+sub rex {
+ local *opcode=shift;
+ my ($dst,$src)=@_;
+
+ if ($dst>=8 || $src>=8) {
+ $rex=0x40;
+ $rex|=0x04 if($dst>=8);
+ $rex|=0x01 if($src>=8);
+ push @opcode,$rex;
+ }
+}
+
+# older gas and ml64 don't handle SSE>2 instructions
+my %regrm = ( "%eax"=>0, "%ecx"=>1, "%edx"=>2, "%ebx"=>3,
+ "%esp"=>4, "%ebp"=>5, "%esi"=>6, "%edi"=>7 );
+
+my $pextrd = sub {
+ if (shift =~ /\$([0-9]+),%xmm([0-9]+),(%\w+)/) {
+ my @opcode=(0x66);
+ $imm=$1;
+ $src=$2;
+ $dst=$3;
+ if ($dst =~ /%r([0-9]+)d/) { $dst = $1; }
+ elsif ($dst =~ /%e/) { $dst = $regrm{$dst}; }
+ rex(\@opcode,$src,$dst);
+ push @opcode,0x0f,0x3a,0x16;
+ push @opcode,0xc0|(($src&7)<<3)|($dst&7); # ModR/M
+ push @opcode,$imm;
+ @opcode;
+ } else {
+ ();
+ }
+};
+
+my $pinsrd = sub {
+ if (shift =~ /\$([0-9]+),(%\w+),%xmm([0-9]+)/) {
+ my @opcode=(0x66);
+ $imm=$1;
+ $src=$2;
+ $dst=$3;
+ if ($src =~ /%r([0-9]+)d/) { $src = $1; }
+ elsif ($src =~ /%e/) { $src = $regrm{$src}; }
+ rex(\@opcode,$dst,$src);
+ push @opcode,0x0f,0x3a,0x22;
+ push @opcode,0xc0|(($dst&7)<<3)|($src&7); # ModR/M
+ push @opcode,$imm;
+ @opcode;
+ } else {
+ ();
+ }
+};
+
+my $pshufb = sub {
+ if (shift =~ /%xmm([0-9]+),%xmm([0-9]+)/) {
+ my @opcode=(0x66);
+ rex(\@opcode,$2,$1);
+ push @opcode,0x0f,0x38,0x00;
+ push @opcode,0xc0|($1&7)|(($2&7)<<3); # ModR/M
+ @opcode;
+ } else {
+ ();
+ }
+};
+