R600/SI: Initial local memory support

Enough for the radeonsi driver to use it for calculating derivatives.

Reviewed-by: Tom Stellard <thomas.stellard@amd.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@186012 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 66 insertions, 52 deletions

diff --git a/test/CodeGen/R600/local-memory-two-objects.ll b/test/CodeGen/R600/local-memory-two-objects.ll
new file mode 100644
index 0000000..6d3610e
--- /dev/null
+++ b/test/CodeGen/R600/local-memory-two-objects.ll
@@ -0,0 +1,51 @@
+; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+; TODO: Add RUN and CHECK lines for SI once this test works there
+
+@local_memory_two_objects.local_mem0 = internal addrspace(3) unnamed_addr global [4 x i32] zeroinitializer, align 4
+@local_memory_two_objects.local_mem1 = internal addrspace(3) unnamed_addr global [4 x i32] zeroinitializer, align 4
+
+; CHECK: @local_memory_two_objects
+
+; Check that the LDS size emitted correctly
+; CHECK: .long 166120
+; CHECK-NEXT: .long 8
+
+; Make sure the lds writes are using different addresses.
+; CHECK: LDS_WRITE {{[*]*}} {{PV|T}}[[ADDRW:[0-9]*\.[XYZW]]]
+; CHECK-NOT: LDS_WRITE {{[*]*}} T[[ADDRW]]
+
+; GROUP_BARRIER must be the last instruction in a clause
+; CHECK: GROUP_BARRIER
+; CHECK-NEXT: ALU clause
+
+; Make sure the lds reads are using different addresses.
+; CHECK: LDS_READ_RET {{[*]*}} OQAP, {{PV|T}}[[ADDRR:[0-9]*\.[XYZW]]]
+; CHECK-NOT: LDS_READ_RET {{[*]*}} OQAP, T[[ADDRR]]
+
+define void @local_memory_two_objects(i32 addrspace(1)* %out) {
+entry:
+  %x.i = call i32 @llvm.r600.read.tidig.x() #0
+  %arrayidx = getelementptr inbounds [4 x i32] addrspace(3)* @local_memory_two_objects.local_mem0, i32 0, i32 %x.i
+  store i32 %x.i, i32 addrspace(3)* %arrayidx, align 4
+  %mul = shl nsw i32 %x.i, 1
+  %arrayidx1 = getelementptr inbounds [4 x i32] addrspace(3)* @local_memory_two_objects.local_mem1, i32 0, i32 %x.i
+  store i32 %mul, i32 addrspace(3)* %arrayidx1, align 4
+  %sub = sub nsw i32 3, %x.i
+  call void @llvm.AMDGPU.barrier.local()
+  %arrayidx2 = getelementptr inbounds [4 x i32] addrspace(3)* @local_memory_two_objects.local_mem0, i32 0, i32 %sub
+  %0 = load i32 addrspace(3)* %arrayidx2, align 4
+  %arrayidx3 = getelementptr inbounds i32 addrspace(1)* %out, i32 %x.i
+  store i32 %0, i32 addrspace(1)* %arrayidx3, align 4
+  %arrayidx4 = getelementptr inbounds [4 x i32] addrspace(3)* @local_memory_two_objects.local_mem1, i32 0, i32 %sub
+  %1 = load i32 addrspace(3)* %arrayidx4, align 4
+  %add = add nsw i32 %x.i, 4
+  %arrayidx5 = getelementptr inbounds i32 addrspace(1)* %out, i32 %add
+  store i32 %1, i32 addrspace(1)* %arrayidx5, align 4
+  ret void
+}
+
+declare i32 @llvm.r600.read.tidig.x() #0
+declare void @llvm.AMDGPU.barrier.local()
+
+attributes #0 = { readnone }
diff --git a/test/CodeGen/R600/local-memory.ll b/test/CodeGen/R600/local-memory.ll
index 0ff3848..5458fb9 100644
--- a/test/CodeGen/R600/local-memory.ll
+++ b/test/CodeGen/R600/local-memory.ll
@@ -1,21 +1,27 @@
-; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
-
+; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck --check-prefix=EG-CHECK %s
+; RUN: llc < %s -march=r600 -mcpu=verde | FileCheck --check-prefix=SI-CHECK %s
 
 @local_memory.local_mem = internal addrspace(3) unnamed_addr global [16 x i32] zeroinitializer, align 4
 
-; CHECK: @local_memory
+; EG-CHECK: @local_memory
+; SI-CHECK: @local_memory
 
 ; Check that the LDS size emitted correctly
-; CHECK: .long 166120
-; CHECK-NEXT: .long 16
+; EG-CHECK: .long 166120
+; EG-CHECK-NEXT: .long 16
+; SI-CHECK: .long 47180
+; SI-CHECK-NEXT: .long 32768
 
-; CHECK: LDS_WRITE
+; EG-CHECK: LDS_WRITE
+; SI-CHECK: DS_WRITE_B32
 
 ; GROUP_BARRIER must be the last instruction in a clause
-; CHECK: GROUP_BARRIER
-; CHECK-NEXT: ALU clause
+; EG-CHECK: GROUP_BARRIER
+; EG-CHECK-NEXT: ALU clause
+; SI-CHECK: S_BARRIER
 
-; CHECK: LDS_READ_RET
+; EG-CHECK: LDS_READ_RET
+; SI-CHECK: DS_READ_B32
 
 define void @local_memory(i32 addrspace(1)* %out) {
 entry:
@@ -33,49 +39,6 @@ entry:
   ret void
 }
 
-@local_memory_two_objects.local_mem0 = internal addrspace(3) unnamed_addr global [4 x i32] zeroinitializer, align 4
-@local_memory_two_objects.local_mem1 = internal addrspace(3) unnamed_addr global [4 x i32] zeroinitializer, align 4
-
-; CHECK: @local_memory_two_objects
-
-; Check that the LDS size emitted correctly
-; CHECK: .long 166120
-; CHECK-NEXT: .long 8
-
-; Make sure the lds writes are using different addresses.
-; CHECK: LDS_WRITE {{[*]*}} {{PV|T}}[[ADDRW:[0-9]*\.[XYZW]]]
-; CHECK-NOT: LDS_WRITE {{[*]*}} T[[ADDRW]]
-
-; GROUP_BARRIER must be the last instruction in a clause
-; CHECK: GROUP_BARRIER
-; CHECK-NEXT: ALU clause
-
-; Make sure the lds reads are using different addresses.
-; CHECK: LDS_READ_RET {{[*]*}} OQAP, {{PV|T}}[[ADDRR:[0-9]*\.[XYZW]]]
-; CHECK-NOT: LDS_READ_RET {{[*]*}} OQAP, T[[ADDRR]]
-
-define void @local_memory_two_objects(i32 addrspace(1)* %out) {
-entry:
-  %x.i = call i32 @llvm.r600.read.tidig.x() #0
-  %arrayidx = getelementptr inbounds [4 x i32] addrspace(3)* @local_memory_two_objects.local_mem0, i32 0, i32 %x.i
-  store i32 %x.i, i32 addrspace(3)* %arrayidx, align 4
-  %mul = shl nsw i32 %x.i, 1
-  %arrayidx1 = getelementptr inbounds [4 x i32] addrspace(3)* @local_memory_two_objects.local_mem1, i32 0, i32 %x.i
-  store i32 %mul, i32 addrspace(3)* %arrayidx1, align 4
-  %sub = sub nsw i32 3, %x.i
-  call void @llvm.AMDGPU.barrier.local()
-  %arrayidx2 = getelementptr inbounds [4 x i32] addrspace(3)* @local_memory_two_objects.local_mem0, i32 0, i32 %sub
-  %0 = load i32 addrspace(3)* %arrayidx2, align 4
-  %arrayidx3 = getelementptr inbounds i32 addrspace(1)* %out, i32 %x.i
-  store i32 %0, i32 addrspace(1)* %arrayidx3, align 4
-  %arrayidx4 = getelementptr inbounds [4 x i32] addrspace(3)* @local_memory_two_objects.local_mem1, i32 0, i32 %sub
-  %1 = load i32 addrspace(3)* %arrayidx4, align 4
-  %add = add nsw i32 %x.i, 4
-  %arrayidx5 = getelementptr inbounds i32 addrspace(1)* %out, i32 %add
-  store i32 %1, i32 addrspace(1)* %arrayidx5, align 4
-  ret void
-}
-
 declare i32 @llvm.r600.read.tidig.x() #0
 declare void @llvm.AMDGPU.barrier.local()