[llvm] [RISCV][GISel] Don't custom legalize load/store of vector of pointers if ELEN < XLEN. (PR #101565)

[Craig Topper via llvm-commits]

================
@@ -302,38 +317,28 @@ RISCVLegalizerInfo::RISCVLegalizerInfo(const RISCVSubtarget &ST)
                                                {nxv8s32, p0, nxv8s32, 32},
                                                {nxv16s32, p0, nxv16s32, 32}});
 
-  auto &ExtLoadActions =
-      getActionDefinitionsBuilder({G_SEXTLOAD, G_ZEXTLOAD})
-          .legalForTypesWithMemDesc({{s32, p0, s8, 8}, {s32, p0, s16, 16}});
-  if (XLen == 64) {
-    LoadStoreActions.legalForTypesWithMemDesc({{s64, p0, s8, 8},
-                                               {s64, p0, s16, 16},
-                                               {s64, p0, s32, 32},
-                                               {s64, p0, s64, 64}});
-    ExtLoadActions.legalForTypesWithMemDesc(
-        {{s64, p0, s8, 8}, {s64, p0, s16, 16}, {s64, p0, s32, 32}});
-  } else if (ST.hasStdExtD()) {
-    LoadStoreActions.legalForTypesWithMemDesc({{s64, p0, s64, 64}});
-  }
-  if (ST.hasVInstructions() && ST.getELen() == 64)
-    LoadStoreActions.legalForTypesWithMemDesc({{nxv1s8, p0, nxv1s8, 8},
-                                               {nxv1s16, p0, nxv1s16, 16},
-                                               {nxv1s32, p0, nxv1s32, 32}});
+    if (ST.getELen() == 64)
+      LoadStoreActions.legalForTypesWithMemDesc({{nxv1s8, p0, nxv1s8, 8},
+                                                 {nxv1s16, p0, nxv1s16, 16},
+                                                 {nxv1s32, p0, nxv1s32, 32}});
+
+    if (ST.hasVInstructionsI64())
+      LoadStoreActions.legalForTypesWithMemDesc({{nxv1s64, p0, nxv1s64, 64},
+                                                 {nxv2s64, p0, nxv2s64, 64},
+                                                 {nxv4s64, p0, nxv4s64, 64},
+                                                 {nxv8s64, p0, nxv8s64, 64}});
 
-  if (ST.hasVInstructionsI64())
-    LoadStoreActions.legalForTypesWithMemDesc({{nxv1s64, p0, nxv1s64, 64},
+    // we will take the custom lowering logic if we have scalable vector types
+    // with non-standard alignments
+    LoadStoreActions.customIf(typeIsLegalIntOrFPVec(0, IntOrFPVecTys, ST));
 
-                                               {nxv2s64, p0, nxv2s64, 64},
-                                               {nxv4s64, p0, nxv4s64, 64},
-                                               {nxv8s64, p0, nxv8s64, 64}});
+    // Pointers require that XLen sized elements are legal.
+    if (XLen <= ST.getELen())
+      LoadStoreActions.customIf(typeIsLegalPtrVec(0, PtrVecTys, ST));
----------------
topperc wrote:

Maybe. It would make the custom handler simpler, we wouldn't need to call allowsMemoryAccessForAlignment since we wouldn't get to the custom handler if we had already check the alignment with legalIf.

On the other hand calling allowsMemoryAccessForAlignment allows us to handle `FeatureUnalignedVectorMem` correctly. So maybe that's a vote for removing the legalIf and just letting the custom handle always handle it?

https://github.com/llvm/llvm-project/pull/101565