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

[Farhana Aleen via Phabricator via llvm-commits]

FarhanaAleen added inline comments.


================
Comment at: lib/Target/AMDGPU/SIISelLowering.cpp:5434
+    // Use ds_read_b128 if possible.
+    if (Subtarget->useDS128(EnableDS128) && Load->getAlignment() >= 16 &&
+        MemVT.getStoreSize() == 16)
----------------
rampitec wrote:
> You only have pattern for v4i32, but enable operation for all 128 bit. Will it work with v8i16 for example?
Yes, it works for i16/i8. 

During dag combine, AMDGPU loadCombiner combines vector types of 8/16/64 to vector types of 32 bit type.


https://reviews.llvm.org/D44210