[llvm] [InstCombine] Remove OneUse clause from `((A+B)+B) → A+2*B` (PR #80705)

[Ricardo Jesus via llvm-commits]

https://github.com/rj-jesus created https://github.com/llvm/llvm-project/pull/80705

This patch removes the `m_OneUse` check that was being done for the following patterns:
  (A + RHS) + RHS --> A + (RHS << 1)
  LHS + (A + LHS) --> A + (LHS << 1)

This may increase the instruction count if (A + RHS) or (A + LHS) are used more than once but will usually help with LICM later on if A is an induction variable.

This is exemplified here: https://godbolt.org/z/q94vbf5e9 - the first instruction of the second iteration ([1,0]) can start much earlier on with the patch, allowing for significantly higher IPC.

I've benchmarked this with RAJAPerf on a Grace system and see an overall performance uplift of 2% across the whole benchmark which comes mainly from ~45% uplifts in three kernels. Worst I saw was a 1.2% slowdown on a single kernel which I look into later.

Please let me know if I should move this elsewhere!

>From ee897e69393452cde3ce5174ccd6029953eddd91 Mon Sep 17 00:00:00 2001
From: Ricardo Jesus <rjj at ed.ac.uk>
Date: Mon, 5 Feb 2024 13:53:06 +0000
Subject: [PATCH] =?UTF-8?q?[InstCombine]=20Remove=20OneUse=20clause=20from?=
 =?UTF-8?q?=20`((A+B)+B)=20=E2=86=92=20A+2*B`?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

This patch removes the `m_OneUse` check that was being done for the
following patterns:
  (A + RHS) + RHS --> A + (RHS << 1)
  LHS + (A + LHS) --> A + (LHS << 1)

This may increase the instruction count if (A + RHS) or (A + LHS) are
used more than once but usually will help with LICM later on if A is an
induction variable.
---
 llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
index 8a00b75a1f740..984f7a619d4ab 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
@@ -1521,11 +1521,14 @@ Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
     return BinaryOperator::CreateSub(A, B);
 
   // (A + RHS) + RHS --> A + (RHS << 1)
-  if (match(LHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(RHS)))))
+  // This may increase instruction count if (A + RHS) is used more than once
+  // but helps make (RHS << 1) suitable for potential LICM.
+  if (match(LHS, m_c_Add(m_Value(A), m_Specific(RHS))))
     return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));
 
   // LHS + (A + LHS) --> A + (LHS << 1)
-  if (match(RHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(LHS)))))
+  // Same as above.
+  if (match(RHS, m_c_Add(m_Value(A), m_Specific(LHS))))
     return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));
 
   {