Re: [PATCH] tcg/aarch64: Fix constant subtraction in tcg_out_addsub2

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

On 3/5/21 3:45 AM, Peter Maydell wrote:
> On Sun, 28 Feb 2021 at 23:55, Richard Henderson
> <richard.henderson@linaro.org> wrote:
>>
>> An hppa guest executing
>>
>> 0x000000000000e05c:  ldil L%10000,r4
>> 0x000000000000e060:  ldo 0(r4),r4
>> 0x000000000000e064:  sub r3,r4,sp
>>
>> produces
>>
>>   ---- 000000000000e064 000000000000e068
>>   sub2_i32 tmp0,tmp4,r3,$0x1,$0x10000,$0x0
>>
>> after folding and constant propagation.  Then we hit
>>
>> tcg-target.c.inc:640: tcg_out_insn_3401: Assertion `aimm <= 0xfff' failed.
>>
>> because aimm is in fact -16, but unsigned.
>>
>> The ((bl < 0) ^ sub) condition which negates bl is incorrect and will
>> always lead to this abort.  If the constant is positive, sub will make
>> it negative; if the constant is negative, sub will keep it negative.
>>
>> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>> ---
>> -static inline void tcg_out_addsub2(TCGContext *s, TCGType ext, TCGReg rl,
>> -                                   TCGReg rh, TCGReg al, TCGReg ah,
>> -                                   tcg_target_long bl, tcg_target_long bh,
>> -                                   bool const_bl, bool const_bh, bool sub)
>> +static void tcg_out_addsub2(TCGContext *s, TCGType ext, TCGReg rl,
>> +                            TCGReg rh, TCGReg al, TCGReg ah,
>> +                            tcg_target_long bl, tcg_target_long bh,
>> +                            bool const_bl, bool const_bh, bool sub)
>>   {
>>       TCGReg orig_rl = rl;
>>       AArch64Insn insn;
> 
> Seems like an unrelated change ?

Mm.  Perhaps I should remove all of the inline markers all at once.

r~

> 
>> @@ -1423,11 +1423,13 @@ static inline void tcg_out_addsub2(TCGContext *s, TCGType ext, TCGReg rl,
>>       }
>>
>>       if (const_bl) {
>> -        insn = I3401_ADDSI;
>> -        if ((bl < 0) ^ sub) {
>> -            insn = I3401_SUBSI;
>> +        if (bl < 0) {
>>               bl = -bl;
>> +            insn = sub ? I3401_ADDSI : I3401_SUBSI;
>> +        } else {
>> +            insn = sub ? I3401_SUBSI : I3401_ADDSI;
>>           }
>> +
>>           if (unlikely(al == TCG_REG_XZR)) {
>>               /* ??? We want to allow al to be zero for the benefit of
>>                  negation via subtraction.  However, that leaves open the
>> --
>> 2.25.1
> 
> Otherwise
> Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
> 
> thanks
> -- PMM
>