1 %************************************************************************
3 \section[Driver-asm-fiddling]{Fiddling with assembler files}
5 %************************************************************************
10 Utterly stomp out C functions' prologues and epilogues; i.e., the
11 stuff to do with the C stack.
13 Any other required tidying up.
16 General note [chak]: Many regexps are very fragile because they rely on white
17 space being in the right place. This caused trouble with gcc 2.95 (at least
18 on Linux), where the use of white space in .s files generated by gcc suddenly
19 changed. To guarantee compatibility across different versions of gcc, make
20 sure (at least on i386-.*-linux) that regexps tolerate varying amounts of white
21 space between an assembler statement and its arguments as well as after a the
22 comma separating multiple arguments.
24 \emph{For the time being, I have corrected the regexps for i386-.*-linux. I
25 didn't touch all the regexps for other i386 platforms, as I don't have
26 a box to test these changes.}
31 The HP linker is very picky about symbols being in the appropriate
32 space (code vs. data). When we mangle the threaded code to put the
33 info tables just prior to the code, they wind up in code space
34 rather than data space. This means that references to *_info from
35 un-mangled parts of the RTS (e.g. unthreaded GC code) get
36 unresolved symbols. Solution: mini-mangler for .c files on HP. I
37 think this should really be triggered in the driver by a new -rts
38 option, so that user code doesn't get mangled inappropriately.
40 With reversed tables, jumps are to the _info label rather than to
41 the _entry label. The _info label is just an address in code
42 space, rather than an entry point with the descriptive blob we
43 talked about yesterday. As a result, you can't use the call-style
44 JMP_ macro. However, some JMP_ macros take _info labels as targets
45 and some take code entry points within the RTS. The latter won't
46 work with the goto-style JMP_ macro. Sigh. Solution: Use the goto
47 style JMP_ macro, and mangle some more assembly, changing all
48 "RP'literal" and "LP'literal" references to "R'literal" and
49 "L'literal," so that you get the real address of the code, rather
50 than the descriptive blob. Also change all ".word P%literal"
51 entries in info tables and vector tables to just ".word literal,"
52 for the same reason. Advantage: No more ridiculous call sequences.
55 %************************************************************************
57 \subsection{Top-level code}
59 %************************************************************************
62 $TargetPlatform = $TARGETPLATFORM;
64 ($Pgm = $0) =~ s|.*/||;
68 if ( $TargetPlatform =~ /^i386-/ ) {
72 $StolenX86Regs = $ARGV[2];
76 &mangle_asm($ifile,$ofile);
81 %************************************************************************
83 \subsection{Constants for various architectures}
85 %************************************************************************
88 sub init_TARGET_STUFF {
90 #--------------------------------------------------------#
91 if ( $TargetPlatform =~ /^alpha-.*-.*/ ) {
93 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
94 $T_US = ''; # _ if symbols have an underscore on the front
95 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
96 $T_CONST_LBL = '^\$L?C(\d+):$'; # regexp for what such a lbl looks like
99 $T_MOVE_DIRVS = '^(\s*(\$.*\.\.ng:|\.align\s+\d+|\.(globl|ent)\s+\S+|\#.*|\.(file|loc)\s+\S+\s+\S+|\.text|\.r?data)\n)';
100 $T_COPY_DIRVS = '^\s*(\$.*\.\.ng:|\#|\.(file|globl|ent|loc))';
102 $T_DOT_WORD = '\.(long|quad|byte|word)';
103 $T_DOT_GLOBAL = '^\t\.globl';
104 $T_HDR_literal = "\.rdata\n\t\.align 3\n";
105 $T_HDR_misc = "\.text\n\t\.align 3\n";
106 $T_HDR_data = "\.data\n\t\.align 3\n";
107 $T_HDR_rodata = "\.rdata\n\t\.align 3\n";
108 $T_HDR_closure = "\.data\n\t\.align 3\n";
109 $T_HDR_info = "\.text\n\t\.align 3\n";
110 $T_HDR_entry = "\.text\n\t\.align 3\n";
111 $T_HDR_vector = "\.text\n\t\.align 3\n";
113 #--------------------------------------------------------#
114 } elsif ( $TargetPlatform =~ /^hppa/ ) {
116 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
117 $T_US = ''; # _ if symbols have an underscore on the front
118 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
119 $T_CONST_LBL = '^L\$C(\d+)$'; # regexp for what such a lbl looks like
122 $T_MOVE_DIRVS = '^((\s+\.(IMPORT|EXPORT|PARAM).*|\s+\.align\s+\d+|\s+\.(SPACE|SUBSPA)\s+\S+|\s*)\n)';
123 $T_COPY_DIRVS = '^\s+\.(IMPORT|EXPORT)';
125 $T_DOT_WORD = '\.(blockz|word|half|byte)';
126 $T_DOT_GLOBAL = '^\s+\.EXPORT';
127 $T_HDR_literal = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
128 $T_HDR_misc = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
129 $T_HDR_data = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
130 $T_HDR_rodata = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
131 $T_HDR_closure = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
132 $T_HDR_info = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
133 $T_HDR_entry = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
134 $T_HDR_vector = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
136 #--------------------------------------------------------#
137 } elsif ( $TargetPlatform =~ /^i386-.*-(linuxaout|freebsd2|nextstep3|cygwin32|mingw32)$/ ) {
138 # NeXT added but not tested. CaS
140 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
141 $T_US = '_'; # _ if symbols have an underscore on the front
142 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
143 $T_CONST_LBL = '^LC(\d+):$';
145 $T_X86_PRE_LLBL_PAT = 'L';
146 $T_X86_PRE_LLBL = 'L';
147 $T_X86_BADJMP = '^\tjmp [^L\*]';
149 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s.*|\.globl\s+\S+|\.text|\.data|\.stab[^n].*|\.type\s+.*|\.size\s+.*|\.lcomm.*)\n)';
150 $T_COPY_DIRVS = '\.(globl|stab|lcomm)';
151 $T_DOT_WORD = '\.(long|word|value|byte|space)';
152 $T_DOT_GLOBAL = '\.globl';
153 $T_HDR_literal = "\.text\n\t\.align 2\n";
154 $T_HDR_misc = "\.text\n\t\.align 2,0x90\n";
155 $T_HDR_data = "\.data\n\t\.align 2\n";
156 $T_HDR_rodata = "\.text\n\t\.align 2\n";
157 $T_HDR_closure = "\.data\n\t\.align 2\n";
158 $T_HDR_info = "\.text\n\t\.align 2\n"; # NB: requires padding
159 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
160 $T_HDR_vector = "\.text\n\t\.align 2\n"; # NB: requires padding
162 #--------------------------------------------------------#
163 } elsif ( $TargetPlatform =~ /^i386-.*-(solaris2|linux|gnu|freebsd|netbsd|openbsd|kfreebsdgnu)$/ ) {
165 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
166 $T_US = ''; # _ if symbols have an underscore on the front
167 $T_PRE_APP = # regexp that says what comes before APP/NO_APP
168 ($TargetPlatform =~ /-(linux|gnu|freebsd|netbsd|openbsd)$/) ? '#' : '/' ;
169 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
171 $T_X86_PRE_LLBL_PAT = '\.L';
172 $T_X86_PRE_LLBL = '.L';
173 $T_X86_BADJMP = '^\tjmp\s+[^\.\*]';
175 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s.*|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.size\s+\S+\s*,\s*\d+|\.ident.*|\.local.*)\n)';
176 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
178 $T_DOT_WORD = '\.(long|value|word|byte|zero)';
179 $T_DOT_GLOBAL = '\.globl';
180 $T_HDR_literal = "\.section\t\.rodata\n"; # or just use .text??? (WDP 95/11)
181 $T_HDR_misc = "\.text\n\t\.align 4\n";
182 $T_HDR_data = "\.data\n\t\.align 4\n";
183 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 4\n";
184 $T_HDR_closure = "\.data\n\t\.align 4\n";
185 $T_HDR_info = "\.text\n\t\.align 4\n";
186 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
187 $T_HDR_vector = "\.text\n\t\.align 4\n"; # NB: requires padding
189 #--------------------------------------------------------#
190 } elsif ( $TargetPlatform =~ /^ia64-.*-linux$/ ) {
192 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
193 $T_US = ''; # _ if symbols have an underscore on the front
195 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
198 $T_MOVE_DIRVS = '^(\s*\.(global|proc|pred\.safe_across_calls|text|data|section|subsection|align|size|type|ident)\s+.*\n)';
199 $T_COPY_DIRVS = '\.(global|proc)';
201 $T_DOT_WORD = '\.(long|value|byte|zero)';
202 $T_DOT_GLOBAL = '\.global';
203 $T_HDR_literal = "\.section\t\.rodata\n";
204 $T_HDR_misc = "\.text\n\t\.align 16\n"; # May contain code; align like 'entry'
205 $T_HDR_data = "\.data\n\t\.align 8\n";
206 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 8\n";
207 $T_HDR_closure = "\.data\n\t\.align 8\n";
208 $T_HDR_info = "\.text\n\t\.align 8\n";
209 $T_HDR_entry = "\.text\n\t\.align 16\n";
210 $T_HDR_vector = "\.text\n\t\.align 8\n";
212 #--------------------------------------------------------#
213 } elsif ( $TargetPlatform =~ /^x86_64-.*-(linux|openbsd)$/ ) {
215 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
216 $T_US = ''; # _ if symbols have an underscore on the front
218 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
221 $T_MOVE_DIRVS = '^(\s*\.(globl|text|data|section|align|size|type|ident|local)\s+.*\n)';
222 $T_COPY_DIRVS = '\.(globl|type|size|local)';
224 $T_DOT_WORD = '\.(quad|long|value|byte|zero)';
225 $T_DOT_GLOBAL = '\.global';
227 $T_HDR_literal16 = "\.section\t\.rodata.cst16\n\t.align 16\n";
228 $T_HDR_literal = "\.section\t\.rodata\n";
230 $T_HDR_misc = "\.text\n\t\.align 8\n";
231 $T_HDR_data = "\.data\n\t\.align 8\n";
232 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 8\n";
234 # the assembler on x86_64/Linux refuses to generate code for
236 # where x is in the text section and y in the rodata section.
237 # It works if y is in the text section, though. This is probably
238 # going to cause difficulties for PIC, I imagine.
239 $T_HDR_relrodata= "\.text\n\t\.align 8\n";
241 $T_HDR_closure = "\.data\n\t\.align 8\n";
242 $T_HDR_info = "\.text\n\t\.align 8\n";
243 $T_HDR_entry = "\.text\n\t\.align 8\n";
244 $T_HDR_vector = "\.text\n\t\.align 8\n";
246 #--------------------------------------------------------#
247 } elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) {
249 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
250 $T_US = '_'; # _ if symbols have an underscore on the front
251 $T_PRE_APP = '^# MAY NOT APPLY'; # regexp that says what comes before APP/NO_APP
252 $T_CONST_LBL = '^LC(\d+):$';
255 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.proc\s+\d+|\.const|\.cstring|\.globl\s+\S+|\.text|\.data|\.even|\.stab[^n].*)\n)';
256 $T_COPY_DIRVS = '\.(globl|proc|stab)';
258 $T_DOT_WORD = '\.long';
259 $T_DOT_GLOBAL = '\.globl';
260 $T_HDR_literal = "\.text\n\t\.even\n";
261 $T_HDR_misc = "\.text\n\t\.even\n";
262 $T_HDR_data = "\.data\n\t\.even\n";
263 $T_HDR_rodata = "\.text\n\t\.even\n";
264 $T_HDR_closure = "\.data\n\t\.even\n";
265 $T_HDR_info = "\.text\n\t\.even\n";
266 $T_HDR_entry = "\.text\n\t\.even\n";
267 $T_HDR_vector = "\.text\n\t\.even\n";
269 #--------------------------------------------------------#
270 } elsif ( $TargetPlatform =~ /^mips-.*/ ) {
272 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
273 $T_US = ''; # _ if symbols have an underscore on the front
274 $T_PRE_APP = '^\s*#'; # regexp that says what comes before APP/NO_APP
275 $T_CONST_LBL = '^\$LC(\d+):$'; # regexp for what such a lbl looks like
278 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.(globl|ent)\s+\S+|\.text|\.r?data)\n)';
279 $T_COPY_DIRVS = '\.(globl|ent)';
281 $T_DOT_WORD = '\.word';
282 $T_DOT_GLOBAL = '^\t\.globl';
283 $T_HDR_literal = "\t\.rdata\n\t\.align 2\n";
284 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
285 $T_HDR_data = "\t\.data\n\t\.align 2\n";
286 $T_HDR_rodata = "\t\.rdata\n\t\.align 2\n";
287 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
288 $T_HDR_info = "\t\.text\n\t\.align 2\n";
289 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
290 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
292 #--------------------------------------------------------#
293 } elsif ( $TargetPlatform =~ /^powerpc-apple-darwin.*/ ) {
294 # Apple PowerPC Darwin/MacOS X.
295 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
296 $T_US = '_'; # _ if symbols have an underscore on the front
297 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
298 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
301 $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)';
302 $T_COPY_DIRVS = '\.(globl|lcomm)';
304 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
305 $T_DOT_GLOBAL = '\.globl';
306 $T_HDR_toc = "\.toc\n";
307 $T_HDR_literal = "\t\.const\n\t\.align 2\n";
308 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
309 $T_HDR_data = "\t\.data\n\t\.align 2\n";
310 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
311 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
312 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
313 $T_HDR_info = "\t\.text\n\t\.align 2\n";
314 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
315 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
317 #--------------------------------------------------------#
318 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ ) {
319 # Apple PowerPC Darwin/MacOS X.
320 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
321 $T_US = '_'; # _ if symbols have an underscore on the front
322 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
323 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
325 $T_X86_PRE_LLBL_PAT = 'L';
326 $T_X86_PRE_LLBL = 'L';
327 $T_X86_BADJMP = '^\tjmp [^L\*]';
329 $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)';
330 $T_COPY_DIRVS = '\.(globl|lcomm)';
332 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
333 $T_DOT_GLOBAL = '\.globl';
334 $T_HDR_toc = "\.toc\n";
335 $T_HDR_literal16= "\t\.literal8\n\t\.align 4\n";
336 $T_HDR_literal = "\t\.const\n\t\.align 4\n";
337 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
338 $T_HDR_data = "\t\.data\n\t\.align 2\n";
339 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
340 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
341 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
342 $T_HDR_info = "\t\.text\n\t\.align 2\n";
343 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
344 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
346 #--------------------------------------------------------#
347 } elsif ( $TargetPlatform =~ /^x86_64-apple-darwin.*/ ) {
348 # Apple PowerPC Darwin/MacOS X.
349 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
350 $T_US = '_'; # _ if symbols have an underscore on the front
351 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
352 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
355 $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)';
356 $T_COPY_DIRVS = '\.(globl|lcomm)';
358 $T_DOT_WORD = '\.(quad|long|short|byte|fill|space)';
359 $T_DOT_GLOBAL = '\.globl';
360 $T_HDR_toc = "\.toc\n";
361 $T_HDR_literal16= "\t\.literal8\n\t\.align 4\n";
362 $T_HDR_literal = "\t\.const\n\t\.align 4\n";
363 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
364 $T_HDR_data = "\t\.data\n\t\.align 2\n";
365 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
366 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
367 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
368 $T_HDR_info = "\t\.text\n\t\.align 2\n";
369 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
370 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
372 #--------------------------------------------------------#
373 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ ) {
375 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
376 $T_US = ''; # _ if symbols have an underscore on the front
377 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
378 $T_CONST_LBL = '^\.LC\d+:'; # regexp for what such a lbl looks like
381 $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)';
382 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
384 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
385 $T_DOT_GLOBAL = '\.globl';
386 $T_HDR_toc = "\.toc\n";
387 $T_HDR_literal = "\t\.section\t.rodata\n\t\.align 2\n";
388 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
389 $T_HDR_data = "\t\.data\n\t\.align 2\n";
390 $T_HDR_rodata = "\t\.section\t.rodata\n\t\.align 2\n";
391 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
392 $T_HDR_info = "\t\.text\n\t\.align 2\n";
393 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
394 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
396 #--------------------------------------------------------#
397 } elsif ( $TargetPlatform =~ /^powerpc64-.*-linux/ ) {
399 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
400 $T_US = '\.'; # _ if symbols have an underscore on the front
401 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
402 $T_CONST_LBL = '^\.LC\d+:'; # regexp for what such a lbl looks like
405 $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)';
406 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
408 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
409 $T_DOT_GLOBAL = '\.globl';
410 $T_HDR_toc = "\.toc\n";
411 $T_HDR_literal = "\t\.section\t\".toc\",\"aw\"\n";
412 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
413 $T_HDR_data = "\t\.data\n\t\.align 2\n";
414 $T_HDR_rodata = "\t\.section\t.rodata\n\t\.align 2\n";
415 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
416 $T_HDR_info = "\t\.text\n\t\.align 2\n";
417 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
418 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
420 #--------------------------------------------------------#
421 } elsif ( $TargetPlatform =~ /^sparc-.*-(solaris2|openbsd)/ ) {
423 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
424 $T_US = ''; # _ if symbols have an underscore on the front
425 $T_PRE_APP = 'DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
426 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
429 $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)';
430 $T_COPY_DIRVS = '\.(global|local|proc|stab)';
432 $T_DOT_WORD = '\.(long|word|byte|half|skip|uahalf|uaword)';
433 $T_DOT_GLOBAL = '^\t\.global';
434 $T_HDR_literal = "\.text\n\t\.align 8\n";
435 $T_HDR_misc = "\.text\n\t\.align 4\n";
436 $T_HDR_data = "\.data\n\t\.align 8\n";
437 $T_HDR_rodata = "\.text\n\t\.align 4\n";
438 $T_HDR_closure = "\.data\n\t\.align 4\n";
439 $T_HDR_info = "\.text\n\t\.align 4\n";
440 $T_HDR_entry = "\.text\n\t\.align 4\n";
441 $T_HDR_vector = "\.text\n\t\.align 4\n";
443 #--------------------------------------------------------#
444 } elsif ( $TargetPlatform =~ /^sparc-.*-sunos4/ ) {
446 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
447 $T_US = '_'; # _ if symbols have an underscore on the front
448 $T_PRE_APP = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
449 $T_CONST_LBL = '^LC(\d+):$';
452 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*)\n)';
453 $T_COPY_DIRVS = '\.(global|proc|stab)';
455 $T_DOT_WORD = '\.word';
456 $T_DOT_GLOBAL = '^\t\.global';
457 $T_HDR_literal = "\.text\n\t\.align 8\n";
458 $T_HDR_misc = "\.text\n\t\.align 4\n";
459 $T_HDR_data = "\.data\n\t\.align 8\n";
460 $T_HDR_rodata = "\.text\n\t\.align 4\n";
461 $T_HDR_closure = "\.data\n\t\.align 4\n";
462 $T_HDR_info = "\.text\n\t\.align 4\n";
463 $T_HDR_entry = "\.text\n\t\.align 4\n";
464 $T_HDR_vector = "\.text\n\t\.align 4\n";
466 #--------------------------------------------------------#
467 } elsif ( $TargetPlatform =~ /^sparc-.*-linux/ ) {
468 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
469 $T_US = ''; # _ if symbols have an underscore on the front
470 $T_PRE_APP = '#'; # regexp that says what comes before APP/NO_APP
471 # Probably doesn't apply anyway
472 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
475 $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)';
476 $T_COPY_DIRVS = '\.(global|local|globl|proc|stab)';
478 $T_DOT_WORD = '\.(long|word|nword|xword|byte|half|short|skip|uahalf|uaword)';
479 $T_DOT_GLOBAL = '^\t\.global';
480 $T_HDR_literal = "\.text\n\t\.align 8\n";
481 $T_HDR_misc = "\.text\n\t\.align 4\n";
482 $T_HDR_data = "\.data\n\t\.align 8\n";
483 $T_HDR_rodata = "\.text\n\t\.align 4\n";
484 $T_HDR_closure = "\.data\n\t\.align 4\n";
485 $T_HDR_info = "\.text\n\t\.align 4\n";
486 $T_HDR_entry = "\.text\n\t\.align 4\n";
487 $T_HDR_vector = "\.text\n\t\.align 4\n";
489 #--------------------------------------------------------#
491 print STDERR "$Pgm: don't know how to mangle assembly language for: $TargetPlatform\n";
495 if($T_HDR_relrodata eq "") {
497 # relrodata defaults to rodata.
498 $T_HDR_relrodata = $T_HDR_rodata;
502 print STDERR "T_STABBY: $T_STABBY\n";
503 print STDERR "T_US: $T_US\n";
504 print STDERR "T_PRE_APP: $T_PRE_APP\n";
505 print STDERR "T_CONST_LBL: $T_CONST_LBL\n";
506 print STDERR "T_POST_LBL: $T_POST_LBL\n";
507 if ( $TargetPlatform =~ /^i386-/ ) {
508 print STDERR "T_X86_PRE_LLBL_PAT: $T_X86_PRE_LLBL_PAT\n";
509 print STDERR "T_X86_PRE_LLBL: $T_X86_PRE_LLBL\n";
510 print STDERR "T_X86_BADJMP: $T_X86_BADJMP\n";
512 print STDERR "T_MOVE_DIRVS: $T_MOVE_DIRVS\n";
513 print STDERR "T_COPY_DIRVS: $T_COPY_DIRVS\n";
514 print STDERR "T_DOT_WORD: $T_DOT_WORD\n";
515 print STDERR "T_HDR_literal: $T_HDR_literal\n";
516 print STDERR "T_HDR_misc: $T_HDR_misc\n";
517 print STDERR "T_HDR_data: $T_HDR_data\n";
518 print STDERR "T_HDR_rodata: $T_HDR_rodata\n";
519 print STDERR "T_HDR_closure: $T_HDR_closure\n";
520 print STDERR "T_HDR_info: $T_HDR_info\n";
521 print STDERR "T_HDR_entry: $T_HDR_entry\n";
522 print STDERR "T_HDR_vector: $T_HDR_vector\n";
528 %************************************************************************
530 \subsection{Mangle away}
532 %************************************************************************
536 local($in_asmf, $out_asmf) = @_;
538 # multi-line regexp matching:
543 &init_TARGET_STUFF();
544 &init_FUNNY_THINGS();
546 open(INASM, "< $in_asmf")
547 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
548 open(OUTASM,"> $out_asmf")
549 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
551 # read whole file, divide into "chunks":
552 # record some info about what we've found...
554 @chk = (); # contents of the chunk
555 $numchks = 0; # number of them
556 @chkcat = (); # what category of thing in each chunk
557 @chksymb = (); # what symbol(base) is defined in this chunk
558 %entrychk = (); # ditto, its entry code
559 %closurechk = (); # ditto, the (static) closure
560 %srtchk = (); # ditto, its SRT (for top-level things)
561 %infochk = (); # given a symbol base, say what chunk its info tbl is in
562 %vectorchk = (); # ditto, return vector table
563 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
565 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
568 tr/\r//d if $TargetPlatform =~ /-mingw32$/; # In case Perl doesn't convert line endings
569 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
570 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
571 next if /^\t\.def.*endef$/;
572 next if /${T_PRE_APP}(NO_)?APP/o;
573 next if /^;/ && $TargetPlatform =~ /^hppa/;
575 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|ia64)-/;
577 if ( $TargetPlatform =~ /^mips-/
578 && /^\t\.(globl\S+\.text|comm\t)/ ) {
579 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
580 # Treat .comm variables as data. These show up in two (known) places:
582 # - the module_registered variable used in the __stginit fragment.
583 # even though these are declared static and initialised, gcc 3.3
584 # likes to make them .comm, presumably to save space in the
587 # - global variables used to pass arguments from C to STG in
588 # a foreign export. (is this still true? --SDM)
590 } elsif ( /^\t\.comm.*$/ ) {
592 $chkcat[$i] = 'data';
595 # Labels ending "_str": these are literal strings.
596 } elsif ( /^${T_US}([A-Za-z0-9_]+)_str${T_POST_LBL}$/ ) {
598 $chkcat[$i] = 'rodata';
600 } elsif ( $TargetPlatform =~ /-darwin/
601 && (/^\s*\.subsections_via_symbols/
602 ||/^\s*\.no_dead_strip.*/)) {
603 # Don't allow Apple's linker to do any dead-stripping of symbols
604 # in this file, because it will mess up info-tables in mangled
606 # The .no_dead_strip directives are actually put there by
607 # the gcc3 "used" attribute on entry points.
609 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && (
610 /^\s*\.picsymbol_stub/
611 || /^\s*\.section __TEXT,__picsymbol_stub\d,.*/
612 || /^\s*\.section __TEXT,__picsymbolstub\d,.*/
613 || /^\s*\.symbol_stub/
614 || /^\s*\.section __TEXT,__symbol_stub\d,.*/
615 || /^\s*\.section __TEXT,__symbolstub\d,.*/
616 || /^\s*\.lazy_symbol_pointer/
617 || /^\s*\.non_lazy_symbol_pointer/
618 || /^\s*\.section __IMPORT.*/))
621 $chkcat[$i] = 'dyld';
625 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.data/)
626 { # non_lazy_symbol_ptrs that point to local symbols
628 $chkcat[$i] = 'dyld';
631 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.align/)
632 { # non_lazy_symbol_ptrs that point to local symbols
634 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^L_.*:$/)
635 { # non_lazy_symbol_ptrs that point to local symbols
636 $chk[++$i] = $dyld_section . $_;
637 $chkcat[$i] = 'dyld';
640 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
641 # duplicated from the bottom
645 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
646 # Alphas: Local labels not to be confused with new chunks
648 # NB: all the rest start with a non-space
650 } elsif ( $TargetPlatform =~ /^mips-/
651 && /^\d+:/ ) { # a funny-looking very-local label
654 } elsif ( /$T_CONST_LBL/o ) {
656 $chkcat[$i] = 'literal';
659 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
661 $chkcat[$i] = 'splitmarker';
664 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
667 $chkcat[$i] = 'infotbl';
668 $chksymb[$i] = $symb;
670 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
672 $infochk{$symb} = $i;
674 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
676 $chkcat[$i] = 'entry';
681 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
683 $chkcat[$i] = 'closure';
686 $closurechk{$1} = $i;
688 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
695 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
697 $chkcat[$i] = 'data';
700 } elsif ( /^${T_US}(stg_ap_stack_entries|stg_stack_save_entries|stg_arg_bitmaps)${T_POST_LBL}$/o ) {
702 $chkcat[$i] = 'data';
705 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
708 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
709 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
710 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
711 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
714 $chkcat[$i] = 'data';
717 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
722 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
723 # all CC_ symbols go in the data section...
725 $chkcat[$i] = 'data';
728 } elsif ( /^${T_US}([A-Za-z0-9_]+)_hpc${T_POST_LBL}$/o ) {
729 # hpc shares tick boxes across modules
731 $chkcat[$i] = 'data';
734 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
736 $chkcat[$i] = 'misc';
738 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
740 $chkcat[$i] = 'vector';
745 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
746 && /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
747 # Some Solaris system headers contain function definitions (as
748 # opposed to mere prototypes), which end up in the .hc file when
749 # a Haskell module foreign imports the corresponding system
750 # functions (most notably stat()). We put them into the text
751 # segment. Note that this currently does not extend to function
752 # names starting with an underscore.
755 $chkcat[$i] = 'misc';
758 } elsif ( $TargetPlatform =~ /^i386-apple-darwin/ && /^(___i686\.get_pc_thunk\.[abcd]x):/o) {
759 # To handle PIC on Darwin/x86, we need to appropriately pass through
760 # the get_pc_thunk functions. The need to be put into a special section
761 # marked as coalesced (otherwise the .weak_definition doesn't work
764 $chkcat[$i] = 'get_pc_thunk';
767 } elsif ( /^${T_US}[A-Za-z0-9_]/o
768 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
770 && ( $TargetPlatform !~ /^powerpc64/ # we need to avoid local labels in this case
771 || ! /^\.L\d+:$/ ) ) {
778 /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
779 || /^${T_US}__stg_.*${T_POST_LBL}$/o # more RTS internals
780 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
781 || /^${T_US}.*_slow${T_POST_LBL}$/o # slow entry
782 || /^${T_US}__stginit.*${T_POST_LBL}$/o # __stginit<module>
783 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
784 || /^${T_US}.*_fast${T_POST_LBL}$/o # primops
785 || /^_uname:/o # x86/Solaris2
788 $chkcat[$i] = 'misc';
790 /^${T_US}.*_srtd${T_POST_LBL}$/o # large bitmaps
791 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
794 $chkcat[$i] = 'relrodata';
797 print STDERR "Warning: retaining unknown function \`$thing' in output from C compiler\n";
798 $chkcat[$i] = 'unknown';
801 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ && /^\.LCTOC1 = /o ) {
802 # PowerPC Linux's large-model PIC (-fPIC) generates a gobal offset
803 # table "by hand". Be sure to copy it over.
804 # Note that this label and all entries in the table should actually
805 # go into the .got2 section, but it isn't easy to distinguish them
806 # from other constant literals (.LC\d+), so we just put everything
809 $chkcat[$i] = 'literal';
810 $chksymb[$i] = 'LCTOC1';
811 } else { # simple line (duplicated at the top)
816 $numchks = $#chk + 1;
817 $chk[$numchks] = ''; # We might push .note.GNU-stack into this
818 $chkcat[$numchks] = 'verbatim'; # If we do, write it straight back out
820 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
821 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
824 # the division into chunks is imperfect;
825 # we throw some things over the fence into the next
828 # also, there are things we would like to know
829 # about the whole module before we start spitting
832 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
833 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
835 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
837 # Alphas: NB: we start meddling at chunk 1, not chunk 0
838 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
839 # spits a ".quad 0" in after the very first ".rdata"; we
840 # detect this special case (tossing the ".quad 0")!
841 local($magic_rdata_seen) = 0;
843 # HPPAs, MIPSen: also start medding at chunk 1
845 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
846 $c = $chk[$i]; # convenience copy
848 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
850 # toss all prologue stuff; HPPA is pretty weird
852 $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
854 # be slightly paranoid to make sure there's
855 # nothing surprising in there
856 if ( $c =~ /--- BEGIN ---/ ) {
857 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
859 # remove junk whitespace around the split point
863 if ($TargetPlatform =~ /^i386-/) {
864 if ($p =~ /^\tsubl\s+\$(\d+),\s*\%esp\n/) {
866 die "Error: reserved stack space exceeded!\n Possible workarounds: compile with -fasm, or try another version of gcc.\n"
870 # gcc 3.4.3 puts this kind of stuff in the prologue, eg.
871 # when compiling PrimOps.cmm with -optc-O2:
874 # movl %ecx, 16(%esp)
875 # movl %edx, 20(%esp)
876 # but then the code of the function doesn't assume
877 # anything about the contnets of these stack locations.
878 # I think it's to do with the use of inline functions for
879 # PK_Word64() and friends, where gcc is initialising the
880 # contents of the struct to zero, and failing to optimise
881 # away the initialisation. Let's live dangerously and
882 # discard these initalisations.
884 $p =~ s/^\tpushl\s+\%e(di|si|bx)\n//g;
885 $p =~ s/^\txorl\s+\%e(ax|cx|dx),\s*\%e(ax|cx|dx)\n//g;
886 $p =~ s/^\tmovl\s+\%e(ax|cx|dx|si|di),\s*\d*\(\%esp\)\n//g;
887 $p =~ s/^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n//g;
888 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
889 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-(cygwin32|mingw32)/);
891 if ($TargetPlatform =~ /^i386-apple-darwin/) {
893 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
895 $pcrel_reg =~ s/(.|\n)*.*___i686\.get_pc_thunk\.([abcd]x)\n(.|\n)*/$2/ or $pcrel_reg = "";
896 $p =~ s/^\s+call\s+___i686\.get_pc_thunk\..x//;
897 $p =~ s/^\"?L\d+\$pb\"?:\n//;
899 if ($pcrel_reg eq "bx") {
900 # Bad gcc. Goes and uses %ebx, our BaseReg, for PIC. Bad gcc.
901 die "Darwin/x86: -fPIC -via-C doesn't work yet, use -fasm. Aborting."
905 } elsif ($TargetPlatform =~ /^x86_64-/) {
906 $p =~ s/^\tpushq\s+\%r(bx|bp|12|13|14)\n//g;
907 $p =~ s/^\tmovq\s+\%r(bx|bp|12|13|14),\s*\d*\(\%rsp\)\n//g;
908 $p =~ s/^\tsubq\s+\$\d+,\s*\%rsp\n//;
910 } elsif ($TargetPlatform =~ /^ia64-/) {
911 $p =~ s/^\t\.prologue .*\n//;
912 $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, 3[12], \d+, 0\n//;
913 $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
914 $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
916 # Remove .proc and .body directives
917 $p =~ s/^\t\.proc [a-zA-Z0-9_#.]+\n//;
918 $p =~ s/^\t\.body\n//;
919 # If there's a label, move it to the body
920 if ($p =~ /^[a-zA-Z0-9#.]+:\n/) {
924 # Remove floating-point spill instructions. This is actually a bad
925 # thing to remove, because we will be putting junk into the floating-point
926 # registers and this will be visible to the caller.
927 # Only fp registers 2-5 and 16-31 may be spilled.
928 if ($p =~ s/^\tstf\.spill \[r1[4-9]\] = f([2-5]|1[6-9]|2[0-9]|30|31)(, [0-9]+)?\n//g) {
929 # Being paranoid, only try to remove these if we saw a spill
931 $p =~ s/^\tmov r1[4-9] = r12\n//;
932 $p =~ s/^\tadds r1[4-9] = -[0-9]+, r12\n//g;
933 $p =~ s/^\t\.save\.f 0x[0-9a-fA-F]\n//g;
936 $p =~ s/^\tnop\.[mifb]\s+0\n//g; # remove nop instructions
937 $p =~ s/^\t\.(mii|mmi|mfi)\n//g; # bundling is no longer sensible
938 $p =~ s/^\t;;\n//g; # discard stops
939 $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
941 # GCC 3.3 saves r1 in the prologue, move this to the body
942 if ($p =~ /^\tmov r\d+ = r1\n/) {
946 # GCC 3.2 saves pr in the prologue, move this to the body
947 if ($p =~ /^\tmov r\d+ = pr\n/) {
951 } elsif ($TargetPlatform =~ /^m68k-/) {
952 $p =~ s/^\tlink a6,#-?\d.*\n//;
953 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
954 # The above showed up in the asm code,
955 # so I added it here.
956 # I hope it's correct.
958 $p =~ s/^\tmovel d2,sp\@-\n//;
959 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
960 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
961 } elsif ($TargetPlatform =~ /^mips-/) {
962 # the .frame/.mask/.fmask that we use is the same
963 # as that produced by GCC for miniInterpret; this
964 # gives GDB some chance of figuring out what happened
965 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
966 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
967 $p =~ s/^\t\.(mask|fmask).*\n//g;
968 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
969 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
970 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
971 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
972 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
973 $p =~ s/__FRAME__/$FRAME/;
974 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
976 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
978 $p =~ s/^\tmflr r0\n//;
979 $p =~ s/^\tbl saveFP # f\d+\n//;
980 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
981 $p =~ s/^\"?L\d+\$pb\"?:\n//;
982 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
983 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
984 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
985 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
986 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
987 $p =~ s/^\tbcl 20,31,\"?L\d+\$pb\"?\n//;
988 $p =~ s/^\"?L\d+\$pb\"?:\n//;
989 $p =~ s/^\tmflr r31\n//;
991 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
992 # under some circumstances, only when generating position dependent code.
993 # I have no idea why, and I don't think it is necessary, so let's toss it.
994 $p =~ s/^\tli r\d+,0\n//g;
995 $p =~ s/^\tstw r\d+,\d+\(r1\)\n//g;
996 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/) {
997 $p =~ s/^\tmflr 0\n//;
998 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
999 $p =~ s/^\tstfd \d+,\d+\(1\)\n//g;
1000 $p =~ s/^\tstw r0,8\(1\)\n//;
1001 $p =~ s/^\tstwu 1,-\d+\(1\)\n//;
1002 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1004 # GCC's "large-model" PIC (-fPIC)
1006 $pcrel_label =~ s/(.|\n)*^.LCF(\d+):\n(.|\n)*/$2/ or $pcrel_label = "";
1008 $p =~ s/^\tbcl 20,31,.LCF\d+\n//;
1009 $p =~ s/^.LCF\d+:\n//;
1010 $p =~ s/^\tmflr 30\n//;
1011 $p =~ s/^\tlwz 0,\.LCL\d+-\.LCF\d+\(30\)\n//;
1012 $p =~ s/^\tadd 30,0,30\n//;
1014 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1015 # under some circumstances, only when generating position dependent code.
1016 # I have no idea why, and I don't think it is necessary, so let's toss it.
1017 $p =~ s/^\tli \d+,0\n//g;
1018 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1019 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1020 $p =~ s/^\tmr 31,1\n//;
1021 $p =~ s/^\tmflr 0\n//;
1022 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
1023 $p =~ s/^\tstfd \d+,-?\d+\(1\)\n//g;
1024 $p =~ s/^\tstd r0,8\(1\)\n//;
1025 $p =~ s/^\tstdu 1,-\d+\(1\)\n//;
1026 $p =~ s/^\tstd \d+,-?\d+\(1\)\n//g;
1028 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1029 # under some circumstances, only when generating position dependent code.
1030 # I have no idea why, and I don't think it is necessary, so let's toss it.
1031 $p =~ s/^\tli \d+,0\n//g;
1032 $p =~ s/^\tstd \d+,\d+\(1\)\n//g;
1034 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
1037 # HWL HACK: dont die, just print a warning
1038 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
1039 die "Prologue junk?: $p\n" if $p =~ /^\s+[^\s\.]/;
1041 # For PIC, we want to keep part of the prologue
1042 if ($TargetPlatform =~ /^powerpc-apple-darwin.*/ && $pcrel_label ne "") {
1043 # Darwin: load the current instruction pointer into register r31
1044 $p .= "bcl 20,31,$pcrel_label\n";
1045 $p .= "$pcrel_label:\n";
1046 $p .= "\tmflr r31\n";
1047 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/ && $pcrel_label ne "") {
1048 # Linux: load the GOT pointer into register 30
1049 $p .= "\tbcl 20,31,.LCF$pcrel_label\n";
1050 $p .= ".LCF$pcrel_label:\n";
1051 $p .= "\tmflr 30\n";
1052 $p .= "\tlwz 0,.LCL$pcrel_label-.LCF$pcrel_label(30)\n";
1053 $p .= "\tadd 30,0,30\n";
1054 } elsif ($TargetPlatform =~ /^i386-apple-darwin.*/ && $pcrel_label ne "") {
1055 $p .= "\tcall ___i686.get_pc_thunk.$pcrel_reg\n";
1056 $p .= "$pcrel_label:\n";
1059 # glue together what's left
1064 if ( $TargetPlatform =~ /^mips-/ ) {
1065 # MIPS: first, this basic sequence may occur "--- END ---" or not
1066 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
1069 # toss all epilogue stuff; again, paranoidly
1070 if ( $c =~ /--- END ---/ ) {
1071 if (($r, $e) = split(/--- END ---/, $c)) {
1072 # rtail holds code that is after the epilogue in the assembly-code
1073 # layout and should not be filtered as part of the epilogue.
1075 if ($TargetPlatform =~ /^i386-/) {
1077 $e =~ s/^\tpopl\s+\%edi\n//;
1078 $e =~ s/^\tpopl\s+\%esi\n//;
1079 $e =~ s/^\tpopl\s+\%edx\n//;
1080 $e =~ s/^\tpopl\s+\%ecx\n//;
1081 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
1082 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
1083 } elsif ($TargetPlatform =~ /^ia64-/) {
1084 # GCC may have put the function's epilogue code in the _middle_
1085 # of the function. We try to detect that here and extract the
1086 # code that belongs to the body of the function. We'll put that
1087 # code back after cleaning up the epilogue.
1088 # The epilogue is first split into:
1089 # $e, the epilogue code (up to the return instruction)
1090 # $rtail, the rest of the function body
1091 # $edir, the directives following the function
1092 # (everything starting with .endp)
1093 # The return instruction and endp directive are stripped in the
1095 if (!(($e, $rtail) = split(/^\tbr\.ret\.sptk\.many b0\n/, $e))) {
1096 die "Epilogue doesn't seem to have one return instruction: $e\n";
1098 if (!(($rtail, $edir) = split(/^\t\.endp [a-zA-Z0-9_#.]+\n/, $rtail))) {
1099 die "Epilogue doesn't seem to have one endp directive: $e\n";
1101 # print STDERR "Epilogue: $e\n";
1102 # print STDERR "Code tail: $rtail\n";
1103 # print STDERR "Directives: $edir\n";
1105 # If a return value is saved here, move it to the function body
1106 if ($e =~ /^\tmov r8 = r14\n/) {
1111 # Remove floating-point fill instructions. This is actually a bad
1112 # thing to remove, because we will be putting junk into the
1113 # floating-point registers and this will be visible to the caller.
1114 # Only fp registers 2-5 and 16-31 may be restored.
1115 if ($e =~ s/^\tldf\.fill f([2-5]|1[6-9]|2[0-9]|30|31) = \[r1[4-9]\](, [0-9]+)?\n//g) {
1116 # Being paranoid, only try to remove this if we saw a fill
1118 $e =~ s/^\tadds r1[4-9] = [0-9]+, r12//g;
1121 $e =~ s/^\tnop\.[mifb]\s+0\n//g; # remove nop instructions
1123 $e =~ s/^\tmov ar\.pfs = r\d+\n//;
1124 $e =~ s/^\tmov b0 = r\d+\n//;
1125 $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
1126 #$e =~ s/^\tbr\.ret\.sptk\.many b0\n//; # already removed
1127 $e =~ s/^\t\.(mii|mmi|mfi|mib)\n//g; # bundling is no longer sensible
1128 $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
1129 $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
1131 # Tack edir onto the end of rtail. Some of the directives in
1132 # edir are relevant to the next chunk.
1134 } elsif ($TargetPlatform =~ /^m68k-/) {
1135 $e =~ s/^\tunlk a6\n//;
1137 } elsif ($TargetPlatform =~ /^mips-/) {
1138 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
1139 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
1140 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
1141 $e =~ s/^\tj\t\$31\n//;
1142 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
1143 $e =~ s/^\taddi r1,r1,\d+\n//;
1144 $e =~ s/^\tlwz r\d+,\d+\(r1\)\n//;
1145 $e =~ s/^\tlmw r\d+,-\d+\(r1\)\n//;
1146 $e =~ s/^\tmtlr r0\n//;
1148 $e =~ s/^\tb restFP ;.*\n//;
1149 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1150 $e =~ s/^\tmr 3,0\n//;
1151 $e =~ s/^\taddi 1,1,\d+\n//;
1152 $e =~ s/^\tld 0,16\(1\)\n//;
1153 $e =~ s/^\tmtlr 0\n//;
1155 # callee-save registers
1156 $e =~ s/^\tld \d+,-?\d+\(1\)\n//g;
1157 $e =~ s/^\tlfd \d+,-?\d+\(1\)\n//g;
1159 # get rid of the debug junk along with the blr
1160 $e =~ s/^\tblr\n\t.long .*\n\t.byte .*\n//;
1162 # incase we missed it with the last one get the blr alone
1165 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
1168 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
1170 # glue together what's left
1171 $c = $r . $e . $rtail;
1172 $c =~ s/\n\t\n/\n/; # junk blank line
1176 # On SPARCs, we don't do --- BEGIN/END ---, we just
1177 # toss the register-windowing save/restore/ret* instructions
1178 # directly unless they've been generated by function definitions in header
1180 if ( $TargetPlatform =~ /^sparc-/ ) {
1181 if ( ! ( $TargetPlatform =~ /solaris2$/ && $chkcat[$i] eq 'unknown' )) {
1182 $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
1184 # throw away PROLOGUE comments
1185 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
1188 # On Alphas, the prologue mangling is done a little later (below)
1190 # toss all calls to __DISCARD__
1191 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
1192 $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
1193 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-darwin.*/;
1194 $c =~ s/^\tbl\s+__DISCARD__(\@plt)?\n//go if $TargetPlatform =~ /^powerpc-.*-linux/;
1195 $c =~ s/^\tbl\s+\.__DISCARD__\n\s+nop\n//go if $TargetPlatform =~ /^powerpc64-.*-linux/;
1196 $c =~ s/^\tcall\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /i386-apple-darwin.*/;
1198 # IA64: mangle tailcalls into jumps here
1199 if ($TargetPlatform =~ /^ia64-/) {
1200 # Example of what is mangled:
1201 # br.call.sptk.many b0 = b6
1213 while ($c =~ s/^\tbr\.call\.sptk\.many b0 = (.*)\n(?:^\.L([0-9]*):\n)?(?:\t;;\n)?(?:\t\.(?:mii|mmi|mfi|mfb)\n)?(?:\tmov r1 = r\d+\n)?(?:\t;;\n)?(?:\tnop\.[mifb] \d+\n)*\t--- TAILCALL ---\n(?:\t;;\n\tbr \.L\d+\n)?/\tbr\.few $1\n/) {
1214 # Eek, the gcc optimiser is getting smarter... if we see a jump to the --- TAILCALL ---
1215 # marker then we reapply the substitution at the source sites
1216 $c =~ s/^\tbr \.L$2\n/\t--- TAILCALL ---\n/g if ($2);
1219 # Verify that all instances of TAILCALL were processed
1220 if ($c =~ /^\t--- TAILCALL ---\n/) {
1221 die "Unmangled TAILCALL tokens remain after mangling"
1225 # MIPS: that may leave some gratuitous asm macros around
1226 # (no harm done; but we get rid of them to be tidier)
1227 $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/
1228 if $TargetPlatform =~ /^mips-/;
1230 # toss stack adjustment after DoSparks
1231 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
1232 if $TargetPlatform =~ /^m68k-/; # this looks old...
1234 if ( $TargetPlatform =~ /^alpha-/ &&
1235 ! $magic_rdata_seen &&
1236 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
1237 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
1238 $magic_rdata_seen = 1;
1241 # pick some end-things and move them to the next chunk
1243 # pin a funny end-thing on (for easier matching):
1244 $c .= 'FUNNY#END#THING';
1246 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
1250 # on x86 we try not to copy any directives into a literal
1251 # chunk, rather we keep looking for the next real chunk. This
1252 # is because we get things like
1254 # .globl blah_closure
1260 if ( $TargetPlatform =~ /^(i386|sparc|powerpc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
1262 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
1265 if ( $j < $numchks ) {
1266 $chk[$j] = $to_move . $chk[$j];
1270 elsif ( ( $i < ($numchks - 1)
1271 && ( $to_move =~ /${T_COPY_DIRVS}/
1272 || ( $TargetPlatform =~ /^hppa/
1273 && $to_move =~ /align/
1274 && $chkcat[$i+1] eq 'literal')
1277 || ($to_move =~ /^[ \t]*\.section[ \t]+\.note\.GNU-stack,/)
1279 $chk[$i + 1] = $to_move . $chk[$i + 1];
1280 # otherwise they're tossed
1283 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
1286 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
1288 # toss all prologue stuff, except for loading gp, and the ..ng address
1289 unless ($c =~ /\.ent.*\n\$.*\.\.ng:/) {
1290 if (($p, $r) = split(/^\t\.prologue/, $c)) {
1291 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
1292 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
1293 $keep =~ s/^\t\.(mask|fmask).*\n//g;
1294 $c = $keep . "..ng:\n";
1296 print STDERR "malformed code block ($ent)?\n"
1299 $c .= "\t.prologue" . $r;
1303 $c =~ s/FUNNY#END#THING//;
1305 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
1307 $chk[$i] = $c; # update w/ convenience copy
1310 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
1311 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
1314 if ( $TargetPlatform =~ /^alpha-/ ) {
1315 # print out the header stuff first
1316 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
1317 print OUTASM $chk[0];
1319 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1320 print OUTASM $chk[0];
1322 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1323 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
1325 # get rid of horrible "<dollar>Revision: .*$" strings
1326 local(@lines0) = split(/\n/, $chk[0]);
1328 while ( $z <= $#lines0 ) {
1329 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
1332 while ( $z <= $#lines0 ) {
1334 last if $lines0[$z] =~ /[,\t]0x0$/;
1340 $chk[0] = join("\n", @lines0);
1341 $chk[0] =~ s/\n\n+/\n/;
1342 print OUTASM $chk[0];
1345 # print out all the literal strings next
1346 for ($i = 0; $i < $numchks; $i++) {
1347 if ( $chkcat[$i] eq 'literal' ) {
1349 # HACK: try to detect 16-byte constants and align them
1350 # on a 16-byte boundary. x86_64 sometimes needs 128-bit
1351 # aligned constants, and so does Darwin/x86.
1352 if ( $TargetPlatform =~ /^x86_64/
1353 || $TargetPlatform =~ /^i386-apple-darwin/ ) {
1355 if ($z =~ /(\.long.*\n.*\.long.*\n.*\.long.*\n.*\.long|\.quad.*\n.*\.quad)/) {
1356 print OUTASM $T_HDR_literal16;
1358 print OUTASM $T_HDR_literal;
1361 print OUTASM $T_HDR_literal;
1364 print OUTASM $chk[$i];
1365 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
1367 $chkcat[$i] = 'DONE ALREADY';
1371 # on the HPPA, print out all the bss next
1372 if ( $TargetPlatform =~ /^hppa/ ) {
1373 for ($i = 1; $i < $numchks; $i++) {
1374 if ( $chkcat[$i] eq 'bss' ) {
1375 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
1376 print OUTASM $chk[$i];
1378 $chkcat[$i] = 'DONE ALREADY';
1383 # $numchks + 1 as we have the extra one for .note.GNU-stack
1384 for ($i = $FIRST_MANGLABLE; $i < $numchks + 1; $i++) {
1385 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
1387 next if $chkcat[$i] eq 'DONE ALREADY';
1389 if ( $chkcat[$i] eq 'misc' || $chkcat[$i] eq 'unknown' ) {
1390 if ($chk[$i] ne '') {
1391 print OUTASM $T_HDR_misc;
1392 &print_doctored($chk[$i], 0);
1395 } elsif ( $chkcat[$i] eq 'verbatim' ) {
1396 print OUTASM $chk[$i];
1398 } elsif ( $chkcat[$i] eq 'toss' ) {
1399 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
1401 } elsif ( $chkcat[$i] eq 'data' ) {
1402 if ($chk[$i] ne '') {
1403 print OUTASM $T_HDR_data;
1404 print OUTASM $chk[$i];
1407 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1408 # we can just re-constitute this one...
1409 # NB: we emit _three_ underscores no matter what,
1410 # so ghc-split doesn't have to care.
1411 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1413 } elsif ( $chkcat[$i] eq 'closure'
1414 || $chkcat[$i] eq 'srt'
1415 || $chkcat[$i] eq 'infotbl'
1416 || $chkcat[$i] eq 'entry') { # do them in that order
1417 $symb = $chksymb[$i];
1420 if ( defined($closurechk{$symb}) ) {
1421 print OUTASM $T_HDR_closure;
1422 print OUTASM $chk[$closurechk{$symb}];
1423 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1427 if ( defined($srtchk{$symb}) ) {
1428 print OUTASM $T_HDR_relrodata;
1429 print OUTASM $chk[$srtchk{$symb}];
1430 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1434 if ( defined($infochk{$symb}) ) {
1436 print OUTASM $T_HDR_info;
1437 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1439 # entry code will be put here!
1441 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1445 if ( defined($entrychk{$symb}) ) {
1447 $c = $chk[$entrychk{$symb}];
1449 # If this is an entry point with an info table,
1450 # eliminate the entry symbol and all directives involving it.
1451 if (defined($infochk{$symb}) && $TargetPlatform !~ /^ia64-/) {
1453 foreach $l (split(/\n/,$c)) {
1454 next if $l =~ /^.*$symb_(entry|ret)${T_POST_LBL}/;
1456 # If we have .type/.size direrctives involving foo_entry,
1457 # then make them refer to foo_info instead. The information
1458 # in these directives is used by the cachegrind annotator,
1459 # so it is worthwhile keeping.
1460 if ($l =~ /^\s*\.(type|size).*$symb_(entry|ret)/) {
1461 $l =~ s/$symb(_entry|_ret)/${symb}_info/g;
1465 next if $l =~ /^\s*\..*$symb.*\n?/;
1468 $c = join("\n",@o) . "\n";
1471 print OUTASM $T_HDR_entry;
1473 &print_doctored($c, 1); # NB: the 1!!!
1475 $chkcat[$entrychk{$symb}] = 'DONE ALREADY';
1478 } elsif ( $chkcat[$i] eq 'vector' ) {
1479 $symb = $chksymb[$i];
1482 if ( defined($vectorchk{$symb}) ) {
1483 print OUTASM $T_HDR_vector;
1484 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1486 # direct return code will be put here!
1487 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1489 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1490 # Alphas: the commented nop is for the splitter, to ensure
1491 # that no module ends with a label as the very last
1492 # thing. (The linker will adjust the label to point
1493 # to the first code word of the next module linked in,
1494 # even if alignment constraints cause the label to move!)
1496 print OUTASM "\t# nop\n";
1499 } elsif ( $chkcat[$i] eq 'rodata' ) {
1500 print OUTASM $T_HDR_rodata;
1501 print OUTASM $chk[$i];
1502 $chkcat[$i] = 'DONE ALREADY';
1503 } elsif ( $chkcat[$i] eq 'relrodata' ) {
1504 print OUTASM $T_HDR_relrodata;
1505 print OUTASM $chk[$i];
1506 $chkcat[$i] = 'DONE ALREADY';
1507 } elsif ( $chkcat[$i] eq 'toc' ) {
1508 # silly optimisation to print tocs, since they come in groups...
1509 print OUTASM $T_HDR_toc;
1511 while ($chkcat[$j] eq 'toc')
1512 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1515 print OUTASM $chk[$j];
1517 $chkcat[$j] = 'DONE ALREADY';
1521 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' ) {
1522 # apple-darwin: dynamic linker stubs
1523 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1524 { # print them out unchanged, but remove the stubs for __DISCARD__
1525 print OUTASM $chk[$i];
1527 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ && $chkcat[$i] eq 'get_pc_thunk' ) {
1528 # i386-apple-darwin: __i686.get_pc_thunk.[abcd]x
1529 print OUTASM ".section __TEXT,__textcoal_nt,coalesced,no_toc\n";
1530 print OUTASM $chk[$i];
1532 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1536 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1539 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1540 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1545 sub hppa_mash_prologue { # OK, epilogue, too
1548 # toss all prologue stuff
1549 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1551 # Lie about our .CALLINFO
1552 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1559 # toss all epilogue stuff
1560 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1562 # Sorry; we moved the _info stuff to the code segment.
1563 s/_info,DATA/_info,CODE/g;
1570 sub print_doctored {
1571 local($_, $need_fallthru_patch) = @_;
1573 if ( $TargetPlatform =~ /^x86_64-/ ) {
1576 # movq -4(%ebp), %rax
1581 s/^\tmovq\s+(-?\d*\(\%r(bx|bp|13)\)),\s*(\%r(ax|cx|dx|10|11))\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1582 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e(ax|cx|si|di))\n\tjmp\s+\*\%r\3/\tjmp\t$T_US$1/g;
1585 if ( $TargetPlatform !~ /^i386-/
1586 || ! /^\t[a-z]/ # no instructions in here, apparently
1587 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1592 # OK, must do some x86 **HACKING**
1594 local($entry_patch) = '';
1595 local($exit_patch) = '';
1597 # gotta watch out for weird instructions that
1598 # invisibly smash various regs:
1599 # rep* %ecx used for counting
1600 # scas* %edi used for destination index
1601 # cmps* %e[sd]i used for indices
1602 # loop* %ecx used for counting
1607 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1609 # * GCC used an "STG reg" for its own purposes
1611 # * some secret uses of machine reg, requiring STG reg
1612 # to be saved/restored
1614 # The most dangerous "GCC uses" of an "STG reg" are when
1615 # the reg holds the target of a jmp -- it's tricky to
1616 # insert the patch-up code before we get to the target!
1617 # So here we change the jmps:
1619 # --------------------------------------------------------
1620 # it can happen that we have jumps of the form...
1621 # jmp *<something involving %esp>
1623 # jmp <something involving another naughty register...>
1625 # a reasonably-common case is:
1627 # movl $_blah,<bad-reg>
1630 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[acd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1634 # movl -4(%ebx), %eax
1639 s/^\tmovl\s+(-?\d*\(\%e(bx|si)\)),\s*(\%e[acd]x)\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1641 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1642 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1643 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1644 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1645 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1646 if /(jmp|call)\s+.*\%esi/;
1648 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1649 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1650 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1651 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1652 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1653 if /(jmp|call)\s+.*\%edi/;
1656 # OK, now we can decide what our patch-up code is going to
1659 # Offsets into register table - you'd better update these magic
1660 # numbers should you change its contents!
1661 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1662 local($OFFSET_Hp)=88;
1664 # Note funky ".=" stuff; we're *adding* to these _patch guys
1665 if ( $StolenX86Regs <= 2
1666 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1667 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1668 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1670 # nothing for call_{entry,exit} because %esi is callee-save
1672 if ( $StolenX86Regs <= 3
1673 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1674 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1675 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1677 # nothing for call_{entry,exit} because %edi is callee-save
1680 # --------------------------------------------------------
1681 # next, here we go with non-%esp patching!
1683 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1685 # Before calling GC we must set up the exit condition before the call
1686 # and entry condition when we come back
1688 # fix _all_ non-local jumps:
1690 if ( $TargetPlatform =~ /^.*-apple-darwin.*/ ) {
1691 # On Darwin, we've got local-looking jumps that are
1692 # actually global (i.e. jumps to Lfoo$stub or via
1693 # Lfoo$non_lazy_ptr), so we fix those first.
1694 # In fact, we just fix everything that contains a dollar
1695 # because false positives don't hurt here.
1697 s/^(\tjmp\s+\*?L.*\$.*\n)/$exit_patch$1/g;
1700 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1701 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1703 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1705 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1706 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1708 if ($StolenX86Regs == 2 ) {
1709 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1710 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1711 } elsif ($StolenX86Regs == 3 ) {
1712 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1713 if /^\t(jmp|call)\s+.*\%edi/;
1716 # --------------------------------------------------------
1717 # that's it -- print it
1719 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1723 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1724 print OUTASM $exit_patch;
1725 # ToDo: make it not print if there is a "jmp" at the end
1731 sub init_FUNNY_THINGS {
1732 %KNOWN_FUNNY_THING = (
1734 # "${T_US}stg_.*{T_POST_LBL}", 1,
1739 The following table reversal is used for both info tables and return
1740 vectors. In both cases, we remove the first entry from the table,
1741 reverse the table, put the label at the end, and paste some code
1742 (that which is normally referred to by the first entry in the table)
1743 right after the table itself. (The code pasting is done elsewhere.)
1747 local($symb, $tbl, $discard1) = @_;
1749 return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1751 local($before) = '';
1753 local(@imports) = (); # hppa only
1756 local(@lines) = split(/\n/, $tbl);
1759 # Deal with the header...
1760 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1761 $label .= $lines[$i] . "\n",
1762 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1763 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1764 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1766 $before .= $lines[$i] . "\n"; # otherwise...
1770 $infoname =~ s/(.|\n)*^([A-Za-z0-9_]+_info)${T_POST_LBL}$(.|\n)*/\2/;
1772 # Grab the table data...
1773 if ( $TargetPlatform !~ /^hppa/ ) {
1774 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1776 # Convert addresses of SRTs, slow entrypoints and large bitmaps
1777 # to offsets (relative to the info label),
1778 # in order to support position independent code.
1779 $line =~ s/$infoname/0/
1780 || $line =~ s/([A-Za-z0-9_]+_srtd)$/\1 - $infoname/
1781 || $line =~ s/([A-Za-z0-9_]+_srt(\+\d+)?)$/\1 - $infoname/
1782 || $line =~ s/([A-Za-z0-9_]+_str)$/\1 - $infoname/
1783 || $line =~ s/([A-Za-z0-9_]+_slow)$/\1 - $infoname/
1784 || $line =~ s/([A-Za-z0-9_]+_btm)$/\1 - $infoname/
1785 || $line =~ s/([A-Za-z0-9_]+_alt)$/\1 - $infoname/
1786 || $line =~ s/([A-Za-z0-9_]+_dflt)$/\1 - $infoname/
1787 || $line =~ s/([A-Za-z0-9_]+_ret)$/\1 - $infoname/;
1788 push(@words, $line);
1790 } else { # hppa weirdness
1791 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1792 # FIXME: the RTS now expects offsets instead of addresses
1793 # for all labels in info tables.
1794 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1795 push(@imports, $lines[$i]);
1797 # We don't use HP's ``function pointers''
1798 # We just use labels in code space, like normal people
1799 $lines[$i] =~ s/P%//;
1800 push(@words, $lines[$i]);
1805 # Now throw away any initial zero word from the table. This is a hack
1806 # that lets us reduce the size of info tables when the SRT field is not
1807 # needed: see comments StgFunInfoTable in InfoTables.h.
1809 # The .zero business is for Linux/ELF.
1810 # The .skip business is for Sparc/Solaris/ELF.
1811 # The .blockz business is for HPPA.
1813 # if ($words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1818 for (; $i <= $#lines; $i++) {
1819 $after .= $lines[$i] . "\n";
1822 # Alphas: If we have anonymous text (not part of a procedure), the
1823 # linker may complain about missing exception information. Bleh.
1824 # To suppress this, we place a .ent/.end pair around the code.
1825 # At the same time, we have to be careful and not enclose any leading
1826 # .file/.loc directives.
1827 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1828 local ($ident) = $1;
1829 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1830 $after .= "\t.end $ident\n";
1833 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1834 # assembler (!) wherein .quad constants inside .text sections are
1835 # first narrowed to 32 bits then sign-extended back to 64 bits.
1836 # This obviously screws up our 64-bit bitmaps, so we work around
1837 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1838 if ( $TargetPlatform =~ /^alpha-/ ) {
1840 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1841 local ($number) = $1;
1842 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1843 local ($sign, $base, $digits) = ($1, $2, $3);
1844 $base = (10, 8, 16)[length $base];
1845 local ($hi, $lo) = (0, 0);
1846 foreach $i (split(//, $digits)) {
1847 $j = $lo * $base + $i;
1848 $lo = $j % 4294967296;
1849 $hi = $hi * $base + ($j - $lo) / 4294967296;
1851 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1853 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1854 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1856 print STDERR "Cannot handle \".quad $number\" in info table\n";
1863 if ( $TargetPlatform =~ /x86_64-apple-darwin/ ) {
1864 # Tack a label to the front of the info table, too.
1865 # For now, this just serves to work around a crash in Apple's new
1866 # 64-bit linker (it seems to assume that there is no data before the
1867 # first label in a section).
1869 # The plan for the future is to do this on all Darwin platforms, and
1870 # to add a reference to this label after the entry code, just as the
1871 # NCG does, so we can enable dead-code-stripping in the linker without
1872 # losing our info tables. (Hence the name _dsp, for dead-strip preventer)
1874 $before .= "\n${infoname}_dsp:\n";
1878 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
1879 . join("\n", @words) . "\n"
1882 # print STDERR "before=$before\n";
1883 # print STDERR "label=$label\n";
1884 # print STDERR "words=",(reverse @words),"\n";
1885 # print STDERR "after=$after\n";
1891 The HP is a major nuisance. The threaded code mangler moved info
1892 tables from data space to code space, but unthreaded code in the RTS
1893 still has references to info tables in data space. Since the HP
1894 linker is very precise about where symbols live, we need to patch the
1895 references in the unthreaded RTS as well.
1898 sub mini_mangle_asm_hppa {
1899 local($in_asmf, $out_asmf) = @_;
1901 open(INASM, "< $in_asmf")
1902 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1903 open(OUTASM,"> $out_asmf")
1904 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1907 s/_info,DATA/_info,CODE/; # Move _info references to code space
1913 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1914 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1920 sub tidy_up_and_die {
1921 local($return_val, $msg) = @_;
1923 exit (($return_val == 0) ? 0 : 1);