Re: [PATCH 04/12] target/arm: Fill in TCGCPUOps.pointer_wrap

[Richard Henderson <richard.henderson@linaro.org>]

On 5/26/25 19:21, Philippe Mathieu-Daudé wrote:
> +Gustavo
> 
> On 4/5/25 22:57, Richard Henderson wrote:
>> For a-profile, check A32 vs A64 state.
>> For m-profile, use cpu_pointer_wrap_uint32.
>>
>> Cc: qemu-arm@nongnu.org
>> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>> ---
>>   target/arm/cpu.c         | 24 ++++++++++++++++++++++++
>>   target/arm/tcg/cpu-v7m.c |  1 +
>>   2 files changed, 25 insertions(+)
>>
>> diff --git a/target/arm/cpu.c b/target/arm/cpu.c
>> index 45cb6fd7ee..18edcf49c6 100644
>> --- a/target/arm/cpu.c
>> +++ b/target/arm/cpu.c
>> @@ -2710,6 +2710,29 @@ static const struct SysemuCPUOps arm_sysemu_ops = {
>>   #endif
>>   #ifdef CONFIG_TCG
>> +#ifndef CONFIG_USER_ONLY
>> +static vaddr aprofile_pointer_wrap(CPUState *cs, int mmu_idx,
>> +                                   vaddr result, vaddr base)
>> +{
>> +    /*
>> +     * The Stage2 and Phys indexes are only used for ptw on arm32,
>> +     * and all pte's are aligned, so we never produce a wrap for these.
>> +     * Double check that we're not truncating a 40-bit physical address.
>> +     */
>> +    assert((unsigned)mmu_idx < (ARMMMUIdx_Stage2_S & ARM_MMU_IDX_COREIDX_MASK));
>> +
>> +    if (!is_a64(cpu_env(cs))) {
>> +        return (uint32_t)result;
>> +    }
>> +
>> +    /*
>> +     * TODO: For FEAT_CPA2, decide how to we want to resolve
>> +     * Unpredictable_CPACHECK in AddressIncrement.
>> +     */
>> +    return result;
>> +}
>> +#endif /* !CONFIG_USER_ONLY */
>> +
>>   static const TCGCPUOps arm_tcg_ops = {
>>       .mttcg_supported = true,
>>       /* ARM processors have a weak memory model */
>> @@ -2729,6 +2752,7 @@ static const TCGCPUOps arm_tcg_ops = {
>>       .untagged_addr = aarch64_untagged_addr,
>>   #else
>>       .tlb_fill_align = arm_cpu_tlb_fill_align,
>> +    .pointer_wrap = aprofile_pointer_wrap,
> 
> IIUC this is also used by non A-profiles (R-profiles and
> non Cortex cores).

Yes, r-profile is mostly a-profile.  Those non-cortex cores are also a-profile: armv[456].

The point is the separation between m-profile and not. In particular, the mmu indexes are 
different between A and M (see ARM_MMU_IDX_TYPE_MASK). The assert would not be valid for 
m-profile. We can avoid a check vs ARM_FEATURE_M by only using this function for 
not-m-profile.


r~