[PATCH] D44210: [AMDGPU] Supported ds_read_b128 generation; Widened vector length for local address-space

[Stanislav Mekhanoshin via Phabricator via llvm-commits]

rampitec added a comment.

Given performance benefit is somewhat unclear can you put it under an option?



================
Comment at: lib/Target/AMDGPU/SIISelLowering.cpp:5428
+    if (Subtarget->getGeneration() >= SISubtarget::SEA_ISLANDS &&
+        isAligned16(Alignment) && MemVT.getStoreSize() == 16)
+        return SDValue();
----------------
arsenm wrote:
> You don't need the isAligned16 helper. You just need to check that the alignment is >= 16, not % 16
Second to that.


https://reviews.llvm.org/D44210