R600: Increase nearby load scheduling threshold.

This partially fixes weird looking load scheduling
in memcpy test. The load clustering doesn't seem
particularly smart, but this method seems to be partially
deprecated so it might not be worth trying to fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@214943 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/R600/AMDGPUInstrInfo.cpp

@@ -218,15 +218,26 @@ bool AMDGPUInstrInfo::enableClusterLoads() const {
   return true;
 }
 
-bool AMDGPUInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
+// FIXME: This behaves strangely. If, for example, you have 32 load + stores,
-                                             int64_t Offset1, int64_t Offset2,
+// the first 16 loads will be interleaved with the stores, and the next 16 will
-                                             unsigned NumLoads) const {
+// be clustered as expected. It should really split into 2 16 store batches.
-  assert(Offset2 > Offset1
+//
-         && "Second offset should be larger than first offset!");
+// Loads are clustered until this returns false, rather than trying to schedule
-  // If we have less than 16 loads in a row, and the offsets are within 16,
+// groups of stores. This also means we have to deal with saying different
-  // then schedule together.
+// address space loads should be clustered, and ones which might cause bank
-  // TODO: Make the loads schedule near if it fits in a cacheline
+// conflicts.
-  return (NumLoads < 16 && (Offset2 - Offset1) < 16);
+//
+// This might be deprecated so it might not be worth that much effort to fix.
+bool AMDGPUInstrInfo::shouldScheduleLoadsNear(SDNode *Load0, SDNode *Load1,
+                                              int64_t Offset0, int64_t Offset1,
+                                              unsigned NumLoads) const {
+  assert(Offset1 > Offset0 &&
+         "Second offset should be larger than first offset!");
+  // If we have less than 16 loads in a row, and the offsets are within 64
+  // bytes, then schedule together.
+
+  // A cacheline is 64 bytes (for global memory).
+  return (NumLoads <= 16 && (Offset1 - Offset0) < 64);
 }
 
 bool

test/CodeGen/R600/llvm.memcpy.ll

@@ -11,6 +11,11 @@ declare void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* nocapture, i8 addrspace
 ; 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
@@ -31,13 +36,8 @@ declare void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* nocapture, i8 addrspace
 ; 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_READ_U8
@@ -47,6 +47,7 @@ declare void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* nocapture, i8 addrspace
 ; SI: DS_READ_U8
 ; SI: DS_READ_U8
 ; SI: DS_READ_U8
+
 ; SI: DS_READ_U8
 ; SI: DS_READ_U8
 ; SI: DS_READ_U8
@@ -63,6 +64,9 @@ declare void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* nocapture, i8 addrspace
 ; SI: DS_WRITE_B8
 ; SI: DS_WRITE_B8
 ; SI: DS_WRITE_B8
+; SI: DS_WRITE_B8
+; SI: DS_WRITE_B8
+
 ; SI: DS_WRITE_B8
 ; SI: DS_WRITE_B8
 ; SI: DS_WRITE_B8
@@ -83,21 +87,13 @@ define void @test_small_memcpy_i64_lds_to_lds_align1(i64 addrspace(3)* noalias %
 
 ; 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_READ_U16
@@ -117,6 +113,15 @@ define void @test_small_memcpy_i64_lds_to_lds_align1(i64 addrspace(3)* noalias %
 ; SI: DS_WRITE_B16
 ; SI: DS_WRITE_B16
 
+; SI: DS_WRITE_B16
+; SI: DS_WRITE_B16
+; SI: DS_WRITE_B16
+; SI: DS_WRITE_B16
+; SI: DS_WRITE_B16
+; SI: DS_WRITE_B16
+; SI: DS_WRITE_B16
+; 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)*
@@ -278,37 +283,37 @@ define void @test_small_memcpy_i64_global_to_global_align1(i64 addrspace(1)* noa
 
 ; 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_LOAD_USHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_LOAD_USHORT
+; SI-DAG: BUFFER_LOAD_USHORT
 
-; 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_STORE_SHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_STORE_SHORT
+; SI-DAG: BUFFER_STORE_SHORT
 ; SI-DAG: BUFFER_STORE_SHORT
 
 ; SI: S_ENDPGM
@@ -321,9 +326,9 @@ define void @test_small_memcpy_i64_global_to_global_align2(i64 addrspace(1)* noa
 
 ; 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: 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)*
@@ -334,9 +339,9 @@ define void @test_small_memcpy_i64_global_to_global_align4(i64 addrspace(1)* noa
 
 ; 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: 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)*
@@ -347,9 +352,9 @@ define void @test_small_memcpy_i64_global_to_global_align8(i64 addrspace(1)* noa
 
 ; 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: 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)*