[llvm] bc886e9 - [InstCombine] fold mul with decremented "shl -1" factor

[Sanjay Patel via llvm-commits]

Author: Sanjay Patel
Date: 2022-10-31T09:06:55-04:00
New Revision: bc886e9b587b9d009f49b12eaaa9ebc1c71905a1

URL: https://github.com/llvm/llvm-project/commit/bc886e9b587b9d009f49b12eaaa9ebc1c71905a1
DIFF: https://github.com/llvm/llvm-project/commit/bc886e9b587b9d009f49b12eaaa9ebc1c71905a1.diff

LOG: [InstCombine] fold mul with decremented "shl -1" factor

This is a sibling to:
6064e92b0a84
...but we canonicalize the shl+add to shl+xor,
so the pattern is different than I expected:
https://alive2.llvm.org/ce/z/8CX16e

I have not found any patterns that are safe
to propagate no-wrap, so that is not included
here.

Added: 
    

Modified: 
    llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
    llvm/test/Transforms/InstCombine/mul.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
index e4fccda750e6c..45174b752dc64 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
@@ -140,7 +140,7 @@ static Value *foldMulSelectToNegate(BinaryOperator &I,
   return nullptr;
 }
 
-/// Reduce integer multiplication patterns that contain a (1 << Z) factor.
+/// Reduce integer multiplication patterns that contain a (+/-1 << Z) factor.
 /// Callers are expected to call this twice to handle commuted patterns.
 static Value *foldMulShl1(BinaryOperator &Mul, bool CommuteOperands,
                           InstCombiner::BuilderTy &Builder) {
@@ -171,6 +171,17 @@ static Value *foldMulShl1(BinaryOperator &Mul, bool CommuteOperands,
     return Builder.CreateAdd(Shl, FrX, Mul.getName(), HasNUW, PropagateNSW);
   }
 
+  // Similar to above, but a decrement of the shifted value is disguised as
+  // 'not' and becomes a sub:
+  // X * (~(-1 << Z)) --> X * ((1 << Z) - 1) --> (X << Z) - X
+  // This increases uses of X, so it may require a freeze, but that is still
+  // expected to be an improvement because it removes the multiply.
+  if (match(Y, m_OneUse(m_Not(m_OneUse(m_Shl(m_AllOnes(), m_Value(Z))))))) {
+    Value *FrX = Builder.CreateFreeze(X, X->getName() + ".fr");
+    Value *Shl = Builder.CreateShl(FrX, Z, "mulshl");
+    return Builder.CreateAdd(Shl, FrX, Mul.getName());
+  }
+
   return nullptr;
 }
 

diff  --git a/llvm/test/Transforms/InstCombine/mul.ll b/llvm/test/Transforms/InstCombine/mul.ll
index 29508e5e5d2ce..7c3a73b7668b5 100644
--- a/llvm/test/Transforms/InstCombine/mul.ll
+++ b/llvm/test/Transforms/InstCombine/mul.ll
@@ -229,10 +229,10 @@ define i32 @shl1_increment_use(i32 %x, i32 %y) {
 
 define i8 @shl1_decrement(i8 %x, i8 %y) {
 ; CHECK-LABEL: @shl1_decrement(
-; CHECK-NEXT:    [[POW2X:%.*]] = shl i8 -1, [[X:%.*]]
-; CHECK-NEXT:    [[X1:%.*]] = xor i8 [[POW2X]], -1
-; CHECK-NEXT:    [[M:%.*]] = mul i8 [[X1]], [[Y:%.*]]
-; CHECK-NEXT:    ret i8 [[M]]
+; CHECK-NEXT:    [[Y_FR:%.*]] = freeze i8 [[Y:%.*]]
+; CHECK-NEXT:    [[MULSHL:%.*]] = shl i8 [[Y_FR]], [[X:%.*]]
+; CHECK-NEXT:    [[M1:%.*]] = add i8 [[MULSHL]], [[Y_FR]]
+; CHECK-NEXT:    ret i8 [[M1]]
 ;
   %pow2x = shl i8 -1, %x
   %x1 = xor i8 %pow2x, -1
@@ -243,10 +243,9 @@ define i8 @shl1_decrement(i8 %x, i8 %y) {
 define i8 @shl1_decrement_commute(i8 %x, i8 noundef %p) {
 ; CHECK-LABEL: @shl1_decrement_commute(
 ; CHECK-NEXT:    [[Y:%.*]] = ashr i8 [[P:%.*]], 1
-; CHECK-NEXT:    [[NOTMASK:%.*]] = shl nsw i8 -1, [[X:%.*]]
-; CHECK-NEXT:    [[X1:%.*]] = xor i8 [[NOTMASK]], -1
-; CHECK-NEXT:    [[M:%.*]] = mul i8 [[Y]], [[X1]]
-; CHECK-NEXT:    ret i8 [[M]]
+; CHECK-NEXT:    [[MULSHL:%.*]] = shl i8 [[Y]], [[X:%.*]]
+; CHECK-NEXT:    [[M1:%.*]] = add i8 [[MULSHL]], [[Y]]
+; CHECK-NEXT:    ret i8 [[M1]]
 ;
   %y = ashr i8 %p, 1 ; thwart complexity-based canonicalization
   %pow2x = shl i8 1, %x
@@ -257,10 +256,10 @@ define i8 @shl1_decrement_commute(i8 %x, i8 noundef %p) {
 
 define i8 @shl1_nuw_decrement(i8 %x, i8 %y) {
 ; CHECK-LABEL: @shl1_nuw_decrement(
-; CHECK-NEXT:    [[POW2X:%.*]] = shl i8 -1, [[X:%.*]]
-; CHECK-NEXT:    [[X1:%.*]] = xor i8 [[POW2X]], -1
-; CHECK-NEXT:    [[M:%.*]] = mul nuw i8 [[X1]], [[Y:%.*]]
-; CHECK-NEXT:    ret i8 [[M]]
+; CHECK-NEXT:    [[Y_FR:%.*]] = freeze i8 [[Y:%.*]]
+; CHECK-NEXT:    [[MULSHL:%.*]] = shl i8 [[Y_FR]], [[X:%.*]]
+; CHECK-NEXT:    [[M1:%.*]] = add i8 [[MULSHL]], [[Y_FR]]
+; CHECK-NEXT:    ret i8 [[M1]]
 ;
   %pow2x = shl i8 -1, %x
   %x1 = xor i8 %pow2x, -1
@@ -270,10 +269,10 @@ define i8 @shl1_nuw_decrement(i8 %x, i8 %y) {
 
 define i8 @shl1_nsw_decrement(i8 %x, i8 %y) {
 ; CHECK-LABEL: @shl1_nsw_decrement(
-; CHECK-NEXT:    [[POW2X:%.*]] = shl nsw i8 -1, [[X:%.*]]
-; CHECK-NEXT:    [[X1:%.*]] = xor i8 [[POW2X]], -1
-; CHECK-NEXT:    [[M:%.*]] = mul nsw i8 [[X1]], [[Y:%.*]]
-; CHECK-NEXT:    ret i8 [[M]]
+; CHECK-NEXT:    [[Y_FR:%.*]] = freeze i8 [[Y:%.*]]
+; CHECK-NEXT:    [[MULSHL:%.*]] = shl i8 [[Y_FR]], [[X:%.*]]
+; CHECK-NEXT:    [[M1:%.*]] = add i8 [[MULSHL]], [[Y_FR]]
+; CHECK-NEXT:    ret i8 [[M1]]
 ;
   %pow2x = shl nsw i8 -1, %x
   %x1 = xor i8 %pow2x, -1