[RFC] RISC-V: Support Zibi extension.

[Jiawei]

Hi Jan,

Thanks for your review, will update in next version.

Best regards,
Jiawei

在 2026/3/13 20:40, Jan Beulich 写道:
> On 13.03.2026 13:30, Jiawei wrote:
>> --- /dev/null
>> +++ b/gas/testsuite/gas/riscv/zibi.d
>> @@ -0,0 +1,16 @@
>> +#as: -march=rv64ic_zibi
>> +#source: zibi.s
>> +#objdump: -drw
>> +
>> +.*:[ 	]+file format .*
>> +
>> +
>> +Disassembly of section .text:
>> +
>> +0+ <target>:
>> +[ 	]+[0-9a-f]+:[ 	]+00042063[ 	]+beqi[ 	]+s0,0x0,0 <target>[ 	]+0: R_RISCV_BRANCH	.*
> While the immediate looks correct here, ...
>
>> +[ 	]+[0-9a-f]+:[ 	]+fe442ee3[ 	]+beqi[ 	]+s0,0x24,0 <target>[ 	]+4: R_RISCV_BRANCH	.*
>> +[ 	]+[0-9a-f]+:[ 	]+fe043ce3[ 	]+bnei[ 	]+s0,0x20,0 <target>[ 	]+8: R_RISCV_BRANCH	.*
>> +[ 	]+[0-9a-f]+:[ 	]+fe443ae3[ 	]+bnei[ 	]+s0,0x24,0 <target>[ 	]+c: R_RISCV_BRANCH	.*
> ... these three don't match ...
>
>> --- /dev/null
>> +++ b/gas/testsuite/gas/riscv/zibi.s
>> @@ -0,0 +1,7 @@
>> +	.text
>> +target:
>> +	beqi	x8, 0, target
>> +	beqi	x8, 4, target
>> +	bnei	x8, 0, target
>> +	bnei	x8, 4, target
> ... the input.
>
>> --- a/opcodes/riscv-opc.c
>> +++ b/opcodes/riscv-opc.c
>> @@ -527,6 +527,7 @@ const struct riscv_opcode riscv_opcodes[] =
>>   {"beqz",        0, INSN_CLASS_I, "s,p",       MATCH_BEQ, MASK_BEQ|MASK_RS2, match_opcode, INSN_ALIAS|INSN_CONDBRANCH },
>>   {"beq",         0, INSN_CLASS_ZCA, "Cs,Cz,Cp",  MATCH_C_BEQZ, MASK_C_BEQZ, match_opcode, INSN_ALIAS|INSN_CONDBRANCH },
>>   {"beq",         0, INSN_CLASS_I, "s,t,p",     MATCH_BEQ, MASK_BEQ, match_opcode, INSN_CONDBRANCH },
>> +{"beqi",        0, INSN_CLASS_ZIBI, "s,>,p",     MATCH_BEQI, MASK_BEQI, match_opcode, INSN_CONDBRANCH },
>>   {"blez",        0, INSN_CLASS_I, "t,p",       MATCH_BGE, MASK_BGE|MASK_RS1, match_opcode, INSN_ALIAS|INSN_CONDBRANCH },
>>   {"bgez",        0, INSN_CLASS_I, "s,p",       MATCH_BGE, MASK_BGE|MASK_RS2, match_opcode, INSN_ALIAS|INSN_CONDBRANCH },
>>   {"bge",         0, INSN_CLASS_I, "s,t,p",     MATCH_BGE, MASK_BGE, match_opcode, INSN_CONDBRANCH },
> The new MATCH_BEQI imo wants to align with the adjacent MATCH_*.
>
>> @@ -543,6 +544,7 @@ const struct riscv_opcode riscv_opcodes[] =
>>   {"bnez",        0, INSN_CLASS_I, "s,p",       MATCH_BNE, MASK_BNE|MASK_RS2, match_opcode, INSN_ALIAS|INSN_CONDBRANCH },
>>   {"bne",         0, INSN_CLASS_ZCA, "Cs,Cz,Cp",  MATCH_C_BNEZ, MASK_C_BNEZ, match_opcode, INSN_ALIAS|INSN_CONDBRANCH },
>>   {"bne",         0, INSN_CLASS_I, "s,t,p",     MATCH_BNE, MASK_BNE, match_opcode, INSN_CONDBRANCH },
>> +{"bnei",        0, INSN_CLASS_ZIBI, "s,>,p",     MATCH_BNEI, MASK_BNEI, match_opcode, INSN_CONDBRANCH },
>>   {"addi",        0, INSN_CLASS_ZCA, "Ct,Cc,CK",  MATCH_C_ADDI4SPN, MASK_C_ADDI4SPN, match_c_addi4spn, INSN_ALIAS },
>>   {"addi",        0, INSN_CLASS_ZCA, "d,CU,Cj",   MATCH_C_ADDI, MASK_C_ADDI, match_rd_nonzero, INSN_ALIAS },
>>   {"addi",        0, INSN_CLASS_ZCA, "d,CU,z",    MATCH_C_NOP, MASK_C_ADDI|MASK_RVC_IMM, match_c_nop, INSN_ALIAS },
> Here it wants to align with at least one of the two.
>
> And then, just like one can write ADD instead of ADDI, shouldn't this similarly
> be possible for these two new insns (possibly requiring a spec amendment)?
>
> Jan