void created this revision.
void added reviewers: nickdesaulniers, danielkiss, MaskRay.
Herald added subscribers: StephenFan, pengfei, hiraditya, kristof.beyls.
Herald added a project: All.
void requested review of this revision.
Herald added projects: clang, LLVM.
Herald added subscribers: llvm-commits, cfe-commits.
Support the "-fzero-call-used-regs" option on AArch64. This involves much less
specialized code than the X86 version. Most of the checks can be done with
TableGen.
Repository:
rG LLVM Github Monorepo
https://reviews.llvm.org/D124836
Files:
clang/include/clang/Driver/Options.td
clang/lib/Driver/ToolChains/Clang.cpp
llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
llvm/lib/Target/AArch64/AArch64FrameLowering.h
llvm/lib/Target/AArch64/AArch64RegisterInfo.td
llvm/lib/Target/X86/X86RegisterInfo.cpp
llvm/test/CodeGen/AArch64/zero-call-used-regs.ll
llvm/utils/TableGen/RegisterInfoEmitter.cpp
Index: llvm/utils/TableGen/RegisterInfoEmitter.cpp
===================================================================
--- llvm/utils/TableGen/RegisterInfoEmitter.cpp
+++ llvm/utils/TableGen/RegisterInfoEmitter.cpp
@@ -1188,6 +1188,8 @@
<< "MCRegister) const override;\n"
<< " bool isFixedRegister(const MachineFunction &, "
<< "MCRegister) const override;\n"
+ << " bool isArgumentRegister(const MachineFunction &, "
+ << "MCRegister) const override;\n"
<< " /// Devirtualized TargetFrameLowering.\n"
<< " static const " << TargetName << "FrameLowering *getFrameLowering(\n"
<< " const MachineFunction &MF);\n"
@@ -1662,6 +1664,20 @@
OS << " false;\n";
OS << "}\n\n";
+ OS << "bool " << ClassName << "::\n"
+ << "isArgumentRegister(const MachineFunction &MF, "
+ << "MCRegister PhysReg) const {\n"
+ << " return\n";
+ for (const CodeGenRegisterCategory &Category : RegCategories)
+ if (Category.getName() == "ArgumentRegisters") {
+ for (const CodeGenRegisterClass *RC : Category.getClasses())
+ OS << " " << RC->getQualifiedName()
+ << "RegClass.contains(PhysReg) ||\n";
+ break;
+ }
+ OS << " false;\n";
+ OS << "}\n\n";
+
OS << "ArrayRef<const char *> " << ClassName
<< "::getRegMaskNames() const {\n";
if (!CSRSets.empty()) {
Index: llvm/test/CodeGen/AArch64/zero-call-used-regs.ll
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/AArch64/zero-call-used-regs.ll
@@ -0,0 +1,233 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=aarch64-unknown-unknown | FileCheck %s
+
+@result = dso_local global i32 0, align 4
+
+define dso_local i32 @skip(i32 noundef %a, i32 noundef %b, i32 noundef %c) local_unnamed_addr #0 "zero-call-used-regs"="skip" {
+; CHECK-LABEL: skip:
+; CHECK: // %bb.0: // %entry
+; CHECK-NEXT: mul w8, w1, w0
+; CHECK-NEXT: orr w0, w8, w2
+; CHECK-NEXT: ret
+
+entry:
+ %mul = mul nsw i32 %b, %a
+ %or = or i32 %mul, %c
+ ret i32 %or
+}
+
+define dso_local i32 @used_gpr_arg(i32 noundef %a, i32 noundef %b, i32 noundef %c) local_unnamed_addr #0 noinline optnone "zero-call-used-regs"="used-gpr-arg" {
+; CHECK-LABEL: used_gpr_arg:
+; CHECK: // %bb.0: // %entry
+; CHECK-NEXT: mul w8, w1, w0
+; CHECK-NEXT: orr w0, w8, w2
+; CHECK-NEXT: mov x1, #0
+; CHECK-NEXT: mov x2, #0
+; CHECK-NEXT: ret
+
+entry:
+ %mul = mul nsw i32 %b, %a
+ %or = or i32 %mul, %c
+ ret i32 %or
+}
+
+define dso_local i32 @used_gpr(i32 noundef %a, i32 noundef %b, i32 noundef %c) local_unnamed_addr #0 noinline optnone "zero-call-used-regs"="used-gpr" {
+; CHECK-LABEL: used_gpr:
+; CHECK: // %bb.0: // %entry
+; CHECK-NEXT: mul w8, w1, w0
+; CHECK-NEXT: orr w0, w8, w2
+; CHECK-NEXT: mov x1, #0
+; CHECK-NEXT: mov x2, #0
+; CHECK-NEXT: mov x8, #0
+; CHECK-NEXT: ret
+
+entry:
+ %mul = mul nsw i32 %b, %a
+ %or = or i32 %mul, %c
+ ret i32 %or
+}
+
+define dso_local i32 @used_arg(i32 noundef %a, i32 noundef %b, i32 noundef %c) local_unnamed_addr #0 noinline optnone "zero-call-used-regs"="used-arg" {
+; CHECK-LABEL: used_arg:
+; CHECK: // %bb.0: // %entry
+; CHECK-NEXT: mul w8, w1, w0
+; CHECK-NEXT: orr w0, w8, w2
+; CHECK-NEXT: mov x1, #0
+; CHECK-NEXT: mov x2, #0
+; CHECK-NEXT: ret
+
+entry:
+ %mul = mul nsw i32 %b, %a
+ %or = or i32 %mul, %c
+ ret i32 %or
+}
+
+define dso_local i32 @used(i32 noundef %a, i32 noundef %b, i32 noundef %c) local_unnamed_addr #0 noinline optnone "zero-call-used-regs"="used" {
+; CHECK-LABEL: used:
+; CHECK: // %bb.0: // %entry
+; CHECK-NEXT: mul w8, w1, w0
+; CHECK-NEXT: orr w0, w8, w2
+; CHECK-NEXT: mov x1, #0
+; CHECK-NEXT: mov x2, #0
+; CHECK-NEXT: mov x8, #0
+; CHECK-NEXT: ret
+
+entry:
+ %mul = mul nsw i32 %b, %a
+ %or = or i32 %mul, %c
+ ret i32 %or
+}
+
+define dso_local i32 @all_gpr_arg(i32 noundef %a, i32 noundef %b, i32 noundef %c) local_unnamed_addr #0 "zero-call-used-regs"="all-gpr-arg" {
+; CHECK-LABEL: all_gpr_arg:
+; CHECK: // %bb.0: // %entry
+; CHECK-NEXT: mul w8, w1, w0
+; CHECK-NEXT: mov x1, #0
+; CHECK-NEXT: mov x3, #0
+; CHECK-NEXT: mov x4, #0
+; CHECK-NEXT: orr w0, w8, w2
+; CHECK-NEXT: mov x2, #0
+; CHECK-NEXT: mov x5, #0
+; CHECK-NEXT: mov x6, #0
+; CHECK-NEXT: mov x7, #0
+; CHECK-NEXT: ret
+
+entry:
+ %mul = mul nsw i32 %b, %a
+ %or = or i32 %mul, %c
+ ret i32 %or
+}
+
+define dso_local i32 @all_gpr(i32 noundef %a, i32 noundef %b, i32 noundef %c) local_unnamed_addr #0 "zero-call-used-regs"="all-gpr" {
+; CHECK-LABEL: all_gpr:
+; CHECK: // %bb.0: // %entry
+; CHECK-NEXT: mul w8, w1, w0
+; CHECK-NEXT: mov x1, #0
+; CHECK-NEXT: mov x3, #0
+; CHECK-NEXT: mov x4, #0
+; CHECK-NEXT: orr w0, w8, w2
+; CHECK-NEXT: mov x2, #0
+; CHECK-NEXT: mov x5, #0
+; CHECK-NEXT: mov x6, #0
+; CHECK-NEXT: mov x7, #0
+; CHECK-NEXT: mov x8, #0
+; CHECK-NEXT: mov x9, #0
+; CHECK-NEXT: mov x10, #0
+; CHECK-NEXT: mov x11, #0
+; CHECK-NEXT: mov x12, #0
+; CHECK-NEXT: mov x13, #0
+; CHECK-NEXT: mov x14, #0
+; CHECK-NEXT: mov x15, #0
+; CHECK-NEXT: mov x16, #0
+; CHECK-NEXT: mov x17, #0
+; CHECK-NEXT: mov x18, #0
+; CHECK-NEXT: ret
+
+entry:
+ %mul = mul nsw i32 %b, %a
+ %or = or i32 %mul, %c
+ ret i32 %or
+}
+
+define dso_local i32 @all_arg(i32 noundef %a, i32 noundef %b, i32 noundef %c) local_unnamed_addr #0 "zero-call-used-regs"="all-arg" {
+; CHECK-LABEL: all_arg:
+; CHECK: // %bb.0: // %entry
+; CHECK-NEXT: mul w8, w1, w0
+; CHECK-NEXT: mov x1, #0
+; CHECK-NEXT: mov x3, #0
+; CHECK-NEXT: mov x4, #0
+; CHECK-NEXT: orr w0, w8, w2
+; CHECK-NEXT: mov x2, #0
+; CHECK-NEXT: mov x5, #0
+; CHECK-NEXT: mov x6, #0
+; CHECK-NEXT: mov x7, #0
+; CHECK-NEXT: movi q0, #0000000000000000
+; CHECK-NEXT: movi q1, #0000000000000000
+; CHECK-NEXT: movi q2, #0000000000000000
+; CHECK-NEXT: movi q3, #0000000000000000
+; CHECK-NEXT: movi q4, #0000000000000000
+; CHECK-NEXT: movi q5, #0000000000000000
+; CHECK-NEXT: movi q6, #0000000000000000
+; CHECK-NEXT: ret
+
+entry:
+ %mul = mul nsw i32 %b, %a
+ %or = or i32 %mul, %c
+ ret i32 %or
+}
+
+define dso_local i32 @all(i32 noundef %a, i32 noundef %b, i32 noundef %c) local_unnamed_addr #0 "zero-call-used-regs"="all" {
+; CHECK-LABEL: all:
+; CHECK: // %bb.0: // %entry
+; CHECK-NEXT: mul w8, w1, w0
+; CHECK-NEXT: mov x1, #0
+; CHECK-NEXT: mov x3, #0
+; CHECK-NEXT: mov x4, #0
+; CHECK-NEXT: orr w0, w8, w2
+; CHECK-NEXT: mov x2, #0
+; CHECK-NEXT: mov x5, #0
+; CHECK-NEXT: mov x6, #0
+; CHECK-NEXT: mov x7, #0
+; CHECK-NEXT: mov x8, #0
+; CHECK-NEXT: mov x9, #0
+; CHECK-NEXT: mov x10, #0
+; CHECK-NEXT: mov x11, #0
+; CHECK-NEXT: mov x12, #0
+; CHECK-NEXT: mov x13, #0
+; CHECK-NEXT: mov x14, #0
+; CHECK-NEXT: mov x15, #0
+; CHECK-NEXT: mov x16, #0
+; CHECK-NEXT: mov x17, #0
+; CHECK-NEXT: mov x18, #0
+; CHECK-NEXT: movi q0, #0000000000000000
+; CHECK-NEXT: movi q1, #0000000000000000
+; CHECK-NEXT: movi q2, #0000000000000000
+; CHECK-NEXT: movi q3, #0000000000000000
+; CHECK-NEXT: movi q4, #0000000000000000
+; CHECK-NEXT: movi q5, #0000000000000000
+; CHECK-NEXT: movi q6, #0000000000000000
+; CHECK-NEXT: movi q7, #0000000000000000
+; CHECK-NEXT: movi q8, #0000000000000000
+; CHECK-NEXT: movi q9, #0000000000000000
+; CHECK-NEXT: movi q10, #0000000000000000
+; CHECK-NEXT: movi q11, #0000000000000000
+; CHECK-NEXT: movi q12, #0000000000000000
+; CHECK-NEXT: movi q13, #0000000000000000
+; CHECK-NEXT: movi q14, #0000000000000000
+; CHECK-NEXT: movi q15, #0000000000000000
+; CHECK-NEXT: movi q16, #0000000000000000
+; CHECK-NEXT: movi q17, #0000000000000000
+; CHECK-NEXT: movi q18, #0000000000000000
+; CHECK-NEXT: movi q19, #0000000000000000
+; CHECK-NEXT: movi q20, #0000000000000000
+; CHECK-NEXT: movi q21, #0000000000000000
+; CHECK-NEXT: movi q22, #0000000000000000
+; CHECK-NEXT: movi q23, #0000000000000000
+; CHECK-NEXT: movi q24, #0000000000000000
+; CHECK-NEXT: movi q25, #0000000000000000
+; CHECK-NEXT: movi q26, #0000000000000000
+; CHECK-NEXT: movi q27, #0000000000000000
+; CHECK-NEXT: movi q28, #0000000000000000
+; CHECK-NEXT: movi q29, #0000000000000000
+; CHECK-NEXT: movi q30, #0000000000000000
+; CHECK-NEXT: movi q31, #0000000000000000
+; CHECK-NEXT: ret
+
+entry:
+ %mul = mul nsw i32 %b, %a
+ %or = or i32 %mul, %c
+ ret i32 %or
+}
+
+; Don't emit zeroing registers in "main" function.
+define dso_local i32 @main() local_unnamed_addr #0 {
+; CHECK-LABEL: main:
+; CHECK: // %bb.0: // %entry
+; CHECK-NEXT: mov w0, wzr
+; CHECK-NEXT: ret
+
+entry:
+ ret i32 0
+}
+
+attributes #0 = { mustprogress nofree norecurse nosync nounwind readnone willreturn uwtable "frame-pointer"="non-leaf" "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="generic" "target-features"="+neon,+v8a" }
Index: llvm/lib/Target/X86/X86RegisterInfo.cpp
===================================================================
--- llvm/lib/Target/X86/X86RegisterInfo.cpp
+++ llvm/lib/Target/X86/X86RegisterInfo.cpp
@@ -656,7 +656,7 @@
[&](MCRegister &RegA) { return IsSubReg(RegA, Reg); }))
return true;
- return false;
+ return X86GenRegisterInfo::isArgumentRegister(MF, Reg);
}
bool X86RegisterInfo::isFixedRegister(const MachineFunction &MF,
Index: llvm/lib/Target/AArch64/AArch64RegisterInfo.td
===================================================================
--- llvm/lib/Target/AArch64/AArch64RegisterInfo.td
+++ llvm/lib/Target/AArch64/AArch64RegisterInfo.td
@@ -1385,3 +1385,33 @@
return AArch64SVCR::lookupSVCRByEncoding(MCOp.getImm()) != nullptr;
}];
}
+
+//===----------------------------------------------------------------------===//
+// Register categories.
+//
+
+def GeneralPurposeRegisters : RegisterCategory<[GPR64, GPR32]>;
+
+def GPR_ARG : RegisterClass<"AArch64", [i64, i32], 64,
+ (add (sequence "X%u", 0, 7),
+ (sequence "W%u", 0, 7))>;
+def FPR8_ARG : RegisterClass<"AArch64", [untyped], 8, (trunc FPR8, 7)> {
+ let Size = 8;
+}
+def FPR16_ARG : RegisterClass<"AArch64", [f16, bf16], 16, (trunc FPR16, 7)> {
+ let Size = 16;
+}
+def FPR32_ARG : RegisterClass<"AArch64", [f32, i32], 32, (trunc FPR32, 7)>;
+def FPR64_ARG : RegisterClass<"AArch64",
+ [f64, i64, v2f32, v1f64, v8i8, v4i16, v2i32,
+ v1i64, v4f16, v4bf16], 64,
+ (trunc FPR64, 7)>;
+def FPR128_ARG : RegisterClass<"AArch64",
+ [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64,
+ f128, v8f16, v8bf16], 128,
+ (trunc FPR128, 7)>;
+def ArgumentRegisters : RegisterCategory<[GPR_ARG, FPR8_ARG, FPR16_ARG,
+ FPR32_ARG, FPR64_ARG, FPR128_ARG]>;
+
+def FIXED_REGS : RegisterClass<"AArch64", [i64], 64, (add FP, SP, VG, FFR)>;
+def FixedRegisters : RegisterCategory<[CCR, FIXED_REGS]>;
Index: llvm/lib/Target/AArch64/AArch64FrameLowering.h
===================================================================
--- llvm/lib/Target/AArch64/AArch64FrameLowering.h
+++ llvm/lib/Target/AArch64/AArch64FrameLowering.h
@@ -153,6 +153,10 @@
MachineBasicBlock::iterator MBBI) const;
void emitCalleeSavedSVERestores(MachineBasicBlock &MBB,
MachineBasicBlock::iterator MBBI) const;
+
+ /// Emit target zero call-used regs.
+ void emitZeroCallUsedRegs(BitVector RegsToZero,
+ MachineBasicBlock &MBB) const override;
};
} // End llvm namespace
Index: llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
===================================================================
--- llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
+++ llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
@@ -675,6 +675,123 @@
emitCalleeSavedRestores(MBB, MBBI, true);
}
+static MCRegister getRegisterOrZero(MCRegister Reg) {
+ switch (Reg.id()) {
+ default:
+ // The called routine is expected to preserve r19-r28
+ // r29 and r30 are used as frame pointer and link register resp.
+ return 0;
+
+ // GPRs
+#define CASE(n) \
+ case AArch64::W##n: \
+ case AArch64::X##n: \
+ return AArch64::X##n
+ CASE(0);
+ CASE(1);
+ CASE(2);
+ CASE(3);
+ CASE(4);
+ CASE(5);
+ CASE(6);
+ CASE(7);
+ CASE(8);
+ CASE(9);
+ CASE(10);
+ CASE(11);
+ CASE(12);
+ CASE(13);
+ CASE(14);
+ CASE(15);
+ CASE(16);
+ CASE(17);
+ CASE(18);
+#undef CASE
+
+ // FPRs
+#define CASE(n) \
+ case AArch64::B##n: \
+ case AArch64::H##n: \
+ case AArch64::S##n: \
+ case AArch64::D##n: \
+ case AArch64::Q##n: \
+ return AArch64::Q##n
+ CASE(0);
+ CASE(1);
+ CASE(2);
+ CASE(3);
+ CASE(4);
+ CASE(5);
+ CASE(6);
+ CASE(7);
+ CASE(8);
+ CASE(9);
+ CASE(10);
+ CASE(11);
+ CASE(12);
+ CASE(13);
+ CASE(14);
+ CASE(15);
+ CASE(16);
+ CASE(17);
+ CASE(18);
+ CASE(19);
+ CASE(20);
+ CASE(21);
+ CASE(22);
+ CASE(23);
+ CASE(24);
+ CASE(25);
+ CASE(26);
+ CASE(27);
+ CASE(28);
+ CASE(29);
+ CASE(30);
+ CASE(31);
+#undef CASE
+ }
+}
+
+void AArch64FrameLowering::emitZeroCallUsedRegs(BitVector RegsToZero,
+ MachineBasicBlock &MBB) const {
+ const MachineFunction &MF = *MBB.getParent();
+
+ // Insertion point.
+ MachineBasicBlock::iterator MBBI = MBB.getFirstTerminator();
+
+ // Fake a debug loc.
+ DebugLoc DL;
+ if (MBBI != MBB.end())
+ DL = MBBI->getDebugLoc();
+
+ const TargetSubtargetInfo &STI = MF.getSubtarget();
+ const TargetRegisterInfo &TRI = *STI.getRegisterInfo();
+ const AArch64InstrInfo *TII =
+ static_cast<const AArch64InstrInfo *>(MF.getSubtarget().getInstrInfo());
+
+ BitVector GPRsToZero(TRI.getNumRegs());
+ BitVector FPRsToZero(TRI.getNumRegs());
+ for (MCRegister Reg : RegsToZero.set_bits()) {
+ if (TRI.isGeneralPurposeRegister(MF, Reg)) {
+ // For GPRs, we only care to clear out the 64-bit register.
+ if (MCRegister XReg = getRegisterOrZero(Reg))
+ GPRsToZero.set(XReg);
+ } else if (AArch64::FPR128RegClass.contains(Reg)) {
+ // For FPRs,
+ if (MCRegister XReg = getRegisterOrZero(Reg))
+ FPRsToZero.set(XReg);
+ }
+ }
+
+ // Zero out GPRs.
+ for (MCRegister Reg : GPRsToZero.set_bits())
+ BuildMI(MBB, MBBI, DL, TII->get(AArch64::MOVi64imm), Reg).addImm(0);
+
+ // Zero out FP/vector registers.
+ for (MCRegister Reg : FPRsToZero.set_bits())
+ BuildMI(MBB, MBBI, DL, TII->get(AArch64::MOVID), Reg).addImm(0);
+}
+
// Find a scratch register that we can use at the start of the prologue to
// re-align the stack pointer. We avoid using callee-save registers since they
// may appear to be free when this is called from canUseAsPrologue (during
Index: clang/lib/Driver/ToolChains/Clang.cpp
===================================================================
--- clang/lib/Driver/ToolChains/Clang.cpp
+++ clang/lib/Driver/ToolChains/Clang.cpp
@@ -5984,7 +5984,7 @@
// FIXME: There's no reason for this to be restricted to X86. The backend
// code needs to be changed to include the appropriate function calls
// automatically.
- if (!Triple.isX86())
+ if (!Triple.isX86() && !Triple.isAArch64())
D.Diag(diag::err_drv_unsupported_opt_for_target)
<< A->getAsString(Args) << TripleStr;
}
Index: clang/include/clang/Driver/Options.td
===================================================================
--- clang/include/clang/Driver/Options.td
+++ clang/include/clang/Driver/Options.td
@@ -2977,7 +2977,7 @@
def fzero_call_used_regs_EQ
: Joined<["-"], "fzero-call-used-regs=">, Group<f_Group>, Flags<[CC1Option]>,
- HelpText<"Clear call-used registers upon function return.">,
+ HelpText<"Clear call-used registers upon function return (AArch64/x86 only)">,
Values<"skip,used-gpr-arg,used-gpr,used-arg,used,all-gpr-arg,all-gpr,all-arg,all">,
NormalizedValues<["Skip", "UsedGPRArg", "UsedGPR", "UsedArg", "Used",
"AllGPRArg", "AllGPR", "AllArg", "All"]>,
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
