LLVM IR related to the issue

Files:

• minimized.ll: correct IR generated by rustc -Z merge-functions=disabled
• minimized-opt.ll: wrong IR after applying opt -mergefunc

Generating the wrong IR:

opt -O0 -mergefunc -S -o minimized-opt.ll minimized.ll

Writing the PTX:

llc -march=nvptx64 -mcpu=sm_35 -O3 -o minimized-llc.ptx minimized.ll
llc -march=nvptx64 -mcpu=sm_35 -O3 -o minimized-opt-llc.ptx minimized-opt.ll

previous readme

Executive summary: NVPTX + MergeFunctions = badness

The test case nocore.rs is not minimal but is still small and demonstrates the actual usage case that helped me to find this issue (namely, writing tests for NVPTX intrinsics from stdsimd).

I got nocore-mergefunc-nousealiases-bad.ptx and nocore-nomergefunc-ok.ptx by directly running rustc from a custom sysroot/toolchain that supports extra target options to control the MergeFunctions pass:

RUST_TARGET_PATH=`pwd` rustc --target nvptx64-nvidia-cuda.json -C opt-level=3 nocore.rs --emit asm

First, if you use a vanilla target json, currently rustc is known to complain:

LLVM ERROR: Module has aliases, which NVPTX does not support.

On the other hand, either disabling MergeFunctions outright, or keeping MergeFunctions but disabling the -mergefunc-use-aliases LLVM option, will successfully generate a PTX file. However the correctness of the PTX is not guaranteed!

Running ptxas -arch sm_35 -o nocore-bad.o nocore-mergefunc-nousealiases-bad.ptx yields the following error (I am using ptxas from CUDA 9.2):

ptxas ./nocore-mergefunc-nousealiases-bad.ptx, line 38; error   : Illegal call target, device function expected
ptxas fatal   : Ptx assembly aborted due to errors

Running ptxas -arch sm_35 -o nocore-ok.o nocore-nomergefunc-ok.ptx succeeds without error.

todo

Running llc:

llc -march=nvptx64 -mcpu=sm_35 -O3 -o=minimal-llc.s minimal.ll