R600: Treat CONSTANT_ADDRESS loads like GLOBAL_ADDRESS loads when necessary

These are really the same address space in hardware.  The only
difference is that CONSTANT_ADDRESS uses a special cache for faster
access.  When we are unable to use the constant kcache for some reason
(e.g. smaller types or lack of indirect addressing) then the instruction
selector must use GLOBAL_ADDRESS loads instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@187006 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/R600/AMDGPUISelDAGToDAG.cpp b/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
index 696910911f9e07cb1eb0465d0675f1e8479b4432..307b80483aabe66d3960167b34fcad7b13300f06 100644 (file)
--- a/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
+++ b/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
@@ -620,6 +620,13 @@ bool AMDGPUDAGToDAGISel::isConstantLoad(const LoadSDNode *N, int CbId) const {
 }
 
 bool AMDGPUDAGToDAGISel::isGlobalLoad(const LoadSDNode *N) const {
+  if (N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS) {
+    const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
+    if (ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
+        N->getMemoryVT().bitsLT(MVT::i32)) {
+      return true;
+    }
+  }
   return checkType(N->getSrcValue(), AMDGPUAS::GLOBAL_ADDRESS);
 }
 

diff --git a/lib/Target/R600/R600Instructions.td b/lib/Target/R600/R600Instructions.td
index 3511dcd1ead36af6b2a326ca2c14690237f2c2c4..9ff389701e63e8bfdad0d03114f42e436c596e79 100644 (file)
--- a/lib/Target/R600/R600Instructions.td
+++ b/lib/Target/R600/R600Instructions.td
@@ -1418,16 +1418,6 @@ def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
   [(set v4i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
 >;
 
-//===----------------------------------------------------------------------===//
-// Constant Loads
-// XXX: We are currently storing all constants in the global address space.
-//===----------------------------------------------------------------------===//
-
-def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
-  [(set i32:$dst_gpr, (constant_load ADDRVTX_READ:$src_gpr))]
->;
-
-
 } // End Predicates = [isEG]
 
 //===----------------------------------------------------------------------===//
@@ -1883,15 +1873,6 @@ def VTX_READ_GLOBAL_128_cm : VTX_READ_128_cm <1,
   [(set v4i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
 >;
 
-//===----------------------------------------------------------------------===//
-// Constant Loads
-// XXX: We are currently storing all constants in the global address space.
-//===----------------------------------------------------------------------===//
-
-def CONSTANT_LOAD_cm : VTX_READ_32_cm <1,
-  [(set i32:$dst_gpr, (constant_load ADDRVTX_READ:$src_gpr))]
->;
-
 } // End isCayman
 
 //===----------------------------------------------------------------------===//

diff --git a/test/CodeGen/R600/load.ll b/test/CodeGen/R600/load.ll
index f36f20c63d552191ab2cb64d0b9c61c9ed2d9867..6eef7c7966e50e06e6f80a8dea67748081882689 100644 (file)
--- a/test/CodeGen/R600/load.ll
+++ b/test/CodeGen/R600/load.ll
@@ -2,6 +2,10 @@
 ; RUN: llc < %s -march=r600 -mcpu=cayman | FileCheck --check-prefix=R600-CHECK %s
 ; RUN: llc < %s -march=r600 -mcpu=SI | FileCheck --check-prefix=SI-CHECK  %s
 
+;===------------------------------------------------------------------------===;
+; GLOBAL ADDRESS SPACE
+;===------------------------------------------------------------------------===;
+
 ; Load an i8 value from the global address space.
 ; R600-CHECK: @load_i8
 ; R600-CHECK: VTX_READ_8 T{{[0-9]+\.X, T[0-9]+\.X}}
@@ -100,31 +104,6 @@ entry:
   ret void
 }
 
-; Load an i32 value from the constant address space.
-; R600-CHECK: @load_const_addrspace_i32
-; R600-CHECK: VTX_READ_32 T{{[0-9]+}}.X, T{{[0-9]+}}.X, 0
-
-; SI-CHECK: @load_const_addrspace_i32
-; SI-CHECK: S_LOAD_DWORD SGPR{{[0-9]+}}
-define void @load_const_addrspace_i32(i32 addrspace(1)* %out, i32 addrspace(2)* %in) {
-entry:
-  %0 = load i32 addrspace(2)* %in
-  store i32 %0, i32 addrspace(1)* %out
-  ret void
-}
-
-; Load a f32 value from the constant address space.
-; R600-CHECK: @load_const_addrspace_f32
-; R600-CHECK: VTX_READ_32 T{{[0-9]+}}.X, T{{[0-9]+}}.X, 0
-
-; SI-CHECK: @load_const_addrspace_f32
-; SI-CHECK: S_LOAD_DWORD SGPR{{[0-9]+}}
-define void @load_const_addrspace_f32(float addrspace(1)* %out, float addrspace(2)* %in) {
-  %1 = load float addrspace(2)* %in
-  store float %1, float addrspace(1)* %out
-  ret void
-}
-
 ; R600-CHECK: @load_i64
 ; R600-CHECK: RAT
 ; R600-CHECK: RAT
@@ -166,3 +145,121 @@ entry:
   store i64 %1, i64 addrspace(1)* %out
   ret void
 }
+
+;===------------------------------------------------------------------------===;
+; CONSTANT ADDRESS SPACE
+;===------------------------------------------------------------------------===;
+
+; Load a sign-extended i8 value
+; R600-CHECK: @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: @load_const_i8_sext
+; SI-CHECK: BUFFER_LOAD_SBYTE VGPR{{[0-9]+}},
+define void @load_const_i8_sext(i32 addrspace(1)* %out, i8 addrspace(2)* %in) {
+entry:
+  %0 = load i8 addrspace(2)* %in
+  %1 = sext i8 %0 to i32
+  store i32 %1, i32 addrspace(1)* %out
+  ret void
+}
+
+; Load an aligned i8 value
+; R600-CHECK: @load_const_i8_aligned
+; R600-CHECK: VTX_READ_8 T{{[0-9]+\.X, T[0-9]+\.X}}
+; SI-CHECK: @load_const_i8_aligned
+; SI-CHECK: BUFFER_LOAD_UBYTE VGPR{{[0-9]+}},
+define void @load_const_i8_aligned(i32 addrspace(1)* %out, i8 addrspace(2)* %in) {
+entry:
+  %0 = load i8 addrspace(2)* %in
+  %1 = zext i8 %0 to i32
+  store i32 %1, i32 addrspace(1)* %out
+  ret void
+}
+
+; Load an un-aligned i8 value
+; R600-CHECK: @load_const_i8_unaligned
+; R600-CHECK: VTX_READ_8 T{{[0-9]+\.X, T[0-9]+\.X}}
+; SI-CHECK: @load_const_i8_unaligned
+; SI-CHECK: BUFFER_LOAD_UBYTE VGPR{{[0-9]+}},
+define void @load_const_i8_unaligned(i32 addrspace(1)* %out, i8 addrspace(2)* %in) {
+entry:
+  %0 = getelementptr i8 addrspace(2)* %in, i32 1
+  %1 = load i8 addrspace(2)* %0
+  %2 = zext i8 %1 to i32
+  store i32 %2, i32 addrspace(1)* %out
+  ret void
+}
+
+; Load a sign-extended i16 value
+; R600-CHECK: @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: @load_const_i16_sext
+; SI-CHECK: BUFFER_LOAD_SSHORT
+define void @load_const_i16_sext(i32 addrspace(1)* %out, i16 addrspace(2)* %in) {
+entry:
+  %0 = load i16 addrspace(2)* %in
+  %1 = sext i16 %0 to i32
+  store i32 %1, i32 addrspace(1)* %out
+  ret void
+}
+
+; Load an aligned i16 value
+; R600-CHECK: @load_const_i16_aligned
+; R600-CHECK: VTX_READ_16 T{{[0-9]+\.X, T[0-9]+\.X}}
+; SI-CHECK: @load_const_i16_aligned
+; SI-CHECK: BUFFER_LOAD_USHORT
+define void @load_const_i16_aligned(i32 addrspace(1)* %out, i16 addrspace(2)* %in) {
+entry:
+  %0 = load i16 addrspace(2)* %in
+  %1 = zext i16 %0 to i32
+  store i32 %1, i32 addrspace(1)* %out
+  ret void
+}
+
+; Load an un-aligned i16 value
+; R600-CHECK: @load_const_i16_unaligned
+; R600-CHECK: VTX_READ_16 T{{[0-9]+\.X, T[0-9]+\.X}}
+; SI-CHECK: @load_const_i16_unaligned
+; SI-CHECK: BUFFER_LOAD_USHORT
+define void @load_const_i16_unaligned(i32 addrspace(1)* %out, i16 addrspace(2)* %in) {
+entry:
+  %0 = getelementptr i16 addrspace(2)* %in, i32 1
+  %1 = load i16 addrspace(2)* %0
+  %2 = zext i16 %1 to i32
+  store i32 %2, i32 addrspace(1)* %out
+  ret void
+}
+
+; Load an i32 value from the constant address space.
+; R600-CHECK: @load_const_addrspace_i32
+; R600-CHECK: VTX_READ_32 T{{[0-9]+}}.X, T{{[0-9]+}}.X, 0
+
+; SI-CHECK: @load_const_addrspace_i32
+; SI-CHECK: S_LOAD_DWORD SGPR{{[0-9]+}}
+define void @load_const_addrspace_i32(i32 addrspace(1)* %out, i32 addrspace(2)* %in) {
+entry:
+  %0 = load i32 addrspace(2)* %in
+  store i32 %0, i32 addrspace(1)* %out
+  ret void
+}
+
+; Load a f32 value from the constant address space.
+; R600-CHECK: @load_const_addrspace_f32
+; R600-CHECK: VTX_READ_32 T{{[0-9]+}}.X, T{{[0-9]+}}.X, 0
+
+; SI-CHECK: @load_const_addrspace_f32
+; SI-CHECK: S_LOAD_DWORD SGPR{{[0-9]+}}
+define void @load_const_addrspace_f32(float addrspace(1)* %out, float addrspace(2)* %in) {
+  %1 = load float addrspace(2)* %in
+  store float %1, float addrspace(1)* %out
+  ret void
+}
+