Avoid broadcast/extract when implementing memset

[martien-de-jong]

The generic G_MEMSET legalizer helper would tweak the alignment of stack objects to make them amenable for vector implementations. However, the vector store that it creates doesn't have that alignment info available, and our legalization scalarizes it. With that scalarization it is not better than the original code that uses 32 bit stores.

I have disabled the cleverness which gives good results, also in stack size on the reduced example that I have added as a regression test.

DRAFT DRAFT DRAFT
This would be a draft PR, but I want to see whether it comes through standard CI testing

[andcarminati]

I think we should extend the target hook to have something like:

 LLT AIEBaseTargetLowering::getOptimalMemOpLLT(
     const MemOp &Op, const AttributeList &FuncAttributes) const {
 
+  bool AllowVectors =  && Op.isFixedDstAlign();
   if (Subtarget.isAIE2P()) {
-    if (AllowVecRegMemOps && Op.size() >= 64 && Op.isAligned(Align(64)))
+    if (AllowVectors && Op.size() >= 64 && Op.isAligned(Align(64)))
       return LLT::fixed_vector(16, 32);
   }
+  
   if (Subtarget.isAIE2() || Subtarget.isAIE2P()) {
-    if (AllowVecRegMemOps && Op.size() >= 32 && Op.isAligned(Align(32)))
+    if (AllowVectors && Op.size() >= 32 && Op.isAligned(Align(32)))
       return LLT::fixed_vector(8, 32);
-    if (AllowVecRegMemOps && Op.size() >= 16 && Op.isAligned(Align(16)))
+    if (AllowVectors && Op.size() >= 16 && Op.isAligned(Align(16)))
       return LLT::fixed_vector(4, 32);
     if (Op.size() >= 4 && Op.isAligned(Align(4)))
       return LLT::scalar(32);

It is risky to remove all that code.

[andcarminati]

I think we should extend the target hook to have something like:

 LLT AIEBaseTargetLowering::getOptimalMemOpLLT(
     const MemOp &Op, const AttributeList &FuncAttributes) const {
 
+  bool AllowVectors =  && Op.isFixedDstAlign();
   if (Subtarget.isAIE2P()) {
-    if (AllowVecRegMemOps && Op.size() >= 64 && Op.isAligned(Align(64)))
+    if (AllowVectors && Op.size() >= 64 && Op.isAligned(Align(64)))
       return LLT::fixed_vector(16, 32);
   }
+  
   if (Subtarget.isAIE2() || Subtarget.isAIE2P()) {
-    if (AllowVecRegMemOps && Op.size() >= 32 && Op.isAligned(Align(32)))
+    if (AllowVectors && Op.size() >= 32 && Op.isAligned(Align(32)))
       return LLT::fixed_vector(8, 32);
-    if (AllowVecRegMemOps && Op.size() >= 16 && Op.isAligned(Align(16)))
+    if (AllowVectors && Op.size() >= 16 && Op.isAligned(Align(16)))
       return LLT::fixed_vector(4, 32);
     if (Op.size() >= 4 && Op.isAligned(Align(4)))
       return LLT::scalar(32);

It is risky to remove all that code.

Discussed offline: we have an alignment here, we are just spotting the wrong one in the legalizer.

[martien-de-jong]

Yes. blunt fix. We cut away an optimization and don't meet it in PR testing.
However, this is not the way. Instead I propose we change the lowering code to get the appropriate alignment in the MMO's of the store instructions that are generated. We need to have a less lazy constructor of that derived MMO. I think that should be perfectly acceptable to upstream.