The story begun when i`ve started to make indentation fixes in the DAC960 driver.
And in order to ensure i didn`t broke anything i was checking a diff between the
resulting object files.
Surprisingly enough i`ve realised soon that indeed some indentation changes
give gcc a reason to produce different code.
One of the cases is below, all three of them are in the attached .tar.gz file.
The code in question is the 2.5.72-bk1 kernel, however there was no changes
in the related code for some time, so plain .72 should be safe.
The examples are in the form of pairs of a C diff, and a "objdump -d" output diff.
The C diff:
diff -X scripts/Xrule -urN 25/drivers/block/DAC960.c 25dac/drivers/block/DAC960.c
--- 25/drivers/block/DAC960.c 2003-06-17 01:09:50.000000000 +0400
+++ 25dac/drivers/block/DAC960.c 2003-06-29 22:11:01.000000000 +0400
@@ -272,8 +272,7 @@
dma_addr_t RequestSenseDMA;
struct pci_pool *RequestSensePool = NULL;
- if (Controller->FirmwareType == DAC960_V1_Controller)
- {
+ if (Controller->FirmwareType == DAC960_V1_Controller) {
CommandAllocationLength = offsetof(DAC960_Command_T, V1.EndMarker);
CommandAllocationGroupSize = DAC960_V1_CommandAllocationGroupSize;
ScatterGatherPool = pci_pool_create("DAC960_V1_ScatterGather",
--- ./origDAC960.o.d 2003-06-29 21:02:55.000000000 +0400
+++ ./newDAC960.o.d 2003-06-29 22:13:46.000000000 +0400
@@ -1,5 +1,5 @@
-origDAC960.o: file format elf32-i386
+./newDAC960.o: file format elf32-i386
Disassembly of section .text:
@@ -5837,7 +5837,7 @@
52a8: 84 c0 test %al,%al
52aa: 75 14 jne 52c0 <DAC960_V1_ProcessCompletedCommand+0x80>
52ac: 0f 0b ud2a
- 52ae: 7d 0d jge 52bd <DAC960_V1_ProcessCompletedCommand+0x7d>
+ 52ae: 7c 0d jl 52bd <DAC960_V1_ProcessCompletedCommand+0x7d>
52b0: 27 daa
52b1: 00 00 add %al,(%eax)
52b3: 00 8d b6 00 00 00 add %cl,0xb6(%ebp)
@@ -5951,7 +5951,7 @@
5421: 84 c0 test %al,%al
5423: 0f 85 97 fe ff ff jne 52c0 <DAC960_V1_ProcessCompletedCommand+0x80>
5429: 0f 0b ud2a
- 542b: 8f 0d 27 00 00 00 popl 0x27
+ 542b: 8e 0d 27 00 00 00 movl 0x27,%cs
5431: e9 8a fe ff ff jmp 52c0 <DAC960_V1_ProcessCompletedCommand+0x80>
5436: 89 1c 24 mov %ebx,(%esp,1)
5439: e8 fc ff ff ff call 543a <DAC960_V1_ProcessCompletedCommand+0x1fa>
@@ -7414,7 +7414,7 @@
6ba2: 84 c0 test %al,%al
6ba4: 75 0a jne 6bb0 <DAC960_V2_ProcessCompletedCommand+0xa0>
6ba6: 0f 0b ud2a
- 6ba8: bc 11 27 00 00 mov $0x2711,%esp
+ 6ba8: bb 11 27 00 00 mov $0x2711,%ebx
6bad: 00 89 f6 83 bc 24 add %cl,0x24bc83f6(%ecx)
6bb3: 84 00 test %al,(%eax)
6bb5: 00 00 add %al,(%eax)
Thats it.
The point is i thought and hoped that gcc abstract syntax tree constructor is
indentation invariant, and that is seemingly not true.
regards, Samium Gromoff
On Mon, 30 Jun 2003, Samium Gromoff wrote:
>
> - if (Controller->FirmwareType == DAC960_V1_Controller)
> - {
> + if (Controller->FirmwareType == DAC960_V1_Controller) {
> -origDAC960.o: file format elf32-i386
> +./newDAC960.o: file format elf32-i386
>
> Disassembly of section .text:
>
> @@ -5837,7 +5837,7 @@
> 52a8: 84 c0 test %al,%al
> 52aa: 75 14 jne 52c0 <DAC960_V1_ProcessCompletedCommand+0x80>
> 52ac: 0f 0b ud2a
> - 52ae: 7d 0d jge 52bd <DAC960_V1_ProcessCompletedCommand+0x7d>
> + 52ae: 7c 0d jl 52bd <DAC960_V1_ProcessCompletedCommand+0x7d>
> 52b0: 27 daa
> 52b1: 00 00 add %al,(%eax)
> 52b3: 00 8d b6 00 00 00 add %cl,0xb6(%ebp)
> @@ -5951,7 +5951,7 @@
> 5421: 84 c0 test %al,%al
> 5423: 0f 85 97 fe ff ff jne 52c0 <DAC960_V1_ProcessCompletedCommand+0x80>
> 5429: 0f 0b ud2a
> - 542b: 8f 0d 27 00 00 00 popl 0x27
> + 542b: 8e 0d 27 00 00 00 movl 0x27,%cs
> 5431: e9 8a fe ff ff jmp 52c0 <DAC960_V1_ProcessCompletedCommand+0x80>
> 5436: 89 1c 24 mov %ebx,(%esp,1)
> 5439: e8 fc ff ff ff call 543a <DAC960_V1_ProcessCompletedCommand+0x1fa>
> @@ -7414,7 +7414,7 @@
> 6ba2: 84 c0 test %al,%al
> 6ba4: 75 0a jne 6bb0 <DAC960_V2_ProcessCompletedCommand+0xa0>
> 6ba6: 0f 0b ud2a
> - 6ba8: bc 11 27 00 00 mov $0x2711,%esp
> + 6ba8: bb 11 27 00 00 mov $0x2711,%ebx
> 6bad: 00 89 f6 83 bc 24 add %cl,0x24bc83f6(%ecx)
> 6bb3: 84 00 test %al,(%eax)
> 6bb5: 00 00 add %al,(%eax)
>
> Thats it.
> The point is i thought and hoped that gcc abstract syntax tree constructor is
> indentation invariant, and that is seemingly not true.
It's okay, no need to worry. See the "ud2a"s just above the differences?
Those are BUG()s, and they're going to be followed by a short __LINE__
then __FILE__ pointer. Your indentation change removed one line, so the
BUG()'s __LINE__ numbers have gone down one. (And it takes a while for
the disassembly to get back to sanity with the instructions thereafter.)
Hugh
On Mon, 30 Jun 2003 08:04:03 +0100 (BST)
Hugh Dickins <[email protected]> wrote:
> On Mon, 30 Jun 2003, Samium Gromoff wrote:
> >
> > - if (Controller->FirmwareType == DAC960_V1_Controller)
> > - {
> > + if (Controller->FirmwareType == DAC960_V1_Controller) {
>
> > -origDAC960.o: file format elf32-i386
> > +./newDAC960.o: file format elf32-i386
> >
> > Disassembly of section .text:
> >
> > @@ -5837,7 +5837,7 @@
> > 52a8: 84 c0 test %al,%al
> > 52aa: 75 14 jne 52c0 <DAC960_V1_ProcessCompletedCommand+0x80>
> > 52ac: 0f 0b ud2a
> > - 52ae: 7d 0d jge 52bd <DAC960_V1_ProcessCompletedCommand+0x7d>
> > + 52ae: 7c 0d jl 52bd <DAC960_V1_ProcessCompletedCommand+0x7d>
> > 52b0: 27 daa
> > 52b1: 00 00 add %al,(%eax)
> > 52b3: 00 8d b6 00 00 00 add %cl,0xb6(%ebp)
> > @@ -5951,7 +5951,7 @@
> > 5421: 84 c0 test %al,%al
> > 5423: 0f 85 97 fe ff ff jne 52c0 <DAC960_V1_ProcessCompletedCommand+0x80>
> > 5429: 0f 0b ud2a
> > - 542b: 8f 0d 27 00 00 00 popl 0x27
> > + 542b: 8e 0d 27 00 00 00 movl 0x27,%cs
> > 5431: e9 8a fe ff ff jmp 52c0 <DAC960_V1_ProcessCompletedCommand+0x80>
> > 5436: 89 1c 24 mov %ebx,(%esp,1)
> > 5439: e8 fc ff ff ff call 543a <DAC960_V1_ProcessCompletedCommand+0x1fa>
> > @@ -7414,7 +7414,7 @@
> > 6ba2: 84 c0 test %al,%al
> > 6ba4: 75 0a jne 6bb0 <DAC960_V2_ProcessCompletedCommand+0xa0>
> > 6ba6: 0f 0b ud2a
> > - 6ba8: bc 11 27 00 00 mov $0x2711,%esp
> > + 6ba8: bb 11 27 00 00 mov $0x2711,%ebx
> > 6bad: 00 89 f6 83 bc 24 add %cl,0x24bc83f6(%ecx)
> > 6bb3: 84 00 test %al,(%eax)
> > 6bb5: 00 00 add %al,(%eax)
> >
> > Thats it.
> > The point is i thought and hoped that gcc abstract syntax tree constructor is
> > indentation invariant, and that is seemingly not true.
>
> It's okay, no need to worry. See the "ud2a"s just above the differences?
> Those are BUG()s, and they're going to be followed by a short __LINE__
> then __FILE__ pointer. Your indentation change removed one line, so the
> BUG()'s __LINE__ numbers have gone down one. (And it takes a while for
> the disassembly to get back to sanity with the instructions thereafter.)
Uhhuh, i see now... those by-one differences looked strange for me... :-)
>
> Hugh
>
>
--
regards, Samium Gromoff
Samium Gromoff <[email protected]> said:
[...]
> Surprisingly enough i`ve realised soon that indeed some indentation changes
> give gcc a reason to produce different code.
[...]
> --- ./origDAC960.o.d 2003-06-29 21:02:55.000000000 +0400
> +++ ./newDAC960.o.d 2003-06-29 22:13:46.000000000 +0400
> - 52ae: 7d 0d jge 52bd <DAC960_V1_ProcessCompletedCommand+0x7d>
> + 52ae: 7c 0d jl 52bd <DAC960_V1_ProcessCompletedCommand+0x7d>
[...]
> - 542b: 8f 0d 27 00 00 00 popl 0x27
> + 542b: 8e 0d 27 00 00 00 movl 0x27,%cs
[...]
> - 6ba8: bc 11 27 00 00 mov $0x2711,%esp
> + 6ba8: bb 11 27 00 00 mov $0x2711,%ebx
This looks like 1-bit errors to my eye...
--
Dr. Horst H. von Brand User #22616 counter.li.org
Departamento de Informatica Fono: +56 32 654431
Universidad Tecnica Federico Santa Maria +56 32 654239
Casilla 110-V, Valparaiso, Chile Fax: +56 32 797513