llvm-D42262-jumpthreading-not-i1
authorLLVM Packaging Team <pkg-llvm-team@lists.alioth.debian.org>
Fri, 15 May 2020 16:25:16 +0000 (17:25 +0100)
committerScott Kitterman <scott@kitterman.com>
Fri, 15 May 2020 16:25:16 +0000 (17:25 +0100)
commitd7fe2634259d6d0c3b7636f424d7a4d386663b16
tree77367fc14374f63508eef30e76389e274b67d141
parentbfefc43a96a01422831de02f0a742d9cc0c0929c
llvm-D42262-jumpthreading-not-i1

commit 6a311a7a804831fea43cfb2f61322adcb407a1af
Author: Keno Fischer <keno@juliacomputing.com>
Date:   Thu Jan 18 15:57:05 2018 -0500

    [JumpThreading] Don't restrict cast-traversal to i1

    Summary:
    In D17663, JumpThreading learned to look trough simple cast instructions,
    but only if the source of those cast instructions was a phi/cmp i1
    (in an effort to limit compile time effects). I think this condition
    is too restrictive. For switches with limited value range, InstCombine
    will readily introduce an extra `trunc` instruction to a smaller
    integer type (e.g. from i8 to i2), leaving us in the somewhat perverse
    situation that jump-threading would work before running instcombine,
    but not after. Since instcombine produces this pattern, I think we
    need to consider it canonical and support it in JumpThreading.
    In general, for limiting recursion, I think the existing restriction
    to phi and cmp nodes should be sufficient to avoid looking through
    unprofitable chains of instructions.

    Reviewers: haicheng, gberry, bmakam, mcrosier

    Subscribers: llvm-commits

    Differential Revision: https://reviews.llvm.org/D42262

Gbp-Pq: Topic julia
Gbp-Pq: Name llvm-D42262-jumpthreading-not-i1.patch
lib/Transforms/Scalar/JumpThreading.cpp
test/Transforms/JumpThreading/basic.ll