[llvm] [InstCombine] Fix missed opportunity to fold 'or' into 'mul' operand. (PR #74225)

[Craig Topper via llvm-commits]

https://github.com/topperc created https://github.com/llvm/llvm-project/pull/74225

We were able to fold
or (mul X, Y), X --> mul X, (add Y, 1) (when the multiply has no common bits with X)
    
This patch makes the transform work if the mul operands are commuted.

>From 8fbaeb51258b7a06d205cc2853d11bca68676a69 Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper at sifive.com>
Date: Sat, 2 Dec 2023 20:22:50 -0800
Subject: [PATCH 1/2] [InstCombine] Add test for missed opportunity to fold
 'or' into 'mul' operand. NFC

We are able to fold

or (mul X, Y), X --> mul X, (add Y, 1) (when the multiply has no common bits with X)

but we miss it if the mul operands are commuted.
---
 llvm/test/Transforms/InstCombine/or.ll | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/llvm/test/Transforms/InstCombine/or.ll b/llvm/test/Transforms/InstCombine/or.ll
index 8bf8c6fcd928b..95427d2d24d3b 100644
--- a/llvm/test/Transforms/InstCombine/or.ll
+++ b/llvm/test/Transforms/InstCombine/or.ll
@@ -1511,6 +1511,21 @@ define <2 x i12> @mul_no_common_bits_commute(<2 x i12> %p) {
   ret <2 x i12> %r
 }
 
+define i32 @mul_no_common_bits_commute2(i32 %p1, i32 %p2) {
+; CHECK-LABEL: @mul_no_common_bits_commute2(
+; CHECK-NEXT:    [[X:%.*]] = and i32 [[P1:%.*]], 7
+; CHECK-NEXT:    [[Y:%.*]] = shl i32 [[P2:%.*]], 3
+; CHECK-NEXT:    [[M:%.*]] = mul i32 [[Y]], [[X]]
+; CHECK-NEXT:    [[R:%.*]] = or disjoint i32 [[M]], [[X]]
+; CHECK-NEXT:    ret i32 [[R]]
+;
+  %x = and i32 %p1, 7
+  %y = shl i32 %p2, 3
+  %m = mul i32 %y, %x
+  %r = or i32 %m, %x
+  ret i32 %r
+}
+
 ; negative test - extra use requires extra instructions
 
 define i32 @mul_no_common_bits_uses(i32 %p1, i32 %p2) {

>From 181008731488badc2281248a7283524efb3e0580 Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper at sifive.com>
Date: Sat, 2 Dec 2023 20:25:39 -0800
Subject: [PATCH 2/2] [InstCombine] Fix missed opportunity to fold 'or' into
 'mul' operand.

We were able to fold
or (mul X, Y), X --> mul X, (add Y, 1) (when the multiply has no common bits with X)

This patch makes the transform work if the mul operands are commuted.
---
 llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp | 2 +-
 llvm/test/Transforms/InstCombine/or.ll                  | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index 7379bdf93169e..50d20eb2f97b2 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -3394,7 +3394,7 @@ Instruction *InstCombinerImpl::visitOr(BinaryOperator &I) {
   // If the operands have no common bits set:
   // or (mul X, Y), X --> add (mul X, Y), X --> mul X, (Y + 1)
   if (match(&I,
-            m_c_Or(m_OneUse(m_Mul(m_Value(X), m_Value(Y))), m_Deferred(X))) &&
+            m_c_Or(m_Value(X), m_OneUse(m_c_Mul(m_Deferred(X), m_Value(Y))))) &&
       haveNoCommonBitsSet(Op0, Op1, DL)) {
     Value *IncrementY = Builder.CreateAdd(Y, ConstantInt::get(Ty, 1));
     return BinaryOperator::CreateMul(X, IncrementY);
diff --git a/llvm/test/Transforms/InstCombine/or.ll b/llvm/test/Transforms/InstCombine/or.ll
index 95427d2d24d3b..81ef38d43e8c3 100644
--- a/llvm/test/Transforms/InstCombine/or.ll
+++ b/llvm/test/Transforms/InstCombine/or.ll
@@ -1515,8 +1515,8 @@ define i32 @mul_no_common_bits_commute2(i32 %p1, i32 %p2) {
 ; CHECK-LABEL: @mul_no_common_bits_commute2(
 ; CHECK-NEXT:    [[X:%.*]] = and i32 [[P1:%.*]], 7
 ; CHECK-NEXT:    [[Y:%.*]] = shl i32 [[P2:%.*]], 3
-; CHECK-NEXT:    [[M:%.*]] = mul i32 [[Y]], [[X]]
-; CHECK-NEXT:    [[R:%.*]] = or disjoint i32 [[M]], [[X]]
+; CHECK-NEXT:    [[TMP1:%.*]] = or disjoint i32 [[Y]], 1
+; CHECK-NEXT:    [[R:%.*]] = mul i32 [[X]], [[TMP1]]
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %x = and i32 %p1, 7