[llvm-branch-commits] [llvm] [LV] Add ExtNegatedMulAccReduction expression type (PR #160154)

[Sander de Smalen via llvm-branch-commits]

================
@@ -3541,17 +3547,24 @@ tryToMatchAndCreateMulAccumulateReduction(VPReductionRecipe *Red,
         IsMulAccValidAndClampRange(RecipeA->getOpcode() ==
                                        Instruction::CastOps::ZExt,
                                    Mul, RecipeA, RecipeB, nullptr)) {
+      if (Sub)
+        return new VPExpressionRecipe(
+            RecipeA, RecipeB, Mul,
+            cast<VPWidenRecipe>(Sub->getDefiningRecipe()), Red);
       return new VPExpressionRecipe(RecipeA, RecipeB, Mul, Red);
     }
     // Match reduce.add(mul).
-    if (IsMulAccValidAndClampRange(true, Mul, nullptr, nullptr, nullptr))
+    // TODO: Add an expression type for this variant with a negated mul
+    if (!Sub &&
+        IsMulAccValidAndClampRange(true, Mul, nullptr, nullptr, nullptr))
       return new VPExpressionRecipe(Mul, Red);
   }
   // Match reduce.add(ext(mul(ext(A), ext(B)))).
   // All extend recipes must have same opcode or A == B
   // which can be transform to reduce.add(zext(mul(sext(A), sext(B)))).
-  if (match(VecOp, m_ZExtOrSExt(m_Mul(m_ZExtOrSExt(m_VPValue()),
-                                      m_ZExtOrSExt(m_VPValue()))))) {
+  // TODO: Add an expression type for this variant with a negated mul
+  if (!Sub && match(VecOp, m_ZExtOrSExt(m_Mul(m_ZExtOrSExt(m_VPValue()),
----------------
sdesmalen-arm wrote:

nit: rather than doing `!Sub &&` here, I would just have an explicit `if (Sub) return nullptr;` in case other cases get added in the future that may also not support the `sub`.

https://github.com/llvm/llvm-project/pull/160154