[llvm-dev] Canonicalizing conditional loads

[Jason Eckhardt via llvm-dev]

>Is there a way to tell clang to always generate the code as the first function, or perhaps a llvm pass that converts the conditional branch/load into a select or hoists the load before/after the branch, also using a select?

Have you tried enabling either the EarlyIfConverter pass (for selects/conditional-moves) or EarlyIfPredicator pass (for predicated instructions) in the LLVM backend? If you are writing a downstream target, you will need to implement a few TargetInstrInfo APIs to get either pass up and running. Some in-tree backends use the former pass (e.g., AArch64, PPC). I use the latter in a downstream backend with reasonable results.

There is also an IR-level speculator in SimplifyCFG that may catch some of your opportunities (see caveats):

/// Speculate a conditional basic block flattening the CFG.
///
/// Note that this is a very risky transform currently. Speculating
/// instructions like this is most often not desirable. Instead, there is an MI
/// pass which can do it with full awareness of the resource constraints.
/// However, some cases are "obvious" and we should do directly. An example of
/// this is speculating a single, reasonably cheap instruction.
///
/// There is only one distinct advantage to flattening the CFG at the IR level:
/// it makes very common but simplistic optimizations such as are common in
/// instcombine and the DAG combiner more powerful by removing CFG edges and
/// modeling their effects with easier to reason about SSA value graphs.
///
///
/// An illustration of this transform is turning this IR:
/// \code
///   BB:
///     %cmp = icmp ult %x, %y
///     br i1 %cmp, label %EndBB, label %ThenBB
///   ThenBB:
///     %sub = sub %x, %y
///     br label BB2
///   EndBB:
///     %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
///     ...
/// \endcode
///
/// Into this IR:
/// \code
///   BB:
///     %cmp = icmp ult %x, %y
///     %sub = sub %x, %y
///     %cond = select i1 %cmp, 0, %sub
///     ...
/// \endcode
///
/// \returns true if the conditional block is removed.
bool SimplifyCFGOpt::SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
                                            const TargetTransformInfo
________________________________
From: llvm-dev <llvm-dev-bounces at lists.llvm.org> on behalf of Diogo Sampaio via llvm-dev <llvm-dev at lists.llvm.org>
Sent: Wednesday, August 18, 2021 3:45 PM
To: cfe-dev at lists.llvm.org <cfe-dev at lists.llvm.org>; llvm-dev <llvm-dev at lists.llvm.org>
Subject: [llvm-dev] Canonicalizing conditional loads

External email: Use caution opening links or attachments


Hi,
our target machine has conditional (predicated) load instructions. However it is hard to generate at dag-select
as clang emits conditional branches to avoid performing the load in most cases.
e.g.:https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgodbolt.org%2Fz%2FohKTnqeqz&data=04%7C01%7Cjeckhardt%40nvidia.com%7C6b309028a3f646dc312c08d962892a54%7C43083d15727340c1b7db39efd9ccc17a%7C0%7C0%7C637649163524599564%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=f9eSwLPRl7pZ1%2FzXAL5i%2B8dzPkwifT9Fk6FvIS7X6WA%3D&reserved=0
All these functions perform the same thing, but only the first one does not add a conditional branch instruction.

Is there a way to tell clang to always generate the code as the first function, or perhaps a llvm pass that converts
the conditional branch/load into a select or hoists the load before/after the branch, also using a select?

If not, would anyone else be interested in such thing? Any ideas for which would be the best solution?


Regards,
Diogo Sampaio




_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.llvm.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Fllvm-dev&data=04%7C01%7Cjeckhardt%40nvidia.com%7C6b309028a3f646dc312c08d962892a54%7C43083d15727340c1b7db39efd9ccc17a%7C0%7C0%7C637649163524599564%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=Qdl1GdBO9QXdAGG6cQu6kvKRD2WELiVHa9eMIQ4Ubeo%3D&reserved=0
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20210818/6fddc0e4/attachment.html>