Comment 2 for bug 850099

Thank you for the bug report. I've confirmed this with the gcc-linaro-4.6-2011.09 RC on ARM:

michaelh@leo1:~/linaro/bugs$ /tools/toolchains/arch/armv7l/gcc-linaro-4.6-2011.09-armv7l-natty-cbuild181-ursa4-cortexa9r1/bin/gcc -O2 -marm -S lp850099.c
michaelh@leo1:~/linaro/bugs$ cat lp850099.s
test:
 stmfd sp!, {r3, r4, r5, lr}
 rsbs r4, r0, #0
 rsc r5, r1, #0
 mov r3, r5
 ldrd r0, [r2]
 mov r2, r4
 bl __aeabi_lcmp
 cmp r0, #0
 movgt r0, #0
 movle r0, #1
 ldmfd sp!, {r3, r4, r5, pc}

Due to the memory reference, the compiler generates a call to the 64
bit signed compare helper function. Previous versions inlined the
compare:

test:
 stmfd sp!, {r4, r5}
 ldrd r4, [r2]
 rsbs r2, r0, #0
 rsc r3, r1, #0
 cmp r2, r4
 sbcs r1, r3, r5
 movlt r0, #0
 movge r0, #1
 ldmfd sp!, {r4, r5}
 bx lr

The cmp r2, r4; sbcs r1, r3, r5 sequence is from the *arm_cmpdi_insn
pattern.

This is a performance regression. It is exposed in the kernel as the
kernel doesn't have a implementation of __aeabi_lcmp, although it does
have __aeabi_ulcmp.

Changing the flags to -O2, -O3, or -mthumb doesn't work around the
problem. The issue does not exist in gcc-4.5.3, gcc-4.6.1,
gcc-linaro-4.6-2011.08, or upstream r178624.

I've set it to medium priority as it is a ftbfs and a performance
regression.