AMDGPU: Report extractelement as free in cost model

The cost for scalarized operations is computed as N * (scalar operation
cost + 1 extractelement + 1 insertelement). This partially fixes
inflating the cost of scalarized operations since every operation is
scalarized and free. I don't think we want any cost asociated with
scalarization, but for now insertelement is still counted. I'm not sure
if we should pretend that insertelement is also free, or add a way
to compute a custom scalarization cost.

llvm-svn: 254438

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

@@ -80,3 +80,14 @@ unsigned AMDGPUTTIImpl::getMaxInterleaveFactor(unsigned VF) {
   // Semi-arbitrary large amount.
   return 64;
 }
+
+int AMDGPUTTIImpl::getVectorInstrCost(unsigned Opcode, Type *ValTy,
+                                      unsigned Index) {
+  switch (Opcode) {
+  case Instruction::ExtractElement:
+    // Dynamic indexing isn't free and is best avoided.
+    return Index == ~0u ? 2 : 0;
+  default:
+    return BaseT::getVectorInstrCost(Opcode, ValTy, Index);
+  }
+}

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h

@@ -60,6 +60,8 @@ public:
   unsigned getNumberOfRegisters(bool Vector);
   unsigned getRegisterBitWidth(bool Vector);
   unsigned getMaxInterleaveFactor(unsigned VF);
+
+  int getVectorInstrCost(unsigned Opcode, Type *ValTy, unsigned Index);
 };
 
 } // end namespace llvm

llvm/test/Analysis/CostModel/AMDGPU/extractelement.ll

@@ -0,0 +1,110 @@
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa < %s | FileCheck %s
+
+; CHECK: 'extractelement_v2i32'
+; CHECK: estimated cost of 0 for {{.*}} extractelement <2 x i32>
+define void @extractelement_v2i32(i32 addrspace(1)* %out, <2 x i32> addrspace(1)* %vaddr) {
+  %vec = load <2 x i32>, <2 x i32> addrspace(1)* %vaddr
+  %elt = extractelement <2 x i32> %vec, i32 1
+  store i32 %elt, i32 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'extractelement_v2f32'
+; CHECK: estimated cost of 0 for {{.*}} extractelement <2 x float>
+define void @extractelement_v2f32(float addrspace(1)* %out, <2 x float> addrspace(1)* %vaddr) {
+  %vec = load <2 x float>, <2 x float> addrspace(1)* %vaddr
+  %elt = extractelement <2 x float> %vec, i32 1
+  store float %elt, float addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'extractelement_v3i32'
+; CHECK: estimated cost of 0 for {{.*}} extractelement <3 x i32>
+define void @extractelement_v3i32(i32 addrspace(1)* %out, <3 x i32> addrspace(1)* %vaddr) {
+  %vec = load <3 x i32>, <3 x i32> addrspace(1)* %vaddr
+  %elt = extractelement <3 x i32> %vec, i32 1
+  store i32 %elt, i32 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'extractelement_v4i32'
+; CHECK: estimated cost of 0 for {{.*}} extractelement <4 x i32>
+define void @extractelement_v4i32(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %vaddr) {
+  %vec = load <4 x i32>, <4 x i32> addrspace(1)* %vaddr
+  %elt = extractelement <4 x i32> %vec, i32 1
+  store i32 %elt, i32 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'extractelement_v8i32'
+; CHECK: estimated cost of 0 for {{.*}} extractelement <8 x i32>
+define void @extractelement_v8i32(i32 addrspace(1)* %out, <8 x i32> addrspace(1)* %vaddr) {
+  %vec = load <8 x i32>, <8 x i32> addrspace(1)* %vaddr
+  %elt = extractelement <8 x i32> %vec, i32 1
+  store i32 %elt, i32 addrspace(1)* %out
+  ret void
+}
+
+; FIXME: Should be non-0
+; CHECK: 'extractelement_v8i32_dynindex'
+; CHECK: estimated cost of 2 for {{.*}} extractelement <8 x i32>
+define void @extractelement_v8i32_dynindex(i32 addrspace(1)* %out, <8 x i32> addrspace(1)* %vaddr, i32 %idx) {
+  %vec = load <8 x i32>, <8 x i32> addrspace(1)* %vaddr
+  %elt = extractelement <8 x i32> %vec, i32 %idx
+  store i32 %elt, i32 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'extractelement_v2i64'
+; CHECK: estimated cost of 0 for {{.*}} extractelement <2 x i64>
+define void @extractelement_v2i64(i64 addrspace(1)* %out, <2 x i64> addrspace(1)* %vaddr) {
+  %vec = load <2 x i64>, <2 x i64> addrspace(1)* %vaddr
+  %elt = extractelement <2 x i64> %vec, i64 1
+  store i64 %elt, i64 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'extractelement_v3i64'
+; CHECK: estimated cost of 0 for {{.*}} extractelement <3 x i64>
+define void @extractelement_v3i64(i64 addrspace(1)* %out, <3 x i64> addrspace(1)* %vaddr) {
+  %vec = load <3 x i64>, <3 x i64> addrspace(1)* %vaddr
+  %elt = extractelement <3 x i64> %vec, i64 1
+  store i64 %elt, i64 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'extractelement_v4i64'
+; CHECK: estimated cost of 0 for {{.*}} extractelement <4 x i64>
+define void @extractelement_v4i64(i64 addrspace(1)* %out, <4 x i64> addrspace(1)* %vaddr) {
+  %vec = load <4 x i64>, <4 x i64> addrspace(1)* %vaddr
+  %elt = extractelement <4 x i64> %vec, i64 1
+  store i64 %elt, i64 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'extractelement_v8i64'
+; CHECK: estimated cost of 0 for {{.*}} extractelement <8 x i64>
+define void @extractelement_v8i64(i64 addrspace(1)* %out, <8 x i64> addrspace(1)* %vaddr) {
+  %vec = load <8 x i64>, <8 x i64> addrspace(1)* %vaddr
+  %elt = extractelement <8 x i64> %vec, i64 1
+  store i64 %elt, i64 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'extractelement_v4i8'
+; CHECK: estimated cost of 0 for {{.*}} extractelement <4 x i8>
+define void @extractelement_v4i8(i8 addrspace(1)* %out, <4 x i8> addrspace(1)* %vaddr) {
+  %vec = load <4 x i8>, <4 x i8> addrspace(1)* %vaddr
+  %elt = extractelement <4 x i8> %vec, i8 1
+  store i8 %elt, i8 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'extractelement_v2i16'
+; CHECK: estimated cost of 0 for {{.*}} extractelement <2 x i16>
+define void @extractelement_v2i16(i16 addrspace(1)* %out, <2 x i16> addrspace(1)* %vaddr) {
+  %vec = load <2 x i16>, <2 x i16> addrspace(1)* %vaddr
+  %elt = extractelement <2 x i16> %vec, i16 1
+  store i16 %elt, i16 addrspace(1)* %out
+  ret void
+}

llvm/test/Analysis/CostModel/AMDGPU/lit.local.cfg

@@ -0,0 +1,2 @@
+if not 'AMDGPU' in config.root.targets:
+    config.unsupported = True