rv64 implement muldi3 intrinsic

[johannst]

This fixes #459.

Implement the __muldi3 intrinsic to prevent infinite recursion during
multiplication on rv64 without the 'm' extension.

[Amanieu]

In the case of rv32im we would want to use the existing implementation which builds a 64-bit multiply on top of 32-bit multiplies. Could this be made conditional to just targets without m?

[johannst]

For rv32 the rv64 specific __muldi3 intrinsic is not active.
But I agree for the rv64im case. I'll see if we can only enable this in case the m extension is not available.

[Amanieu]

Even though LLVM doesn't actually use __muldi3 for 64-bit multiply on rv32im, we should still make it available because it may be used when linking with C code generated by GCC. Same with __mulsi3.

[johannst]

Okay thanks for the clarification.

Let me try to summarize to check if I understood you correctly

• rv32i: riscv:__mulsi3, riscv:__muldi3
• rv32im: riscv:__mulsi3, int/mul:__muldi3
• rv64i: riscv:__mulsi3, riscv:__muldi3
• rv64im: riscv:__mulsi3, int/mul:__muldi3

Please correct me if I got you wrong.

[Amanieu]

That's correct.

[johannst]

I adapted the conditions accordingly.
The one for int/mul:__muldi3 became somewhat lengthy, if you see a better way I am happy to make it more readable.