AMDGPU: Split LDS vector loads

[arsenm]

Summary

If properly aligned this could allow using ds_read_b64.

[• tstellarAMD]

LGTM.

[arsenm]

r253975

lib/Target/AMDGPU/AMDGPUISelLowering.cpp

if (Result.getNode()) {
     return Result;
   }
 
   StoreSDNode *Store = cast<StoreSDNode>(Op);
   SDValue Chain = Store->getChain();
   if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
        Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
       Store->getValue().getValueType().isVector()) {
-    return ScalarizeVectorStore(Op, DAG);
+    return SplitVectorStore(Op, DAG);
   }
 
   EVT MemVT = Store->getMemoryVT();
   if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
       MemVT.bitsLT(MVT::i32)) {
     unsigned Mask = 0;
     if (Store->getMemoryVT() == MVT::i8) {
       Mask = 0xff;

lib/Target/AMDGPU/SIISelLowering.cpp

switch (Load->getAddressSpace()) {
         if (NumElements >= 8)
           return SplitVectorLoad(Op, DAG);
 
         // v4 loads are supported for private and global memory.
         if (NumElements <= 4)
           break;
         // fall-through
       case AMDGPUAS::LOCAL_ADDRESS:
-        return ScalarizeVectorLoad(Op, DAG);
+        // If properly aligned, if we split we might be able to use ds_read_b64.
+        return SplitVectorLoad(Op, DAG);
     }
   }
 
   return AMDGPUTargetLowering::LowerLOAD(Op, DAG);
 }
 
 SDValue SITargetLowering::LowerSampleIntrinsic(unsigned Opcode,
                                                const SDValue &Op,

test/CodeGen/AMDGPU/ds_read2_superreg.ll

define void @simple_read2_v3f32_superreg_align4(float addrspace(1)* %out) #0 {
   %add1 = fadd float %add0, %elt1
 
   %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
   store float %add1, float addrspace(1)* %out.gep
   ret void
 }
 
 ; CI-LABEL: {{^}}simple_read2_v4f32_superreg_align8:
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_W:[0-9]+]]:[[REG_Z:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:3 offset1:2{{$}}
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_X:[0-9]+]]:[[REG_Y:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:1{{$}}
+; CI-DAG: ds_read2_b64 v{{\[}}[[REG_W:[0-9]+]]:[[REG_Z:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:1{{$}}
+
+; FIXME: These moves shouldn't be necessary, it should be able to
+; store the same register if offset1 was the non-zero offset.
+
+; CI: v_mov_b32
+; CI: v_mov_b32
 ; CI: buffer_store_dwordx4
 ; CI: s_endpgm
 define void @simple_read2_v4f32_superreg_align8(<4 x float> addrspace(1)* %out) #0 {
   %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
   %arrayidx0 = getelementptr inbounds [512 x <4 x float>], [512 x <4 x float>] addrspace(3)* @lds.v4, i32 0, i32 %x.i
   %val0 = load <4 x float>, <4 x float> addrspace(3)* %arrayidx0, align 8
   %out.gep = getelementptr inbounds <4 x float>, <4 x float> addrspace(1)* %out, i32 %x.i
   store <4 x float> %val0, <4 x float> addrspace(1)* %out.gep
   ret void
 }
 
 ; CI-LABEL: {{^}}simple_read2_v4f32_superreg:
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_W:[0-9]+]]:[[REG_Z:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:3 offset1:2{{$}}
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_X:[0-9]+]]:[[REG_Y:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:1{{$}}
+; CI: ds_read2_b64 v{{\[}}[[REG_W:[0-9]+]]:[[REG_Z:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:1{{$}}
+; CI: v_mov_b32
+; CI: v_mov_b32
 ; CI: buffer_store_dwordx4
 ; CI: s_endpgm
 define void @simple_read2_v4f32_superreg(<4 x float> addrspace(1)* %out) #0 {
   %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
   %arrayidx0 = getelementptr inbounds [512 x <4 x float>], [512 x <4 x float>] addrspace(3)* @lds.v4, i32 0, i32 %x.i
   %val0 = load <4 x float>, <4 x float> addrspace(3)* %arrayidx0
   %out.gep = getelementptr inbounds <4 x float>, <4 x float> addrspace(1)* %out, i32 %x.i
   store <4 x float> %val0, <4 x float> addrspace(1)* %out.gep
   ret void
 }
 
 ; FIXME: Extra moves shuffling superregister
 ; CI-LABEL: {{^}}simple_read2_v8f32_superreg:
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_ELT3:[0-9]+]]:[[REG_ELT7:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:3 offset1:7{{$}}
+; CI: ds_read2_b64 v{{\[}}[[REG_ELT3:[0-9]+]]:[[REG_ELT7:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:1 offset1:3{{$}}
 ; CI: v_mov_b32
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_ELT6:[0-9]+]]:[[REG_ELT5:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:2 offset1:1{{$}}
 ; CI: v_mov_b32
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_ELT4:[0-9]+]]:[[REG_ELT2:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:6 offset1:5{{$}}
+; CI: ds_read2_b64 v{{\[}}[[REG_ELT6:[0-9]+]]:[[REG_ELT5:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:2{{$}}
 ; CI: v_mov_b32
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_ELT1:[0-9]+]]:[[REG_ELT0:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:4{{$}}
 ; CI: v_mov_b32
 ; CI: buffer_store_dwordx4
 ; CI: buffer_store_dwordx4
 ; CI: s_endpgm
 define void @simple_read2_v8f32_superreg(<8 x float> addrspace(1)* %out) #0 {
   %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
   %arrayidx0 = getelementptr inbounds [512 x <8 x float>], [512 x <8 x float>] addrspace(3)* @lds.v8, i32 0, i32 %x.i
   %val0 = load <8 x float>, <8 x float> addrspace(3)* %arrayidx0
   %out.gep = getelementptr inbounds <8 x float>, <8 x float> addrspace(1)* %out, i32 %x.i
   store <8 x float> %val0, <8 x float> addrspace(1)* %out.gep
   ret void
 }
 
 ; FIXME: Extra moves shuffling superregister
 ; CI-LABEL: {{^}}simple_read2_v16f32_superreg:
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_ELT11:[0-9]+]]:[[REG_ELT15:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:3 offset1:7{{$}}
-; CI: v_mov_b32
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_ELT14:[0-9]+]]:[[REG_ELT13:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:11 offset1:15{{$}}
-; CI: v_mov_b32
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_ELT12:[0-9]+]]:[[REG_ELT10:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:2 offset1:1{{$}}
+; CI: ds_read2_b64 v{{\[}}[[REG_ELT11:[0-9]+]]:[[REG_ELT15:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:3 offset1:5{{$}}
 ; CI: v_mov_b32
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_ELT9:[0-9]+]]:[[REG_ELT8:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:6 offset1:5{{$}}
 ; CI: v_mov_b32
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_ELT7:[0-9]+]]:[[REG_ELT6:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:10 offset1:9{{$}}
+; CI: ds_read2_b64 v{{\[}}[[REG_ELT14:[0-9]+]]:[[REG_ELT13:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:7 offset1:2{{$}}
+; CI: ds_read2_b64 v{{\[}}[[REG_ELT14:[0-9]+]]:[[REG_ELT13:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:6 offset1:4{{$}}
 ; CI: v_mov_b32
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_ELT5:[0-9]+]]:[[REG_ELT4:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:14 offset1:13{{$}}
 ; CI: v_mov_b32
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_ELT3:[0-9]+]]:[[REG_ELT2:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:12 offset1:8{{$}}
+; CI: ds_read2_b64 v{{\[}}[[REG_ELT12:[0-9]+]]:[[REG_ELT10:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:1{{$}}
 ; CI: v_mov_b32
-; CI-DAG: ds_read2_b32 v{{\[}}[[REG_ELT1:[0-9]+]]:[[REG_ELT0:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:4{{$}}
 ; CI: v_mov_b32
 
 ; CI: s_waitcnt lgkmcnt(0)
 ; CI: buffer_store_dwordx4
 ; CI: buffer_store_dwordx4
 ; CI: buffer_store_dwordx4
 ; CI: buffer_store_dwordx4
 ; CI: s_endpgm

test/CodeGen/AMDGPU/indirect-private-64.ll

define void @private_access_f64_alloca(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in, i32 %b) nounwind {
   ret void
 }
 
 ; SI-LABEL: {{^}}private_access_v2f64_alloca:
 
 ; SI-ALLOCA: buffer_store_dwordx4
 ; SI-ALLOCA: buffer_load_dwordx4
 
-; SI-PROMOTE: ds_write_b32
-; SI-PROMOTE: ds_write_b32
-; SI-PROMOTE: ds_write_b32
-; SI-PROMOTE: ds_write_b32
-; SI-PROMOTE: ds_read_b32
-; SI-PROMOTE: ds_read_b32
-; SI-PROMOTE: ds_read_b32
-; SI-PROMOTE: ds_read_b32
+; SI-PROMOTE: ds_write_b64
+; SI-PROMOTE: ds_write_b64
+; SI-PROMOTE: ds_read_b64
+; SI-PROMOTE: ds_read_b64
 define void @private_access_v2f64_alloca(<2 x double> addrspace(1)* noalias %out, <2 x double> addrspace(1)* noalias %in, i32 %b) nounwind {
   %val = load <2 x double>, <2 x double> addrspace(1)* %in, align 16
   %array = alloca <2 x double>, i32 16, align 16
   %ptr = getelementptr <2 x double>, <2 x double>* %array, i32 %b
   store <2 x double> %val, <2 x double>* %ptr, align 16
   call void @llvm.AMDGPU.barrier.local() noduplicate nounwind
   %result = load <2 x double>, <2 x double>* %ptr, align 16
   store <2 x double> %result, <2 x double> addrspace(1)* %out, align 16
define void @private_access_i64_alloca(i64 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in, i32 %b) nounwind {
   ret void
 }
 
 ; SI-LABEL: {{^}}private_access_v2i64_alloca:
 
 ; SI-ALLOCA: buffer_store_dwordx4
 ; SI-ALLOCA: buffer_load_dwordx4
 
-; SI-PROMOTE: ds_write_b32
-; SI-PROMOTE: ds_write_b32
-; SI-PROMOTE: ds_write_b32
-; SI-PROMOTE: ds_write_b32
-; SI-PROMOTE: ds_read_b32
-; SI-PROMOTE: ds_read_b32
-; SI-PROMOTE: ds_read_b32
-; SI-PROMOTE: ds_read_b32
+; SI-PROMOTE: ds_write_b64
+; SI-PROMOTE: ds_write_b64
+; SI-PROMOTE: ds_read_b64
+; SI-PROMOTE: ds_read_b64
 define void @private_access_v2i64_alloca(<2 x i64> addrspace(1)* noalias %out, <2 x i64> addrspace(1)* noalias %in, i32 %b) nounwind {
   %val = load <2 x i64>, <2 x i64> addrspace(1)* %in, align 16
   %array = alloca <2 x i64>, i32 16, align 16
   %ptr = getelementptr <2 x i64>, <2 x i64>* %array, i32 %b
   store <2 x i64> %val, <2 x i64>* %ptr, align 16
   call void @llvm.AMDGPU.barrier.local() noduplicate nounwind
   %result = load <2 x i64>, <2 x i64>* %ptr, align 16
   store <2 x i64> %result, <2 x i64> addrspace(1)* %out, align 16
   ret void
 }

test/CodeGen/AMDGPU/llvm.memcpy.ll

 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: ds_read2_b32
+; SI: ds_read2_b32
+; SI: ds_read2_b32
+; SI: ds_read2_b32
+
+; SI: ds_write2_b32
+; SI: ds_write2_b32
+; SI: ds_write2_b32
+; SI: ds_write2_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: ds_read_b64
+; SI: ds_read_b64
+; SI: ds_read_b64
+; SI: ds_read_b64
+
+; SI: ds_write_b64
+; SI: ds_write_b64
+; SI: ds_write_b64
+; SI: ds_write_b64
 
 ; 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
 }

test/CodeGen/AMDGPU/merge-stores.ll

define void @merge_local_store_2_constants_i32(i32 addrspace(3)* %out) #0 {
   %out.gep.1 = getelementptr i32, i32 addrspace(3)* %out, i32 1
 
   store i32 123, i32 addrspace(3)* %out.gep.1
   store i32 456, i32 addrspace(3)* %out
   ret void
 }
 
 ; GCN-LABEL: {{^}}merge_local_store_4_constants_i32:
-; GCN: ds_write_b32
-; GCN: ds_write_b32
-; GCN: ds_write_b32
-; GCN: ds_write_b32
+; GCN-DAG: v_mov_b32_e32 [[K2:v[0-9]+]], 0x1c8
+; GCN-DAG: v_mov_b32_e32 [[K3:v[0-9]+]], 0x14d
+; GCN-DAG: ds_write2_b32 v{{[0-9]+}}, [[K2]], [[K3]] offset0:2 offset1:3
+
+; GCN-DAG: v_mov_b32_e32 [[K0:v[0-9]+]], 0x4d2
+; GCN-DAG: v_mov_b32_e32 [[K1:v[0-9]+]], 0x7b
+; GCN-DAG: ds_write2_b32 v{{[0-9]+}}, [[K0]], [[K1]] offset1:1
+
+; GCN: s_endpgm
 define void @merge_local_store_4_constants_i32(i32 addrspace(3)* %out) #0 {
   %out.gep.1 = getelementptr i32, i32 addrspace(3)* %out, i32 1
   %out.gep.2 = getelementptr i32, i32 addrspace(3)* %out, i32 2
   %out.gep.3 = getelementptr i32, i32 addrspace(3)* %out, i32 3
 
   store i32 123, i32 addrspace(3)* %out.gep.1
   store i32 456, i32 addrspace(3)* %out.gep.2
   store i32 333, i32 addrspace(3)* %out.gep.3

test/CodeGen/AMDGPU/store.ll

 ; EG: LDS_WRITE
 ; EG: LDS_WRITE
 
 ; CM: LDS_WRITE
 ; CM: LDS_WRITE
 ; CM: LDS_WRITE
 ; CM: LDS_WRITE
 
-; SI: ds_write_b32
-; SI: ds_write_b32
-; SI: ds_write_b32
-; SI: ds_write_b32
+; SI: ds_write_b64
+; SI: ds_write_b64
 define void @store_local_v4i32(<4 x i32> addrspace(3)* %out, <4 x i32> %in) {
 entry:
   store <4 x i32> %in, <4 x i32> addrspace(3)* %out
   ret void
 }
 
+; FUNC-LABEL: {{^}}store_local_v4i32_align4:
+; EG: LDS_WRITE
+; EG: LDS_WRITE
+; EG: LDS_WRITE
+; EG: LDS_WRITE
+
+; CM: LDS_WRITE
+; CM: LDS_WRITE
+; CM: LDS_WRITE
+; CM: LDS_WRITE
+
+; SI: ds_write2_b32
+; SI: ds_write2_b32
+define void @store_local_v4i32_align4(<4 x i32> addrspace(3)* %out, <4 x i32> %in) {
+entry:
+  store <4 x i32> %in, <4 x i32> addrspace(3)* %out, align 4
+  ret void
+}
+
 ; FUNC-LABEL: {{^}}store_local_i64_i8:
 ; EG: LDS_BYTE_WRITE
 ; SI: ds_write_b8
 define void @store_local_i64_i8(i8 addrspace(3)* %out, i64 %in) {
 entry:
   %0 = trunc i64 %in to i8
   store i8 %0, i8 addrspace(3)* %out
   ret void