Disable indvar widening if arithmetics on the wider type are more expensive
authorJingyue Wu <jingyue@google.com>
Wed, 12 Nov 2014 06:58:45 +0000 (06:58 +0000)
committerJingyue Wu <jingyue@google.com>
Wed, 12 Nov 2014 06:58:45 +0000 (06:58 +0000)
commit635a9b14fa91dcd6aedadb37aa232c24817e76cd
treedce9946fbf1faad094e29f38bdb80476b24f16b4
parent83a63877dc4ebfc2db71539a7be559205f6f547c
Disable indvar widening if arithmetics on the wider type are more expensive

Summary:
IndVarSimplify should not widen an indvar if arithmetics on the wider
indvar are more expensive than those on the narrower indvar. For
instance, although NVPTX64 treats i64 as a legal type, an ADD on i64 is
twice as expensive as that on i32, because the hardware needs to
simulate a 64-bit integer using two 32-bit integers.

Split from D6188, and based on D6195 which adds NVPTXTargetTransformInfo.

Fixes PR21148.

Test Plan:
Added @indvar_32_bit that verifies we do not widen an indvar if the arithmetics
on the wider type are more expensive.

Reviewers: jholewinski, eliben, meheff, atrick

Reviewed By: atrick

Subscribers: jholewinski, llvm-commits

Differential Revision: http://reviews.llvm.org/D6196

llvm-svn: 221772
llvm/lib/Target/NVPTX/NVPTXTargetTransformInfo.cpp
llvm/lib/Transforms/Scalar/IndVarSimplify.cpp
llvm/test/Transforms/IndVarSimplify/no-widen-expensive.ll [new file with mode: 0644]