R600/SI: Split global vector loads with more than 4 elements

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@201368 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 98 insertions, 88 deletions

diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
index f19aa8efbb..0d9f34fc4e 100644
--- a/lib/Target/R600/SIISelLowering.cpp
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -452,9 +452,11 @@ SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::BRCOND: return LowerBRCOND(Op, DAG);
   case ISD::LOAD: {
     LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
-    if ((Load->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
-         Load->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
-        Op.getValueType().isVector()) {
+    if (Op.getValueType().isVector() &&
+        (Load->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
+         Load->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS ||
+         (Load->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS &&
+          Op.getValueType().getVectorNumElements() > 4))) {
       SDValue MergedValues[2] = {
         SplitVectorLoad(Op, DAG),
         Load->getChain()
diff --git a/test/CodeGen/R600/load.ll b/test/CodeGen/R600/load.ll
index 0153524d13..1486c4dbff 100644
--- a/test/CodeGen/R600/load.ll
+++ b/test/CodeGen/R600/load.ll
@@ -1,16 +1,15 @@
-; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck --check-prefix=R600-CHECK %s
-; RUN: llc < %s -march=r600 -mcpu=cayman | FileCheck --check-prefix=R600-CHECK %s
-; RUN: llc < %s -march=r600 -mcpu=SI -verify-machineinstrs | FileCheck --check-prefix=SI-CHECK  %s
+; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck --check-prefix=R600-CHECK --check-prefix=FUNC %s
+; RUN: llc < %s -march=r600 -mcpu=cayman | FileCheck --check-prefix=R600-CHECK --check-prefix=FUNC %s
+; RUN: llc < %s -march=r600 -mcpu=SI -verify-machineinstrs | FileCheck --check-prefix=SI-CHECK --check-prefix=FUNC %s
 
 ;===------------------------------------------------------------------------===;
 ; GLOBAL ADDRESS SPACE
 ;===------------------------------------------------------------------------===;
 
 ; Load an i8 value from the global address space.
-; R600-CHECK-LABEL: @load_i8
+; FUNC-LABEL: @load_i8
 ; R600-CHECK: VTX_READ_8 T{{[0-9]+\.X, T[0-9]+\.X}}
 
-; SI-CHECK-LABEL: @load_i8
 ; SI-CHECK: BUFFER_LOAD_UBYTE v{{[0-9]+}},
 define void @load_i8(i32 addrspace(1)* %out, i8 addrspace(1)* %in) {
   %1 = load i8 addrspace(1)* %in
@@ -19,13 +18,12 @@ define void @load_i8(i32 addrspace(1)* %out, i8 addrspace(1)* %in) {
   ret void
 }
 
-; R600-CHECK-LABEL: @load_i8_sext
+; FUNC-LABEL: @load_i8_sext
 ; R600-CHECK: VTX_READ_8 [[DST:T[0-9]\.[XYZW]]], [[DST]]
 ; R600-CHECK: LSHL {{[* ]*}}T{{[0-9]}}.[[LSHL_CHAN:[XYZW]]], [[DST]]
 ; R600-CHECK: 24
 ; R600-CHECK: ASHR {{[* ]*}}T{{[0-9]\.[XYZW]}}, PV.[[LSHL_CHAN]]
 ; R600-CHECK: 24
-; SI-CHECK-LABEL: @load_i8_sext
 ; SI-CHECK: BUFFER_LOAD_SBYTE
 define void @load_i8_sext(i32 addrspace(1)* %out, i8 addrspace(1)* %in) {
 entry:
@@ -35,10 +33,9 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v2i8
+; FUNC-LABEL: @load_v2i8
 ; R600-CHECK: VTX_READ_8
 ; R600-CHECK: VTX_READ_8
-; SI-CHECK-LABEL: @load_v2i8
 ; SI-CHECK: BUFFER_LOAD_UBYTE
 ; SI-CHECK: BUFFER_LOAD_UBYTE
 define void @load_v2i8(<2 x i32> addrspace(1)* %out, <2 x i8> addrspace(1)* %in) {
@@ -49,7 +46,7 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v2i8_sext
+; FUNC-LABEL: @load_v2i8_sext
 ; R600-CHECK-DAG: VTX_READ_8 [[DST_X:T[0-9]\.[XYZW]]], [[DST_X]]
 ; R600-CHECK-DAG: VTX_READ_8 [[DST_Y:T[0-9]\.[XYZW]]], [[DST_Y]]
 ; R600-CHECK-DAG: LSHL {{[* ]*}}T{{[0-9]}}.[[LSHL_X_CHAN:[XYZW]]], [[DST_X]]
@@ -60,7 +57,6 @@ entry:
 ; R600-CHECK-DAG: 24
 ; R600-CHECK-DAG: ASHR {{[* ]*}}T{{[0-9]\.[XYZW]}}, PV.[[LSHL_Y_CHAN]]
 ; R600-CHECK-DAG: 24
-; SI-CHECK-LABEL: @load_v2i8_sext
 ; SI-CHECK: BUFFER_LOAD_SBYTE
 ; SI-CHECK: BUFFER_LOAD_SBYTE
 define void @load_v2i8_sext(<2 x i32> addrspace(1)* %out, <2 x i8> addrspace(1)* %in) {
@@ -71,12 +67,11 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v4i8
+; FUNC-LABEL: @load_v4i8
 ; R600-CHECK: VTX_READ_8
 ; R600-CHECK: VTX_READ_8
 ; R600-CHECK: VTX_READ_8
 ; R600-CHECK: VTX_READ_8
-; SI-CHECK-LABEL: @load_v4i8
 ; SI-CHECK: BUFFER_LOAD_UBYTE
 ; SI-CHECK: BUFFER_LOAD_UBYTE
 ; SI-CHECK: BUFFER_LOAD_UBYTE
@@ -89,7 +84,7 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v4i8_sext
+; FUNC-LABEL: @load_v4i8_sext
 ; R600-CHECK-DAG: VTX_READ_8 [[DST_X:T[0-9]\.[XYZW]]], [[DST_X]]
 ; R600-CHECK-DAG: VTX_READ_8 [[DST_Y:T[0-9]\.[XYZW]]], [[DST_Y]]
 ; R600-CHECK-DAG: VTX_READ_8 [[DST_Z:T[0-9]\.[XYZW]]], [[DST_Z]]
@@ -110,7 +105,6 @@ entry:
 ; R600-CHECK-DAG: 24
 ; R600-CHECK-DAG: ASHR {{[* ]*}}T{{[0-9]\.[XYZW]}}, PV.[[LSHL_W_CHAN]]
 ; R600-CHECK-DAG: 24
-; SI-CHECK-LABEL: @load_v4i8_sext
 ; SI-CHECK: BUFFER_LOAD_SBYTE
 ; SI-CHECK: BUFFER_LOAD_SBYTE
 ; SI-CHECK: BUFFER_LOAD_SBYTE
@@ -124,9 +118,8 @@ entry:
 }
 
 ; Load an i16 value from the global address space.
-; R600-CHECK-LABEL: @load_i16
+; FUNC-LABEL: @load_i16
 ; R600-CHECK: VTX_READ_16 T{{[0-9]+\.X, T[0-9]+\.X}}
-; SI-CHECK-LABEL: @load_i16
 ; SI-CHECK: BUFFER_LOAD_USHORT
 define void @load_i16(i32 addrspace(1)* %out, i16 addrspace(1)* %in) {
 entry:
@@ -136,13 +129,12 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_i16_sext
+; FUNC-LABEL: @load_i16_sext
 ; R600-CHECK: VTX_READ_16 [[DST:T[0-9]\.[XYZW]]], [[DST]]
 ; R600-CHECK: LSHL {{[* ]*}}T{{[0-9]}}.[[LSHL_CHAN:[XYZW]]], [[DST]]
 ; R600-CHECK: 16
 ; R600-CHECK: ASHR {{[* ]*}}T{{[0-9]\.[XYZW]}}, PV.[[LSHL_CHAN]]
 ; R600-CHECK: 16
-; SI-CHECK-LABEL: @load_i16_sext
 ; SI-CHECK: BUFFER_LOAD_SSHORT
 define void @load_i16_sext(i32 addrspace(1)* %out, i16 addrspace(1)* %in) {
 entry:
@@ -152,10 +144,9 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v2i16
+; FUNC-LABEL: @load_v2i16
 ; R600-CHECK: VTX_READ_16
 ; R600-CHECK: VTX_READ_16
-; SI-CHECK-LABEL: @load_v2i16
 ; SI-CHECK: BUFFER_LOAD_USHORT
 ; SI-CHECK: BUFFER_LOAD_USHORT
 define void @load_v2i16(<2 x i32> addrspace(1)* %out, <2 x i16> addrspace(1)* %in) {
@@ -166,7 +157,7 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v2i16_sext
+; FUNC-LABEL: @load_v2i16_sext
 ; R600-CHECK-DAG: VTX_READ_16 [[DST_X:T[0-9]\.[XYZW]]], [[DST_X]]
 ; R600-CHECK-DAG: VTX_READ_16 [[DST_Y:T[0-9]\.[XYZW]]], [[DST_Y]]
 ; R600-CHECK-DAG: LSHL {{[* ]*}}T{{[0-9]}}.[[LSHL_X_CHAN:[XYZW]]], [[DST_X]]
@@ -177,7 +168,6 @@ entry:
 ; R600-CHECK-DAG: 16
 ; R600-CHECK-DAG: ASHR {{[* ]*}}T{{[0-9]\.[XYZW]}}, PV.[[LSHL_Y_CHAN]]
 ; R600-CHECK-DAG: 16
-; SI-CHECK-LABEL: @load_v2i16_sext
 ; SI-CHECK: BUFFER_LOAD_SSHORT
 ; SI-CHECK: BUFFER_LOAD_SSHORT
 define void @load_v2i16_sext(<2 x i32> addrspace(1)* %out, <2 x i16> addrspace(1)* %in) {
@@ -188,12 +178,11 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v4i16
+; FUNC-LABEL: @load_v4i16
 ; R600-CHECK: VTX_READ_16
 ; R600-CHECK: VTX_READ_16
 ; R600-CHECK: VTX_READ_16
 ; R600-CHECK: VTX_READ_16
-; SI-CHECK-LABEL: @load_v4i16
 ; SI-CHECK: BUFFER_LOAD_USHORT
 ; SI-CHECK: BUFFER_LOAD_USHORT
 ; SI-CHECK: BUFFER_LOAD_USHORT
@@ -206,7 +195,7 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v4i16_sext
+; FUNC-LABEL: @load_v4i16_sext
 ; R600-CHECK-DAG: VTX_READ_16 [[DST_X:T[0-9]\.[XYZW]]], [[DST_X]]
 ; R600-CHECK-DAG: VTX_READ_16 [[DST_Y:T[0-9]\.[XYZW]]], [[DST_Y]]
 ; R600-CHECK-DAG: VTX_READ_16 [[DST_Z:T[0-9]\.[XYZW]]], [[DST_Z]]
@@ -227,7 +216,6 @@ entry:
 ; R600-CHECK-DAG: 16
 ; R600-CHECK-DAG: ASHR {{[* ]*}}T{{[0-9]\.[XYZW]}}, PV.[[LSHL_W_CHAN]]
 ; R600-CHECK-DAG: 16
-; SI-CHECK-LABEL: @load_v4i16_sext
 ; SI-CHECK: BUFFER_LOAD_SSHORT
 ; SI-CHECK: BUFFER_LOAD_SSHORT
 ; SI-CHECK: BUFFER_LOAD_SSHORT
@@ -241,10 +229,9 @@ entry:
 }
 
 ; load an i32 value from the global address space.
-; R600-CHECK-LABEL: @load_i32
+; FUNC-LABEL: @load_i32
 ; R600-CHECK: VTX_READ_32 T{{[0-9]+}}.X, T{{[0-9]+}}.X, 0
 
-; SI-CHECK-LABEL: @load_i32
 ; SI-CHECK: BUFFER_LOAD_DWORD v{{[0-9]+}}
 define void @load_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
 entry:
@@ -254,10 +241,9 @@ entry:
 }
 
 ; load a f32 value from the global address space.
-; R600-CHECK-LABEL: @load_f32
+; FUNC-LABEL: @load_f32
 ; R600-CHECK: VTX_READ_32 T{{[0-9]+}}.X, T{{[0-9]+}}.X, 0
 
-; SI-CHECK-LABEL: @load_f32
 ; SI-CHECK: BUFFER_LOAD_DWORD v{{[0-9]+}}
 define void @load_f32(float addrspace(1)* %out, float addrspace(1)* %in) {
 entry:
@@ -267,10 +253,9 @@ entry:
 }
 
 ; load a v2f32 value from the global address space
-; R600-CHECK-LABEL: @load_v2f32
+; FUNC-LABEL: @load_v2f32
 ; R600-CHECK: VTX_READ_64
 
-; SI-CHECK-LABEL: @load_v2f32
 ; SI-CHECK: BUFFER_LOAD_DWORDX2
 define void @load_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)* %in) {
 entry:
@@ -279,11 +264,10 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_i64
+; FUNC-LABEL: @load_i64
 ; R600-CHECK: MEM_RAT
 ; R600-CHECK: MEM_RAT
 
-; SI-CHECK-LABEL: @load_i64
 ; SI-CHECK: BUFFER_LOAD_DWORDX2
 define void @load_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %in) {
 entry:
@@ -292,13 +276,12 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_i64_sext
+; FUNC-LABEL: @load_i64_sext
 ; R600-CHECK: MEM_RAT
 ; R600-CHECK: MEM_RAT
 ; R600-CHECK: ASHR {{[* ]*}}T{{[0-9]\.[XYZW]}}, T{{[0-9]\.[XYZW]}},  literal.x
 ; R600-CHECK: 31
-; SI-CHECK-LABEL: @load_i64_sext
-; SI-CHECK: BUFFER_LOAD_DWORDX2 [[VAL:v\[[0-9]:[0-9]\]]]
+; SI-CHECK: BUFFER_LOAD_DWORD
 
 define void @load_i64_sext(i64 addrspace(1)* %out, i32 addrspace(1)* %in) {
 entry:
@@ -308,7 +291,7 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_i64_zext
+; FUNC-LABEL: @load_i64_zext
 ; R600-CHECK: MEM_RAT
 ; R600-CHECK: MEM_RAT
 define void @load_i64_zext(i64 addrspace(1)* %out, i32 addrspace(1)* %in) {
@@ -319,18 +302,65 @@ entry:
   ret void
 }
 
+; FUNC-LABEL: @load_v8i32
+; R600-CHECK: VTX_READ_128
+; R600-CHECK: VTX_READ_128
+; XXX: We should be using DWORDX4 instructions on SI.
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+define void @load_v8i32(<8 x i32> addrspace(1)* %out, <8 x i32> addrspace(1)* %in) {
+entry:
+  %0 = load <8 x i32> addrspace(1)* %in
+  store <8 x i32> %0, <8 x i32> addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: @load_v16i32
+; R600-CHECK: VTX_READ_128
+; R600-CHECK: VTX_READ_128
+; R600-CHECK: VTX_READ_128
+; R600-CHECK: VTX_READ_128
+; XXX: We should be using DWORDX4 instructions on SI.
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+; SI-CHECK: BUFFER_LOAD_DWORD
+define void @load_v16i32(<16 x i32> addrspace(1)* %out, <16 x i32> addrspace(1)* %in) {
+entry:
+  %0 = load <16 x i32> addrspace(1)* %in
+  store <16 x i32> %0, <16 x i32> addrspace(1)* %out
+  ret void
+}
+
 ;===------------------------------------------------------------------------===;
 ; CONSTANT ADDRESS SPACE
 ;===------------------------------------------------------------------------===;
 
 ; Load a sign-extended i8 value
-; R600-CHECK-LABEL: @load_const_i8_sext
+; FUNC-LABEL: @load_const_i8_sext
 ; R600-CHECK: VTX_READ_8 [[DST:T[0-9]\.[XYZW]]], [[DST]]
 ; R600-CHECK: LSHL {{[* ]*}}T{{[0-9]}}.[[LSHL_CHAN:[XYZW]]], [[DST]]
 ; R600-CHECK: 24
 ; R600-CHECK: ASHR {{[* ]*}}T{{[0-9]\.[XYZW]}}, PV.[[LSHL_CHAN]]
 ; R600-CHECK: 24
-; SI-CHECK-LABEL: @load_const_i8_sext
 ; SI-CHECK: BUFFER_LOAD_SBYTE v{{[0-9]+}},
 define void @load_const_i8_sext(i32 addrspace(1)* %out, i8 addrspace(2)* %in) {
 entry:
@@ -341,9 +371,8 @@ entry:
 }
 
 ; Load an aligned i8 value
-; R600-CHECK-LABEL: @load_const_i8_aligned
+; FUNC-LABEL: @load_const_i8_aligned
 ; R600-CHECK: VTX_READ_8 T{{[0-9]+\.X, T[0-9]+\.X}}
-; SI-CHECK-LABEL: @load_const_i8_aligned
 ; SI-CHECK: BUFFER_LOAD_UBYTE v{{[0-9]+}},
 define void @load_const_i8_aligned(i32 addrspace(1)* %out, i8 addrspace(2)* %in) {
 entry:
@@ -354,9 +383,8 @@ entry:
 }
 
 ; Load an un-aligned i8 value
-; R600-CHECK-LABEL: @load_const_i8_unaligned
+; FUNC-LABEL: @load_const_i8_unaligned
 ; R600-CHECK: VTX_READ_8 T{{[0-9]+\.X, T[0-9]+\.X}}
-; SI-CHECK-LABEL: @load_const_i8_unaligned
 ; SI-CHECK: BUFFER_LOAD_UBYTE v{{[0-9]+}},
 define void @load_const_i8_unaligned(i32 addrspace(1)* %out, i8 addrspace(2)* %in) {
 entry:
@@ -368,13 +396,12 @@ entry:
 }
 
 ; Load a sign-extended i16 value
-; R600-CHECK-LABEL: @load_const_i16_sext
+; FUNC-LABEL: @load_const_i16_sext
 ; R600-CHECK: VTX_READ_16 [[DST:T[0-9]\.[XYZW]]], [[DST]]
 ; R600-CHECK: LSHL {{[* ]*}}T{{[0-9]}}.[[LSHL_CHAN:[XYZW]]], [[DST]]
 ; R600-CHECK: 16
 ; R600-CHECK: ASHR {{[* ]*}}T{{[0-9]\.[XYZW]}}, PV.[[LSHL_CHAN]]
 ; R600-CHECK: 16
-; SI-CHECK-LABEL: @load_const_i16_sext
 ; SI-CHECK: BUFFER_LOAD_SSHORT
 define void @load_const_i16_sext(i32 addrspace(1)* %out, i16 addrspace(2)* %in) {
 entry:
@@ -385,9 +412,8 @@ entry:
 }
 
 ; Load an aligned i16 value
-; R600-CHECK-LABEL: @load_const_i16_aligned
+; FUNC-LABEL: @load_const_i16_aligned
 ; R600-CHECK: VTX_READ_16 T{{[0-9]+\.X, T[0-9]+\.X}}
-; SI-CHECK-LABEL: @load_const_i16_aligned
 ; SI-CHECK: BUFFER_LOAD_USHORT
 define void @load_const_i16_aligned(i32 addrspace(1)* %out, i16 addrspace(2)* %in) {
 entry:
@@ -398,9 +424,8 @@ entry:
 }
 
 ; Load an un-aligned i16 value
-; R600-CHECK-LABEL: @load_const_i16_unaligned
+; FUNC-LABEL: @load_const_i16_unaligned
 ; R600-CHECK: VTX_READ_16 T{{[0-9]+\.X, T[0-9]+\.X}}
-; SI-CHECK-LABEL: @load_const_i16_unaligned
 ; SI-CHECK: BUFFER_LOAD_USHORT
 define void @load_const_i16_unaligned(i32 addrspace(1)* %out, i16 addrspace(2)* %in) {
 entry:
@@ -412,10 +437,9 @@ entry:
 }
 
 ; Load an i32 value from the constant address space.
-; R600-CHECK-LABEL: @load_const_addrspace_i32
+; FUNC-LABEL: @load_const_addrspace_i32
 ; R600-CHECK: VTX_READ_32 T{{[0-9]+}}.X, T{{[0-9]+}}.X, 0
 
-; SI-CHECK-LABEL: @load_const_addrspace_i32
 ; SI-CHECK: S_LOAD_DWORD s{{[0-9]+}}
 define void @load_const_addrspace_i32(i32 addrspace(1)* %out, i32 addrspace(2)* %in) {
 entry:
@@ -425,10 +449,9 @@ entry:
 }
 
 ; Load a f32 value from the constant address space.
-; R600-CHECK-LABEL: @load_const_addrspace_f32
+; FUNC-LABEL: @load_const_addrspace_f32
 ; R600-CHECK: VTX_READ_32 T{{[0-9]+}}.X, T{{[0-9]+}}.X, 0
 
-; SI-CHECK-LABEL: @load_const_addrspace_f32
 ; SI-CHECK: S_LOAD_DWORD s{{[0-9]+}}
 define void @load_const_addrspace_f32(float addrspace(1)* %out, float addrspace(2)* %in) {
   %1 = load float addrspace(2)* %in
@@ -441,9 +464,8 @@ define void @load_const_addrspace_f32(float addrspace(1)* %out, float addrspace(
 ;===------------------------------------------------------------------------===;
 
 ; Load an i8 value from the local address space.
-; R600-CHECK-LABEL: @load_i8_local
+; FUNC-LABEL: @load_i8_local
 ; R600-CHECK: LDS_UBYTE_READ_RET
-; SI-CHECK-LABEL: @load_i8_local
 ; SI-CHECK-NOT: S_WQM_B64
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_U8
@@ -454,10 +476,9 @@ define void @load_i8_local(i32 addrspace(1)* %out, i8 addrspace(3)* %in) {
   ret void
 }
 
-; R600-CHECK-LABEL: @load_i8_sext_local
+; FUNC-LABEL: @load_i8_sext_local
 ; R600-CHECK: LDS_UBYTE_READ_RET
 ; R600-CHECK: ASHR
-; SI-CHECK-LABEL: @load_i8_sext_local
 ; SI-CHECK-NOT: S_WQM_B64
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_I8
@@ -469,10 +490,9 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v2i8_local
+; FUNC-LABEL: @load_v2i8_local
 ; R600-CHECK: LDS_UBYTE_READ_RET
 ; R600-CHECK: LDS_UBYTE_READ_RET
-; SI-CHECK-LABEL: @load_v2i8_local
 ; SI-CHECK-NOT: S_WQM_B64
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_U8
@@ -485,12 +505,11 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v2i8_sext_local
+; FUNC-LABEL: @load_v2i8_sext_local
 ; R600-CHECK-DAG: LDS_UBYTE_READ_RET
 ; R600-CHECK-DAG: LDS_UBYTE_READ_RET
 ; R600-CHECK-DAG: ASHR
 ; R600-CHECK-DAG: ASHR
-; SI-CHECK-LABEL: @load_v2i8_sext_local
 ; SI-CHECK-NOT: S_WQM_B64
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_I8
@@ -503,12 +522,11 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v4i8_local
+; FUNC-LABEL: @load_v4i8_local
 ; R600-CHECK: LDS_UBYTE_READ_RET
 ; R600-CHECK: LDS_UBYTE_READ_RET
 ; R600-CHECK: LDS_UBYTE_READ_RET
 ; R600-CHECK: LDS_UBYTE_READ_RET
-; SI-CHECK-LABEL: @load_v4i8_local
 ; SI-CHECK-NOT: S_WQM_B64
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_U8
@@ -523,7 +541,7 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v4i8_sext_local
+; FUNC-LABEL: @load_v4i8_sext_local
 ; R600-CHECK-DAG: LDS_UBYTE_READ_RET
 ; R600-CHECK-DAG: LDS_UBYTE_READ_RET
 ; R600-CHECK-DAG: LDS_UBYTE_READ_RET
@@ -532,7 +550,6 @@ entry:
 ; R600-CHECK-DAG: ASHR
 ; R600-CHECK-DAG: ASHR
 ; R600-CHECK-DAG: ASHR
-; SI-CHECK-LABEL: @load_v4i8_sext_local
 ; SI-CHECK-NOT: S_WQM_B64
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_I8
@@ -548,9 +565,8 @@ entry:
 }
 
 ; Load an i16 value from the local address space.
-; R600-CHECK-LABEL: @load_i16_local
+; FUNC-LABEL: @load_i16_local
 ; R600-CHECK: LDS_USHORT_READ_RET
-; SI-CHECK-LABEL: @load_i16_local
 ; SI-CHECK-NOT: S_WQM_B64
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_U16
@@ -562,10 +578,9 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_i16_sext_local
+; FUNC-LABEL: @load_i16_sext_local
 ; R600-CHECK: LDS_USHORT_READ_RET
 ; R600-CHECK: ASHR
-; SI-CHECK-LABEL: @load_i16_sext_local
 ; SI-CHECK-NOT: S_WQM_B64
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_I16
@@ -577,10 +592,9 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v2i16_local
+; FUNC-LABEL: @load_v2i16_local
 ; R600-CHECK: LDS_USHORT_READ_RET
 ; R600-CHECK: LDS_USHORT_READ_RET
-; SI-CHECK-LABEL: @load_v2i16_local
 ; SI-CHECK-NOT: S_WQM_B64
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_U16
@@ -593,12 +607,11 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v2i16_sext_local
+; FUNC-LABEL: @load_v2i16_sext_local
 ; R600-CHECK-DAG: LDS_USHORT_READ_RET
 ; R600-CHECK-DAG: LDS_USHORT_READ_RET
 ; R600-CHECK-DAG: ASHR
 ; R600-CHECK-DAG: ASHR
-; SI-CHECK-LABEL: @load_v2i16_sext_local
 ; SI-CHECK-NOT: S_WQM_B64
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_I16
@@ -611,12 +624,11 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v4i16_local
+; FUNC-LABEL: @load_v4i16_local
 ; R600-CHECK: LDS_USHORT_READ_RET
 ; R600-CHECK: LDS_USHORT_READ_RET
 ; R600-CHECK: LDS_USHORT_READ_RET
 ; R600-CHECK: LDS_USHORT_READ_RET
-; SI-CHECK-LABEL: @load_v4i16_local
 ; SI-CHECK-NOT: S_WQM_B64
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_U16
@@ -631,7 +643,7 @@ entry:
   ret void
 }
 
-; R600-CHECK-LABEL: @load_v4i16_sext_local
+; FUNC-LABEL: @load_v4i16_sext_local
 ; R600-CHECK-DAG: LDS_USHORT_READ_RET
 ; R600-CHECK-DAG: LDS_USHORT_READ_RET
 ; R600-CHECK-DAG: LDS_USHORT_READ_RET
@@ -640,7 +652,6 @@ entry:
 ; R600-CHECK-DAG: ASHR
 ; R600-CHECK-DAG: ASHR
 ; R600-CHECK-DAG: ASHR
-; SI-CHECK-LABEL: @load_v4i16_sext_local
 ; SI-CHECK-NOT: S_WQM_B64
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_I16
@@ -656,9 +667,8 @@ entry:
 }
 
 ; load an i32 value from the local address space.
-; R600-CHECK-LABEL: @load_i32_local
+; FUNC-LABEL: @load_i32_local
 ; R600-CHECK: LDS_READ_RET
-; SI-CHECK-LABEL: @load_i32_local
 ; SI-CHECK-NOT: S_WQM_B64
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_B32
@@ -670,9 +680,8 @@ entry:
 }
 
 ; load a f32 value from the local address space.
-; R600-CHECK-LABEL: @load_f32_local
+; FUNC-LABEL: @load_f32_local
 ; R600-CHECK: LDS_READ_RET
-; SI-CHECK-LABEL: @load_f32_local
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_B32
 define void @load_f32_local(float addrspace(1)* %out, float addrspace(3)* %in) {
@@ -683,10 +692,9 @@ entry:
 }
 
 ; load a v2f32 value from the local address space
-; R600-CHECK-LABEL: @load_v2f32_local
+; FUNC-LABEL: @load_v2f32_local
 ; R600-CHECK: LDS_READ_RET
 ; R600-CHECK: LDS_READ_RET
-; SI-CHECK-LABEL: @load_v2f32_local
 ; SI-CHECK: S_MOV_B32 m0
 ; SI-CHECK: DS_READ_B32
 ; SI-CHECK: DS_READ_B32