[Mesa-dev] [PATCH 3/4] ac/nir: fix 64-bit shifts

Connor Abbott cwabbott0 at gmail.com
Thu Aug 17 00:58:29 UTC 2017


I didn't enable the Int64 capability for radv until the next patch in
this series, so the problem this fixes was never exposed to users at
all. I just ran into it when enabling VK_EXT_shader_ballot -- the
extension doesn't require it, but the only way to generate SPIR-V
using it with glslang is with the ARB_shader_ballot GLSL extension
that does use 64-bit integers.

P.S. it would be nice to have some Int64 tests in the Vulkan CTS so we
can catch stuff like this easier.

On Wed, Aug 16, 2017 at 8:49 PM, Andres Gomez <agomez at igalia.com> wrote:
> Connor, it looks like we could want this patch in 17.1 (?)
>
> On Fri, 2017-06-30 at 19:56 -0700, Connor Abbott wrote:
>> From: Connor Abbott <cwabbott0 at gmail.com>
>>
>> NIR always makes the shift amount 32 bits, but LLVM asserts if the two
>> sources aren't the same type. Zero-extend the shift amount to make LLVM
>> happy.
>>
>> Signed-off-by: Connor Abbott <cwabbott0 at gmail.com>
>> ---
>>  src/amd/common/ac_nir_to_llvm.c | 15 ++++++++++++---
>>  1 file changed, 12 insertions(+), 3 deletions(-)
>>
>> diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
>> index 88f3f44..e72747a 100644
>> --- a/src/amd/common/ac_nir_to_llvm.c
>> +++ b/src/amd/common/ac_nir_to_llvm.c
>> @@ -1621,13 +1621,22 @@ static void visit_alu(struct nir_to_llvm_context *ctx, const nir_alu_instr *inst
>>               result = LLVMBuildXor(ctx->builder, src[0], src[1], "");
>>               break;
>>       case nir_op_ishl:
>> -             result = LLVMBuildShl(ctx->builder, src[0], src[1], "");
>> +             result = LLVMBuildShl(ctx->builder, src[0],
>> +                                   LLVMBuildZExt(ctx->builder, src[1],
>> +                                                 LLVMTypeOf(src[0]), ""),
>> +                                   "");
>>               break;
>>       case nir_op_ishr:
>> -             result = LLVMBuildAShr(ctx->builder, src[0], src[1], "");
>> +             result = LLVMBuildAShr(ctx->builder, src[0],
>> +                                    LLVMBuildZExt(ctx->builder, src[1],
>> +                                                  LLVMTypeOf(src[0]), ""),
>> +                                    "");
>>               break;
>>       case nir_op_ushr:
>> -             result = LLVMBuildLShr(ctx->builder, src[0], src[1], "");
>> +             result = LLVMBuildLShr(ctx->builder, src[0],
>> +                                    LLVMBuildZExt(ctx->builder, src[1],
>> +                                                  LLVMTypeOf(src[0]), ""),
>> +                                    "");
>>               break;
>>       case nir_op_ilt:
>>               result = emit_int_cmp(&ctx->ac, LLVMIntSLT, src[0], src[1]);
> --
> Br,
>
> Andres


More information about the mesa-dev mailing list