA-Z0-9_.])/s, $c)) {
1659 if ($fragment =~ /^r(\d+)((?:[^a-zA-Z0-9_.].*)?)$/s) {
1662 if ($regnum < $first_out_reg) {
1663 # This is a local or fixed register
1665 # Local registers 32 and 33 (r64 and r65) are
1666 # used to hold saved state; they shouldn't be touched
1667 if ($regnum == 64 || $regnum == 65) {
1668 die "Reserved register $regnum is in use";
1672 # This is an output register
1673 $regnum = $regnum - $first_out_reg + $STG_FIRST_OUT_REG;
1674 if ($regnum > $STG_LAST_OUT_REG) {
1675 die "Register number ($regnum) is out of expected range";
1679 # Update this fragment
1680 $fragment = "r" . $regnum . $2;
1691 sub hppa_mash_prologue { # OK, epilogue, too
1694 # toss all prologue stuff
1695 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1697 # Lie about our .CALLINFO
1698 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1705 # toss all epilogue stuff
1706 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1708 # Sorry; we moved the _info stuff to the code segment.
1709 s/_info,DATA/_info,CODE/g;
1716 sub print_doctored {
1717 local($_, $need_fallthru_patch) = @_;
1719 if ( $TargetPlatform =~ /^x86_64-/ ) {
1722 # movq -4(%ebp), %rax
1727 s/^\tmovq\s+(-?\d*\(\%r(bx|bp|13)\)),\s*(\%r(ax|cx|dx|10|11))\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1728 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e(ax|cx|si|di))\n\tjmp\s+\*\%r\3/\tjmp\t$T_US$1/g;
1731 if ( $TargetPlatform !~ /^i386-/
1732 || ! /^\t[a-z]/ # no instructions in here, apparently
1733 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1738 # OK, must do some x86 **HACKING**
1740 local($entry_patch) = '';
1741 local($exit_patch) = '';
1743 # gotta watch out for weird instructions that
1744 # invisibly smash various regs:
1745 # rep* %ecx used for counting
1746 # scas* %edi used for destination index
1747 # cmps* %e[sd]i used for indices
1748 # loop* %ecx used for counting
1753 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1755 # * GCC used an "STG reg" for its own purposes
1757 # * some secret uses of machine reg, requiring STG reg
1758 # to be saved/restored
1760 # The most dangerous "GCC uses" of an "STG reg" are when
1761 # the reg holds the target of a jmp -- it's tricky to
1762 # insert the patch-up code before we get to the target!
1763 # So here we change the jmps:
1765 # --------------------------------------------------------
1766 # it can happen that we have jumps of the form...
1767 # jmp *<something involving %esp>
1769 # jmp <something involving another naughty register...>
1771 # a reasonably-common case is:
1773 # movl $_blah,<bad-reg>
1776 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[acd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1780 # movl -4(%ebx), %eax
1785 s/^\tmovl\s+(-?\d*\(\%e(bx|si)\)),\s*(\%e[acd]x)\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1787 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1788 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1789 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1790 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1791 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1792 if /(jmp|call)\s+.*\%esi/;
1794 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1795 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1796 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1797 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1798 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1799 if /(jmp|call)\s+.*\%edi/;
1802 # OK, now we can decide what our patch-up code is going to
1805 # Offsets into register table - you'd better update these magic
1806 # numbers should you change its contents!
1807 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1808 local($OFFSET_Hp)=88;
1810 # Note funky ".=" stuff; we're *adding* to these _patch guys
1811 if ( $StolenX86Regs <= 2
1812 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1813 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1814 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1816 # nothing for call_{entry,exit} because %esi is callee-save
1818 if ( $StolenX86Regs <= 3
1819 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1820 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1821 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1823 # nothing for call_{entry,exit} because %edi is callee-save
1826 # --------------------------------------------------------
1827 # next, here we go with non-%esp patching!
1829 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1831 # Before calling GC we must set up the exit condition before the call
1832 # and entry condition when we come back
1834 # fix _all_ non-local jumps:
1836 if ( $TargetPlatform =~ /^.*-apple-darwin.*/ ) {
1837 # On Darwin, we've got local-looking jumps that are
1838 # actually global (i.e. jumps to Lfoo$stub or via
1839 # Lfoo$non_lazy_ptr), so we fix those first.
1840 # In fact, we just fix everything that contains a dollar
1841 # because false positives don't hurt here.
1843 s/^(\tjmp\s+\*?L.*\$.*\n)/$exit_patch$1/g;
1846 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1847 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1849 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1851 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1852 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1854 if ($StolenX86Regs == 2 ) {
1855 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1856 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1857 } elsif ($StolenX86Regs == 3 ) {
1858 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1859 if /^\t(jmp|call)\s+.*\%edi/;
1862 # --------------------------------------------------------
1863 # that's it -- print it
1865 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1869 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1870 print OUTASM $exit_patch;
1871 # ToDo: make it not print if there is a "jmp" at the end
1877 sub init_FUNNY_THINGS {
1878 %KNOWN_FUNNY_THING = (
1880 # "${T_US}stg_.*{T_POST_LBL}", 1,
1885 The following table reversal is used for both info tables and return
1886 vectors. In both cases, we remove the first entry from the table,
1887 reverse the table, put the label at the end, and paste some code
1888 (that which is normally referred to by the first entry in the table)
1889 right after the table itself. (The code pasting is done elsewhere.)
1893 local($symb, $tbl, $discard1) = @_;
1895 return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1897 local($before) = '';
1899 local(@imports) = (); # hppa only
1902 local(@lines) = split(/\n/, $tbl);
1905 # Deal with the header...
1906 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1907 $label .= $lines[$i] . "\n",
1908 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1909 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1910 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1912 $before .= $lines[$i] . "\n"; # otherwise...
1916 $infoname =~ s/(.|\n)*^([A-Za-z0-9_]+_info)${T_POST_LBL}$(.|\n)*/\2/;
1918 # Grab the table data...
1919 if ( $TargetPlatform !~ /^hppa/ ) {
1920 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1922 # Convert addresses of SRTs, slow entrypoints and large bitmaps
1923 # to offsets (relative to the info label),
1924 # in order to support position independent code.
1925 $line =~ s/$infoname/0/
1926 || $line =~ s/([A-Za-z0-9_]+_srtd)$/\1 - $infoname/
1927 || $line =~ s/([A-Za-z0-9_]+_srt(\+\d+)?)$/\1 - $infoname/
1928 || $line =~ s/([A-Za-z0-9_]+_str)$/\1 - $infoname/
1929 || $line =~ s/([A-Za-z0-9_]+_slow)$/\1 - $infoname/
1930 || $line =~ s/([A-Za-z0-9_]+_btm)$/\1 - $infoname/
1931 || $line =~ s/([A-Za-z0-9_]+_alt)$/\1 - $infoname/
1932 || $line =~ s/([A-Za-z0-9_]+_dflt)$/\1 - $infoname/
1933 || $line =~ s/([A-Za-z0-9_]+_ret)$/\1 - $infoname/;
1934 push(@words, $line);
1936 } else { # hppa weirdness
1937 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1938 # FIXME: the RTS now expects offsets instead of addresses
1939 # for all labels in info tables.
1940 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1941 push(@imports, $lines[$i]);
1943 # We don't use HP's ``function pointers''
1944 # We just use labels in code space, like normal people
1945 $lines[$i] =~ s/P%//;
1946 push(@words, $lines[$i]);
1951 # Now throw away any initial zero word from the table. This is a hack
1952 # that lets us reduce the size of info tables when the SRT field is not
1953 # needed: see comments StgFunInfoTable in InfoTables.h.
1955 # The .zero business is for Linux/ELF.
1956 # The .skip business is for Sparc/Solaris/ELF.
1957 # The .blockz business is for HPPA.
1959 # if ($words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1964 for (; $i <= $#lines; $i++) {
1965 $after .= $lines[$i] . "\n";
1968 # Alphas: If we have anonymous text (not part of a procedure), the
1969 # linker may complain about missing exception information. Bleh.
1970 # To suppress this, we place a .ent/.end pair around the code.
1971 # At the same time, we have to be careful and not enclose any leading
1972 # .file/.loc directives.
1973 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1974 local ($ident) = $1;
1975 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1976 $after .= "\t.end $ident\n";
1979 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1980 # assembler (!) wherein .quad constants inside .text sections are
1981 # first narrowed to 32 bits then sign-extended back to 64 bits.
1982 # This obviously screws up our 64-bit bitmaps, so we work around
1983 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1984 if ( $TargetPlatform =~ /^alpha-/ ) {
1986 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1987 local ($number) = $1;
1988 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1989 local ($sign, $base, $digits) = ($1, $2, $3);
1990 $base = (10, 8, 16)[length $base];
1991 local ($hi, $lo) = (0, 0);
1992 foreach $i (split(//, $digits)) {
1993 $j = $lo * $base + $i;
1994 $lo = $j % 4294967296;
1995 $hi = $hi * $base + ($j - $lo) / 4294967296;
1997 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1999 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
2000 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
2002 print STDERR "Cannot handle \".quad $number\" in info table\n";
2009 if ( $TargetPlatform =~ /x86_64-apple-darwin/ ) {
2010 # Tack a label to the front of the info table, too.
2011 # For now, this just serves to work around a crash in Apple's new
2012 # 64-bit linker (it seems to assume that there is no data before the
2013 # first label in a section).
2015 # The plan for the future is to do this on all Darwin platforms, and
2016 # to add a reference to this label after the entry code, just as the
2017 # NCG does, so we can enable dead-code-stripping in the linker without
2018 # losing our info tables. (Hence the name _dsp, for dead-strip preventer)
2020 $before .= "\n${infoname}_dsp:\n";
2024 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
2025 . join("\n", @words) . "\n"
2028 # print STDERR "before=$before\n";
2029 # print STDERR "label=$label\n";
2030 # print STDERR "words=",(reverse @words),"\n";
2031 # print STDERR "after=$after\n";
2037 The HP is a major nuisance. The threaded code mangler moved info
2038 tables from data space to code space, but unthreaded code in the RTS
2039 still has references to info tables in data space. Since the HP
2040 linker is very precise about where symbols live, we need to patch the
2041 references in the unthreaded RTS as well.
2044 sub mini_mangle_asm_hppa {
2045 local($in_asmf, $out_asmf) = @_;
2047 open(INASM, "< $in_asmf")
2048 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
2049 open(OUTASM,"> $out_asmf")
2050 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
2053 s/_info,DATA/_info,CODE/; # Move _info references to code space
2059 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
2060 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
2066 sub tidy_up_and_die {
2067 local($return_val, $msg) = @_;
2069 exit (($return_val == 0) ? 0 : 1);