[llvm] AMDGPU: Improve getShuffleCost accuracy for 8- and 16-bit shuffles (PR #168818)

[Matt Arsenault via llvm-commits]

Nicolai =?utf-8?q?Hähnle?= <nicolai.haehnle at amd.com>,
Nicolai =?utf-8?q?Hähnle?= <nicolai.haehnle at amd.com>,
Nicolai =?utf-8?q?Hähnle?= <nicolai.haehnle at amd.com>
Message-ID:
In-Reply-To: <llvm.org/llvm/llvm-project/pull/168818 at github.com>


================
@@ -1241,46 +1241,108 @@ InstructionCost GCNTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
       (ScalarSize == 16 || ScalarSize == 8)) {
     // Larger vector widths may require additional instructions, but are
     // typically cheaper than scalarized versions.
-    unsigned NumVectorElts = cast<FixedVectorType>(SrcTy)->getNumElements();
-    unsigned RequestedElts =
-        count_if(Mask, [](int MaskElt) { return MaskElt != -1; });
-    unsigned EltsPerReg = 32 / ScalarSize;
-    if (RequestedElts == 0)
+    //
+    // We assume that shuffling at a register granularity can be done for free.
+    // This is not true for vectors fed into memory instructions, but it is
+    // effectively true for all other shuffling. The emphasis of the logic here
+    // is to assist generic transform in cleaning up / canonicalizing those
+    // shuffles.
+    unsigned NumDstElts = cast<FixedVectorType>(DstTy)->getNumElements();
+    unsigned NumSrcElts = cast<FixedVectorType>(SrcTy)->getNumElements();
----------------
arsenm wrote:

Can you keep this in terms of ElementCount to avoid crashing on scalable vectors 

https://github.com/llvm/llvm-project/pull/168818