[llvm-branch-commits] [llvm] [SLP] Cancel the phantom load saving on fadd reductions that lose an fma (PR #228903)

[Dmitry Sidorov via llvm-branch-commits]

================
@@ -17591,6 +17610,19 @@ BoUpSLP::getEntryCost(const TreeEntry *E, ArrayRef<Value *> VectorizedVals,
               Instruction::Load, VecTy, LI0->getAlign(),
               LI0->getPointerAddressSpace(), CostKind,
               TTI::getOperandInfo(LI0->getPointerOperand()));
+          // The vector load of a bundle saves nothing over scalar loads the
+          // target coalesces as well. A reduction that loses its fmas pays
+          // that saving back.
+          if (E->ReuseShuffleIndices.empty() && E->ReorderIndices.empty() &&
+              It == MinBWs.end() &&
+              reductionLosesFMAs(RdxKind, RdxFMF, VectorizedVals)) {
+            Align BestAlign = LI0->getAlign();
+            for (Value *V : VL)
+              BestAlign = std::max(BestAlign, cast<LoadInst>(V)->getAlign());
+            VecLdCost += TTI->getLoadCoalescingSaving(
+                LI0->getType(), VL.size(), BestAlign,
+                LI0->getPointerAddressSpace(), CostKind);
+          }
----------------
MrSidims wrote:

I'm not saying, that there is nothing to fix on AMDGPU TTI side - https://github.com/llvm/llvm-project/blob/3a829ed35f2231fe456fbda8df51920b1016c0ec/llvm/test/Analysis/CostModel/AMDGPU/load.ll is jsut wrong, yet SLP would still get something like this patch in the end.

https://github.com/llvm/llvm-project/pull/228903