On Wed, 28 Jun 2017, Ralf Baechle wrote: > > Wouldn't the more correct thing be to modify the memcpy loop such that > > prefetches do not fetch the larger offset? Just turning off prefetch > > altogether in memcpy for r6 seems like a heavy hammer to me... > > It's the same heavy hammer we've been using for ages to deal with this > problem for other configurations as well. So while you're right I don't > want to force Aleksandar to come up with The One Perfect Fix. Though > that would be lovely :) The The One Perfect Fix is to switch to the `at' assembly mode as with PREFE (which suffers from the same limitation and uses the same internal operand code even), as GAS does the right thing automagically if allowed, even for R6: $ cat pref.s .text foo: pref 0, 32767($2) $ as -32 -mips32 -o pref.o pref.s $ as -32 -mips32r6 -o pref6.o pref.s $ objdump -d pref*.o pref.o: file format elf32-tradbigmips Disassembly of section .text: 00000000 <foo>: 0: cc407fff pref 0x0,32767(v0) ... pref6.o: file format elf32-tradbigmips Disassembly of section .text: 00000000 <foo>: 0: 24417fff addiu at,v0,32767 4: 7c200035 pref 0x0,0(at) ... Implementing it should be a trivial exercise given the existing example. Maciej
