https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96428
Bug ID: 96428
Summary: [nvptx] nvptx_gen_shuffle does not handle V2DI mode –
Fails with an ICE
Product: gcc
Version: 10.0
Status: UNCONFIRMED
Keywords: ice-on-valid-code
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: burnus at gcc dot gnu.org
CC: vries at gcc dot gnu.org
Target Milestone: ---
Target: nvptx
Created attachment 48986
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=48986&action=edit
Test case (Fortran, use 'gfortran -fopenacc" with nvptx offloading)
This is a peudoregression due to
https://gcc.gnu.org/g:2ff0f48819c8a7ed5d7c03e2bfc02e5907e2ff1a – however, that
change looks fine. But for this harmlessly looking code, one now gets an ICE.
The ICE is due to nvptx_gen_shuffle not handling V2DI mode but now it is
produced for the testcase.
Additionally, I observe that shfl.mode.b32 (which is used for the implemented
shufflings) the Nvidia documention states: "Deprecated in PTX ISA version 6.0
in favor of shfl.sync." / "Not supported in PTX ISA version 6.4 for .target
sm_70 or higher."
[I find: "Volta (CUDA 9 and later): SM70 or SM_70, compute_70 – Tesla V100, GTX
1180 (GV104) [...]" / "Turing (CUDA 10 and later) SM75 or SM_75, compute_75 –
RTX 2080, Titan RTX, Quadro R8000".]