Extended Description

I initially spotted this issue with AOMP but it seems from upstream clang.
ROCm/aomp#24

reproducer:
git clone https://github.com/ye-luo/miniqmc
cd miniqmc/build
cmake -DCMAKE_CXX_COMPILER=clang++ -DENABLE_OFFLOAD=ON
-DUSE_OBJECT_TARGET=ON -DCMAKE_EXE_LINKER_FLAGS="-v" ..
make -j32 check_spo_batched

all the 6 kernels use 254 registers.

Then I comment out "target teams" at 159, 311, 405.
make -j32 check_spo_batched
now all the 3 kernels left use 243 registers.

If I add
-DCMAKE_CXX_FLAGS="-Xcuda-ptxas -v" to cmake and print out register usage reported by ptxas. The three kernels take 146, 30, 30 registers when compiled.

I think the register usage is fine when kernels are compiled individually.
Somehow at linking, all the assembled kernels get the worst register usage among all the individual kernels.

It destroys performance completely.