[llvm] [IVDescriptors] Don't require nsz/nnan for (min|max)num. (PR #137003)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: Florian Hahn (fhahn)

<details>
<summary>Changes</summary>

We do not need to require nsz or nnan for FP reductions with minnum and maxnum. NaNs and nsz are handled consistently with the vector versions of minnum and maxnum. vector.reduce.(fmin|fmax) also matches the minnum/maxnum semantics for comparisons.

IIUC this was also the conclusion when support for minimum/maximum was added (https://reviews.llvm.org/D151482?id=531021#inline-1481555).

Alive2 agrees that maxnum/minnum can be re-ordered (scalar version: https://alive2.llvm.org/ce/z/GVmgBX) and also verifies the vectorized code end-to-end, with a tweak to replace llvm.reduce.fmax with a scalar maxnum, as Alive2 doesn't support llvm.reduce.fmax: https://alive2.llvm.org/ce/z/EwJKeJ . Note that verification requires a higher timeout than available in the online version.

---
Full diff: https://github.com/llvm/llvm-project/pull/137003.diff


2 Files Affected:

- (modified) llvm/lib/Analysis/IVDescriptors.cpp (+6-3) 
- (modified) llvm/test/Transforms/LoopVectorize/minmax_reduction.ll (+4-2) 


``````````diff
diff --git a/llvm/lib/Analysis/IVDescriptors.cpp b/llvm/lib/Analysis/IVDescriptors.cpp
index 94c347b01bbfb..4a6efd0ca821a 100644
--- a/llvm/lib/Analysis/IVDescriptors.cpp
+++ b/llvm/lib/Analysis/IVDescriptors.cpp
@@ -892,10 +892,13 @@ RecurrenceDescriptor::InstDesc RecurrenceDescriptor::isRecurrenceInstr(
        return true;
      if (isa<FPMathOperator>(I) && I->hasNoNaNs() && I->hasNoSignedZeros())
        return true;
-     // minimum and maximum intrinsics do not require nsz and nnan flags since
-     // NaN and signed zeroes are propagated in the intrinsic implementation.
+     // minimum/minnum and maximum/maxnum intrinsics do not require nsz and nnan
+     // flags since NaN and signed zeroes are propagated in the intrinsic
+     // implementation.
      return match(I, m_Intrinsic<Intrinsic::minimum>(m_Value(), m_Value())) ||
-            match(I, m_Intrinsic<Intrinsic::maximum>(m_Value(), m_Value()));
+            match(I, m_Intrinsic<Intrinsic::maximum>(m_Value(), m_Value())) ||
+            match(I, m_Intrinsic<Intrinsic::minnum>(m_Value(), m_Value())) ||
+            match(I, m_Intrinsic<Intrinsic::maxnum>(m_Value(), m_Value()));
     };
     if (isIntMinMaxRecurrenceKind(Kind) ||
         (HasRequiredFMF() && isFPMinMaxRecurrenceKind(Kind)))
diff --git a/llvm/test/Transforms/LoopVectorize/minmax_reduction.ll b/llvm/test/Transforms/LoopVectorize/minmax_reduction.ll
index 85a90f2e04c5e..97b65e3435b5a 100644
--- a/llvm/test/Transforms/LoopVectorize/minmax_reduction.ll
+++ b/llvm/test/Transforms/LoopVectorize/minmax_reduction.ll
@@ -1002,7 +1002,8 @@ for.body:                                         ; preds = %entry, %for.body
 }
 
 ; CHECK-LABEL: @fmin_intrinsic_nofast(
-; CHECK-NOT: <2 x float> @llvm.minnum.v2f32
+; CHECK: call <2 x float> @llvm.minnum.v2f32
+; CHECK: call float @llvm.vector.reduce.fmin.v2f32
 define float @fmin_intrinsic_nofast(ptr nocapture readonly %x) {
 entry:
   br label %for.body
@@ -1022,7 +1023,8 @@ for.body:                                         ; preds = %entry, %for.body
 }
 
 ; CHECK-LABEL: @fmax_intrinsic_nofast(
-; CHECK-NOT: <2 x float> @llvm.maxnum.v2f32
+; CHECK: call <2 x float> @llvm.maxnum.v2f32
+; CHECK: call float @llvm.vector.reduce.fmax.v2f32
 define float @fmax_intrinsic_nofast(ptr nocapture readonly %x) {
 entry:
   br label %for.body

``````````

</details>


https://github.com/llvm/llvm-project/pull/137003