[llvm] AMDGPU/GlobalISel: Switch to extended LLTs (PR #196522)

Mon Jun 22 06:41:48 PDT 2026

================
@@ -2054,7 +2071,12 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
           .legalForCartesianProduct(AllS64Vectors, {S64})
           .clampNumElements(0, V16S32, V32S32)
           .clampNumElements(0, V2S64, V16S64)
-          .fewerElementsIf(isWideVec16(0), changeTo(0, V2S16))
+          .fewerElementsIf(isWideVec16(0),
+                           [](const LegalityQuery &Query) {
+                             LLT EltTy = Query.Types[0].getElementType();
+                             return std::make_pair(0,
+                                                   LLT::fixed_vector(2, EltTy));
+                           })
----------------
mbrkusanin wrote:

`changeElementCountTo(0, ElementCount::getFixed(2)))`

https://github.com/llvm/llvm-project/pull/196522
