[Mesa-dev] [PATCH] avoid crashing in mod by 0 with llvmpipe

[Jeff Muizelaar]

Do you mind fixing up the language to match what you're looking for
and committing it?

Thanks,

-Jeff

On Fri, Jan 15, 2016 at 3:00 PM, Roland Scheidegger <sroland at vmware.com> wrote:
> Am 15.01.2016 um 20:13 schrieb Jeff Muizelaar:
>> This adds code that is basically the same as the code in smod, div and idiv.
> That would be umod, not smod, above (and udiv).
>
>> However, unlike idiv we return -1.
>>
>> diff --git a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_action.c
>> b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_action.c
>> index 0ad78b0..d17ef31 100644
>> --- a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_action.c
>> +++ b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_action.c
>> @@ -1524,18 +1524,32 @@ min_emit_cpu(
>>
>>  /* TGSI_OPCODE_MOD (CPU Only) */
>>  static void
>>  mod_emit_cpu(
>>     const struct lp_build_tgsi_action * action,
>>     struct lp_build_tgsi_context * bld_base,
>>     struct lp_build_emit_data * emit_data)
>>  {
>> -   emit_data->output[emit_data->chan] = lp_build_mod(&bld_base->int_bld,
>> -                                   emit_data->args[0], emit_data->args[1]);
>> +   LLVMBuilderRef builder = bld_base->base.gallivm->builder;
>> +   LLVMValueRef div_mask = lp_build_cmp(&bld_base->uint_bld,
>> +                    PIPE_FUNC_EQUAL, emit_data->args[1],
>> +                    bld_base->uint_bld.zero);
>> +   /* We want to make sure that we never divide/mod by zero to not
>> +    * generate sigfpe. We don't want to crash just because the
>> +    * shader is doing something weird. */
>> +   LLVMValueRef divisor = LLVMBuildOr(builder,
>> +                      div_mask,
>> +                      emit_data->args[1], "");
>> +   LLVMValueRef result = lp_build_mod(&bld_base->int_bld,
>> +                      emit_data->args[0], divisor);
>> +   /* umod by zero is guaranteed to return 0xffffffff */
> I think this should use language along the lines of what idiv does, i.e.
> no guaranteed value, pick -1. Because we don't say anything about that
> in the gallium tgsi docs section about that (unlike udiv/umod).
> (AFAIK we only really care for udiv/umod due to d3d10, which doesn't
> have the signed versions - glsl doesn't care either way.)
>
>> +   emit_data->output[emit_data->chan] = LLVMBuildOr(builder,
>> +                            div_mask,
>> +                            result, "");
>>  }
>>
>>  /* TGSI_OPCODE_NOT */
>>  static void
>>  not_emit_cpu(
>>     const struct lp_build_tgsi_action * action,
>>     struct lp_build_tgsi_context * bld_base,
>>     struct lp_build_emit_data * emit_data)
>
> Subject should probably say: gallivm: ...
>
> With that fixed: Reviewed-by: Roland Scheidegger <sroland at vmware.com>
>