[AMDGPU] Supported ds_read_b128 generation; Widened vector length for local address-space.

Summary: Starting from GCN 2nd generation, ISA supports ds_read_b128 on top of ds_read_b64.
         This patch supports ds_read_b128 instruction pattern and generation of this instruction.
         In the vectorizer, this patch also widen the vector length so that vectorizer generates
         128 bit loads for local address-space which gets translated to ds_read_b128.
         Since the performance benefit is not clear; compiler generates ds_read_b128 under -amdgpu-ds128.

Author: FarhanaAleen

Reviewed By: rampitec, arsenm

Subscribers: llvm-commits, AMDGPU

Differential Revision: https://reviews.llvm.org/D44210

llvm-svn: 327153
diff --git a/llvm/test/CodeGen/AMDGPU/load-local-f32.ll b/llvm/test/CodeGen/AMDGPU/load-local-f32.ll
index f035d22..3d87208 100644
--- a/llvm/test/CodeGen/AMDGPU/load-local-f32.ll
+++ b/llvm/test/CodeGen/AMDGPU/load-local-f32.ll
@@ -2,6 +2,11 @@
 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,SICIVI,FUNC %s
 ; RUN: llc -march=r600 -mcpu=redwood < %s | FileCheck -check-prefixes=EG,FUNC %s
 
+; Testing for ds_read_128
+; RUN: llc -march=amdgcn -mcpu=tahiti -amdgpu-ds128 < %s | FileCheck -check-prefixes=SI,FUNC %s
+; RUN: llc -march=amdgcn -mcpu=tonga -amdgpu-ds128 < %s | FileCheck -check-prefixes=CIVI,FUNC %s
+; RUN: llc -march=amdgcn -mcpu=gfx900 -amdgpu-ds128 < %s | FileCheck -check-prefixes=CIVI,FUNC %s
+
 ; FUNC-LABEL: {{^}}load_f32_local:
 ; SICIVI: s_mov_b32 m0
 ; GFX9-NOT: m0
@@ -122,4 +127,18 @@
   ret void
 }
 
+; Tests if ds_read_b128 gets generated for the 16 byte aligned load.
+; FUNC-LABEL: {{^}}local_v4f32_to_128:
+; SI-NOT: ds_read_b128
+; CIVI: ds_read_b128
+; EG: LDS_READ_RET
+; EG: LDS_READ_RET
+; EG: LDS_READ_RET
+; EG: LDS_READ_RET
+define amdgpu_kernel void @local_v4f32_to_128(<4 x float> addrspace(3)* %out, <4 x float> addrspace(3)* %in) {
+  %ld = load <4 x float>, <4 x float> addrspace(3)* %in, align 16
+  store <4 x float> %ld, <4 x float> addrspace(3)* %out
+  ret void
+}
+
 attributes #0 = { nounwind }