1 files changed, 9 insertions, 18 deletions
diff --git a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
index 941f781f2bf..3273d3233b6 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
@@ -73,6 +73,12 @@ static LegalityPredicate vectorWiderThan(unsigned TypeIdx, unsigned Size) {
   };
 }
 
+static LegalityPredicate numElementsNotEven(unsigned TypeIdx) {
+  return [=](const LegalityQuery &Query) {
+    const LLT QueryTy = Query.Types[TypeIdx];
+    return QueryTy.isVector() && QueryTy.getNumElements() % 2 != 0;
+  };
+}
 
 AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST,
                                          const GCNTargetMachine &TM) {
@@ -453,28 +459,13 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST,
           GlobalPtr, LocalPtr, FlatPtr, PrivatePtr,
           LLT::vector(2, LocalPtr), LLT::vector(2, PrivatePtr)}, {S1})
     .clampScalar(0, S32, S64)
-    .fewerElementsIf(
-      [=](const LegalityQuery &Query) {
-        if (Query.Types[1].isVector())
-          return true;
-
-        LLT Ty = Query.Types[0];
-
-        // FIXME: Hack until odd splits handled
-        return Ty.isVector() &&
-          (Ty.getScalarSizeInBits() > 32 || Ty.getNumElements() % 2 != 0);
-      },
-      scalarize(0))
-    // FIXME: Handle 16-bit vectors better
-    .fewerElementsIf(
-      [=](const LegalityQuery &Query) {
-        return Query.Types[0].isVector() &&
-               Query.Types[0].getElementType().getSizeInBits() < 32;},
-      scalarize(0))
+    .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
+    .fewerElementsIf(numElementsNotEven(0), scalarize(0))
     .scalarize(1)
     .clampMaxNumElements(0, S32, 2)
     .clampMaxNumElements(0, LocalPtr, 2)
     .clampMaxNumElements(0, PrivatePtr, 2)
+    .scalarize(0)
     .legalIf(all(isPointer(0), typeIs(1, S1)));
 
   // TODO: Only the low 4/5/6 bits of the shift amount are observed, so we can