The code generated by the JIT for loading 16 bit and 32 bit constants is inefficient.

[markshannon]

This affects both x86-64 and AArch64.

Most guard micro-ops are checking against 16 or 32 bit version numbers, so this is an important use case.

The code we generate is too large and, on AArch64, performs extra memory accesses.

Constant	x86-64 now	x86-64 ideal	AArch64 now	AArch64 ideal
16 bit	1 10 byte instruction	1 5 byte instruction	2 instructions + 1 memory load	1 instruction
32 bit	1 10 byte instruction	1 5 byte instruction	2 instructions + 1 memory load	2 instructions

[brandtbucher]

Is the plan to handle this in the textual assembly transformation step? Or by relaxing the relocations at runtime?

[markshannon]

I think this needs to handled all the way through the tools chain, including the code generator to avoid Clang generating the load instructions. Or it might be possible to fix that with different compiler flags. I'll need to experiment. Have you given this any thought?