[IPRA][ARM] Disable no-CSR optimisation for ARM

This optimisation isn't generally profitable for ARM, because we can
save/restore many registers in the prologue and epilogue using the PUSH
and POP instructions, but mostly use individual LDR/STR instructions for
other spills.

Differential revision: https://reviews.llvm.org/D64910

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@367670 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/CodeGen/TargetFrameLowering.h b/include/llvm/CodeGen/TargetFrameLowering.h
index 878c9ffd2b510f23880b4e887216ef1f0c98a918..284f7ba64dbaa7c0c9c0ad046a66695ca65c60f8 100644 (file)
--- a/include/llvm/CodeGen/TargetFrameLowering.h
+++ b/include/llvm/CodeGen/TargetFrameLowering.h
@@ -378,6 +378,11 @@ public:
     return true;
   }
 
+  /// Check if the no-CSR optimisation is profitable for the given function.
+  virtual bool isProfitableForNoCSROpt(const Function &F) const {
+    return true;
+  }
+
   /// Return initial CFA offset value i.e. the one valid at the beginning of the
   /// function (before any stack operations).
   virtual int getInitialCFAOffset(const MachineFunction &MF) const;

diff --git a/lib/CodeGen/RegUsageInfoCollector.cpp b/lib/CodeGen/RegUsageInfoCollector.cpp
index 64552b58e2ec9c12f4ec6fdf30aa75976e6d7220..757ff0e44953dcaf0366c8a65e57c6c2385aab09 100644 (file)
--- a/lib/CodeGen/RegUsageInfoCollector.cpp
+++ b/lib/CodeGen/RegUsageInfoCollector.cpp
@@ -171,7 +171,8 @@ bool RegUsageInfoCollector::runOnMachineFunction(MachineFunction &MF) {
       SetRegAsDefined(PReg);
   }
 
-  if (TargetFrameLowering::isSafeForNoCSROpt(F)) {
+  if (TargetFrameLowering::isSafeForNoCSROpt(F) &&
+      MF.getSubtarget().getFrameLowering()->isProfitableForNoCSROpt(F)) {
     ++NumCSROpt;
     LLVM_DEBUG(dbgs() << MF.getName()
                       << " function optimized for not having CSR.\n");

diff --git a/lib/CodeGen/TargetFrameLoweringImpl.cpp b/lib/CodeGen/TargetFrameLoweringImpl.cpp
index 9c4483cb240d70d88ed7734e15e1d3c76d016ba6..c5cd87b1481bf2893ff05a763e99ed9f6937024a 100644 (file)
--- a/lib/CodeGen/TargetFrameLoweringImpl.cpp
+++ b/lib/CodeGen/TargetFrameLoweringImpl.cpp
@@ -71,7 +71,9 @@ void TargetFrameLowering::determineCalleeSaves(MachineFunction &MF,
 
   // When interprocedural register allocation is enabled caller saved registers
   // are preferred over callee saved registers.
-  if (MF.getTarget().Options.EnableIPRA && isSafeForNoCSROpt(MF.getFunction()))
+  if (MF.getTarget().Options.EnableIPRA &&
+      isSafeForNoCSROpt(MF.getFunction()) &&
+      isProfitableForNoCSROpt(MF.getFunction()))
     return;
 
   // Get the callee saved register list...

diff --git a/lib/Target/ARM/ARMFrameLowering.h b/lib/Target/ARM/ARMFrameLowering.h
index 7544ca3c38d6ef9492c1b3d225e1b8f9376b5bac..6d8aee597945954e2c37a7d6a1bb901bf348d0e4 100644 (file)
--- a/lib/Target/ARM/ARMFrameLowering.h
+++ b/lib/Target/ARM/ARMFrameLowering.h
@@ -63,6 +63,11 @@ public:
   bool enableShrinkWrapping(const MachineFunction &MF) const override {
     return true;
   }
+  bool isProfitableForNoCSROpt(const Function &F) const override {
+    // The no-CSR optimisation is bad for code size on ARM, because we can save
+    // many registers with a single PUSH/POP pair.
+    return false;
+  }
 
 private:
   void emitPushInst(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,

diff --git a/test/CodeGen/ARM/ipra-no-csr.ll b/test/CodeGen/ARM/ipra-no-csr.ll
new file mode 100644 (file)
index 0000000..8070573
--- /dev/null
+++ b/test/CodeGen/ARM/ipra-no-csr.ll
@@ -0,0 +1,22 @@
+; RUN: llc -mtriple armv7a--none-eabi < %s              | FileCheck %s
+; RUN: llc -mtriple armv7a--none-eabi < %s -enable-ipra | FileCheck %s
+
+; Other targets disable callee-saved registers for internal functions when
+; using IPRA, but that isn't profitable for ARM because the PUSH/POP
+; instructions can more efficiently save registers than using individual
+; LDR/STRs in the caller.
+
+define internal void @callee() norecurse {
+; CHECK-LABEL: callee:
+entry:
+; CHECK: push {r4, lr}
+; CHECK: pop {r4, pc}
+  tail call void asm sideeffect "", "~{r4}"()
+  ret void
+}
+
+define void @caller() {
+entry:
+  call void @callee()
+  ret void
+}