On Thu, Apr 06, 2006 at 11:23:32AM -0700, [EMAIL PROTECTED] wrote:
> Author: bernhard
> Date: Thu Apr 6 11:23:31 2006
> New Revision: 12129
>
> Modified:
> trunk/src/spf_render.c
>
> Log:
> Make some string formating test failures go away under
> Linux on i686.
> However I have no what had caused the failures.
>
>
> Modified: trunk/src/spf_render.c
> ==============================================================================
> --- trunk/src/spf_render.c (original)
> +++ trunk/src/spf_render.c Thu Apr 6 11:23:31 2006
> @@ -163,7 +163,14 @@
> }
>
> if ((info->flags & FLAG_WIDTH) && info->width > len) {
> - STRING *fill = CONST_STRING(interpreter, info->flags & FLAG_ZERO ?
> "0" : " ");
> + STRING *fill;
> +
> + if (info->flags & FLAG_ZERO) {
> + fill = CONST_STRING(interpreter, "0");
> + }
> + else {
> + fill = CONST_STRING(interpreter, " ");
> + }
>
> fill = string_repeat(interpreter, fill, info->width - len, NULL);
I think that this change is masking the true bug, and suspect that the true
bug will return when C compiler's optimiser is turned on. I have no knowledge
of x86 assembly language, so can't follow the logic of what the compiler is
generating, but I can see that the assembly code generated for handle_flags
differs with the application of this patch. Specifically, it appears that
when the patch is added, something extra is added to the stack. Specifically:
@@ -396,246 +396,262 @@ handle_flags:
testl %eax, %eax
je .L27
movl 12(%ebp), %eax
movl (%eax), %eax
cmpl -8(%ebp), %eax
jbe .L27
.LBB5:
- .loc 1 166 0
+ .loc 1 168 0
+ movl 12(%ebp), %eax
+ movl 8(%eax), %eax
+ shrl $2, %eax
+ andl $1, %eax
+ testl %eax, %eax
+ je .L28
+ .loc 1 169 0
movl 8(%ebp), %eax
movl 168(%eax), %eax
- addl $32, %eax
+ addl $132, %eax
movl (%eax), %eax
movl %eax, -12(%ebp)
- .loc 1 168 0
+ jmp .L29
+.L28:
+ .loc 1 172 0
+ movl 8(%ebp), %eax
+ movl 168(%eax), %eax
+ addl $124, %eax
+ movl (%eax), %eax
+ movl %eax, -12(%ebp)
+.L29:
+ .loc 1 175 0
pushl $0
movl 12(%ebp), %eax
movl -8(%ebp), %edx
movl (%eax), %eax
subl %edx, %eax
pushl %eax
pushl -12(%ebp)
pushl 8(%ebp)
call [EMAIL PROTECTED]
I assume that the thing added to the stack is a pointer to the generated
string "0" or " ", and that the bug goes away because there happens to be
a GC run triggered inside string_repeat, and with the temporary on the
stack it doesn't get garbage collected.
This is a hunch. But as far as I can make out there is no semantic difference
in the change you made to the C, so it should not have changed anything.
Nicholas Clark
