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:
542 # ia64-specific information for code chunks
546 &init_TARGET_STUFF();
547 &init_FUNNY_THINGS();
549 open(INASM, "< $in_asmf")
550 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
551 open(OUTASM,"> $out_asmf")
552 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
554 # read whole file, divide into "chunks":
555 # record some info about what we've found...
557 @chk = (); # contents of the chunk
558 $numchks = 0; # number of them
559 @chkcat = (); # what category of thing in each chunk
560 @chksymb = (); # what symbol(base) is defined in this chunk
561 %entrychk = (); # ditto, its entry code
562 %closurechk = (); # ditto, the (static) closure
563 %srtchk = (); # ditto, its SRT (for top-level things)
564 %infochk = (); # given a symbol base, say what chunk its info tbl is in
565 %vectorchk = (); # ditto, return vector table
566 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
568 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
571 tr/\r//d if $TargetPlatform =~ /-mingw32$/; # In case Perl doesn't convert line endings
572 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
573 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
574 next if /^\t\.def.*endef$/;
575 next if /${T_PRE_APP}(NO_)?APP/o;
576 next if /^;/ && $TargetPlatform =~ /^hppa/;
578 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|ia64)-/;
580 if ( $TargetPlatform =~ /^mips-/
581 && /^\t\.(globl\S+\.text|comm\t)/ ) {
582 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
583 # Treat .comm variables as data. These show up in two (known) places:
585 # - the module_registered variable used in the __stginit fragment.
586 # even though these are declared static and initialised, gcc 3.3
587 # likes to make them .comm, presumably to save space in the
590 # - global variables used to pass arguments from C to STG in
591 # a foreign export. (is this still true? --SDM)
593 } elsif ( /^\t\.comm.*$/ ) {
595 $chkcat[$i] = 'data';
598 # Labels ending "_str": these are literal strings.
599 } elsif ( /^${T_US}([A-Za-z0-9_]+)_str${T_POST_LBL}$/ ) {
601 $chkcat[$i] = 'rodata';
603 } elsif ( $TargetPlatform =~ /-darwin/
604 && (/^\s*\.subsections_via_symbols/
605 ||/^\s*\.no_dead_strip.*/)) {
606 # Don't allow Apple's linker to do any dead-stripping of symbols
607 # in this file, because it will mess up info-tables in mangled
609 # The .no_dead_strip directives are actually put there by
610 # the gcc3 "used" attribute on entry points.
612 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && (
613 /^\s*\.picsymbol_stub/
614 || /^\s*\.section __TEXT,__picsymbol_stub\d,.*/
615 || /^\s*\.section __TEXT,__picsymbolstub\d,.*/
616 || /^\s*\.symbol_stub/
617 || /^\s*\.section __TEXT,__symbol_stub\d,.*/
618 || /^\s*\.section __TEXT,__symbolstub\d,.*/
619 || /^\s*\.lazy_symbol_pointer/
620 || /^\s*\.non_lazy_symbol_pointer/
621 || /^\s*\.section __IMPORT.*/))
624 $chkcat[$i] = 'dyld';
628 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.data/)
629 { # non_lazy_symbol_ptrs that point to local symbols
631 $chkcat[$i] = 'dyld';
634 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.align/)
635 { # non_lazy_symbol_ptrs that point to local symbols
637 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^L_.*:$/)
638 { # non_lazy_symbol_ptrs that point to local symbols
639 $chk[++$i] = $dyld_section . $_;
640 $chkcat[$i] = 'dyld';
643 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
644 # duplicated from the bottom
648 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
649 # Alphas: Local labels not to be confused with new chunks
651 # NB: all the rest start with a non-space
653 } elsif ( $TargetPlatform =~ /^mips-/
654 && /^\d+:/ ) { # a funny-looking very-local label
657 } elsif ( /$T_CONST_LBL/o ) {
659 $chkcat[$i] = 'literal';
662 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
664 $chkcat[$i] = 'splitmarker';
667 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
670 $chkcat[$i] = 'infotbl';
671 $chksymb[$i] = $symb;
673 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
675 $infochk{$symb} = $i;
677 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
679 $chkcat[$i] = 'entry';
684 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
686 $chkcat[$i] = 'closure';
689 $closurechk{$1} = $i;
691 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
698 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
700 $chkcat[$i] = 'data';
703 } elsif ( /^${T_US}(stg_ap_stack_entries|stg_stack_save_entries|stg_arg_bitmaps)${T_POST_LBL}$/o ) {
705 $chkcat[$i] = 'data';
708 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
711 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
712 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
713 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
714 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
717 $chkcat[$i] = 'data';
720 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
725 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
726 # all CC_ symbols go in the data section...
728 $chkcat[$i] = 'data';
731 } elsif ( /^${T_US}([A-Za-z0-9_]+)_hpc${T_POST_LBL}$/o ) {
732 # hpc shares tick boxes across modules
734 $chkcat[$i] = 'data';
737 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
739 $chkcat[$i] = 'misc';
741 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
743 $chkcat[$i] = 'vector';
748 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
749 && /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
750 # Some Solaris system headers contain function definitions (as
751 # opposed to mere prototypes), which end up in the .hc file when
752 # a Haskell module foreign imports the corresponding system
753 # functions (most notably stat()). We put them into the text
754 # segment. Note that this currently does not extend to function
755 # names starting with an underscore.
758 $chkcat[$i] = 'misc';
761 } elsif ( $TargetPlatform =~ /^i386-apple-darwin/ && /^(___i686\.get_pc_thunk\.[abcd]x):/o) {
762 # To handle PIC on Darwin/x86, we need to appropriately pass through
763 # the get_pc_thunk functions. The need to be put into a special section
764 # marked as coalesced (otherwise the .weak_definition doesn't work
767 $chkcat[$i] = 'get_pc_thunk';
770 } elsif ( /^${T_US}[A-Za-z0-9_]/o
771 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
773 && ( $TargetPlatform !~ /^powerpc64/ # we need to avoid local labels in this case
774 || ! /^\.L\d+:$/ ) ) {
781 /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
782 || /^${T_US}__stg_.*${T_POST_LBL}$/o # more RTS internals
783 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
784 || /^${T_US}.*_slow${T_POST_LBL}$/o # slow entry
785 || /^${T_US}__stginit.*${T_POST_LBL}$/o # __stginit<module>
786 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
787 || /^${T_US}.*_fast${T_POST_LBL}$/o # primops
788 || /^_uname:/o # x86/Solaris2
791 $chkcat[$i] = 'misc';
793 /^${T_US}.*_srtd${T_POST_LBL}$/o # large bitmaps
794 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
797 $chkcat[$i] = 'relrodata';
800 print STDERR "Warning: retaining unknown function \`$thing' in output from C compiler\n";
801 $chkcat[$i] = 'unknown';
804 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ && /^\.LCTOC1 = /o ) {
805 # PowerPC Linux's large-model PIC (-fPIC) generates a gobal offset
806 # table "by hand". Be sure to copy it over.
807 # Note that this label and all entries in the table should actually
808 # go into the .got2 section, but it isn't easy to distinguish them
809 # from other constant literals (.LC\d+), so we just put everything
812 $chkcat[$i] = 'literal';
813 $chksymb[$i] = 'LCTOC1';
814 } else { # simple line (duplicated at the top)
819 $numchks = $#chk + 1;
820 $chk[$numchks] = ''; # We might push .note.GNU-stack into this
821 $chkcat[$numchks] = 'verbatim'; # If we do, write it straight back out
823 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
824 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
827 # the division into chunks is imperfect;
828 # we throw some things over the fence into the next
831 # also, there are things we would like to know
832 # about the whole module before we start spitting
835 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
836 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
838 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
840 # Alphas: NB: we start meddling at chunk 1, not chunk 0
841 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
842 # spits a ".quad 0" in after the very first ".rdata"; we
843 # detect this special case (tossing the ".quad 0")!
844 local($magic_rdata_seen) = 0;
846 # HPPAs, MIPSen: also start medding at chunk 1
848 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
849 $c = $chk[$i]; # convenience copy
851 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
853 # toss all prologue stuff; HPPA is pretty weird
855 $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
860 # be slightly paranoid to make sure there's
861 # nothing surprising in there
862 if ( $c =~ /--- BEGIN ---/ ) {
863 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
865 # remove junk whitespace around the split point
869 if ($TargetPlatform =~ /^i386-/) {
870 if ($p =~ /^\tsubl\s+\$(\d+),\s*\%esp\n/) {
872 die "Error: reserved stack space exceeded!\n Possible workarounds: compile with -fasm, or try another version of gcc.\n"
876 # gcc 3.4.3 puts this kind of stuff in the prologue, eg.
877 # when compiling PrimOps.cmm with -optc-O2:
880 # movl %ecx, 16(%esp)
881 # movl %edx, 20(%esp)
882 # but then the code of the function doesn't assume
883 # anything about the contnets of these stack locations.
884 # I think it's to do with the use of inline functions for
885 # PK_Word64() and friends, where gcc is initialising the
886 # contents of the struct to zero, and failing to optimise
887 # away the initialisation. Let's live dangerously and
888 # discard these initalisations.
890 $p =~ s/^\tpushl\s+\%e(di|si|bx)\n//g;
891 $p =~ s/^\txorl\s+\%e(ax|cx|dx),\s*\%e(ax|cx|dx)\n//g;
892 $p =~ s/^\tmovl\s+\%e(ax|cx|dx|si|di),\s*\d*\(\%esp\)\n//g;
893 $p =~ s/^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n//g;
894 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
895 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-(cygwin32|mingw32)/);
897 if ($TargetPlatform =~ /^i386-apple-darwin/) {
899 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
901 $pcrel_reg =~ s/(.|\n)*.*___i686\.get_pc_thunk\.([abcd]x)\n(.|\n)*/$2/ or $pcrel_reg = "";
902 $p =~ s/^\s+call\s+___i686\.get_pc_thunk\..x//;
903 $p =~ s/^\"?L\d+\$pb\"?:\n//;
905 if ($pcrel_reg eq "bx") {
906 # Bad gcc. Goes and uses %ebx, our BaseReg, for PIC. Bad gcc.
907 die "Darwin/x86: -fPIC -via-C doesn't work yet, use -fasm. Aborting."
911 } elsif ($TargetPlatform =~ /^x86_64-/) {
912 $p =~ s/^\tpushq\s+\%r(bx|bp|12|13|14)\n//g;
913 $p =~ s/^\tmovq\s+\%r(bx|bp|12|13|14),\s*\d*\(\%rsp\)\n//g;
914 $p =~ s/^\tsubq\s+\$\d+,\s*\%rsp\n//;
916 } elsif ($TargetPlatform =~ /^ia64-/) {
917 $p =~ s/^\t\.prologue .*\n//;
919 # Record the number of local and out registers for register relocation later
920 $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, (\d+), (\d+), 0\n//;
924 $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
925 $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
927 # Ignore save/restore of these registers; they're taken
928 # care of in StgRun()
929 $p =~ s/^\t\.save ar\.lc, r\d+\n//;
930 $p =~ s/^\t\.save pr, r\d+\n//;
931 $p =~ s/^\tmov r\d+ = ar\.lc\n//;
932 $p =~ s/^\tmov r\d+ = pr\n//;
934 # Remove .proc and .body directives
935 $p =~ s/^\t\.proc [a-zA-Z0-9_.]+#\n//;
936 $p =~ s/^\t\.body\n//;
938 # If there's a label, move it to the body
939 if ($p =~ /^[a-zA-Z0-9.]+:\n/) {
944 # Remove floating-point spill instructions.
945 # Only fp registers 2-5 and 16-23 are saved by the runtime.
946 if ($p =~ s/^\tstf\.spill \[r1[4-9]\] = f([2-5]|1[6-9]|2[0-3])(, [0-9]+)?\n//g) {
947 # Being paranoid, only try to remove these if we saw a
949 $p =~ s/^\tmov r1[4-9] = r12\n//;
950 $p =~ s/^\tadds r1[4-9] = -[0-9]+, r12\n//g;
951 $p =~ s/^\t\.save\.f 0x[0-9a-fA-F]\n//g;
952 $p =~ s/^\t\.save\.gf 0x0, 0x[0-9a-fA-F]+\n//g;
955 $p =~ s/^\tnop(?:\.[mifb])?\s+\d+\n//g; # remove nop instructions
956 $p =~ s/^\t\.(mii|mmi|mfi)\n//g; # bundling is no longer sensible
957 $p =~ s/^\t;;\n//g; # discard stops
958 $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
960 # GCC 3.3 saves r1 in the prologue, move this to the body
961 # (Does this register get restored anywhere?)
962 if ($p =~ /^\tmov r\d+ = r1\n/) {
966 } elsif ($TargetPlatform =~ /^m68k-/) {
967 $p =~ s/^\tlink a6,#-?\d.*\n//;
968 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
969 # The above showed up in the asm code,
970 # so I added it here.
971 # I hope it's correct.
973 $p =~ s/^\tmovel d2,sp\@-\n//;
974 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
975 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
976 } elsif ($TargetPlatform =~ /^mips-/) {
977 # the .frame/.mask/.fmask that we use is the same
978 # as that produced by GCC for miniInterpret; this
979 # gives GDB some chance of figuring out what happened
980 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
981 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
982 $p =~ s/^\t\.(mask|fmask).*\n//g;
983 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
984 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
985 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
986 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
987 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
988 $p =~ s/__FRAME__/$FRAME/;
989 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
991 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
993 $p =~ s/^\tmflr r0\n//;
994 $p =~ s/^\tbl saveFP # f\d+\n//;
995 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
996 $p =~ s/^\"?L\d+\$pb\"?:\n//;
997 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
998 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
999 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
1000 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
1001 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
1002 $p =~ s/^\tbcl 20,31,\"?L\d+\$pb\"?\n//;
1003 $p =~ s/^\"?L\d+\$pb\"?:\n//;
1004 $p =~ s/^\tmflr r31\n//;
1006 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1007 # under some circumstances, only when generating position dependent code.
1008 # I have no idea why, and I don't think it is necessary, so let's toss it.
1009 $p =~ s/^\tli r\d+,0\n//g;
1010 $p =~ s/^\tstw r\d+,\d+\(r1\)\n//g;
1011 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/) {
1012 $p =~ s/^\tmflr 0\n//;
1013 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
1014 $p =~ s/^\tstfd \d+,\d+\(1\)\n//g;
1015 $p =~ s/^\tstw r0,8\(1\)\n//;
1016 $p =~ s/^\tstwu 1,-\d+\(1\)\n//;
1017 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1019 # GCC's "large-model" PIC (-fPIC)
1021 $pcrel_label =~ s/(.|\n)*^.LCF(\d+):\n(.|\n)*/$2/ or $pcrel_label = "";
1023 $p =~ s/^\tbcl 20,31,.LCF\d+\n//;
1024 $p =~ s/^.LCF\d+:\n//;
1025 $p =~ s/^\tmflr 30\n//;
1026 $p =~ s/^\tlwz 0,\.LCL\d+-\.LCF\d+\(30\)\n//;
1027 $p =~ s/^\tadd 30,0,30\n//;
1029 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1030 # under some circumstances, only when generating position dependent code.
1031 # I have no idea why, and I don't think it is necessary, so let's toss it.
1032 $p =~ s/^\tli \d+,0\n//g;
1033 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1034 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1035 $p =~ s/^\tmr 31,1\n//;
1036 $p =~ s/^\tmflr 0\n//;
1037 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
1038 $p =~ s/^\tstfd \d+,-?\d+\(1\)\n//g;
1039 $p =~ s/^\tstd r0,8\(1\)\n//;
1040 $p =~ s/^\tstdu 1,-\d+\(1\)\n//;
1041 $p =~ s/^\tstd \d+,-?\d+\(1\)\n//g;
1043 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1044 # under some circumstances, only when generating position dependent code.
1045 # I have no idea why, and I don't think it is necessary, so let's toss it.
1046 $p =~ s/^\tli \d+,0\n//g;
1047 $p =~ s/^\tstd \d+,\d+\(1\)\n//g;
1049 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
1052 # HWL HACK: dont die, just print a warning
1053 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
1054 die "Prologue junk?: $p\n" if $p =~ /^\s+[^\s\.]/;
1056 # For PIC, we want to keep part of the prologue
1057 if ($TargetPlatform =~ /^powerpc-apple-darwin.*/ && $pcrel_label ne "") {
1058 # Darwin: load the current instruction pointer into register r31
1059 $p .= "bcl 20,31,$pcrel_label\n";
1060 $p .= "$pcrel_label:\n";
1061 $p .= "\tmflr r31\n";
1062 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/ && $pcrel_label ne "") {
1063 # Linux: load the GOT pointer into register 30
1064 $p .= "\tbcl 20,31,.LCF$pcrel_label\n";
1065 $p .= ".LCF$pcrel_label:\n";
1066 $p .= "\tmflr 30\n";
1067 $p .= "\tlwz 0,.LCL$pcrel_label-.LCF$pcrel_label(30)\n";
1068 $p .= "\tadd 30,0,30\n";
1069 } elsif ($TargetPlatform =~ /^i386-apple-darwin.*/ && $pcrel_label ne "") {
1070 $p .= "\tcall ___i686.get_pc_thunk.$pcrel_reg\n";
1071 $p .= "$pcrel_label:\n";
1074 # glue together what's left
1079 if ( $TargetPlatform =~ /^mips-/ ) {
1080 # MIPS: first, this basic sequence may occur "--- END ---" or not
1081 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
1084 # toss all epilogue stuff; again, paranoidly
1085 if ( $c =~ /--- END ---/ ) {
1086 # Gcc may decide to replicate the function epilogue. We want
1087 # to process all epilogues, so we split the function and then
1089 @fragments = split(/--- END ---/, $c);
1090 $r = shift(@fragments);
1092 # Rebuild `c'; processed fragments will be appended to `c'
1095 foreach $e (@fragments) {
1096 # etail holds code that is after the epilogue in the assembly-code
1097 # layout and should not be filtered as part of the epilogue.
1099 if ($TargetPlatform =~ /^i386-/) {
1101 $e =~ s/^\tpopl\s+\%edi\n//;
1102 $e =~ s/^\tpopl\s+\%esi\n//;
1103 $e =~ s/^\tpopl\s+\%edx\n//;
1104 $e =~ s/^\tpopl\s+\%ecx\n//;
1105 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
1106 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
1107 } elsif ($TargetPlatform =~ /^ia64-/) {
1108 # The epilogue is first split into:
1109 # $e, the epilogue code (up to the return instruction)
1110 # $etail, non-epilogue code (after the return instruction)
1111 # The return instruction is stripped in the process.
1112 if (!(($e, $etail) = split(/^\tbr\.ret\.sptk\.many b0\n/, $e))) {
1113 die "Epilogue doesn't seem to have one return instruction: $e\n";
1115 # Remove 'endp' directive from the tail
1116 $etail =~ s/^\t\.endp [a-zA-Z0-9_.]+#\n//;
1118 # If a return value is saved here, discard it
1119 $e =~ s/^\tmov r8 = r14\n//;
1121 # Remove floating-point fill instructions.
1122 # Only fp registers 2-5 and 16-23 are saved by the runtime.
1123 if ($e =~ s/^\tldf\.fill f([2-5]|1[6-9]|2[0-3]) = \[r1[4-9]\](, [0-9]+)?\n//g) {
1124 # Being paranoid, only try to remove this if we saw a fill
1126 $e =~ s/^\tadds r1[4-9] = [0-9]+, r12//g;
1129 $e =~ s/^\tnop(?:\.[mifb])?\s+\d+\n//g; # remove nop instructions
1130 $e =~ s/^\tmov ar\.pfs = r\d+\n//;
1131 $e =~ s/^\tmov ar\.lc = r\d+\n//;
1132 $e =~ s/^\tmov pr = r\d+, -1\n//;
1133 $e =~ s/^\tmov b0 = r\d+\n//;
1134 $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
1135 #$e =~ s/^\tbr\.ret\.sptk\.many b0\n//; # already removed
1136 $e =~ s/^\t\.(mii|mmi|mfi|mib)\n//g; # bundling is no longer sensible
1137 $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
1138 $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
1139 } elsif ($TargetPlatform =~ /^m68k-/) {
1140 $e =~ s/^\tunlk a6\n//;
1142 } elsif ($TargetPlatform =~ /^mips-/) {
1143 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
1144 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
1145 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
1146 $e =~ s/^\tj\t\$31\n//;
1147 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
1148 $e =~ s/^\taddi r1,r1,\d+\n//;
1149 $e =~ s/^\tlwz r\d+,\d+\(r1\)\n//;
1150 $e =~ s/^\tlmw r\d+,-\d+\(r1\)\n//;
1151 $e =~ s/^\tmtlr r0\n//;
1153 $e =~ s/^\tb restFP ;.*\n//;
1154 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1155 $e =~ s/^\tmr 3,0\n//;
1156 $e =~ s/^\taddi 1,1,\d+\n//;
1157 $e =~ s/^\tld 0,16\(1\)\n//;
1158 $e =~ s/^\tmtlr 0\n//;
1160 # callee-save registers
1161 $e =~ s/^\tld \d+,-?\d+\(1\)\n//g;
1162 $e =~ s/^\tlfd \d+,-?\d+\(1\)\n//g;
1164 # get rid of the debug junk along with the blr
1165 $e =~ s/^\tblr\n\t.long .*\n\t.byte .*\n//;
1167 # incase we missed it with the last one get the blr alone
1170 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
1173 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
1175 # glue together what's left
1178 $c =~ s/\n\t\n/\n/; # junk blank line
1181 if ($TargetPlatform =~ /^ia64-/) {
1182 # On IA64, remove an .endp directive even if no epilogue was found.
1183 # Code optimizations may have removed the "--- END ---" token.
1184 $c =~ s/^\t\.endp [a-zA-Z0-9_.]+#\n//;
1188 # On SPARCs, we don't do --- BEGIN/END ---, we just
1189 # toss the register-windowing save/restore/ret* instructions
1190 # directly unless they've been generated by function definitions in header
1192 if ( $TargetPlatform =~ /^sparc-/ ) {
1193 if ( ! ( $TargetPlatform =~ /solaris2$/ && $chkcat[$i] eq 'unknown' )) {
1194 $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
1196 # throw away PROLOGUE comments
1197 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
1200 # On Alphas, the prologue mangling is done a little later (below)
1202 # toss all calls to __DISCARD__
1203 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
1204 $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
1205 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-darwin.*/;
1206 $c =~ s/^\tbl\s+__DISCARD__(\@plt)?\n//go if $TargetPlatform =~ /^powerpc-.*-linux/;
1207 $c =~ s/^\tbl\s+\.__DISCARD__\n\s+nop\n//go if $TargetPlatform =~ /^powerpc64-.*-linux/;
1208 $c =~ s/^\tcall\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /i386-apple-darwin.*/;
1210 # IA64: fix register allocation; mangle tailcalls into jumps
1211 if ($TargetPlatform =~ /^ia64-/) {
1212 ia64_rename_registers($ia64_locnum, $ia64_outnum) if (defined($ia64_locnum));
1213 ia64_mangle_tailcalls();
1216 # MIPS: that may leave some gratuitous asm macros around
1217 # (no harm done; but we get rid of them to be tidier)
1218 $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/
1219 if $TargetPlatform =~ /^mips-/;
1221 # toss stack adjustment after DoSparks
1222 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
1223 if $TargetPlatform =~ /^m68k-/; # this looks old...
1225 if ( $TargetPlatform =~ /^alpha-/ &&
1226 ! $magic_rdata_seen &&
1227 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
1228 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
1229 $magic_rdata_seen = 1;
1232 # pick some end-things and move them to the next chunk
1234 # pin a funny end-thing on (for easier matching):
1235 $c .= 'FUNNY#END#THING';
1237 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
1241 # on x86 we try not to copy any directives into a literal
1242 # chunk, rather we keep looking for the next real chunk. This
1243 # is because we get things like
1245 # .globl blah_closure
1251 if ( $TargetPlatform =~ /^(i386|sparc|powerpc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
1253 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
1256 if ( $j < $numchks ) {
1257 $chk[$j] = $to_move . $chk[$j];
1261 elsif ( ( $i < ($numchks - 1)
1262 && ( $to_move =~ /${T_COPY_DIRVS}/
1263 || ( $TargetPlatform =~ /^hppa/
1264 && $to_move =~ /align/
1265 && $chkcat[$i+1] eq 'literal')
1268 || ($to_move =~ /^[ \t]*\.section[ \t]+\.note\.GNU-stack,/)
1270 $chk[$i + 1] = $to_move . $chk[$i + 1];
1271 # otherwise they're tossed
1274 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
1277 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
1279 # toss all prologue stuff, except for loading gp, and the ..ng address
1280 unless ($c =~ /\.ent.*\n\$.*\.\.ng:/) {
1281 if (($p, $r) = split(/^\t\.prologue/, $c)) {
1282 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
1283 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
1284 $keep =~ s/^\t\.(mask|fmask).*\n//g;
1285 $c = $keep . "..ng:\n";
1287 print STDERR "malformed code block ($ent)?\n"
1290 $c .= "\t.prologue" . $r;
1294 $c =~ s/FUNNY#END#THING//;
1296 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
1298 $chk[$i] = $c; # update w/ convenience copy
1301 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
1302 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
1305 if ( $TargetPlatform =~ /^alpha-/ ) {
1306 # print out the header stuff first
1307 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
1308 print OUTASM $chk[0];
1310 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1311 print OUTASM $chk[0];
1313 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1314 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
1316 # get rid of horrible "<dollar>Revision: .*$" strings
1317 local(@lines0) = split(/\n/, $chk[0]);
1319 while ( $z <= $#lines0 ) {
1320 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
1323 while ( $z <= $#lines0 ) {
1325 last if $lines0[$z] =~ /[,\t]0x0$/;
1331 $chk[0] = join("\n", @lines0);
1332 $chk[0] =~ s/\n\n+/\n/;
1333 print OUTASM $chk[0];
1336 # print out all the literal strings next
1337 for ($i = 0; $i < $numchks; $i++) {
1338 if ( $chkcat[$i] eq 'literal' ) {
1340 # HACK: try to detect 16-byte constants and align them
1341 # on a 16-byte boundary. x86_64 sometimes needs 128-bit
1342 # aligned constants, and so does Darwin/x86.
1343 if ( $TargetPlatform =~ /^x86_64/
1344 || $TargetPlatform =~ /^i386-apple-darwin/ ) {
1346 if ($z =~ /(\.long.*\n.*\.long.*\n.*\.long.*\n.*\.long|\.quad.*\n.*\.quad)/) {
1347 print OUTASM $T_HDR_literal16;
1349 print OUTASM $T_HDR_literal;
1352 print OUTASM $T_HDR_literal;
1355 print OUTASM $chk[$i];
1356 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
1358 $chkcat[$i] = 'DONE ALREADY';
1362 # on the HPPA, print out all the bss next
1363 if ( $TargetPlatform =~ /^hppa/ ) {
1364 for ($i = 1; $i < $numchks; $i++) {
1365 if ( $chkcat[$i] eq 'bss' ) {
1366 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
1367 print OUTASM $chk[$i];
1369 $chkcat[$i] = 'DONE ALREADY';
1374 # $numchks + 1 as we have the extra one for .note.GNU-stack
1375 for ($i = $FIRST_MANGLABLE; $i < $numchks + 1; $i++) {
1376 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
1378 next if $chkcat[$i] eq 'DONE ALREADY';
1380 if ( $chkcat[$i] eq 'misc' || $chkcat[$i] eq 'unknown' ) {
1381 if ($chk[$i] ne '') {
1382 print OUTASM $T_HDR_misc;
1383 &print_doctored($chk[$i], 0);
1386 } elsif ( $chkcat[$i] eq 'verbatim' ) {
1387 print OUTASM $chk[$i];
1389 } elsif ( $chkcat[$i] eq 'toss' ) {
1390 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
1392 } elsif ( $chkcat[$i] eq 'data' ) {
1393 if ($chk[$i] ne '') {
1394 print OUTASM $T_HDR_data;
1395 print OUTASM $chk[$i];
1398 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1399 # we can just re-constitute this one...
1400 # NB: we emit _three_ underscores no matter what,
1401 # so ghc-split doesn't have to care.
1402 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1404 } elsif ( $chkcat[$i] eq 'closure'
1405 || $chkcat[$i] eq 'srt'
1406 || $chkcat[$i] eq 'infotbl'
1407 || $chkcat[$i] eq 'entry') { # do them in that order
1408 $symb = $chksymb[$i];
1411 if ( defined($closurechk{$symb}) ) {
1412 print OUTASM $T_HDR_closure;
1413 print OUTASM $chk[$closurechk{$symb}];
1414 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1418 if ( defined($srtchk{$symb}) ) {
1419 print OUTASM $T_HDR_relrodata;
1420 print OUTASM $chk[$srtchk{$symb}];
1421 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1425 if ( defined($infochk{$symb}) ) {
1427 print OUTASM $T_HDR_info;
1428 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1430 # entry code will be put here!
1432 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1436 if ( defined($entrychk{$symb}) ) {
1438 $c = $chk[$entrychk{$symb}];
1440 # If this is an entry point with an info table,
1441 # eliminate the entry symbol and all directives involving it.
1442 if (defined($infochk{$symb}) && $TargetPlatform !~ /^ia64-/) {
1444 foreach $l (split(/\n/,$c)) {
1445 next if $l =~ /^.*$symb_(entry|ret)${T_POST_LBL}/;
1447 # If we have .type/.size direrctives involving foo_entry,
1448 # then make them refer to foo_info instead. The information
1449 # in these directives is used by the cachegrind annotator,
1450 # so it is worthwhile keeping.
1451 if ($l =~ /^\s*\.(type|size).*$symb_(entry|ret)/) {
1452 $l =~ s/$symb(_entry|_ret)/${symb}_info/g;
1456 next if $l =~ /^\s*\..*$symb.*\n?/;
1459 $c = join("\n",@o) . "\n";
1462 print OUTASM $T_HDR_entry;
1464 &print_doctored($c, 1); # NB: the 1!!!
1466 $chkcat[$entrychk{$symb}] = 'DONE ALREADY';
1469 } elsif ( $chkcat[$i] eq 'vector' ) {
1470 $symb = $chksymb[$i];
1473 if ( defined($vectorchk{$symb}) ) {
1474 print OUTASM $T_HDR_vector;
1475 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1477 # direct return code will be put here!
1478 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1480 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1481 # Alphas: the commented nop is for the splitter, to ensure
1482 # that no module ends with a label as the very last
1483 # thing. (The linker will adjust the label to point
1484 # to the first code word of the next module linked in,
1485 # even if alignment constraints cause the label to move!)
1487 print OUTASM "\t# nop\n";
1490 } elsif ( $chkcat[$i] eq 'rodata' ) {
1491 print OUTASM $T_HDR_rodata;
1492 print OUTASM $chk[$i];
1493 $chkcat[$i] = 'DONE ALREADY';
1494 } elsif ( $chkcat[$i] eq 'relrodata' ) {
1495 print OUTASM $T_HDR_relrodata;
1496 print OUTASM $chk[$i];
1497 $chkcat[$i] = 'DONE ALREADY';
1498 } elsif ( $chkcat[$i] eq 'toc' ) {
1499 # silly optimisation to print tocs, since they come in groups...
1500 print OUTASM $T_HDR_toc;
1502 while ($chkcat[$j] eq 'toc')
1503 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1506 print OUTASM $chk[$j];
1508 $chkcat[$j] = 'DONE ALREADY';
1512 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' ) {
1513 # apple-darwin: dynamic linker stubs
1514 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1515 { # print them out unchanged, but remove the stubs for __DISCARD__
1516 print OUTASM $chk[$i];
1518 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ && $chkcat[$i] eq 'get_pc_thunk' ) {
1519 # i386-apple-darwin: __i686.get_pc_thunk.[abcd]x
1520 print OUTASM ".section __TEXT,__textcoal_nt,coalesced,no_toc\n";
1521 print OUTASM $chk[$i];
1523 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1527 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1530 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1531 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1535 On IA64, tail calls are converted to branches at this point. The mangler
1536 searches for function calls immediately followed by a '--- TAILCALL ---'
1537 token. Since the compiler can put various combinations of labels, bundling
1538 directives, nop instructions, stops, and a move of the return value
1539 between the branch and the tail call, proper matching of the tail call
1540 gets a little hairy. This subroutine does the mangling.
1542 Here is an example of a tail call before mangling:
1545 br.call.sptk.many b0 = b6
1560 sub ia64_mangle_tailcalls {
1561 # Function input and output are in $c
1563 # Construct the tailcall-mangling expression the first time this function
1565 if (!defined($IA64_MATCH_TAILCALL)) {
1566 # One-line pattern matching constructs. None of these
1567 # should bind references; all parenthesized terms
1568 # should be (?:) terms.
1569 my $stop = q/(?:\t;;\n)/;
1570 my $bundle = q/(?:\t\.(?:mii|mib|mmi|mmb|mfi|mfb|mbb|bbb)\n)/;
1571 my $nop = q/(?:\tnop(?:\.[mifb])?\s+\d+\n)/;
1572 my $movgp = q/(?:\tmov r1 = r\d+\n)/;
1573 my $postbr = q/(?:\tbr \.L\d+\n)/;
1575 my $noeffect = "(?:$stop$bundle?|$nop)*";
1576 my $postbundle = "(?:$bundle?$nop?$nop?$postbr)?";
1578 # Important parts of the pattern match. The branch target
1579 # and subsequent jump label are bound to $1 and $2
1580 # respectively. Sometimes there is no label.
1581 my $callbr = q/^\tbr\.call\.sptk\.many b0 = (.*)\n/;
1582 my $label = q/(?:^\.L([0-9]*):\n)/;
1583 my $tailcall = q/\t--- TAILCALL ---\n/;
1585 $IA64_MATCH_TAILCALL =
1586 $callbr . $label . '?' . $noeffect . $movgp . '?' . $noeffect .
1587 $tailcall . $stop . '?' . '(?:' . $postbundle . ')?';
1590 # Find and mangle tailcalls
1591 while ($c =~ s/$IA64_MATCH_TAILCALL/\tbr\.few $1\n/o) {
1592 # Eek, the gcc optimiser is getting smarter... if we see a jump to the
1593 # --- TAILCALL --- marker then we reapply the substitution at the source sites
1594 $c =~ s/^\tbr \.L$2\n/\t--- TAILCALL ---\n/g if ($2);
1597 # Verify that all instances of TAILCALL were processed
1598 if ($c =~ /^\t--- TAILCALL ---\n/) {
1599 die "Unmangled TAILCALL tokens remain after mangling"
1604 The number of registers allocated on the IA64 register stack is set
1605 upon entry to the runtime with an `alloc' instruction at the entry
1606 point of \verb+StgRun()+. Gcc uses its own `alloc' to allocate
1607 however many registers it likes in each function. When we discard
1608 gcc's alloc, we have to reconcile its register assignment with what
1611 There are three stack areas: fixed registers, input/local registers,
1612 and output registers. We move the output registers to the output
1613 register space and leave the other registers where they are.
1616 sub ia64_rename_registers() {
1617 # The text to be mangled is in $c
1618 # Find number of registers in each stack area
1619 my ($loc, $out) = @_;
1625 # These are the register numbers used in the STG runtime
1626 my $STG_FIRST_OUT_REG = 32 + 34;
1627 my $STG_LAST_OUT_REG = $STG_FIRST_OUT_REG + 7;
1629 $first_out_reg = 32 + $loc;
1631 if ($first_out_reg > $STG_FIRST_OUT_REG) {
1632 die "Too many local registers allocated by gcc";
1635 # Split the string into fragments containing one register name each.
1636 # Rename the register in each fragment and concatenate.
1638 foreach $fragment (split(/(?=r\d+[^a-zA-Z0-9_.])/s, $c)) {
1639 if ($fragment =~ /^r(\d+)((?:[^a-zA-Z0-9_.].*)?)$/s) {
1642 if ($regnum < $first_out_reg) {
1643 # This is a local or fixed register
1645 # Local registers 32 and 33 (r64 and r65) are
1646 # used to hold saved state; they shouldn't be touched
1647 if ($regnum == 64 || $regnum == 65) {
1648 die "Reserved register $regnum is in use";
1652 # This is an output register
1653 $regnum = $regnum - $first_out_reg + $STG_FIRST_OUT_REG;
1654 if ($regnum > $STG_LAST_OUT_REG) {
1655 die "Register number ($regnum) is out of expected range";
1659 # Update this fragment
1660 $fragment = "r" . $regnum . $2;
1671 sub hppa_mash_prologue { # OK, epilogue, too
1674 # toss all prologue stuff
1675 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1677 # Lie about our .CALLINFO
1678 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1685 # toss all epilogue stuff
1686 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1688 # Sorry; we moved the _info stuff to the code segment.
1689 s/_info,DATA/_info,CODE/g;
1696 sub print_doctored {
1697 local($_, $need_fallthru_patch) = @_;
1699 if ( $TargetPlatform =~ /^x86_64-/ ) {
1702 # movq -4(%ebp), %rax
1707 s/^\tmovq\s+(-?\d*\(\%r(bx|bp|13)\)),\s*(\%r(ax|cx|dx|10|11))\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1708 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e(ax|cx|si|di))\n\tjmp\s+\*\%r\3/\tjmp\t$T_US$1/g;
1711 if ( $TargetPlatform !~ /^i386-/
1712 || ! /^\t[a-z]/ # no instructions in here, apparently
1713 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1718 # OK, must do some x86 **HACKING**
1720 local($entry_patch) = '';
1721 local($exit_patch) = '';
1723 # gotta watch out for weird instructions that
1724 # invisibly smash various regs:
1725 # rep* %ecx used for counting
1726 # scas* %edi used for destination index
1727 # cmps* %e[sd]i used for indices
1728 # loop* %ecx used for counting
1733 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1735 # * GCC used an "STG reg" for its own purposes
1737 # * some secret uses of machine reg, requiring STG reg
1738 # to be saved/restored
1740 # The most dangerous "GCC uses" of an "STG reg" are when
1741 # the reg holds the target of a jmp -- it's tricky to
1742 # insert the patch-up code before we get to the target!
1743 # So here we change the jmps:
1745 # --------------------------------------------------------
1746 # it can happen that we have jumps of the form...
1747 # jmp *<something involving %esp>
1749 # jmp <something involving another naughty register...>
1751 # a reasonably-common case is:
1753 # movl $_blah,<bad-reg>
1756 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[acd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1760 # movl -4(%ebx), %eax
1765 s/^\tmovl\s+(-?\d*\(\%e(bx|si)\)),\s*(\%e[acd]x)\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1767 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1768 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1769 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1770 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1771 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1772 if /(jmp|call)\s+.*\%esi/;
1774 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1775 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1776 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1777 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1778 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1779 if /(jmp|call)\s+.*\%edi/;
1782 # OK, now we can decide what our patch-up code is going to
1785 # Offsets into register table - you'd better update these magic
1786 # numbers should you change its contents!
1787 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1788 local($OFFSET_Hp)=88;
1790 # Note funky ".=" stuff; we're *adding* to these _patch guys
1791 if ( $StolenX86Regs <= 2
1792 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1793 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1794 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1796 # nothing for call_{entry,exit} because %esi is callee-save
1798 if ( $StolenX86Regs <= 3
1799 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1800 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1801 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1803 # nothing for call_{entry,exit} because %edi is callee-save
1806 # --------------------------------------------------------
1807 # next, here we go with non-%esp patching!
1809 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1811 # Before calling GC we must set up the exit condition before the call
1812 # and entry condition when we come back
1814 # fix _all_ non-local jumps:
1816 if ( $TargetPlatform =~ /^.*-apple-darwin.*/ ) {
1817 # On Darwin, we've got local-looking jumps that are
1818 # actually global (i.e. jumps to Lfoo$stub or via
1819 # Lfoo$non_lazy_ptr), so we fix those first.
1820 # In fact, we just fix everything that contains a dollar
1821 # because false positives don't hurt here.
1823 s/^(\tjmp\s+\*?L.*\$.*\n)/$exit_patch$1/g;
1826 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1827 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1829 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1831 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1832 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1834 if ($StolenX86Regs == 2 ) {
1835 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1836 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1837 } elsif ($StolenX86Regs == 3 ) {
1838 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1839 if /^\t(jmp|call)\s+.*\%edi/;
1842 # --------------------------------------------------------
1843 # that's it -- print it
1845 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1849 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1850 print OUTASM $exit_patch;
1851 # ToDo: make it not print if there is a "jmp" at the end
1857 sub init_FUNNY_THINGS {
1858 %KNOWN_FUNNY_THING = (
1860 # "${T_US}stg_.*{T_POST_LBL}", 1,
1865 The following table reversal is used for both info tables and return
1866 vectors. In both cases, we remove the first entry from the table,
1867 reverse the table, put the label at the end, and paste some code
1868 (that which is normally referred to by the first entry in the table)
1869 right after the table itself. (The code pasting is done elsewhere.)
1873 local($symb, $tbl, $discard1) = @_;
1875 return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1877 local($before) = '';
1879 local(@imports) = (); # hppa only
1882 local(@lines) = split(/\n/, $tbl);
1885 # Deal with the header...
1886 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1887 $label .= $lines[$i] . "\n",
1888 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1889 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1890 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1892 $before .= $lines[$i] . "\n"; # otherwise...
1896 $infoname =~ s/(.|\n)*^([A-Za-z0-9_]+_info)${T_POST_LBL}$(.|\n)*/\2/;
1898 # Grab the table data...
1899 if ( $TargetPlatform !~ /^hppa/ ) {
1900 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1902 # Convert addresses of SRTs, slow entrypoints and large bitmaps
1903 # to offsets (relative to the info label),
1904 # in order to support position independent code.
1905 $line =~ s/$infoname/0/
1906 || $line =~ s/([A-Za-z0-9_]+_srtd)$/\1 - $infoname/
1907 || $line =~ s/([A-Za-z0-9_]+_srt(\+\d+)?)$/\1 - $infoname/
1908 || $line =~ s/([A-Za-z0-9_]+_str)$/\1 - $infoname/
1909 || $line =~ s/([A-Za-z0-9_]+_slow)$/\1 - $infoname/
1910 || $line =~ s/([A-Za-z0-9_]+_btm)$/\1 - $infoname/
1911 || $line =~ s/([A-Za-z0-9_]+_alt)$/\1 - $infoname/
1912 || $line =~ s/([A-Za-z0-9_]+_dflt)$/\1 - $infoname/
1913 || $line =~ s/([A-Za-z0-9_]+_ret)$/\1 - $infoname/;
1914 push(@words, $line);
1916 } else { # hppa weirdness
1917 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1918 # FIXME: the RTS now expects offsets instead of addresses
1919 # for all labels in info tables.
1920 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1921 push(@imports, $lines[$i]);
1923 # We don't use HP's ``function pointers''
1924 # We just use labels in code space, like normal people
1925 $lines[$i] =~ s/P%//;
1926 push(@words, $lines[$i]);
1931 # Now throw away any initial zero word from the table. This is a hack
1932 # that lets us reduce the size of info tables when the SRT field is not
1933 # needed: see comments StgFunInfoTable in InfoTables.h.
1935 # The .zero business is for Linux/ELF.
1936 # The .skip business is for Sparc/Solaris/ELF.
1937 # The .blockz business is for HPPA.
1939 # if ($words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1944 for (; $i <= $#lines; $i++) {
1945 $after .= $lines[$i] . "\n";
1948 # Alphas: If we have anonymous text (not part of a procedure), the
1949 # linker may complain about missing exception information. Bleh.
1950 # To suppress this, we place a .ent/.end pair around the code.
1951 # At the same time, we have to be careful and not enclose any leading
1952 # .file/.loc directives.
1953 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1954 local ($ident) = $1;
1955 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1956 $after .= "\t.end $ident\n";
1959 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1960 # assembler (!) wherein .quad constants inside .text sections are
1961 # first narrowed to 32 bits then sign-extended back to 64 bits.
1962 # This obviously screws up our 64-bit bitmaps, so we work around
1963 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1964 if ( $TargetPlatform =~ /^alpha-/ ) {
1966 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1967 local ($number) = $1;
1968 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1969 local ($sign, $base, $digits) = ($1, $2, $3);
1970 $base = (10, 8, 16)[length $base];
1971 local ($hi, $lo) = (0, 0);
1972 foreach $i (split(//, $digits)) {
1973 $j = $lo * $base + $i;
1974 $lo = $j % 4294967296;
1975 $hi = $hi * $base + ($j - $lo) / 4294967296;
1977 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1979 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1980 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1982 print STDERR "Cannot handle \".quad $number\" in info table\n";
1989 if ( $TargetPlatform =~ /x86_64-apple-darwin/ ) {
1990 # Tack a label to the front of the info table, too.
1991 # For now, this just serves to work around a crash in Apple's new
1992 # 64-bit linker (it seems to assume that there is no data before the
1993 # first label in a section).
1995 # The plan for the future is to do this on all Darwin platforms, and
1996 # to add a reference to this label after the entry code, just as the
1997 # NCG does, so we can enable dead-code-stripping in the linker without
1998 # losing our info tables. (Hence the name _dsp, for dead-strip preventer)
2000 $before .= "\n${infoname}_dsp:\n";
2004 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
2005 . join("\n", @words) . "\n"
2008 # print STDERR "before=$before\n";
2009 # print STDERR "label=$label\n";
2010 # print STDERR "words=",(reverse @words),"\n";
2011 # print STDERR "after=$after\n";
2017 The HP is a major nuisance. The threaded code mangler moved info
2018 tables from data space to code space, but unthreaded code in the RTS
2019 still has references to info tables in data space. Since the HP
2020 linker is very precise about where symbols live, we need to patch the
2021 references in the unthreaded RTS as well.
2024 sub mini_mangle_asm_hppa {
2025 local($in_asmf, $out_asmf) = @_;
2027 open(INASM, "< $in_asmf")
2028 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
2029 open(OUTASM,"> $out_asmf")
2030 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
2033 s/_info,DATA/_info,CODE/; # Move _info references to code space
2039 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
2040 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
2046 sub tidy_up_and_die {
2047 local($return_val, $msg) = @_;
2049 exit (($return_val == 0) ? 0 : 1);