[llvm] [SLP]Do extra analysis int minbitwidth if some checks return false. (PR #84363)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: Alexey Bataev (alexey-bataev)

<details>
<summary>Changes</summary>

The instruction itself can be considered good for minbitwidth casting,
even if one of the operand checks returns false.


---
Full diff: https://github.com/llvm/llvm-project/pull/84363.diff


2 Files Affected:

- (modified) llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp (+3-3) 
- (modified) llvm/test/Transforms/SLPVectorizer/AArch64/horizontal.ll (+11-9) 


``````````diff
diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index 1889bc09e85028..c6502c9f77f001 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -13942,7 +13942,7 @@ bool BoUpSLP::collectValuesToDemote(
         !collectValuesToDemote(I->getOperand(1), IsProfitableToDemoteRoot,
                                BitWidth, ToDemote, DemotedConsts, Visited,
                                Level2, IsProfitableToDemote))
-      return false;
+      return IsProfitableToDemote && IsPotentiallyTruncated(I, BitWidth);
     MaxDepthLevel = std::max(Level1, Level2);
     break;
   }
@@ -13958,7 +13958,7 @@ bool BoUpSLP::collectValuesToDemote(
         !collectValuesToDemote(SI->getFalseValue(), IsProfitableToDemoteRoot,
                                BitWidth, ToDemote, DemotedConsts, Visited,
                                Level2, IsProfitableToDemote))
-      return false;
+      return IsProfitableToDemote && IsPotentiallyTruncated(I, BitWidth);
     MaxDepthLevel = std::max(Level1, Level2);
     break;
   }
@@ -13971,7 +13971,7 @@ bool BoUpSLP::collectValuesToDemote(
       if (!collectValuesToDemote(IncValue, IsProfitableToDemoteRoot, BitWidth,
                                  ToDemote, DemotedConsts, Visited,
                                  MaxDepthLevel, IsProfitableToDemote))
-        return false;
+        return IsProfitableToDemote && IsPotentiallyTruncated(I, BitWidth);
     break;
   }
 
diff --git a/llvm/test/Transforms/SLPVectorizer/AArch64/horizontal.ll b/llvm/test/Transforms/SLPVectorizer/AArch64/horizontal.ll
index 1986b51ec94828..02d1f9f60d0ca1 100644
--- a/llvm/test/Transforms/SLPVectorizer/AArch64/horizontal.ll
+++ b/llvm/test/Transforms/SLPVectorizer/AArch64/horizontal.ll
@@ -228,7 +228,7 @@ for.end:                                          ; preds = %for.end.loopexit, %
 ; YAML-NEXT: Function:        test_unrolled_select
 ; YAML-NEXT: Args:
 ; YAML-NEXT:   - String:          'Vectorized horizontal reduction with cost '
-; YAML-NEXT:   - Cost:            '-36'
+; YAML-NEXT:   - Cost:            '-40'
 ; YAML-NEXT:   - String:          ' and with tree size '
 ; YAML-NEXT:   - TreeSize:        '10'
 
@@ -246,15 +246,17 @@ define i32 @test_unrolled_select(ptr noalias nocapture readonly %blk1, ptr noali
 ; CHECK-NEXT:    [[P2_045:%.*]] = phi ptr [ [[BLK2:%.*]], [[FOR_BODY_LR_PH]] ], [ [[ADD_PTR88:%.*]], [[IF_END_86]] ]
 ; CHECK-NEXT:    [[P1_044:%.*]] = phi ptr [ [[BLK1:%.*]], [[FOR_BODY_LR_PH]] ], [ [[ADD_PTR:%.*]], [[IF_END_86]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = load <8 x i8>, ptr [[P1_044]], align 1
-; CHECK-NEXT:    [[TMP1:%.*]] = zext <8 x i8> [[TMP0]] to <8 x i32>
+; CHECK-NEXT:    [[TMP1:%.*]] = zext <8 x i8> [[TMP0]] to <8 x i16>
 ; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr [[P2_045]], align 1
-; CHECK-NEXT:    [[TMP3:%.*]] = zext <8 x i8> [[TMP2]] to <8 x i32>
-; CHECK-NEXT:    [[TMP4:%.*]] = sub nsw <8 x i32> [[TMP1]], [[TMP3]]
-; CHECK-NEXT:    [[TMP5:%.*]] = icmp slt <8 x i32> [[TMP4]], zeroinitializer
-; CHECK-NEXT:    [[TMP6:%.*]] = sub nsw <8 x i32> zeroinitializer, [[TMP4]]
-; CHECK-NEXT:    [[TMP7:%.*]] = select <8 x i1> [[TMP5]], <8 x i32> [[TMP6]], <8 x i32> [[TMP4]]
-; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP7]])
-; CHECK-NEXT:    [[OP_RDX]] = add i32 [[TMP8]], [[S_047]]
+; CHECK-NEXT:    [[TMP3:%.*]] = zext <8 x i8> [[TMP2]] to <8 x i16>
+; CHECK-NEXT:    [[TMP4:%.*]] = sub <8 x i16> [[TMP1]], [[TMP3]]
+; CHECK-NEXT:    [[TMP5:%.*]] = trunc <8 x i16> [[TMP4]] to <8 x i1>
+; CHECK-NEXT:    [[TMP6:%.*]] = icmp slt <8 x i1> [[TMP5]], zeroinitializer
+; CHECK-NEXT:    [[TMP7:%.*]] = sub <8 x i16> zeroinitializer, [[TMP4]]
+; CHECK-NEXT:    [[TMP8:%.*]] = select <8 x i1> [[TMP6]], <8 x i16> [[TMP7]], <8 x i16> [[TMP4]]
+; CHECK-NEXT:    [[TMP9:%.*]] = zext <8 x i16> [[TMP8]] to <8 x i32>
+; CHECK-NEXT:    [[TMP10:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP9]])
+; CHECK-NEXT:    [[OP_RDX]] = add i32 [[TMP10]], [[S_047]]
 ; CHECK-NEXT:    [[CMP83:%.*]] = icmp slt i32 [[OP_RDX]], [[LIM:%.*]]
 ; CHECK-NEXT:    br i1 [[CMP83]], label [[IF_END_86]], label [[FOR_END_LOOPEXIT:%.*]]
 ; CHECK:       if.end.86:

``````````

</details>


https://github.com/llvm/llvm-project/pull/84363