Reland: Fix and test inter-procedural register allocation for ARM

Add an explicit construction of the ArrayRef, gcc 5 and earlier don't
seem to select the ArrayRef constructor which takes a C array when the
construction is implicit.

Original commit message:

- Avoid a crash when IPRA calls ARMFrameLowering::determineCalleeSaves
  with a null RegScavenger. Simply not updating the register scavenger
  is fine because IPRA only cares about the SavedRegs vector, the acutal
  code of the function has already been generated at this point.
- Add a new hook to TargetRegisterInfo to get the set of registers which
  can be clobbered inside a call, even if the compiler can see both
  sides, by linker-generated code.

Differential revision: https://reviews.llvm.org/D64908

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@367819 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/CodeGen/TargetRegisterInfo.h b/include/llvm/CodeGen/TargetRegisterInfo.h
index 7fac869b70100875c0f20e4a616196fd201a9f26..11596cb7ee7c45514c0aa3db3d6a35a1275743be 100644 (file)
--- a/include/llvm/CodeGen/TargetRegisterInfo.h
+++ b/include/llvm/CodeGen/TargetRegisterInfo.h
@@ -447,6 +447,14 @@ public:
     llvm_unreachable("target does not provide no preserved mask");
   }
 
+  /// Return a list of all of the registers which are clobbered "inside" a call
+  /// to the given function. For example, these might be needed for PLT
+  /// sequences of long-branch veneers.
+  virtual ArrayRef<MCPhysReg>
+  getIntraCallClobberedRegs(const MachineFunction *MF) const {
+    return {};
+  }
+
   /// Return true if all bits that are set in mask \p mask0 are also set in
   /// \p mask1.
   bool regmaskSubsetEqual(const uint32_t *mask0, const uint32_t *mask1) const;

diff --git a/lib/CodeGen/RegUsageInfoCollector.cpp b/lib/CodeGen/RegUsageInfoCollector.cpp
index c12e51bd3bf8ca3ceb5b63b7d6f608b444a4a71e..757ff0e44953dcaf0366c8a65e57c6c2385aab09 100644 (file)
--- a/lib/CodeGen/RegUsageInfoCollector.cpp
+++ b/lib/CodeGen/RegUsageInfoCollector.cpp
@@ -142,6 +142,13 @@ bool RegUsageInfoCollector::runOnMachineFunction(MachineFunction &MF) {
   auto SetRegAsDefined = [&RegMask] (unsigned Reg) {
     RegMask[Reg / 32] &= ~(1u << Reg % 32);
   };
+
+  // Some targets can clobber registers "inside" a call, typically in
+  // linker-generated code.
+  for (const MCPhysReg Reg : TRI->getIntraCallClobberedRegs(&MF))
+    for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
+      SetRegAsDefined(*AI);
+
   // Scan all the physical registers. When a register is defined in the current
   // function set it and all the aliasing registers as defined in the regmask.
   // FIXME: Rewrite to use regunits.

diff --git a/lib/Target/ARM/ARMBaseRegisterInfo.cpp b/lib/Target/ARM/ARMBaseRegisterInfo.cpp
index e63ea7a444763a2a896587f957a34d66cc853e99..2fcbd7ecf6bb97e45c1a627e9e3d32891a1da401 100644 (file)
--- a/lib/Target/ARM/ARMBaseRegisterInfo.cpp
+++ b/lib/Target/ARM/ARMBaseRegisterInfo.cpp
@@ -174,6 +174,12 @@ ARMBaseRegisterInfo::getThisReturnPreservedMask(const MachineFunction &MF,
                               : CSR_AAPCS_ThisReturn_RegMask;
 }
 
+ArrayRef<MCPhysReg> ARMBaseRegisterInfo::getIntraCallClobberedRegs(
+    const MachineFunction *MF) const {
+  static const MCPhysReg IntraCallClobberedRegs[] = {ARM::R12};
+  return ArrayRef<MCPhysReg>(IntraCallClobberedRegs);
+}
+
 BitVector ARMBaseRegisterInfo::
 getReservedRegs(const MachineFunction &MF) const {
   const ARMSubtarget &STI = MF.getSubtarget<ARMSubtarget>();

diff --git a/lib/Target/ARM/ARMBaseRegisterInfo.h b/lib/Target/ARM/ARMBaseRegisterInfo.h
index 7e2c72b4d712c1bd19c737d39a12ee7248b8c345..37248ac6ecf8cc46faa9a919cb120c4348f84b3f 100644 (file)
--- a/lib/Target/ARM/ARMBaseRegisterInfo.h
+++ b/lib/Target/ARM/ARMBaseRegisterInfo.h
@@ -129,6 +129,9 @@ public:
   const uint32_t *getThisReturnPreservedMask(const MachineFunction &MF,
                                              CallingConv::ID) const;
 
+  ArrayRef<MCPhysReg>
+  getIntraCallClobberedRegs(const MachineFunction *MF) const override;
+
   BitVector getReservedRegs(const MachineFunction &MF) const override;
   bool isAsmClobberable(const MachineFunction &MF,
                        unsigned PhysReg) const override;

diff --git a/lib/Target/ARM/ARMFrameLowering.cpp b/lib/Target/ARM/ARMFrameLowering.cpp
index 7ae0c6ab3cf8ed38856c2466c55052b2a68a5446..d2a5111a4def3337c15408c53ae81c0382c46a0e 100644 (file)
--- a/lib/Target/ARM/ARMFrameLowering.cpp
+++ b/lib/Target/ARM/ARMFrameLowering.cpp
@@ -2080,9 +2080,8 @@ void ARMFrameLowering::determineCalleeSaves(MachineFunction &MF,
             ExtraCSSpill = true;
         }
       }
-      if (!ExtraCSSpill) {
+      if (!ExtraCSSpill && RS) {
         // Reserve a slot closest to SP or frame pointer.
-        assert(RS && "Register scavenging not provided");
         LLVM_DEBUG(dbgs() << "Reserving emergency spill slot\n");
         const TargetRegisterClass &RC = ARM::GPRRegClass;
         unsigned Size = TRI->getSpillSize(RC);

diff --git a/test/CodeGen/ARM/ipra.ll b/test/CodeGen/ARM/ipra.ll
new file mode 100644 (file)
index 0000000..3252c13
--- /dev/null
+++ b/test/CodeGen/ARM/ipra.ll
@@ -0,0 +1,202 @@
+; RUN: llc -mtriple armv7a--none-eabi < %s              | FileCheck %s --check-prefix=CHECK --check-prefix=DISABLED
+; RUN: llc -mtriple armv7a--none-eabi < %s -enable-ipra | FileCheck %s --check-prefix=CHECK --check-prefix=ENABLED
+
+define void @leaf()  {
+entry:
+  tail call void asm sideeffect "", ""()
+  ret void
+}
+
+define void @leaf_r3() {
+entry:
+  tail call void asm sideeffect "", "~{r3}"()
+  ret void
+}
+
+define void @leaf_r4() {
+entry:
+  tail call void asm sideeffect "", "~{r4}"()
+  ret void
+}
+
+define void @leaf_s0() {
+entry:
+  tail call void asm sideeffect "", "~{s0}"()
+  ret void
+}
+
+define void @leaf_d0() {
+entry:
+  tail call void asm sideeffect "", "~{d0}"()
+  ret void
+}
+
+; r3 is normally caller-saved, but with IPRA we can see that it isn't used in
+; the callee, so can leave a live value in it.
+define void @test_r3_presrved() {
+; CHECK-LABEL: test_r3_presrved:
+entry:
+; CHECK: ASM1: r3
+; DISABLED: mov [[TEMP:r[0-9]+]], r3
+; ENABLED-NOT: r3
+; CHECK: bl      leaf
+; DISABLED: mov r3, [[TEMP]]
+; ENABLED-NOT: r3
+; CHECK: ASM2: r3
+  %a = tail call i32 asm sideeffect "// ASM1: $0", "={r3},0"(i32 undef)
+  tail call void @leaf()
+  %b = tail call i32 asm sideeffect "// ASM2: $0", "={r3},0"(i32 %a)
+  ret void
+}
+
+; Same as above, but r3 is clobbered in the callee, so it is clobbered by the
+; call as normal.
+define void @test_r3_clobbered() {
+; CHECK-LABEL: test_r3_clobbered:
+entry:
+; CHECK: ASM1: r3
+; CHECK: mov [[TEMP:r[0-9]+]], r3
+; CHECK: bl      leaf
+; CHECK: mov r3, [[TEMP]]
+; CHECK: ASM2: r3
+  %a = tail call i32 asm sideeffect "// ASM1: $0", "={r3},0"(i32 undef)
+  tail call void @leaf_r3()
+  %b = tail call i32 asm sideeffect "// ASM2: $0", "={r3},0"(i32 %a)
+  ret void
+}
+
+; r4 is a callee-saved register, so IPRA has no effect.
+define void @test_r4_preserved() {
+; CHECK-LABEL: test_r4_preserved:
+entry:
+; CHECK: ASM1: r4
+; CHECK-NOT: r4
+; CHECK: bl      leaf
+; CHECK-NOT: r4
+; CHECK: ASM2: r4
+  %a = tail call i32 asm sideeffect "// ASM1: $0", "={r4},0"(i32 undef)
+  tail call void @leaf()
+  %b = tail call i32 asm sideeffect "// ASM2: $0", "={r4},0"(i32 %a)
+  ret void
+}
+define void @test_r4_clobbered() {
+; CHECK-LABEL: test_r4_clobbered:
+entry:
+; CHECK: ASM1: r4
+; CHECK-NOT: r4
+; CHECK: bl      leaf_r4
+; CHECK-NOT: r4
+; CHECK: ASM2: r4
+  %a = tail call i32 asm sideeffect "// ASM1: $0", "={r4},0"(i32 undef)
+  tail call void @leaf_r4()
+  %b = tail call i32 asm sideeffect "// ASM2: $0", "={r4},0"(i32 %a)
+  ret void
+}
+
+; r12 is the intra-call scratch register, so we have to assume it is clobbered
+; even if we can see that the callee does not touch it.
+define void @test_r12() {
+; CHECK-LABEL: test_r12:
+entry:
+; CHECK: ASM1: r12
+; CHECK: mov [[TEMP:r[0-9]+]], r12
+; CHECK: bl      leaf
+; CHECK: mov r12, [[TEMP]]
+; CHECK: ASM2: r12
+  %a = tail call i32 asm sideeffect "// ASM1: $0", "={r12},0"(i32 undef)
+  tail call void @leaf()
+  %b = tail call i32 asm sideeffect "// ASM2: $0", "={r12},0"(i32 %a)
+  ret void
+}
+
+; s0 and d0 are caller-saved, IPRA allows us to keep them live in the caller if
+; the callee doesn't modify them.
+define void @test_s0_preserved() {
+; CHECK-LABEL: test_s0_preserved:
+entry:
+; CHECK: ASM1: s0
+; DISABLED: vmov.f32 [[TEMP:s[0-9]+]], s0
+; ENABLED-NOT: s0
+; CHECK: bl      leaf
+; DISABLED: vmov.f32 s0, [[TEMP]]
+; ENABLED-NOT: s0
+; CHECK: ASM2: s0
+  %a = tail call float asm sideeffect "// ASM1: $0", "={s0},0"(float undef)
+  tail call void @leaf()
+  %b = tail call float asm sideeffect "// ASM2: $0", "={s0},0"(float %a)
+  ret void
+}
+
+define void @test_s0_clobbered() {
+; CHECK-LABEL: test_s0_clobbered:
+entry:
+; CHECK: ASM1: s0
+; CHECK: vmov.f32 [[TEMP:s[0-9]+]], s0
+; CHECK: bl      leaf_s0
+; CHECK: vmov.f32 s0, [[TEMP]]
+; CHECK: ASM2: s0
+  %a = tail call float asm sideeffect "// ASM1: $0", "={s0},0"(float undef)
+  tail call void @leaf_s0()
+  %b = tail call float asm sideeffect "// ASM2: $0", "={s0},0"(float %a)
+  ret void
+}
+
+define void @test_d0_preserved() {
+; CHECK-LABEL: test_d0_preserved:
+entry:
+; CHECK: ASM1: d0
+; DISABLED: vmov.f64 [[TEMP:d[0-9]+]], d0
+; ENABLED-NOT: d0
+; CHECK: bl      leaf
+; DISABLED: vmov.f64 d0, [[TEMP]]
+; ENABLED-NOT: d0
+; CHECK: ASM2: d0
+  %a = tail call double asm sideeffect "// ASM1: $0", "={d0},0"(double undef)
+  tail call void @leaf()
+  %b = tail call double asm sideeffect "// ASM2: $0", "={d0},0"(double %a)
+  ret void
+}
+
+define void @test_d0_clobbered() {
+; CHECK-LABEL: test_d0_clobbered:
+entry:
+; CHECK: ASM1: d0
+; CHECK: vmov.f64 [[TEMP:d[0-9]+]], d0
+; CHECK: bl      leaf_d0
+; CHECK: vmov.f64 d0, [[TEMP]]
+; CHECK: ASM2: d0
+  %a = tail call double asm sideeffect "// ASM1: $0", "={d0},0"(double undef)
+  tail call void @leaf_d0()
+  %b = tail call double asm sideeffect "// ASM2: $0", "={d0},0"(double %a)
+  ret void
+}
+
+; s0 and d0 overlap, so clobbering one in the callee prevents the other from
+; being kept live across the call.
+define void @test_s0_clobber_d0() {
+; CHECK-LABEL: test_s0_clobber_d0:
+entry:
+; CHECK: ASM1: s0
+; CHECK: vmov.f32 [[TEMP:s[0-9]+]], s0
+; CHECK: bl      leaf_d0
+; CHECK: vmov.f32 s0, [[TEMP]]
+; CHECK: ASM2: s0
+  %a = tail call float asm sideeffect "// ASM1: $0", "={s0},0"(float undef)
+  tail call void @leaf_d0()
+  %b = tail call float asm sideeffect "// ASM2: $0", "={s0},0"(float %a)
+  ret void
+}
+
+define void @test_d0_clobber_s0() {
+; CHECK-LABEL: test_d0_clobber_s0:
+entry:
+; CHECK: ASM1: d0
+; CHECK: vmov.f64 [[TEMP:d[0-9]+]], d0
+; CHECK: bl      leaf_s0
+; CHECK: vmov.f64 d0, [[TEMP]]
+; CHECK: ASM2: d0
+  %a = tail call double asm sideeffect "// ASM1: $0", "={d0},0"(double undef)
+  tail call void @leaf_s0()
+  %b = tail call double asm sideeffect "// ASM2: $0", "={d0},0"(double %a)
+  ret void
+}