[PATCH v1 2/2] LoongArch: Add call and tail pseudo instructions
mengqinggang
mengqinggang@loongson.cn
Fri Dec 1 02:44:32 GMT 2023
在 2023/12/1 上午10:31, Xi Ruoyao 写道:
> On Fri, 2023-12-01 at 10:26 +0800, mengqinggang wrote:
>> We consider generating call and tail by default in the normal code model
>> in the future,
>> and supporting relax to b/bl.
> Should be fine for call, but for tail it may not be a good idea. The
> problem is the compiler have to reserve a temporary register for tail,
> thus increasing register pressure and this cannot be solved by relax.
> So IMO call/tail (or call36/tail36) should still be limited to -
> mcmodel=medium.
This may require some performance testing before make a decision.
>> For the extreme code model, the function call are implemented by no-plt.
>> There are two instruction sequences, the second will use got entry.
>> One call64 macro instruction can't expand to two instruction sequences.
> Naming them different: call64.pcrel and call64.got ?
>
>> int f();
>> pcalau12i $r13,%pc_hi20(f)
>> addi.d $r12,$r0,%pc_lo12(f)
>> lu32i.d $r12,%pc64_lo20(f)
>> lu52i.d $r12,$r12,%pc64_hi12(f)
>> add.d $r12,$r13,$r12
>> jirl $r1,$r12,0
>>
>>
>> extern int f();
>> pcalau12i $r13,%got_pc_hi20(f)
>> addi.d $r12,$r0,%got_pc_lo12(f)
>> lu32i.d $r12,%got64_pc_lo20(f)
>> lu52i.d $r12,$r12,%got64_pc_hi12(f)
>> ldx.d $r12,$r13,$r12
>> jirl $r1,$r12,0
>>
>>
>>
>> 在 2023/12/1 上午3:08, WANG Xuerui 写道:
>>> On 12/1/23 00:10, Xi Ruoyao wrote:
>>>> I'd not use a plain "call" or "tail" as the pseudo name. The problem is
>>>> people not very familiar with LoongArch may use them instead of "bl" or
>>>> "b" everywhere and regress the code.
>>>>
>>>> Esp. the programmers know some RISC-V: in Aug 28 we've noticed several
>>>> RISC-V developers were attending Loongson user conference and they may
>>>> contribute some LoongArch code too in the future. In RISC-V assembly
>>>> call/tail are aliases of jal with rd = ra/r0.
>>>>
>>>> Maybe "call_medium"? "call_36"? "call_far"?
>>>>
>>>> Or if we can relax call/tail into b/bl it would be even better.
>>> I'd suggest making "call"/"tail" context-dependent and/or relaxable,
>>> and maybe providing unambiguous "{call,tail}.{16,36}" for explicit
>>> control.
>>>> On Thu, 2023-11-30 at 19:13 +0800, mengqinggang wrote:
>>>>> ---
>>>>> gas/testsuite/gas/loongarch/medium-call.d | 10 ++++++++--
>>>>> gas/testsuite/gas/loongarch/medium-call.s | 2 ++
>>>>> ld/testsuite/ld-loongarch-elf/medium-call.s | 2 ++
>>>>> opcodes/loongarch-opc.c | 11 +++++++++++
>>>>> 4 files changed, 23 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/gas/testsuite/gas/loongarch/medium-call.d
>>>>> b/gas/testsuite/gas/loongarch/medium-call.d
>>>>> index 4183818cb4f..3491760b96a 100644
>>>>> --- a/gas/testsuite/gas/loongarch/medium-call.d
>>>>> +++ b/gas/testsuite/gas/loongarch/medium-call.d
>>>>> @@ -10,6 +10,12 @@ Disassembly of section .text:
>>>>> [ ]+0:[ ]+1e000001[ ]+pcaddu18i[ ]+\$ra, 0
>>>>> [ ]+0: R_LARCH_CALL36[ ]+a
>>>>> [ ]+4:[ ]+4c000021[ ]+jirl[ ]+\$ra, \$ra, 0
>>>>> -[ ]+8:[ ]+1e00000c[ ]+pcaddu18i[ ]+\$t0, 0
>>>>> +[ ]+8:[ ]+1e000001[ ]+pcaddu18i[ ]+\$ra, 0
>>>>> [ ]+8: R_LARCH_CALL36[ ]+a
>>>>> -[ ]+c:[ ]+4c000180[ ]+jr[ ]+\$t0
>>>>> +[ ]+c:[ ]+4c000021[ ]+jirl[ ]+\$ra, \$ra, 0
>>>>> +[ ]+10:[ ]+1e00000c[ ]+pcaddu18i[ ]+\$t0, 0
>>>>> +[ ]+10: R_LARCH_CALL36[ ]+a
>>>>> +[ ]+14:[ ]+4c000180[ ]+jr[ ]+\$t0
>>>>> +[ ]+18:[ ]+1e00000c[ ]+pcaddu18i[ ]+\$t0, 0
>>>>> +[ ]+18: R_LARCH_CALL36[ ]+a
>>>>> +[ ]+1c:[ ]+4c000180[ ]+jr[ ]+\$t0
>>>>> diff --git a/gas/testsuite/gas/loongarch/medium-call.s
>>>>> b/gas/testsuite/gas/loongarch/medium-call.s
>>>>> index f2977d1c6d7..55a2fc6ffff 100644
>>>>> --- a/gas/testsuite/gas/loongarch/medium-call.s
>>>>> +++ b/gas/testsuite/gas/loongarch/medium-call.s
>>>>> @@ -1,6 +1,8 @@
>>>>> # call .L1, r1(ra) temp register, r1(ra) return register.
>>>>> + call a
>>>>> pcaddu18i $r1, %call36(a)
>>>>> jirl $r1, $r1, 0
>>>>> # tail .L1, r12(t0) temp register, r0(zero) return register.
>>>>> + tail a
>>>>> pcaddu18i $r12, %call36(a)
>>>>> jirl $r0, $r12, 0
>>>>> diff --git a/ld/testsuite/ld-loongarch-elf/medium-call.s
>>>>> b/ld/testsuite/ld-loongarch-elf/medium-call.s
>>>>> index 4d1888b76a0..31d4c2c1c75 100644
>>>>> --- a/ld/testsuite/ld-loongarch-elf/medium-call.s
>>>>> +++ b/ld/testsuite/ld-loongarch-elf/medium-call.s
>>>>> @@ -1,7 +1,9 @@
>>>>> .L1:
>>>>> # call .L1, r1(ra) temp register, r1(ra) return register.
>>>>> + call .L1
>>>>> pcaddu18i $r1, %call36(.L1)
>>>>> jirl $r1, $r1, 0
>>>>> # tail .L1, r12(t0) temp register, r0(zero) return register.
>>>>> + tail .L1
>>>>> pcaddu18i $r12, %call36(.L1)
>>>>> jirl $r0, $r12, 0
>>>>> diff --git a/opcodes/loongarch-opc.c b/opcodes/loongarch-opc.c
>>>>> index 15c7da6340c..1d02f020686 100644
>>>>> --- a/opcodes/loongarch-opc.c
>>>>> +++ b/opcodes/loongarch-opc.c
>>>>> @@ -293,6 +293,15 @@ const char *const loongarch_x_normal_name[32] =
>>>>> &LARCH_opts.ase_lp64, \
>>>>> &LARCH_opts.ase_gpcr
>>>>> +#define INSN_LA_CALL \
>>>>> + "pcaddu18i $ra,%%call36(%1);" \
>>>>> + "jirl $ra,$ra,0;", \
>>>>> + &LARCH_opts.ase_lp64, 0
>>>>> +
>>>>> +#define INSN_LA_TAIL \
>>>>> + "pcaddu18i $t0,%%call36(%1);" \
>>>>> + "jirl $zero,$t0,0;", \
>>>>> + &LARCH_opts.ase_lp64, 0
>>>>> static struct loongarch_opcode loongarch_macro_opcodes[] =
>>>>> {
>>>>> @@ -340,6 +349,8 @@ static struct loongarch_opcode
>>>>> loongarch_macro_opcodes[] =
>>>>> { 0, 0, "la.tls.gd", "r,l", INSN_LA_TLS_GD64, 0 },
>>>>> { 0, 0, "la.tls.gd", "r,l", INSN_LA_TLS_GD64_LARGE_ABS, 0 },
>>>>> { 0, 0, "la.tls.gd", "r,r,l",
>>>>> INSN_LA_TLS_GD64_LARGE_PCREL, 0 },
>>>>> + { 0, 0, "call", "la", INSN_LA_CALL, 0 },
>>>>> + { 0, 0, "tail", "la", INSN_LA_TAIL, 0 },
>>>>> { 0, 0, "pcaddi", "r,la", "pcaddi %1,
>>>>> %%pcrel_20(%2)", &LARCH_opts.ase_ilp32, 0, 0 },
>>>>> { 0, 0, 0, 0, 0, 0, 0, 0 } /* Terminate the list. */
>>>>> };
More information about the Binutils
mailing list