[llvm] r180959 - R600: If previous bundle is dot4, PV valid chan is always X
Vincent Lejeune
vljn at ovi.com
Thu May 2 14:52:55 PDT 2013
Author: vljn
Date: Thu May 2 16:52:55 2013
New Revision: 180959
URL: http://llvm.org/viewvc/llvm-project?rev=180959&view=rev
Log:
R600: If previous bundle is dot4, PV valid chan is always X
Added:
llvm/trunk/test/CodeGen/R600/pv.ll
Modified:
llvm/trunk/lib/Target/R600/R600Packetizer.cpp
Modified: llvm/trunk/lib/Target/R600/R600Packetizer.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/R600Packetizer.cpp?rev=180959&r1=180958&r2=180959&view=diff
==============================================================================
--- llvm/trunk/lib/Target/R600/R600Packetizer.cpp (original)
+++ llvm/trunk/lib/Target/R600/R600Packetizer.cpp Thu May 2 16:52:55 2013
@@ -73,24 +73,56 @@ private:
return TRI.getHWRegChan(MI->getOperand(0).getReg());
}
- std::vector<unsigned> getPreviousVector(MachineBasicBlock::iterator I) const {
- std::vector<unsigned> Result;
+ /// \returns register to PV chan mapping for bundle/single instructions that
+ /// immediatly precedes I.
+ DenseMap<unsigned, unsigned> getPreviousVector(MachineBasicBlock::iterator I)
+ const {
+ DenseMap<unsigned, unsigned> Result;
I--;
if (!TII->isALUInstr(I->getOpcode()) && !I->isBundle())
return Result;
MachineBasicBlock::instr_iterator BI = I.getInstrIterator();
if (I->isBundle())
BI++;
- while (BI->isBundledWithPred() && !TII->isPredicated(BI)) {
+ do {
+ if (TII->isPredicated(BI))
+ continue;
+ if (TII->isTransOnly(BI))
+ continue;
int OperandIdx = TII->getOperandIdx(BI->getOpcode(), R600Operands::WRITE);
- if (OperandIdx > -1 && BI->getOperand(OperandIdx).getImm())
- Result.push_back(BI->getOperand(0).getReg());
- BI++;
- }
+ if (OperandIdx < 0)
+ continue;
+ if (BI->getOperand(OperandIdx).getImm() == 0)
+ continue;
+ unsigned Dst = BI->getOperand(0).getReg();
+ if (BI->getOpcode() == AMDGPU::DOT4_r600_real) {
+ Result[Dst] = AMDGPU::PV_X;
+ continue;
+ }
+ unsigned PVReg = 0;
+ switch (TRI.getHWRegChan(Dst)) {
+ case 0:
+ PVReg = AMDGPU::PV_X;
+ break;
+ case 1:
+ PVReg = AMDGPU::PV_Y;
+ break;
+ case 2:
+ PVReg = AMDGPU::PV_Z;
+ break;
+ case 3:
+ PVReg = AMDGPU::PV_W;
+ break;
+ default:
+ llvm_unreachable("Invalid Chan");
+ }
+ Result[Dst] = PVReg;
+ } while ((++BI)->isBundledWithPred());
return Result;
}
- void substitutePV(MachineInstr *MI, const std::vector<unsigned> &PV) const {
+ void substitutePV(MachineInstr *MI, const DenseMap<unsigned, unsigned> &PVs)
+ const {
R600Operands::Ops Ops[] = {
R600Operands::SRC0,
R600Operands::SRC1,
@@ -101,30 +133,9 @@ private:
if (OperandIdx < 0)
continue;
unsigned Src = MI->getOperand(OperandIdx).getReg();
- for (unsigned j = 0, e = PV.size(); j < e; j++) {
- if (Src == PV[j]) {
- unsigned Chan = TRI.getHWRegChan(Src);
- unsigned PVReg;
- switch (Chan) {
- case 0:
- PVReg = AMDGPU::PV_X;
- break;
- case 1:
- PVReg = AMDGPU::PV_Y;
- break;
- case 2:
- PVReg = AMDGPU::PV_Z;
- break;
- case 3:
- PVReg = AMDGPU::PV_W;
- break;
- default:
- llvm_unreachable("Invalid Chan");
- }
- MI->getOperand(OperandIdx).setReg(PVReg);
- break;
- }
- }
+ const DenseMap<unsigned, unsigned>::const_iterator It = PVs.find(Src);
+ if (It != PVs.end())
+ MI->getOperand(OperandIdx).setReg(It->second);
}
}
public:
@@ -209,7 +220,8 @@ public:
}
dbgs() << "because of Consts read limitations\n";
});
- const std::vector<unsigned> &PV = getPreviousVector(MI);
+ const DenseMap<unsigned, unsigned> &PV =
+ getPreviousVector(CurrentPacketMIs.front());
bool FitsReadPortLimits = fitsReadPortLimitation(CurrentPacketMIs, PV);
DEBUG(
if (!FitsReadPortLimits) {
@@ -236,7 +248,8 @@ public:
}
private:
std::vector<std::pair<int, unsigned> >
- ExtractSrcs(const MachineInstr *MI, const std::vector<unsigned> &PV) const {
+ ExtractSrcs(const MachineInstr *MI, const DenseMap<unsigned, unsigned> &PV)
+ const {
R600Operands::Ops Ops[] = {
R600Operands::SRC0,
R600Operands::SRC1,
@@ -250,7 +263,7 @@ private:
continue;
}
unsigned Src = MI->getOperand(OperandIdx).getReg();
- if (std::find(PV.begin(), PV.end(), Src) != PV.end()) {
+ if (PV.find(Src) != PV.end()) {
Result.push_back(std::pair<int, unsigned>(-1,0));
continue;
}
@@ -294,7 +307,7 @@ private:
bool isLegal(const std::vector<MachineInstr *> &IG,
const std::vector<BankSwizzle> &Swz,
- const std::vector<unsigned> &PV) const {
+ const DenseMap<unsigned, unsigned> &PV) const {
assert (Swz.size() == IG.size());
int Vector[4][3];
memset(Vector, -1, sizeof(Vector));
@@ -316,7 +329,7 @@ private:
bool recursiveFitsFPLimitation(
std::vector<MachineInstr *> IG,
- const std::vector<unsigned> &PV,
+ const DenseMap<unsigned, unsigned> &PV,
std::vector<BankSwizzle> &SwzCandidate,
std::vector<MachineInstr *> CurrentlyChecked)
const {
@@ -345,7 +358,7 @@ private:
bool fitsReadPortLimitation(
std::vector<MachineInstr *> IG,
- const std::vector<unsigned> &PV)
+ const DenseMap<unsigned, unsigned> &PV)
const {
//Todo : support shared src0 - src1 operand
std::vector<BankSwizzle> SwzCandidate;
Added: llvm/trunk/test/CodeGen/R600/pv.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/R600/pv.ll?rev=180959&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/R600/pv.ll (added)
+++ llvm/trunk/test/CodeGen/R600/pv.ll Thu May 2 16:52:55 2013
@@ -0,0 +1,244 @@
+; RUN: llc < %s -march=r600 | FileCheck %s
+
+;CHECK: DOT4 * T{{[0-9]\.W}} (MASKED)
+;CHECK-NEXT: CNDGE T{{[0-9].[XYZW]}}, PV.x
+
+define void @main() #0 {
+main_body:
+ %0 = call float @llvm.R600.load.input(i32 4)
+ %1 = call float @llvm.R600.load.input(i32 5)
+ %2 = call float @llvm.R600.load.input(i32 6)
+ %3 = call float @llvm.R600.load.input(i32 7)
+ %4 = call float @llvm.R600.load.input(i32 8)
+ %5 = call float @llvm.R600.load.input(i32 9)
+ %6 = call float @llvm.R600.load.input(i32 10)
+ %7 = call float @llvm.R600.load.input(i32 11)
+ %8 = call float @llvm.R600.load.input(i32 12)
+ %9 = call float @llvm.R600.load.input(i32 13)
+ %10 = call float @llvm.R600.load.input(i32 14)
+ %11 = call float @llvm.R600.load.input(i32 15)
+ %12 = call float @llvm.R600.load.input(i32 16)
+ %13 = call float @llvm.R600.load.input(i32 17)
+ %14 = call float @llvm.R600.load.input(i32 18)
+ %15 = call float @llvm.R600.load.input(i32 19)
+ %16 = call float @llvm.R600.load.input(i32 20)
+ %17 = call float @llvm.R600.load.input(i32 21)
+ %18 = call float @llvm.R600.load.input(i32 22)
+ %19 = call float @llvm.R600.load.input(i32 23)
+ %20 = call float @llvm.R600.load.input(i32 24)
+ %21 = call float @llvm.R600.load.input(i32 25)
+ %22 = call float @llvm.R600.load.input(i32 26)
+ %23 = call float @llvm.R600.load.input(i32 27)
+ %24 = call float @llvm.R600.load.input(i32 28)
+ %25 = call float @llvm.R600.load.input(i32 29)
+ %26 = call float @llvm.R600.load.input(i32 30)
+ %27 = call float @llvm.R600.load.input(i32 31)
+ %28 = load <4 x float> addrspace(8)* null
+ %29 = extractelement <4 x float> %28, i32 0
+ %30 = fmul float %0, %29
+ %31 = load <4 x float> addrspace(8)* null
+ %32 = extractelement <4 x float> %31, i32 1
+ %33 = fmul float %0, %32
+ %34 = load <4 x float> addrspace(8)* null
+ %35 = extractelement <4 x float> %34, i32 2
+ %36 = fmul float %0, %35
+ %37 = load <4 x float> addrspace(8)* null
+ %38 = extractelement <4 x float> %37, i32 3
+ %39 = fmul float %0, %38
+ %40 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 1)
+ %41 = extractelement <4 x float> %40, i32 0
+ %42 = fmul float %1, %41
+ %43 = fadd float %42, %30
+ %44 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 1)
+ %45 = extractelement <4 x float> %44, i32 1
+ %46 = fmul float %1, %45
+ %47 = fadd float %46, %33
+ %48 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 1)
+ %49 = extractelement <4 x float> %48, i32 2
+ %50 = fmul float %1, %49
+ %51 = fadd float %50, %36
+ %52 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 1)
+ %53 = extractelement <4 x float> %52, i32 3
+ %54 = fmul float %1, %53
+ %55 = fadd float %54, %39
+ %56 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 2)
+ %57 = extractelement <4 x float> %56, i32 0
+ %58 = fmul float %2, %57
+ %59 = fadd float %58, %43
+ %60 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 2)
+ %61 = extractelement <4 x float> %60, i32 1
+ %62 = fmul float %2, %61
+ %63 = fadd float %62, %47
+ %64 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 2)
+ %65 = extractelement <4 x float> %64, i32 2
+ %66 = fmul float %2, %65
+ %67 = fadd float %66, %51
+ %68 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 2)
+ %69 = extractelement <4 x float> %68, i32 3
+ %70 = fmul float %2, %69
+ %71 = fadd float %70, %55
+ %72 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 3)
+ %73 = extractelement <4 x float> %72, i32 0
+ %74 = fmul float %3, %73
+ %75 = fadd float %74, %59
+ %76 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 3)
+ %77 = extractelement <4 x float> %76, i32 1
+ %78 = fmul float %3, %77
+ %79 = fadd float %78, %63
+ %80 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 3)
+ %81 = extractelement <4 x float> %80, i32 2
+ %82 = fmul float %3, %81
+ %83 = fadd float %82, %67
+ %84 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 3)
+ %85 = extractelement <4 x float> %84, i32 3
+ %86 = fmul float %3, %85
+ %87 = fadd float %86, %71
+ %88 = insertelement <4 x float> undef, float %4, i32 0
+ %89 = insertelement <4 x float> %88, float %5, i32 1
+ %90 = insertelement <4 x float> %89, float %6, i32 2
+ %91 = insertelement <4 x float> %90, float 0.000000e+00, i32 3
+ %92 = insertelement <4 x float> undef, float %4, i32 0
+ %93 = insertelement <4 x float> %92, float %5, i32 1
+ %94 = insertelement <4 x float> %93, float %6, i32 2
+ %95 = insertelement <4 x float> %94, float 0.000000e+00, i32 3
+ %96 = call float @llvm.AMDGPU.dp4(<4 x float> %91, <4 x float> %95)
+ %97 = call float @fabs(float %96)
+ %98 = call float @llvm.AMDGPU.rsq(float %97)
+ %99 = fmul float %4, %98
+ %100 = fmul float %5, %98
+ %101 = fmul float %6, %98
+ %102 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 4)
+ %103 = extractelement <4 x float> %102, i32 0
+ %104 = fmul float %103, %8
+ %105 = fadd float %104, %20
+ %106 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 4)
+ %107 = extractelement <4 x float> %106, i32 1
+ %108 = fmul float %107, %9
+ %109 = fadd float %108, %21
+ %110 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 4)
+ %111 = extractelement <4 x float> %110, i32 2
+ %112 = fmul float %111, %10
+ %113 = fadd float %112, %22
+ %114 = call float @llvm.AMDIL.clamp.(float %105, float 0.000000e+00, float 1.000000e+00)
+ %115 = call float @llvm.AMDIL.clamp.(float %109, float 0.000000e+00, float 1.000000e+00)
+ %116 = call float @llvm.AMDIL.clamp.(float %113, float 0.000000e+00, float 1.000000e+00)
+ %117 = call float @llvm.AMDIL.clamp.(float %15, float 0.000000e+00, float 1.000000e+00)
+ %118 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 5)
+ %119 = extractelement <4 x float> %118, i32 0
+ %120 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 5)
+ %121 = extractelement <4 x float> %120, i32 1
+ %122 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 5)
+ %123 = extractelement <4 x float> %122, i32 2
+ %124 = insertelement <4 x float> undef, float %99, i32 0
+ %125 = insertelement <4 x float> %124, float %100, i32 1
+ %126 = insertelement <4 x float> %125, float %101, i32 2
+ %127 = insertelement <4 x float> %126, float 0.000000e+00, i32 3
+ %128 = insertelement <4 x float> undef, float %119, i32 0
+ %129 = insertelement <4 x float> %128, float %121, i32 1
+ %130 = insertelement <4 x float> %129, float %123, i32 2
+ %131 = insertelement <4 x float> %130, float 0.000000e+00, i32 3
+ %132 = call float @llvm.AMDGPU.dp4(<4 x float> %127, <4 x float> %131)
+ %133 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 7)
+ %134 = extractelement <4 x float> %133, i32 0
+ %135 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 7)
+ %136 = extractelement <4 x float> %135, i32 1
+ %137 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 7)
+ %138 = extractelement <4 x float> %137, i32 2
+ %139 = insertelement <4 x float> undef, float %99, i32 0
+ %140 = insertelement <4 x float> %139, float %100, i32 1
+ %141 = insertelement <4 x float> %140, float %101, i32 2
+ %142 = insertelement <4 x float> %141, float 0.000000e+00, i32 3
+ %143 = insertelement <4 x float> undef, float %134, i32 0
+ %144 = insertelement <4 x float> %143, float %136, i32 1
+ %145 = insertelement <4 x float> %144, float %138, i32 2
+ %146 = insertelement <4 x float> %145, float 0.000000e+00, i32 3
+ %147 = call float @llvm.AMDGPU.dp4(<4 x float> %142, <4 x float> %146)
+ %148 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 8)
+ %149 = extractelement <4 x float> %148, i32 0
+ %150 = fmul float %149, %8
+ %151 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 8)
+ %152 = extractelement <4 x float> %151, i32 1
+ %153 = fmul float %152, %9
+ %154 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 8)
+ %155 = extractelement <4 x float> %154, i32 2
+ %156 = fmul float %155, %10
+ %157 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 9)
+ %158 = extractelement <4 x float> %157, i32 0
+ %159 = fmul float %158, %12
+ %160 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 9)
+ %161 = extractelement <4 x float> %160, i32 1
+ %162 = fmul float %161, %13
+ %163 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 9)
+ %164 = extractelement <4 x float> %163, i32 2
+ %165 = fmul float %164, %14
+ %166 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 10)
+ %167 = extractelement <4 x float> %166, i32 0
+ %168 = fmul float %167, %16
+ %169 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 10)
+ %170 = extractelement <4 x float> %169, i32 1
+ %171 = fmul float %170, %17
+ %172 = load <4 x float> addrspace(8)* getelementptr ([1024 x <4 x float>] addrspace(8)* null, i64 0, i32 10)
+ %173 = extractelement <4 x float> %172, i32 2
+ %174 = fmul float %173, %18
+ %175 = fcmp uge float %132, 0.000000e+00
+ %176 = select i1 %175, float %132, float 0.000000e+00
+ %177 = fcmp uge float %147, 0.000000e+00
+ %178 = select i1 %177, float %147, float 0.000000e+00
+ %179 = call float @llvm.pow.f32(float %178, float %24)
+ %180 = fcmp ult float %132, 0.000000e+00
+ %181 = select i1 %180, float 0.000000e+00, float %179
+ %182 = fadd float %150, %105
+ %183 = fadd float %153, %109
+ %184 = fadd float %156, %113
+ %185 = fmul float %176, %159
+ %186 = fadd float %185, %182
+ %187 = fmul float %176, %162
+ %188 = fadd float %187, %183
+ %189 = fmul float %176, %165
+ %190 = fadd float %189, %184
+ %191 = fmul float %181, %168
+ %192 = fadd float %191, %186
+ %193 = fmul float %181, %171
+ %194 = fadd float %193, %188
+ %195 = fmul float %181, %174
+ %196 = fadd float %195, %190
+ %197 = call float @llvm.AMDIL.clamp.(float %192, float 0.000000e+00, float 1.000000e+00)
+ %198 = call float @llvm.AMDIL.clamp.(float %194, float 0.000000e+00, float 1.000000e+00)
+ %199 = call float @llvm.AMDIL.clamp.(float %196, float 0.000000e+00, float 1.000000e+00)
+ %200 = insertelement <4 x float> undef, float %75, i32 0
+ %201 = insertelement <4 x float> %200, float %79, i32 1
+ %202 = insertelement <4 x float> %201, float %83, i32 2
+ %203 = insertelement <4 x float> %202, float %87, i32 3
+ call void @llvm.R600.store.swizzle(<4 x float> %203, i32 60, i32 1)
+ %204 = insertelement <4 x float> undef, float %197, i32 0
+ %205 = insertelement <4 x float> %204, float %198, i32 1
+ %206 = insertelement <4 x float> %205, float %199, i32 2
+ %207 = insertelement <4 x float> %206, float %117, i32 3
+ call void @llvm.R600.store.swizzle(<4 x float> %207, i32 0, i32 2)
+ ret void
+}
+
+; Function Attrs: readnone
+declare float @llvm.R600.load.input(i32) #1
+
+; Function Attrs: readnone
+declare float @llvm.AMDGPU.dp4(<4 x float>, <4 x float>) #1
+
+; Function Attrs: readonly
+declare float @fabs(float) #2
+
+; Function Attrs: readnone
+declare float @llvm.AMDGPU.rsq(float) #1
+
+; Function Attrs: readnone
+declare float @llvm.AMDIL.clamp.(float, float, float) #1
+
+; Function Attrs: nounwind readonly
+declare float @llvm.pow.f32(float, float) #3
+
+declare void @llvm.R600.store.swizzle(<4 x float>, i32, i32)
+
+attributes #0 = { "ShaderType"="1" }
+attributes #1 = { readnone }
+attributes #2 = { readonly }
+attributes #3 = { nounwind readonly }
More information about the llvm-commits
mailing list