$T_create_word = "\t.word";
#--------------------------------------------------------#
+ } elsif ( $TargetPlatform =~ /^ia64-.*-linux$/ ) {
+
+ $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
+ $T_US = ''; # _ if symbols have an underscore on the front
+ $T_PRE_APP = '#';
+ $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
+ $T_POST_LBL = ':';
+
+ $T_MOVE_DIRVS = '^(\s*\.(global|proc|pred\.safe_across_calls|text|data|section|align|size|type|ident)\s+.*\n)';
+ $T_COPY_DIRVS = '\.(global|proc)';
+
+ $T_hsc_cc_PAT = '\.string.*\)(hsc|cc) (.*)\\\\t(.*)"';
+ $T_DOT_WORD = '\.(long|value|byte|zero)';
+ $T_DOT_GLOBAL = '\.global';
+ $T_HDR_literal = "\.section\t\.rodata\n";
+ $T_HDR_misc = "\.text\n\t\.align 8\n";
+ $T_HDR_data = "\.data\n\t\.align 8\n";
+ $T_HDR_consist = "\.text\n";
+ $T_HDR_closure = "\.data\n\t\.align 8\n";
+ $T_HDR_srt = "\.text\n\t\.align 8\n";
+ $T_HDR_info = "\.text\n\t\.align 8\n";
+ $T_HDR_entry = "\.text\n\t\.align 16\n";
+ $T_HDR_fast = "\.text\n\t\.align 16\n";
+ $T_HDR_vector = "\.text\n\t\.align 8\n";
+ $T_HDR_direct = "\.text\n\t\.align 8\n";
+ $T_create_word = "\t.word";
+
+ #--------------------------------------------------------#
} elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) {
$T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
next if /${T_PRE_APP}(NO_)?APP/o;
next if /^;/ && $TargetPlatform =~ /^hppa/;
- next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips)-/;
+ next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|ia64)-/;
if ( $TargetPlatform =~ /^mips-/
&& /^\t\.(globl\S+\.text|comm\t)/ ) {
# toss all prologue stuff; HPPA is pretty weird
# (see elsewhere)
- $c = &mash_hppa_prologue($c) if $TargetPlatform =~ /^hppa/;
+ $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
+
+ # do some register renaming before dropping the prologue
+ $c = &ia64_rename_outputs($c) if $TargetPlatform =~ /^ia64-/;
# be slightly paranoid to make sure there's
# nothing surprising in there
$r = $& . $r;
}
+ } elsif ($TargetPlatform =~ /^ia64-/) {
+ $p =~ s/^\t\.prologue .*\n//;
+ $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, .*\n//;
+ $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
+ $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
+ $p =~ s/^\t\.(mii|mmi)\n//; # bundling is no longer sensible
+ $p =~ s/^\t;;\n//g; # discard stops
+ $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
} elsif ($TargetPlatform =~ /^m68k-/) {
$p =~ s/^\tlink a6,#-?\d.*\n//;
$p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
$e =~ s/^\tpopl\s+\%ecx\n//;
$e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
$e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
+ } elsif ($TargetPlatform =~ /^ia64-/) {
+ $e =~ s/^\tmov ar\.pfs = r\d+\n//;
+ $e =~ s/^\tmov b0 = r\d+\n//;
+ $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
+ $e =~ s/^\tbr\.ret\.sptk\.many b0\n//;
+ $e =~ s/^\t\.(mii|mmi|mib)\n//g; # bundling is no longer sensible
+ $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
+ $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
} elsif ($TargetPlatform =~ /^m68k-/) {
$e =~ s/^\tunlk a6\n//;
$e =~ s/^\trts\n//;
$c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
$c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-.*/;
+ # IA64: mangle tailcalls into jumps here
+ $c =~ s/^\tbr\.call\.sptk\.many b0 = (.*)\n\t;;\n(\tmov r1 = r\d+\n)?\t;;\n\t--- TAILCALL ---\n/\tbr\.few $1\n/g
+ if $TargetPlatform =~ /^ia64-/;
+
# MIPS: that may leave some gratuitous asm macros around
# (no harm done; but we get rid of them to be tidier)
$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/
# The next two only apply if we're not stealing %esi or %edi.
$c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%esi\n\tjmp\s+\*\%esi\n// if ($StolenX86Regs < 3);
$c =~ s/^\tmovl\s+\$${T_US}${symb}_fast\d*,\s*\%edi\n\tjmp\s+\*\%edi\n// if ($StolenX86Regs < 4);
+ } elsif ( $TargetPlatform =~ /^ia64-/ ) {
+ #$c =~ s/^\tbr\.few ${symb}_fast\d*#\n\t;;\n(\t;;\n\t\.endp ${symb}_entry#\n)/$1/;
+ $c =~ s/^\tbr\.few ${symb}_fast\d*#\n(\t;;\n\t\.endp ${symb}_entry#\n)/$1/;
} elsif ( $TargetPlatform =~ /^mips-/ ) {
$c =~ s/^\tjmp \$31,\(\$27\),0\n\t\.align 4\n\t\.end/\t.align 4\n\t.end/;
} elsif ( $TargetPlatform =~ /^m68k-/ ) {
\end{code}
\begin{code}
-sub mash_hppa_prologue { # OK, epilogue, too
+sub hppa_mash_prologue { # OK, epilogue, too
local($_) = @_;
# toss all prologue stuff
}
\end{code}
+On IA64 we use a single register frame throughout STG execution, and delete
+the frame management instructions from the prologue and epilogue - similarly
+to the memory stack. Unfortunately, gcc always uses absolute register names
+instead of logical names like out0. This means that outputs (i.e. inputs to
+other functions) will end up in the wrong registers relative to our "frame".
+Hence this evil register renaming....
+
+\begin{code}
+sub ia64_rename_outputs {
+ local($_) = @_;
+
+ return ($_) if (!/^\talloc r\d+ = ar\.pfs, (\d+), (\d+), (\d+), (\d+)$/);
+
+ local($inputs,$locals,$outputs,$rotating) = ($1,$2,$3,$4);
+ local($oldbase,$newbase,$old,$new,$i);
+ local($LOCALS) = 24; # must correspond to value in StgCRun.c
+
+ # Check everything fits in our standard frame. Only 8 outputs should
+ # ever go in registers.
+
+ die "No inputs allowed: $inputs" if ($inputs > 0);
+ die "Too many locals: $locals" if ($locals > $LOCALS);
+ die "Too many outputs: $outputs" if ($outputs > 8);
+ die "No rotating registers allowed: $rotating" if ($rotating > 0);
+
+ $outbase = 32 + $inputs + $locals;
+ $newbase = 32 + 0 + $LOCALS;
+
+ # Always do renaming from the top to avoid collisions
+ for ($i = $outputs-1; $i >= 0; $i--) {
+ $old = $outbase + $i;
+ $new = $newbase + $i;
+ s/\br$old\b/r$new/g;
+ }
+
+ return ($_);
+}
+\end{code}
+
\begin{code}
sub print_doctored {
local($_, $need_fallthru_patch) = @_;
sub rev_tbl {
local($symb, $tbl, $discard1) = @_;
+ return ($tbl) if ($TargetPlatform =~ /^ia64-/);
+
local($before) = '';
local($label) = '';
local(@imports) = (); # hppa only
/* -----------------------------------------------------------------------------
- * $Id: StgCRun.c,v 1.33 2002/05/21 14:58:49 wolfgang Exp $
+ * $Id: StgCRun.c,v 1.34 2002/06/03 13:08:41 matthewc Exp $
*
* (c) The GHC Team, 1998-2000
*
extern StgThreadReturnCode StgRun(StgFunPtr f, StgRegTable *basereg);
-void StgRunIsImplementedInAssembler()
+void StgRunIsImplementedInAssembler(void)
{
__asm__ volatile (
"\n.globl _StgRun\n"
#endif
+/* -----------------------------------------------------------------------------
+ IA64 architecture
+
+ Again, in assembler - so we can fiddle with the register stack.
+ -------------------------------------------------------------------------- */
+
+#ifdef ia64_TARGET_ARCH
+
+/* the memory stack is rarely used, so 16K is excessive */
+#undef RESERVED_C_STACK_BYTES
+#define RESERVED_C_STACK_BYTES 1024
+
+void StgRunIsImplementedInAssembler(void)
+{
+ __asm__ volatile(
+ ".global StgRun\n"
+ "StgRun:\n"
+ "\talloc r55 = ar.pfs, 0, 24, 8, 0\n" /* setup register frame */
+ "\tmov r54 = b0\n" /* save return address */
+ "\tadds sp = -%0, sp\n" /* setup stack */
+ "\tld8 r16=[r32],8 ;;\n" /* branch to f using descriptor */
+ "\tld8 r1=[r32]\n"
+ "\tmov b6=r16\n"
+ "\tbr.few b6 ;;\n"
+ ".global StgReturn\n"
+ "StgReturn:\n"
+ "\tmov r8 = 0\n" /* return value in r8 */
+ "\tmov ar.pfs = r55\n" /* restore register frame */
+ "\tmov b0 = r54\n" /* restore return address */
+ "\tadds sp = %0, sp\n" /* restore stack */
+ "\tbr.ret.sptk.many b0 ;;\n" /* return */
+ : : "i"(RESERVED_C_STACK_BYTES));
+}
+
+#endif
+
#endif /* !USE_MINIINTERPRETER */