[Mesa-dev] [PATCH] gallivm: don't use vector selects with llvm 3.7

[Roland Scheidegger]

Am 17.04.2016 um 11:06 schrieb Jose Fonseca:
> On 16/04/16 22:30, sroland at vmware.com wrote:
>> From: Roland Scheidegger <sroland at vmware.com>
>>
>> llvm 3.7 sometimes simply miscompiles vector selects.
> 
> Reviewed-by: Jose Fonseca <jfonseca at vmware.com>
> 
>> See https://bugs.freedesktop.org/show_bug.cgi?id=94972
> 
> Great detective work.
> 
> I think you should make a LLVM test case out of the IR on that bug and
> propose on llvm-commits, to guarantee it won't regress again.  See
> test/CodeGen/X86/vselect-avx.ll for example.

After some more digging, there actually IS already a test for this
(pretty much exactly the same as mine). It was in 3.7 as well - of
course not very helpful with the wrong expected result...

This fixed it:

commit 30ead638b8b7ad4d70d25ef3c70f0d6b21522784
Author: Michael Kuperstein <michael.m.kuperstein at intel.com>
Date:   Thu Oct 8 08:13:02 2015 +0000

    [X86] Fix wrong treatment of multi-lane blends in
BUILD_VECTORtoBlendMask()

    This fixes two separate bugs:
    1) The mask for the high lane was not set correctly. That fixes PR24532.
    2) The transformation should bail out if it believes it involves
more than
    2 lanes, as it does not currently do anything sensible in this case.

    Differential Revision: http://reviews.llvm.org/D13505

    git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@249669
91177308-0d34-0410-b5e6-96231b3b80d8

Roland


> 
> Jose
> 
>> ---
>>   src/gallium/auxiliary/gallivm/lp_bld_logic.c | 8 +++++---
>>   1 file changed, 5 insertions(+), 3 deletions(-)
>>
>> diff --git a/src/gallium/auxiliary/gallivm/lp_bld_logic.c
>> b/src/gallium/auxiliary/gallivm/lp_bld_logic.c
>> index 91f316c..620aece 100644
>> --- a/src/gallium/auxiliary/gallivm/lp_bld_logic.c
>> +++ b/src/gallium/auxiliary/gallivm/lp_bld_logic.c
>> @@ -315,14 +315,16 @@ lp_build_select(struct lp_build_context *bld,
>>         mask = LLVMBuildTrunc(builder, mask,
>> LLVMInt1TypeInContext(lc), "");
>>         res = LLVMBuildSelect(builder, mask, a, b, "");
>>      }
>> -   else if (LLVMIsConstant(mask) ||
>> -            LLVMGetInstructionOpcode(mask) == LLVMSExt) {
>> +   else if (!(HAVE_LLVM == 0x0307) &&
>> +            (LLVMIsConstant(mask) ||
>> +             LLVMGetInstructionOpcode(mask) == LLVMSExt)) {
>>         /* Generate a vector select.
>>          *
>>          * Using vector selects should avoid emitting intrinsics hence
>> avoid
>> -       * hidering optimization passes, but vector selects weren't
>> properly
>> +       * hindering optimization passes, but vector selects weren't
>> properly
>>          * supported yet for a long time, and LLVM will generate poor
>> code when
>>          * the mask is not the result of a comparison.
>> +       * Also, llvm 3.7 may miscompile them (bug 94972).
>>          */
>>
>>         /* Convert the mask to a vector of booleans.
>>
> 
> 
> 
>