From: Arnaud A. de Grandmaison Date: Tue, 4 Nov 2014 20:51:29 +0000 (+0000) Subject: [PBQP] Callee saved regs should have a higher cost than scratch regs X-Git-Url: http://demsky.eecs.uci.edu/git/?a=commitdiff_plain;h=e1bec7513461e3710f661b73ba11a76eca032ffc;p=oota-llvm.git [PBQP] Callee saved regs should have a higher cost than scratch regs Registers are not all equal. Some are not allocatable (infinite cost), some have to be preserved but can be used, and some others are just free to use. Ensure there is a cost hierarchy reflecting this fact, so that the allocator will favor scratch registers over callee-saved registers. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@221293 91177308-0d34-0410-b5e6-96231b3b80d8 --- diff --git a/lib/CodeGen/RegAllocPBQP.cpp b/lib/CodeGen/RegAllocPBQP.cpp index fa45dd23654..eb7e563352a 100644 --- a/lib/CodeGen/RegAllocPBQP.cpp +++ b/lib/CodeGen/RegAllocPBQP.cpp @@ -479,6 +479,15 @@ void RegAllocPBQP::findVRegIntervalsToAlloc(const MachineFunction &MF, } } +static bool isACalleeSavedRegister(unsigned reg, const TargetRegisterInfo &TRI, + const MachineFunction &MF) { + const MCPhysReg *CSR = TRI.getCalleeSavedRegs(&MF); + for (unsigned i = 0; CSR[i] != 0; ++i) + if (TRI.regsOverlap(reg, CSR[i])) + return true; + return false; +} + void RegAllocPBQP::initializeGraph(PBQPRAGraph &G) { MachineFunction &MF = G.getMetadata().MF; @@ -523,6 +532,13 @@ void RegAllocPBQP::initializeGraph(PBQPRAGraph &G) { } PBQPRAGraph::RawVector NodeCosts(VRegAllowed.size() + 1, 0); + + // Tweak cost of callee saved registers, as using then force spilling and + // restoring them. This would only happen in the prologue / epilogue though. + for (unsigned i = 0; i != VRegAllowed.size(); ++i) + if (isACalleeSavedRegister(VRegAllowed[i], TRI, MF)) + NodeCosts[1 + i] += 1.0; + PBQPRAGraph::NodeId NId = G.addNode(std::move(NodeCosts)); G.getNodeMetadata(NId).setVReg(VReg); G.getNodeMetadata(NId).setAllowedRegs( diff --git a/test/CodeGen/AArch64/PBQP-csr.ll b/test/CodeGen/AArch64/PBQP-csr.ll new file mode 100644 index 00000000000..64335ae353a --- /dev/null +++ b/test/CodeGen/AArch64/PBQP-csr.ll @@ -0,0 +1,91 @@ +; RUN: llc < %s -verify-machineinstrs -mtriple=aarch64-none-linux-gnu -mcpu=cortex-a57 -mattr=+neon -fp-contract=fast -regalloc=pbqp -pbqp-coalescing | FileCheck %s + +%pl = type { i32, i32, i32, i32, %p*, %l*, double* } +%p = type { i32, %ca*, [27 x %ca*], %v*, %v*, %v*, i32 } +%ca = type { %v, float, i32 } +%v = type { double, double, double } +%l = type opaque +%rs = type { i32, i32, i32, i32, %v*, %v*, [21 x double], %v, %v, %v, double, double, double } + +;CHECK-LABEL: test_csr +define void @test_csr(%pl* nocapture readnone %this, %rs* nocapture %r) align 2 { +;CHECK-NOT: stp {{d[0-9]+}}, {{d[0-9]+}} +entry: + %x.i = getelementptr inbounds %rs* %r, i64 0, i32 7, i32 0 + %y.i = getelementptr inbounds %rs* %r, i64 0, i32 7, i32 1 + %z.i = getelementptr inbounds %rs* %r, i64 0, i32 7, i32 2 + %x.i61 = getelementptr inbounds %rs* %r, i64 0, i32 8, i32 0 + %y.i62 = getelementptr inbounds %rs* %r, i64 0, i32 8, i32 1 + %z.i63 = getelementptr inbounds %rs* %r, i64 0, i32 8, i32 2 + %x.i58 = getelementptr inbounds %rs* %r, i64 0, i32 9, i32 0 + %y.i59 = getelementptr inbounds %rs* %r, i64 0, i32 9, i32 1 + %z.i60 = getelementptr inbounds %rs* %r, i64 0, i32 9, i32 2 + %na = getelementptr inbounds %rs* %r, i64 0, i32 0 + %0 = bitcast double* %x.i to i8* + call void @llvm.memset.p0i8.i64(i8* %0, i8 0, i64 72, i32 8, i1 false) + %1 = load i32* %na, align 4 + %cmp70 = icmp sgt i32 %1, 0 + br i1 %cmp70, label %for.body.lr.ph, label %for.end + +for.body.lr.ph: ; preds = %entry + %fn = getelementptr inbounds %rs* %r, i64 0, i32 4 + %2 = load %v** %fn, align 8 + %fs = getelementptr inbounds %rs* %r, i64 0, i32 5 + %3 = load %v** %fs, align 8 + %4 = sext i32 %1 to i64 + br label %for.body + +for.body: ; preds = %for.body.lr.ph, %for.body + %5 = phi double [ 0.000000e+00, %for.body.lr.ph ], [ %add6.i, %for.body ] + %indvars.iv = phi i64 [ 0, %for.body.lr.ph ], [ %indvars.iv.next, %for.body ] + %6 = phi <2 x double> [ zeroinitializer, %for.body.lr.ph ], [ %17, %for.body ] + %7 = phi <2 x double> [ zeroinitializer, %for.body.lr.ph ], [ %22, %for.body ] + %8 = phi <2 x double> [ zeroinitializer, %for.body.lr.ph ], [ %26, %for.body ] + %9 = phi <2 x double> [ zeroinitializer, %for.body.lr.ph ], [ %28, %for.body ] + %x.i54 = getelementptr inbounds %v* %2, i64 %indvars.iv, i32 0 + %x1.i = getelementptr inbounds %v* %3, i64 %indvars.iv, i32 0 + %y.i56 = getelementptr inbounds %v* %2, i64 %indvars.iv, i32 1 + %10 = bitcast double* %x.i54 to <2 x double>* + %11 = load <2 x double>* %10, align 8 + %y2.i = getelementptr inbounds %v* %3, i64 %indvars.iv, i32 1 + %12 = bitcast double* %x1.i to <2 x double>* + %13 = load <2 x double>* %12, align 8 + %14 = fadd fast <2 x double> %13, %11 + %z.i57 = getelementptr inbounds %v* %2, i64 %indvars.iv, i32 2 + %15 = load double* %z.i57, align 8 + %z4.i = getelementptr inbounds %v* %3, i64 %indvars.iv, i32 2 + %16 = load double* %z4.i, align 8 + %add5.i = fadd fast double %16, %15 + %17 = fadd fast <2 x double> %6, %11 + %18 = bitcast double* %x.i to <2 x double>* + store <2 x double> %17, <2 x double>* %18, align 8 + %19 = load double* %x1.i, align 8 + %20 = insertelement <2 x double> undef, double %15, i32 0 + %21 = insertelement <2 x double> %20, double %19, i32 1 + %22 = fadd fast <2 x double> %7, %21 + %23 = bitcast double* %z.i to <2 x double>* + store <2 x double> %22, <2 x double>* %23, align 8 + %24 = bitcast double* %y2.i to <2 x double>* + %25 = load <2 x double>* %24, align 8 + %26 = fadd fast <2 x double> %8, %25 + %27 = bitcast double* %y.i62 to <2 x double>* + store <2 x double> %26, <2 x double>* %27, align 8 + %28 = fadd fast <2 x double> %14, %9 + %29 = bitcast double* %x.i58 to <2 x double>* + store <2 x double> %28, <2 x double>* %29, align 8 + %add6.i = fadd fast double %add5.i, %5 + store double %add6.i, double* %z.i60, align 8 + %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 + %cmp = icmp slt i64 %indvars.iv.next, %4 + br i1 %cmp, label %for.body, label %for.end.loopexit + +for.end.loopexit: ; preds = %for.body + br label %for.end + +for.end: ; preds = %for.end.loopexit, %entry + ret void +} + +; Function Attrs: nounwind +declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i32, i1) +