[llvm] AMDGPU/GlobalISel: Lower G_EXTRACT in legalizer (PR #181036)

[Petar Avramovic via llvm-commits]

================
@@ -1941,39 +1941,31 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
   for (unsigned Op : {G_EXTRACT, G_INSERT}) {
     unsigned BigTyIdx = Op == G_EXTRACT ? 1 : 0;
     unsigned LitTyIdx = Op == G_EXTRACT ? 0 : 1;
-
-    // FIXME: Doesn't handle extract of illegal sizes.
     getActionDefinitionsBuilder(Op)
-      .lowerIf(all(typeIs(LitTyIdx, S16), sizeIs(BigTyIdx, 32)))
-      .lowerIf([=](const LegalityQuery &Query) {
-          // Sub-vector(or single element) insert and extract.
-          // TODO: verify immediate offset here since lower only works with
-          // whole elements.
-          const LLT BigTy = Query.Types[BigTyIdx];
-          return BigTy.isVector();
-        })
-      // FIXME: Multiples of 16 should not be legal.
-      .legalIf([=](const LegalityQuery &Query) {
+        .widenScalarIf(
+            [=](const LegalityQuery &Query) {
+              const LLT BigTy = Query.Types[BigTyIdx];
+              return (BigTy.getScalarSizeInBits() < 16);
+            },
+            LegalizeMutations::widenScalarOrEltToNextPow2(BigTyIdx, 16))
+        .widenScalarIf(
+            [=](const LegalityQuery &Query) {
+              const LLT LitTy = Query.Types[LitTyIdx];
+              return (LitTy.getScalarSizeInBits() < 16);
+            },
+            LegalizeMutations::widenScalarOrEltToNextPow2(LitTyIdx, 16))
+        .moreElementsIf(isSmallOddVector(BigTyIdx), oneMoreElement(BigTyIdx))
+        .widenScalarToNextPow2(BigTyIdx, 32)
+        .customIf([=](const LegalityQuery &Query) {
+          // Generic lower is not aware of subregs and can produce inefficient
+          // shift+trunc/mask sequences. We can instead use custom lowering
----------------
petar-avramovic wrote:

Need to clarify a bit, not sure what are subregs in this context

https://github.com/llvm/llvm-project/pull/181036