On Fri, Oct 19, 2012 at 02:06:57PM +0100, Peter Maydell wrote: > Two minor patches which inline some operations rather than using helper > functions. The 64 bit negate is a no-brainer since there's a simple > TCG op for it. For abs we implement in terms of movcond: > movi_i32 tmp6,$0x0 > neg_i32 tmp7,tmp5 > movcond_i32 tmp5,tmp5,tmp6,tmp5,tmp7,gt > which the x86-64 backend turns into: > 0x603b53a7: mov %ebp,%ebx > 0x603b53a9: neg %ebx > 0x603b53ab: mov %ebp,%r12d > 0x603b53ae: test %ebp,%ebp > 0x603b53b0: cmovle %ebx,%r12d > > Not sure why it felt the need to use an extra move there, but
Difficult to say without saying the whole TB, but likely because both %ebp and %r12d are not dead after the movcond, and are used separately. -- Aurelien Jarno GPG: 1024D/F1BCDB73 aurel...@aurel32.net http://www.aurel32.net