[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 fa45dd236545865b70b87403f747a8c0913e40c1..eb7e563352ac807ae9e84070e4dd69c9fb7bd082 100644 (file)
--- 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 (file)
index 0000000..64335ae
--- /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)
+