I went through commit 7ffa5b939751 ("scsi: qla2xxx: Fix endianness
annotations in source files") and produced a new version without the
'const' changes and without the new local variables. I also tweaked a few
line breaks so that the line numbers did not change.
I then built the driver both with and without the patch:
mkdir /tmp/patched /tmp/unpatched
rm drivers/scsi/qla2xxx/built-in.a drivers/scsi/qla2xxx/*.o *.[is]
make KBUILD_CFLAGS=-save-temps drivers/scsi/qla2xxx/built-in.a
mv *.s /tmp/patched/
git checkout @^
rm drivers/scsi/qla2xxx/built-in.a drivers/scsi/qla2xxx/*.o *.[is]
make KBUILD_CFLAGS=-save-temps drivers/scsi/qla2xxx/built-in.a
mv *.s /tmp/unpatched/
meld /tmp/patched/ /tmp/unpatched/
This revealed some differences, in both x86_64 and i686 builds.
1. The wordsize changes in qla24xx_els_ct_entry() apparently prevented the
compiler from emitting zero-extension mov (movzwl) instructions.
diff -ru /tmp/unpatched/qla_isr.s /tmp/patched/qla_isr.s
--- /tmp/unpatched/qla_isr.s 2020-05-24 13:59:04.327714528 +1000
+++ /tmp/patched/qla_isr.s 2020-05-24 13:56:58.437529886 +1000
@@ -8947,11 +8947,9 @@
movw %ax, -26(%rbp)
movq -104(%rbp), %rax
movl 36(%rax), %eax
- movzwl %ax, %eax
movl %eax, -76(%rbp)
movq -104(%rbp), %rax
movl 40(%rax), %eax
- movzwl %ax, %eax
movl %eax, -72(%rbp)
cmpl $83, -108(%rbp)
jne .L601
@@ -8991,8 +8989,7 @@
movl $458752, -12(%rbp)
.L603:
movq -104(%rbp), %rax
- movl 32(%rax), %eax
- movzwl %ax, %edi
+ movl 32(%rax), %edi
movl -72(%rbp), %esi
movl -76(%rbp), %ecx
movzwl -26(%rbp), %edx
(and so on.)
2. The get_unaligned_le32() changes produce new pointer offsets in the
assembly code for qla82xx_get_table_desc() and qla82xx_get_data_desc().
diff -ru /tmp/unpatched/qla_target.s /tmp/patched/qla_target.s
--- /tmp/unpatched/qla_target.s 2020-05-24 14:02:32.178019380 +1000
+++ /tmp/patched/qla_target.s 2020-05-24 14:01:43.487947966 +1000
@@ -12884,10 +12884,10 @@
.cfi_offset 6, -16
movq %rsp, %rbp
.cfi_def_cfa_register 6
- subq $32, %rsp
- movq %rdi, -24(%rbp)
- movq %rsi, -32(%rbp)
- movq -32(%rbp), %rax
+ subq $64, %rsp
+ movq %rdi, -56(%rbp)
+ movq %rsi, -64(%rbp)
+ movq -64(%rbp), %rax
movl 52(%rax), %eax
movl %eax, -8(%rbp)
movl $24, -12(%rbp)
@@ -12895,62 +12895,62 @@
cmpl %eax, -12(%rbp)
cmovbe -12(%rbp), %eax
movl %eax, %edx
- movq -32(%rbp), %rax
+ movq -64(%rbp), %rax
movl %edx, 52(%rax)
- movq -24(%rbp), %rax
+ movq -56(%rbp), %rax
(and so on.)
Was this expected? I find it surprising...
FWIW, I think that the .s files could have been validated automatically
had the changes in the patch been confined to annotations.
