Felix Fietkau wrote: > Actually, it emits -16384 in both the working and non-working case, > I think what makes the difference of working vs non-working is this: > > Broken: (0xffffc000 is in v0) > - e8: 00021a02 srl v1,v0,0x8 > - ec: 00021200 sll v0,v0,0x8 > > Working: (0xffffc000 is in v1) > + e8: 7c623a00 ext v0,v1,0x8,0x8 > + ec: 00031a00 sll v1,v1,0x8 > > using ext instead of srl prevents the extra bits from spilling over.
Sure. I was looking at the assembly emitted by the other compilers I tested. None of those even ever use -16384 (0xffffc000). They all simply OR in 0xc000 (as an immediate value). I’m in touch with Linaro about this bug, so maybe we’ll get some traction without having to commit my workaround patch. _______________________________________________ openwrt-devel mailing list openwrt-devel@lists.openwrt.org https://lists.openwrt.org/mailman/listinfo/openwrt-devel
