From 3a5e9cb146a444148b1f40265acadb6a27ee7776 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Mon, 28 Jul 2014 17:49:26 +0000
Subject: [PATCH] R600/SI: Implement getOptimalMemOpType

The default guess uses i32. This needs an address space argument
to really do the right thing in all cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@214104 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/R600/SIISelLowering.cpp |  20 ++
 lib/Target/R600/SIISelLowering.h   |   6 +
 test/CodeGen/R600/llvm.memcpy.ll   | 358 +++++++++++++++++++++++++++++
 3 files changed, 384 insertions(+)
 create mode 100644 test/CodeGen/R600/llvm.memcpy.ll

diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
index a4d4195edde..c3405e10453 100644
--- a/lib/Target/R600/SIISelLowering.cpp
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -274,6 +274,26 @@ bool SITargetLowering::allowsMisalignedMemoryAccesses(EVT  VT,
   return VT.bitsGT(MVT::i32);
 }
 
+EVT SITargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
+                                          unsigned SrcAlign, bool IsMemset,
+                                          bool ZeroMemset,
+                                          bool MemcpyStrSrc,
+                                          MachineFunction &MF) const {
+  // FIXME: Should account for address space here.
+
+  // The default fallback uses the private pointer size as a guess for a type to
+  // use. Make sure we switch these to 64-bit accesses.
+
+  if (Size >= 16 && DstAlign >= 4) // XXX: Should only do for global
+    return MVT::v4i32;
+
+  if (Size >= 8 && DstAlign >= 4)
+    return MVT::v2i32;
+
+  // Use the default.
+  return MVT::Other;
+}
+
 TargetLoweringBase::LegalizeTypeAction
 SITargetLowering::getPreferredVectorAction(EVT VT) const {
   if (VT.getVectorNumElements() != 1 && VT.getScalarType().bitsLE(MVT::i16))
diff --git a/lib/Target/R600/SIISelLowering.h b/lib/Target/R600/SIISelLowering.h
index d03bc864148..93e97c32357 100644
--- a/lib/Target/R600/SIISelLowering.h
+++ b/lib/Target/R600/SIISelLowering.h
@@ -63,6 +63,12 @@ public:
                                       unsigned Align,
                                       bool *IsFast) const override;
 
+  EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
+                          unsigned SrcAlign, bool IsMemset,
+                          bool ZeroMemset,
+                          bool MemcpyStrSrc,
+                          MachineFunction &MF) const override;
+
   TargetLoweringBase::LegalizeTypeAction
   getPreferredVectorAction(EVT VT) const override;
 
diff --git a/test/CodeGen/R600/llvm.memcpy.ll b/test/CodeGen/R600/llvm.memcpy.ll
new file mode 100644
index 00000000000..6e9a51ef4dd
--- /dev/null
+++ b/test/CodeGen/R600/llvm.memcpy.ll
@@ -0,0 +1,358 @@
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+
+declare void @llvm.memcpy.p3i8.p3i8.i32(i8 addrspace(3)* nocapture, i8 addrspace(3)* nocapture, i32, i32, i1) nounwind
+declare void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* nocapture, i8 addrspace(1)* nocapture, i64, i32, i1) nounwind
+
+
+; FUNC-LABEL: @test_small_memcpy_i64_lds_to_lds_align1
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+; SI: DS_READ_U8
+; SI: DS_WRITE_B8
+
+; SI: S_ENDPGM
+define void @test_small_memcpy_i64_lds_to_lds_align1(i64 addrspace(3)* noalias %out, i64 addrspace(3)* noalias %in) nounwind {
+  %bcin = bitcast i64 addrspace(3)* %in to i8 addrspace(3)*
+  %bcout = bitcast i64 addrspace(3)* %out to i8 addrspace(3)*
+  call void @llvm.memcpy.p3i8.p3i8.i32(i8 addrspace(3)* %bcout, i8 addrspace(3)* %bcin, i32 32, i32 1, i1 false) nounwind
+  ret void
+}
+
+; FUNC-LABEL: @test_small_memcpy_i64_lds_to_lds_align2
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+; SI: DS_READ_U16
+; SI: DS_WRITE_B16
+
+; SI: S_ENDPGM
+define void @test_small_memcpy_i64_lds_to_lds_align2(i64 addrspace(3)* noalias %out, i64 addrspace(3)* noalias %in) nounwind {
+  %bcin = bitcast i64 addrspace(3)* %in to i8 addrspace(3)*
+  %bcout = bitcast i64 addrspace(3)* %out to i8 addrspace(3)*
+  call void @llvm.memcpy.p3i8.p3i8.i32(i8 addrspace(3)* %bcout, i8 addrspace(3)* %bcin, i32 32, i32 2, i1 false) nounwind
+  ret void
+}
+
+; FUNC-LABEL: @test_small_memcpy_i64_lds_to_lds_align4
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI: S_ENDPGM
+define void @test_small_memcpy_i64_lds_to_lds_align4(i64 addrspace(3)* noalias %out, i64 addrspace(3)* noalias %in) nounwind {
+  %bcin = bitcast i64 addrspace(3)* %in to i8 addrspace(3)*
+  %bcout = bitcast i64 addrspace(3)* %out to i8 addrspace(3)*
+  call void @llvm.memcpy.p3i8.p3i8.i32(i8 addrspace(3)* %bcout, i8 addrspace(3)* %bcin, i32 32, i32 4, i1 false) nounwind
+  ret void
+}
+
+; FIXME: Use 64-bit ops
+; FUNC-LABEL: @test_small_memcpy_i64_lds_to_lds_align8
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: DS_READ_B32
+; SI-DAG: DS_WRITE_B32
+
+; SI-DAG: S_ENDPGM
+define void @test_small_memcpy_i64_lds_to_lds_align8(i64 addrspace(3)* noalias %out, i64 addrspace(3)* noalias %in) nounwind {
+  %bcin = bitcast i64 addrspace(3)* %in to i8 addrspace(3)*
+  %bcout = bitcast i64 addrspace(3)* %out to i8 addrspace(3)*
+  call void @llvm.memcpy.p3i8.p3i8.i32(i8 addrspace(3)* %bcout, i8 addrspace(3)* %bcin, i32 32, i32 8, i1 false) nounwind
+  ret void
+}
+
+; FUNC-LABEL: @test_small_memcpy_i64_global_to_global_align1
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+; SI-DAG: BUFFER_LOAD_UBYTE
+; SI-DAG: BUFFER_STORE_BYTE
+
+; SI: S_ENDPGM
+define void @test_small_memcpy_i64_global_to_global_align1(i64 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in) nounwind {
+  %bcin = bitcast i64 addrspace(1)* %in to i8 addrspace(1)*
+  %bcout = bitcast i64 addrspace(1)* %out to i8 addrspace(1)*
+  call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %bcout, i8 addrspace(1)* %bcin, i64 32, i32 1, i1 false) nounwind
+  ret void
+}
+
+; FUNC-LABEL: @test_small_memcpy_i64_global_to_global_align2
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_STORE_SHORT
+
+; SI: S_ENDPGM
+define void @test_small_memcpy_i64_global_to_global_align2(i64 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in) nounwind {
+  %bcin = bitcast i64 addrspace(1)* %in to i8 addrspace(1)*
+  %bcout = bitcast i64 addrspace(1)* %out to i8 addrspace(1)*
+  call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %bcout, i8 addrspace(1)* %bcin, i64 32, i32 2, i1 false) nounwind
+  ret void
+}
+
+; FUNC-LABEL: @test_small_memcpy_i64_global_to_global_align4
+; SI: BUFFER_LOAD_DWORDX4
+; SI: BUFFER_STORE_DWORDX4
+; SI: BUFFER_LOAD_DWORDX4
+; SI: BUFFER_STORE_DWORDX4
+; SI: S_ENDPGM
+define void @test_small_memcpy_i64_global_to_global_align4(i64 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in) nounwind {
+  %bcin = bitcast i64 addrspace(1)* %in to i8 addrspace(1)*
+  %bcout = bitcast i64 addrspace(1)* %out to i8 addrspace(1)*
+  call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %bcout, i8 addrspace(1)* %bcin, i64 32, i32 4, i1 false) nounwind
+  ret void
+}
+
+; FUNC-LABEL: @test_small_memcpy_i64_global_to_global_align8
+; SI: BUFFER_LOAD_DWORDX4
+; SI: BUFFER_STORE_DWORDX4
+; SI: BUFFER_LOAD_DWORDX4
+; SI: BUFFER_STORE_DWORDX4
+; SI: S_ENDPGM
+define void @test_small_memcpy_i64_global_to_global_align8(i64 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in) nounwind {
+  %bcin = bitcast i64 addrspace(1)* %in to i8 addrspace(1)*
+  %bcout = bitcast i64 addrspace(1)* %out to i8 addrspace(1)*
+  call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %bcout, i8 addrspace(1)* %bcin, i64 32, i32 8, i1 false) nounwind
+  ret void
+}
+
+; FUNC-LABEL: @test_small_memcpy_i64_global_to_global_align16
+; SI: BUFFER_LOAD_DWORDX4
+; SI: BUFFER_STORE_DWORDX4
+; SI: BUFFER_LOAD_DWORDX4
+; SI: BUFFER_STORE_DWORDX4
+; SI: S_ENDPGM
+define void @test_small_memcpy_i64_global_to_global_align16(i64 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in) nounwind {
+  %bcin = bitcast i64 addrspace(1)* %in to i8 addrspace(1)*
+  %bcout = bitcast i64 addrspace(1)* %out to i8 addrspace(1)*
+  call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %bcout, i8 addrspace(1)* %bcin, i64 32, i32 16, i1 false) nounwind
+  ret void
+}
-- 
2.34.1