This is the mail archive of the
binutils@sourceware.org
mailing list for the binutils project.
Re: Feature request: Set the direction flag on x86/x64 register->register operations
On Thu, Dec 18, 2008 at 9:16 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
> On Tue, Dec 9, 2008 at 4:51 AM, Stefan Dösinger <stefan@codeweavers.com> wrote:
>> Hi,
>>
>> Here is another version of my patch to set the direction flag on
>> register->register operations.
>>
>> The main change is that it is now controlled by a per-instruction suffix
>> instead of a command line switch. Alexandre Julliard didn't like the idea of
>> the command line switch, he prefers gcc to control the bit. I have to modify
>> gcc to make it generate a "movl %edi, %edi" like Windows has it anyway, so
>> going this route is ok for me.
>>
>> I added a test for this feature which specifically tests the instructions
>> that are important to me(mov %edi, %edi and mov %esp, %ebp) and a few more.
>> I ran the existing tests with the flag forced on for all instructions, and
>> they pass. There are a few false positives obviously because the patch
>> changes the encoding, but I checked all of those cases and the code works
>> correctly.
>>
>> I verified the behavior against msvc, and it does set the flag on 8, 16 and
>> 32 bit operations for sure. I didn't check 64 bit ones because I do not have
>> a 64 bit Windows installation to test, but this is moot now anyway because
>> gcc controls the flag. It doesn't affect this patch any longer.
>>
>
> Please use "movl.r" instead of "movlr".
That is what I have in mind:
--- ./tc-i386.c.reverse 2008-12-12 11:24:10.000000000 -0800
+++ ./tc-i386.c 2008-12-18 10:44:52.000000000 -0800
@@ -228,6 +228,10 @@ struct _i386_insn
/* TM holds the template for the insn were currently assembling. */
template tm;
+ /* MATCH holds insn match options. */
+ unsigned int options;
+#define OPTION_REVERSE_DIRECTION_BIT 1
+
/* SUFFIX holds the instruction size suffix for byte, word, dword
or qword, if given. */
char suffix;
@@ -2990,6 +2994,21 @@ parse_insn (char *line, char *mnemonic)
if (!current_templates)
{
+ /* Check match options. */
+ if (mnem_p[-2] == '.')
+ {
+ if (mnem_p[-1] == 'r')
+ i.options |= OPTION_REVERSE_DIRECTION_BIT;
+ else
+ goto check_suffix;
+ mnem_p[-2] = '\0';
+ current_templates = hash_find (op_hash, mnemonic);
+ }
+ }
+
+ if (!current_templates)
+ {
+check_suffix:
/* See if we can get a match by trimming off a suffix. */
switch (mnem_p[-1])
{
@@ -3712,6 +3731,11 @@ match_template (void)
&& operand_type_equal (&i.types [0], &acc32)
&& operand_type_equal (&i.types [1], &acc32))
continue;
+ if ((i.options & OPTION_REVERSE_DIRECTION_BIT) != 0
+ && (t->opcode_modifier.d
+ || t->opcode_modifier.floatd))
+ goto reverse;
+
case 3:
case 4:
case 5:
@@ -3728,6 +3752,7 @@ match_template (void)
if (!t->opcode_modifier.d && !t->opcode_modifier.floatd)
continue;
+reverse:
/* Try reversing direction of operands. */
overlap0 = operand_type_and (i.types[0], operand_types[1]);
overlap1 = operand_type_and (i.types[1], operand_types[0]);
Will it work for you?
--
H.J.