Builtin memcpy in riscV target uses fld/fsd about compcert

[xavierleroy]

The only reason for using fld/fsd (64-bit FP loads and stores) in memcpy is that it transfers 8 bytes in one instruction even if we're on a RISC-V 32-bit machine, which lacks the ld/sd instructions But you're right that on a 64-bit RISC-V, ld/sd works just as well.

Also, the PowerPC and x86 ports avoid generating FP instructions for memcpy if the -fno-fpu flag is given. The RISC-V port should also honor this flag. I'll look into this shortly.

[xavierleroy]

This should be fixed by 35feefd . For 64-bit RISC-V, memcpy now uses X31, like in your patch. For 32-bit RISC-V, a FP register is used unless -fno-fpu is given, in which case the copy is done through X31 by chunks of at most 32 bits.

[xxuejie]

Thank you for the super fast fix! I totally neglected the 32 bit case :) The change deployed to address the issue here is perfect.