[AMDGPU] Improve code size cost model

Summary:
Added estimation for zero size insertelement, extractelement
and llvm.fabs operators.
Updated inline/unroll parameters default values.

Reviewers: rampitec, arsenm

Reviewed By: arsenm

Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D68881

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@375109 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/AMDGPUInline.cpp b/lib/Target/AMDGPU/AMDGPUInline.cpp
index f4df20b8f03ea8e796af4303c310ebda2cbc3f08..a83ec23ec0542e5495b618ec7fe2c8d105b61f06 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUInline.cpp
+++ b/lib/Target/AMDGPU/AMDGPUInline.cpp
@@ -51,7 +51,7 @@ ArgAllocaCutoff("amdgpu-inline-arg-alloca-cutoff", cl::Hidden, cl::init(256),
 
 // Inliner constraint to achieve reasonable compilation time
 static cl::opt<size_t>
-MaxBB("amdgpu-inline-max-bb", cl::Hidden, cl::init(300),
+MaxBB("amdgpu-inline-max-bb", cl::Hidden, cl::init(1100),
       cl::desc("Maximum BB number allowed in a function after inlining"
                " (compile time constraint)"));
 

diff --git a/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp b/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
index b36580cdf2f8b4338eb285ff2766e30f3133a9d1..616196ad5ba37c3f22f85cf1ea9818bb0bf42990 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
+++ b/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
@@ -57,7 +57,7 @@ using namespace llvm;
 static cl::opt<unsigned> UnrollThresholdPrivate(
   "amdgpu-unroll-threshold-private",
   cl::desc("Unroll threshold for AMDGPU if private memory used in a loop"),
-  cl::init(2500), cl::Hidden);
+  cl::init(2000), cl::Hidden);
 
 static cl::opt<unsigned> UnrollThresholdLocal(
   "amdgpu-unroll-threshold-local",
@@ -693,6 +693,39 @@ void GCNTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
   CommonTTI.getUnrollingPreferences(L, SE, UP);
 }
 
+unsigned GCNTTIImpl::getUserCost(const User *U,
+                                 ArrayRef<const Value *> Operands) {
+  // Estimate extractelement elimination
+  if (const ExtractElementInst *EE = dyn_cast<ExtractElementInst>(U)) {
+    ConstantInt *CI = dyn_cast<ConstantInt>(EE->getOperand(1));
+    unsigned Idx = -1;
+    if (CI)
+      Idx = CI->getZExtValue();
+    return getVectorInstrCost(EE->getOpcode(), EE->getOperand(0)->getType(),
+                              Idx);
+  }
+
+  // Estimate insertelement elimination
+  if (const InsertElementInst *IE = dyn_cast<InsertElementInst>(U)) {
+    ConstantInt *CI = dyn_cast<ConstantInt>(IE->getOperand(2));
+    unsigned Idx = -1;
+    if (CI)
+      Idx = CI->getZExtValue();
+    return getVectorInstrCost(IE->getOpcode(), IE->getType(), Idx);
+  }
+
+  // Estimate different intrinsics, e.g. llvm.fabs
+  if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
+    SmallVector<Value *, 4> Args(II->arg_operands());
+    FastMathFlags FMF;
+    if (auto *FPMO = dyn_cast<FPMathOperator>(II))
+      FMF = FPMO->getFastMathFlags();
+    return getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(), Args,
+                                 FMF);
+  }
+  return BaseT::getUserCost(U, Operands);
+}
+
 unsigned R600TTIImpl::getHardwareNumberOfRegisters(bool Vec) const {
   return 4 * 128; // XXX - 4 channels. Should these count as vector instead?
 }

diff --git a/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h b/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h
index c7cd2f2c7bfd096b9968995913eaff0073265239..67f7f9074f10271d0285e6d9f92f28767e83837d 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h
+++ b/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h
@@ -204,7 +204,7 @@ public:
   bool areInlineCompatible(const Function *Caller,
                            const Function *Callee) const;
 
-  unsigned getInliningThresholdMultiplier() { return 7; }
+  unsigned getInliningThresholdMultiplier() { return 9; }
 
   int getInlinerVectorBonusPercent() { return 0; }
 
@@ -214,6 +214,7 @@ public:
   int getMinMaxReductionCost(Type *Ty, Type *CondTy,
                              bool IsPairwiseForm,
                              bool IsUnsigned);
+  unsigned getUserCost(const User *U, ArrayRef<const Value *> Operands);
 };
 
 class R600TTIImpl final : public BasicTTIImplBase<R600TTIImpl> {

diff --git a/test/Analysis/CostModel/AMDGPU/extractelement.ll b/test/Analysis/CostModel/AMDGPU/extractelement.ll
index 889cc8bb8ac8b9a8f381142312325a8d2ae45795..67ce8ffba936e05d3d7cf2b18de780b2a3f1087a 100644 (file)
--- a/test/Analysis/CostModel/AMDGPU/extractelement.ll
+++ b/test/Analysis/CostModel/AMDGPU/extractelement.ll
@@ -1,6 +1,10 @@
 ; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa %s | FileCheck -check-prefixes=GCN,CI %s
-; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=fiji %s | FileCheck -check-prefixes=GCN,VI %s
-; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900 %s | FileCheck -check-prefixes=GCN,GFX9 %s
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=fiji %s | FileCheck -check-prefixes=GCN,GFX89 %s
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900 %s | FileCheck -check-prefixes=GCN,GFX89 %s
+; RUN: opt -cost-model -cost-kind=code-size -analyze -mtriple=amdgcn-unknown-amdhsa %s | FileCheck -check-prefixes=GCN,CI %s
+; RUN: opt -cost-model -cost-kind=code-size -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=fiji %s | FileCheck -check-prefixes=GCN,GFX89 %s
+; RUN: opt -cost-model -cost-kind=code-size -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900 %s | FileCheck -check-prefixes=GCN,GFX89 %s
+
 
 ; GCN: 'extractelement_v2i32'
 ; GCN: estimated cost of 0 for {{.*}} extractelement <2 x i32>
@@ -113,8 +117,7 @@ define amdgpu_kernel void @extractelement_v4i8(i8 addrspace(1)* %out, <4 x i8> a
 
 ; GCN: 'extractelement_0_v2i16':
 ; CI: estimated cost of 1 for {{.*}} extractelement <2 x i16> %vec, i16 0
-; VI: estimated cost of 0 for {{.*}} extractelement <2 x i16>
-; GFX9: estimated cost of 0 for {{.*}} extractelement <2 x i16>
+; GFX89: estimated cost of 0 for {{.*}} extractelement <2 x i16>
 define amdgpu_kernel void @extractelement_0_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 0

diff --git a/test/Analysis/CostModel/AMDGPU/fabs.ll b/test/Analysis/CostModel/AMDGPU/fabs.ll
index 51e65fe914c0e39f0719ce6b473b1735e6777538..de5381c2102ae1d71bef9f486613f54be76d28ee 100644 (file)
--- a/test/Analysis/CostModel/AMDGPU/fabs.ll
+++ b/test/Analysis/CostModel/AMDGPU/fabs.ll
@@ -1,6 +1,7 @@
 ; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa < %s | FileCheck %s
+; RUN: opt -cost-model -cost-kind=code-size -analyze -mtriple=amdgcn-unknown-amdhsa < %s | FileCheck %s
 
-; CHECK: 'fabs_f32'
+; CHECK-LABEL: 'fabs_f32'
 ; CHECK: estimated cost of 0 for {{.*}} call float @llvm.fabs.f32
 define amdgpu_kernel void @fabs_f32(float addrspace(1)* %out, float addrspace(1)* %vaddr) #0 {
   %vec = load float, float addrspace(1)* %vaddr
@@ -9,7 +10,7 @@ define amdgpu_kernel void @fabs_f32(float addrspace(1)* %out, float addrspace(1)
   ret void
 }
 
-; CHECK: 'fabs_v2f32'
+; CHECK-LABEL: 'fabs_v2f32'
 ; CHECK: estimated cost of 0 for {{.*}} call <2 x float> @llvm.fabs.v2f32
 define amdgpu_kernel void @fabs_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)* %vaddr) #0 {
   %vec = load <2 x float>, <2 x float> addrspace(1)* %vaddr
@@ -18,7 +19,7 @@ define amdgpu_kernel void @fabs_v2f32(<2 x float> addrspace(1)* %out, <2 x float
   ret void
 }
 
-; CHECK: 'fabs_v3f32'
+; CHECK-LABEL: 'fabs_v3f32'
 ; CHECK: estimated cost of 0 for {{.*}} call <3 x float> @llvm.fabs.v3f32
 define amdgpu_kernel void @fabs_v3f32(<3 x float> addrspace(1)* %out, <3 x float> addrspace(1)* %vaddr) #0 {
   %vec = load <3 x float>, <3 x float> addrspace(1)* %vaddr
@@ -27,7 +28,7 @@ define amdgpu_kernel void @fabs_v3f32(<3 x float> addrspace(1)* %out, <3 x float
   ret void
 }
 
-; CHECK: 'fabs_v5f32'
+; CHECK-LABEL: 'fabs_v5f32'
 ; CHECK: estimated cost of 0 for {{.*}} call <5 x float> @llvm.fabs.v5f32
 define amdgpu_kernel void @fabs_v5f32(<5 x float> addrspace(1)* %out, <5 x float> addrspace(1)* %vaddr) #0 {
   %vec = load <5 x float>, <5 x float> addrspace(1)* %vaddr
@@ -36,7 +37,7 @@ define amdgpu_kernel void @fabs_v5f32(<5 x float> addrspace(1)* %out, <5 x float
   ret void
 }
 
-; CHECK: 'fabs_f64'
+; CHECK-LABEL: 'fabs_f64'
 ; CHECK: estimated cost of 0 for {{.*}} call double @llvm.fabs.f64
 define amdgpu_kernel void @fabs_f64(double addrspace(1)* %out, double addrspace(1)* %vaddr) #0 {
   %vec = load double, double addrspace(1)* %vaddr
@@ -45,7 +46,7 @@ define amdgpu_kernel void @fabs_f64(double addrspace(1)* %out, double addrspace(
   ret void
 }
 
-; CHECK: 'fabs_v2f64'
+; CHECK-LABEL: 'fabs_v2f64'
 ; CHECK: estimated cost of 0 for {{.*}} call <2 x double> @llvm.fabs.v2f64
 define amdgpu_kernel void @fabs_v2f64(<2 x double> addrspace(1)* %out, <2 x double> addrspace(1)* %vaddr) #0 {
   %vec = load <2 x double>, <2 x double> addrspace(1)* %vaddr
@@ -54,7 +55,7 @@ define amdgpu_kernel void @fabs_v2f64(<2 x double> addrspace(1)* %out, <2 x doub
   ret void
 }
 
-; CHECK: 'fabs_v3f64'
+; CHECK-LABEL: 'fabs_v3f64'
 ; CHECK: estimated cost of 0 for {{.*}} call <3 x double> @llvm.fabs.v3f64
 define amdgpu_kernel void @fabs_v3f64(<3 x double> addrspace(1)* %out, <3 x double> addrspace(1)* %vaddr) #0 {
   %vec = load <3 x double>, <3 x double> addrspace(1)* %vaddr
@@ -63,7 +64,7 @@ define amdgpu_kernel void @fabs_v3f64(<3 x double> addrspace(1)* %out, <3 x doub
   ret void
 }
 
-; CHECK: 'fabs_f16'
+; CHECK-LABEL: 'fabs_f16'
 ; CHECK: estimated cost of 0 for {{.*}} call half @llvm.fabs.f16
 define amdgpu_kernel void @fabs_f16(half addrspace(1)* %out, half addrspace(1)* %vaddr) #0 {
   %vec = load half, half addrspace(1)* %vaddr
@@ -72,7 +73,7 @@ define amdgpu_kernel void @fabs_f16(half addrspace(1)* %out, half addrspace(1)*
   ret void
 }
 
-; CHECK: 'fabs_v2f16'
+; CHECK-LABEL: 'fabs_v2f16'
 ; CHECK: estimated cost of 0 for {{.*}} call <2 x half> @llvm.fabs.v2f16
 define amdgpu_kernel void @fabs_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %vaddr) #0 {
   %vec = load <2 x half>, <2 x half> addrspace(1)* %vaddr
@@ -81,7 +82,7 @@ define amdgpu_kernel void @fabs_v2f16(<2 x half> addrspace(1)* %out, <2 x half>
   ret void
 }
 
-; CHECK: 'fabs_v3f16'
+; CHECK-LABEL: 'fabs_v3f16'
 ; CHECK: estimated cost of 0 for {{.*}} call <3 x half> @llvm.fabs.v3f16
 define amdgpu_kernel void @fabs_v3f16(<3 x half> addrspace(1)* %out, <3 x half> addrspace(1)* %vaddr) #0 {
   %vec = load <3 x half>, <3 x half> addrspace(1)* %vaddr

diff --git a/test/Analysis/CostModel/AMDGPU/insertelement.ll b/test/Analysis/CostModel/AMDGPU/insertelement.ll
index 67ab2607acd5fd6aa7f28d4f65368bbce38c421b..7bd86db270aaa7a958489af6ac9a45653d4fa028 100644 (file)
--- a/test/Analysis/CostModel/AMDGPU/insertelement.ll
+++ b/test/Analysis/CostModel/AMDGPU/insertelement.ll
@@ -1,6 +1,9 @@
 ; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa %s | FileCheck -check-prefixes=GCN,CI %s
-; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=fiji %s | FileCheck -check-prefixes=GCN,VI %s
-; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900 %s | FileCheck -check-prefixes=GCN,GFX9 %s
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=fiji %s | FileCheck -check-prefixes=GCN,GFX89 %s
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900 %s | FileCheck -check-prefixes=GCN,GFX89 %s
+; RUN: opt -cost-model -cost-kind=code-size -analyze -mtriple=amdgcn-unknown-amdhsa %s | FileCheck -check-prefixes=GCN,CI %s
+; RUN: opt -cost-model -cost-kind=code-size -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=fiji %s | FileCheck -check-prefixes=GCN,GFX89 %s
+; RUN: opt -cost-model -cost-kind=code-size -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900 %s | FileCheck -check-prefixes=GCN,GFX89 %s
 
 ; GCN-LABEL: 'insertelement_v2i32'
 ; GCN: estimated cost of 0 for {{.*}} insertelement <2 x i32>
@@ -22,8 +25,7 @@ define amdgpu_kernel void @insertelement_v2i64(<2 x i64> addrspace(1)* %out, <2
 
 ; GCN-LABEL: 'insertelement_0_v2i16'
 ; CI: estimated cost of 1 for {{.*}} insertelement <2 x i16>
-; VI: estimated cost of 0 for {{.*}} insertelement <2 x i16>
-; GFX9: estimated cost of 0 for {{.*}} insertelement <2 x i16>
+; GFX89: estimated cost of 0 for {{.*}} insertelement <2 x i16>
 define amdgpu_kernel void @insertelement_0_v2i16(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(1)* %vaddr) {
   %vec = load <2 x i16>, <2 x i16> addrspace(1)* %vaddr
   %insert = insertelement <2 x i16> %vec, i16 123, i16 0