小立即在 risc-v 中的 lui

Question

For a risc-v instruction like lui x28 123 , venus shows that its machine code is 0x0007BE37 . However lui should load only upper 20 bits of immediate 123 . Since in hex 123 is 0x0000007B , I think the corresponding machine code should be 0x00000E37 .

I am puzzled why venus produces different results, is there a problem with my understanding of lui ?

Answer 1

When you provide a numeric value as a literal to an instruction like LUI , then some assemblers will take that as the exact immediate value you want encoded in the instruction. This is important and useful for compilers generating assembly or assembly programmers who want to use LUI and specify the exact bit pattern for the 20 bits in that instruction, without any interference from the assembler.

If you want it to take only the high 20 bits of your literal, there's compile time functions ¹ like %hi and %lo that some assemblers will accept, eg lui %hi(123) — the %hi is for use in lui and the %lo for use in an instruction that wants the 12 low bits, like addi and sw / lw .

_{¹ These may be evaluated as late as link time.}

---

These functions, ( %hi & %lo ) are complicated due to the traditional choice of pairing with instructions like lw and addi , wherein their immediate makes up the 12-bit rest, but itself is signed . So, the %hi function offsets the value by 1 in the specific cases where the lw 's offset will appear negative and will sign extend with 1's at runtime. So, when these are paired ² the environment makes things work, but if you were to use %hi() in LUI on a random constant without such pairing, you might not get the exact 20 bits you're expecting.

_{² More than one %lo can ~pair with the same %hi as long as the function arguments match. So, for example, a read/modify/write sequence to a global variable might have one lui %hi() for both lw %lo() and sw %lo() .}