On 3/20/25 17:48, Richard Henderson wrote:
On 3/18/25 17:23, Pierrick Bouvier wrote:
uint64_t translator_ldq(CPUArchState *env, DisasContextBase *db, vaddr pc)
{
- uint64_t raw, tgt;
+ uint64_t val;
- if (translator_ld(env, db, &raw, pc, sizeof(raw))) {
- tgt = tswap64(raw);
- } else {
- tgt = cpu_ldq_code(env, pc);
- raw = tswap64(tgt);
- record_save(db, pc, &raw, sizeof(raw));
+ if (!translator_ld(env, db, &val, pc, sizeof(val))) {
+ MemOpIdx oi = make_memop_idx(MO_UQ, db->code_mmuidx);
+ val = cpu_ldq_code_mmu(env, pc, oi, 0);
+ record_save(db, pc, &val, sizeof(val));
}
- return tgt;
+ return tswap64(val);
}
void translator_fake_ld(DisasContextBase *db, const void *data, size_t len)
If I understand correctly, cpu_ldq_code_mmu performs the tswap call we used to
before.
Incorrect: cpu_ldq_code_mmu has no tswap.
It has a conditional bswap, if MO_BSWAP is set, but that's not true for the
MO_UQ used
here. Therefore both the direct load in translator_ld and the cpu_ld*_code_mmu
function
call both produce host-endian values.
Therefore the tswap at the end correctly swaps host to target-endianness.
Oh right, missed that.
r~
