[llvm] [CodeGen] Precompute untyped getMinimalPhysRegClass (PR #193438)

Cullen Rhodes via llvm-commits llvm-commits at lists.llvm.org
Tue May 5 01:47:59 PDT 2026


https://github.com/c-rhodes updated https://github.com/llvm/llvm-project/pull/193438

>From 9a4b2befdd2525ac92ae21139b5d96787ca688a6 Mon Sep 17 00:00:00 2001
From: Cullen Rhodes <cullen.rhodes at arm.com>
Date: Wed, 22 Apr 2026 07:19:31 +0000
Subject: [PATCH 1/5] [AArch64] Avoid getMinimalPhysRegClass in callee-save
 frame lowering

getMinimalPhysRegClass is slow and the callee-save paths in frame
lowering that call it are hot. Which registers are callee-saved is
well-defined, selecting the register classes directly is a -0.22%
geomean improvement on stage1-aarch64-O0-g CTMark.
---
 .../Target/AArch64/AArch64FrameLowering.cpp   | 22 ++++++++++++++++---
 1 file changed, 19 insertions(+), 3 deletions(-)

diff --git a/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp b/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
index 2c68e379205ce..012857356b033 100644
--- a/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
@@ -345,6 +345,23 @@ static unsigned getStackHazardSize(const MachineFunction &MF) {
   return MF.getSubtarget<AArch64Subtarget>().getStreamingHazardSize();
 }
 
+static const TargetRegisterClass *
+getAArch64CalleeSavedRegClass(MCRegister Reg) {
+  if (AArch64::GPR64RegClass.contains(Reg))
+    return &AArch64::GPR64RegClass;
+  if (AArch64::FPR64RegClass.contains(Reg))
+    return &AArch64::FPR64RegClass;
+  if (AArch64::FPR128RegClass.contains(Reg))
+    return &AArch64::FPR128RegClass;
+  if (AArch64::ZPRRegClass.contains(Reg))
+    return &AArch64::ZPRRegClass;
+  if (AArch64::PPRRegClass.contains(Reg))
+    return &AArch64::PPRRegClass;
+  if (Reg == AArch64::VG)
+    return &AArch64::FIXED_REGSRegClass;
+  llvm_unreachable("Unsupported register class.");
+}
+
 StackOffset
 AArch64FrameLowering::getZPRStackSize(const MachineFunction &MF) const {
   const AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
@@ -2623,8 +2640,7 @@ void AArch64FrameLowering::determineCalleeSaves(MachineFunction &MF,
   unsigned PPRCSStackSize = 0;
   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
   for (unsigned Reg : SavedRegs.set_bits()) {
-    auto *RC = TRI->getMinimalPhysRegClass(MCRegister(Reg));
-    assert(RC && "expected register class!");
+    auto *RC = getAArch64CalleeSavedRegClass(MCRegister(Reg));
     auto SpillSize = TRI->getSpillSize(*RC);
     bool IsZPR = AArch64::ZPRRegClass.contains(Reg);
     bool IsPPR = !IsZPR && AArch64::PPRRegClass.contains(Reg);
@@ -2804,7 +2820,7 @@ bool AArch64FrameLowering::assignCalleeSavedSpillSlots(
   int HazardSlotIndex = std::numeric_limits<int>::max();
   for (auto &CS : CSI) {
     MCRegister Reg = CS.getReg();
-    const TargetRegisterClass *RC = RegInfo->getMinimalPhysRegClass(Reg);
+    const TargetRegisterClass *RC = getAArch64CalleeSavedRegClass(Reg);
 
     // Create a hazard slot as we switch between GPR and FPR CSRs.
     if (AFI->isStackHazardIncludedInCalleeSaveArea() &&

>From 2038fc376d322e7c10dd908d823c95db211098dc Mon Sep 17 00:00:00 2001
From: Cullen Rhodes <cullen.rhodes at arm.com>
Date: Mon, 27 Apr 2026 19:18:38 +0000
Subject: [PATCH 2/5] address comments

---
 .../Target/AArch64/AArch64FrameLowering.cpp   | 22 +++----------------
 .../lib/Target/AArch64/AArch64RegisterInfo.td | 15 ++++++++++++-
 2 files changed, 17 insertions(+), 20 deletions(-)

diff --git a/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp b/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
index 012857356b033..5a1ff324f948f 100644
--- a/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
@@ -345,23 +345,6 @@ static unsigned getStackHazardSize(const MachineFunction &MF) {
   return MF.getSubtarget<AArch64Subtarget>().getStreamingHazardSize();
 }
 
-static const TargetRegisterClass *
-getAArch64CalleeSavedRegClass(MCRegister Reg) {
-  if (AArch64::GPR64RegClass.contains(Reg))
-    return &AArch64::GPR64RegClass;
-  if (AArch64::FPR64RegClass.contains(Reg))
-    return &AArch64::FPR64RegClass;
-  if (AArch64::FPR128RegClass.contains(Reg))
-    return &AArch64::FPR128RegClass;
-  if (AArch64::ZPRRegClass.contains(Reg))
-    return &AArch64::ZPRRegClass;
-  if (AArch64::PPRRegClass.contains(Reg))
-    return &AArch64::PPRRegClass;
-  if (Reg == AArch64::VG)
-    return &AArch64::FIXED_REGSRegClass;
-  llvm_unreachable("Unsupported register class.");
-}
-
 StackOffset
 AArch64FrameLowering::getZPRStackSize(const MachineFunction &MF) const {
   const AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
@@ -2640,7 +2623,8 @@ void AArch64FrameLowering::determineCalleeSaves(MachineFunction &MF,
   unsigned PPRCSStackSize = 0;
   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
   for (unsigned Reg : SavedRegs.set_bits()) {
-    auto *RC = getAArch64CalleeSavedRegClass(MCRegister(Reg));
+    auto *RC = TRI->getPhysRegBaseClass(MCRegister(Reg));
+    assert(RC && "expected register class!");
     auto SpillSize = TRI->getSpillSize(*RC);
     bool IsZPR = AArch64::ZPRRegClass.contains(Reg);
     bool IsPPR = !IsZPR && AArch64::PPRRegClass.contains(Reg);
@@ -2820,7 +2804,7 @@ bool AArch64FrameLowering::assignCalleeSavedSpillSlots(
   int HazardSlotIndex = std::numeric_limits<int>::max();
   for (auto &CS : CSI) {
     MCRegister Reg = CS.getReg();
-    const TargetRegisterClass *RC = getAArch64CalleeSavedRegClass(Reg);
+    const TargetRegisterClass *RC = RegInfo->getPhysRegBaseClass(Reg);
 
     // Create a hazard slot as we switch between GPR and FPR CSRs.
     if (AFI->isStackHazardIncludedInCalleeSaveArea() &&
diff --git a/llvm/lib/Target/AArch64/AArch64RegisterInfo.td b/llvm/lib/Target/AArch64/AArch64RegisterInfo.td
index cd94a2590c6d2..d2f03bac88917 100644
--- a/llvm/lib/Target/AArch64/AArch64RegisterInfo.td
+++ b/llvm/lib/Target/AArch64/AArch64RegisterInfo.td
@@ -202,11 +202,13 @@ def GPR64common : RegisterClass<"AArch64", [i64], 64,
 def GPR32 : RegisterClass<"AArch64", [i32], 32, (add GPR32common, WZR)> {
   let AltOrders = [(rotl GPR32, 8)];
   let AltOrderSelect = [{ return 1; }];
+  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::GPR32RegClassID, 0, 32>";
 }
 def GPR64 : RegisterClass<"AArch64", [i64], 64, (add GPR64common, XZR)> {
   let AltOrders = [(rotl GPR64, 8)];
   let AltOrderSelect = [{ return 1; }];
+  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::GPR64RegClassID, 0, 32>";
 }
 
@@ -214,11 +216,13 @@ def GPR64 : RegisterClass<"AArch64", [i64], 64, (add GPR64common, XZR)> {
 def GPR32sp : RegisterClass<"AArch64", [i32], 32, (add GPR32common, WSP)> {
   let AltOrders = [(rotl GPR32sp, 8)];
   let AltOrderSelect = [{ return 1; }];
+  let BaseClassOrder = 1;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::GPR32spRegClassID, 0, 32>";
 }
 def GPR64sp : RegisterClass<"AArch64", [i64], 64, (add GPR64common, SP)> {
   let AltOrders = [(rotl GPR64sp, 8)];
   let AltOrderSelect = [{ return 1; }];
+  let BaseClassOrder = 1;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::GPR64spRegClassID,0, 32>";
 }
 
@@ -500,10 +504,12 @@ def Q31   : AArch64Reg<31, "q31", [D31, D31_HI], ["v31", ""]>, DwarfRegAlias<B31
 
 def FPR8  : RegisterClass<"AArch64", [i8, aarch64mfp8], 8, (sequence "B%u", 0, 31)> {
   let Size = 8;
+  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::FPR8RegClassID, 0, 32>";
 }
 def FPR16 : RegisterClass<"AArch64", [f16, bf16, i16], 16, (sequence "H%u", 0, 31)> {
   let Size = 16;
+  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::FPR16RegClassID, 0, 32>";
 }
 
@@ -511,11 +517,13 @@ def FPR16_lo : RegisterClass<"AArch64", [f16], 16, (trunc FPR16, 16)> {
   let Size = 16;
 }
 def FPR32 : RegisterClass<"AArch64", [f32, i32], 32,(sequence "S%u", 0, 31)> {
+  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::FPR32RegClassID, 0, 32>";
 }
 def FPR64 : RegisterClass<"AArch64", [f64, i64, v2f32, v1f64, v8i8, v4i16, v2i32,
                                       v1i64, v4f16, v4bf16],
                                      64, (sequence "D%u", 0, 31)> {
+  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::FPR64RegClassID, 0, 32>";
 }
 def FPR64_lo : RegisterClass<"AArch64",
@@ -530,6 +538,7 @@ def FPR128 : RegisterClass<"AArch64",
                            [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64, f128,
                             v8f16, v8bf16],
                            128, (sequence "Q%u", 0, 31)> {
+  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::FPR128RegClassID, 0, 32>";
 }
 
@@ -987,6 +996,7 @@ class PPRClass<int firstreg, int lastreg, int step = 1> : RegisterClass<"AArch64
 }
 
 def PPR    : PPRClass<0, 15> {
+  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::PPRRegClassID, 0, 16>";
 }
 def PPR_3b : PPRClass<0, 7> { // Restricted 3 bit SVE predicate register class.
@@ -1190,6 +1200,7 @@ def ZPRMul2_Lo : ZPRClass<0,  14, 2>;
 def ZPRMul2_Hi : ZPRClass<16, 30, 2>;
 
 def ZPR    : ZPRClass<0, 31> {
+  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::ZPRRegClassID, 0, 32>";
 }
 def ZPR_4b : ZPRClass<0, 15> { // Restricted 4 bit SVE vector register class.
@@ -2010,5 +2021,7 @@ let isAllocatable = 0, GeneratePressureSet = 0 in {
 
 def GeneralPurposeRegisters : RegisterCategory<[GPR64, GPR32]>;
 
-def FIXED_REGS : RegisterClass<"AArch64", [i64], 64, (add FP, SP, VG, FFR)>;
+def FIXED_REGS : RegisterClass<"AArch64", [i64], 64, (add FP, SP, VG, FFR)> {
+  let BaseClassOrder = 2;
+}
 def FixedRegisters : RegisterCategory<[CCR, FIXED_REGS]>;

>From 5de08adc0d8b40c957cf4af57158c141d3a8229d Mon Sep 17 00:00:00 2001
From: Cullen Rhodes <cullen.rhodes at arm.com>
Date: Wed, 29 Apr 2026 16:27:13 +0000
Subject: [PATCH 3/5] [CodeGen] Precompute untyped getMinimalPhysRegClass

---
 .../include/llvm/CodeGen/TargetRegisterInfo.h |  8 +++++
 llvm/lib/CodeGen/TargetRegisterInfo.cpp       |  4 +++
 .../Target/AArch64/AArch64FrameLowering.cpp   |  4 +--
 .../lib/Target/AArch64/AArch64RegisterInfo.td | 15 +-------
 llvm/utils/TableGen/RegisterInfoEmitter.cpp   | 36 +++++++++++++++++++
 5 files changed, 51 insertions(+), 16 deletions(-)

diff --git a/llvm/include/llvm/CodeGen/TargetRegisterInfo.h b/llvm/include/llvm/CodeGen/TargetRegisterInfo.h
index df371a7706bb0..06d812ae2ef52 100644
--- a/llvm/include/llvm/CodeGen/TargetRegisterInfo.h
+++ b/llvm/include/llvm/CodeGen/TargetRegisterInfo.h
@@ -829,6 +829,14 @@ class LLVM_ABI TargetRegisterInfo : public MCRegisterInfo {
     return nullptr;
   }
 
+  /// Return the target-defined minimal register class for an untyped physical
+  /// register query.
+  /// Will be nullptr if the register is not in any register class.
+  virtual const TargetRegisterClass *
+  getDefaultMinimalPhysRegClass(MCRegister Reg) const {
+    return nullptr;
+  }
+
 protected:
   /// Overridden by TableGen in targets that have sub-registers.
   virtual unsigned composeSubRegIndicesImpl(unsigned, unsigned) const {
diff --git a/llvm/lib/CodeGen/TargetRegisterInfo.cpp b/llvm/lib/CodeGen/TargetRegisterInfo.cpp
index f7cc512e6eb2b..2bd92582910ee 100644
--- a/llvm/lib/CodeGen/TargetRegisterInfo.cpp
+++ b/llvm/lib/CodeGen/TargetRegisterInfo.cpp
@@ -263,6 +263,10 @@ getCommonMinimalPhysRegClass(const TargetRegisterInfo *TRI, MCRegister Reg1,
 
 const TargetRegisterClass *
 TargetRegisterInfo::getMinimalPhysRegClass(MCRegister Reg, MVT VT) const {
+  if (VT == MVT::Other) {
+    if (const TargetRegisterClass *RC = getDefaultMinimalPhysRegClass(Reg))
+      return RC;
+  }
   return ::getMinimalPhysRegClass(this, Reg, VT);
 }
 
diff --git a/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp b/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
index 5a1ff324f948f..2c68e379205ce 100644
--- a/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
@@ -2623,7 +2623,7 @@ void AArch64FrameLowering::determineCalleeSaves(MachineFunction &MF,
   unsigned PPRCSStackSize = 0;
   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
   for (unsigned Reg : SavedRegs.set_bits()) {
-    auto *RC = TRI->getPhysRegBaseClass(MCRegister(Reg));
+    auto *RC = TRI->getMinimalPhysRegClass(MCRegister(Reg));
     assert(RC && "expected register class!");
     auto SpillSize = TRI->getSpillSize(*RC);
     bool IsZPR = AArch64::ZPRRegClass.contains(Reg);
@@ -2804,7 +2804,7 @@ bool AArch64FrameLowering::assignCalleeSavedSpillSlots(
   int HazardSlotIndex = std::numeric_limits<int>::max();
   for (auto &CS : CSI) {
     MCRegister Reg = CS.getReg();
-    const TargetRegisterClass *RC = RegInfo->getPhysRegBaseClass(Reg);
+    const TargetRegisterClass *RC = RegInfo->getMinimalPhysRegClass(Reg);
 
     // Create a hazard slot as we switch between GPR and FPR CSRs.
     if (AFI->isStackHazardIncludedInCalleeSaveArea() &&
diff --git a/llvm/lib/Target/AArch64/AArch64RegisterInfo.td b/llvm/lib/Target/AArch64/AArch64RegisterInfo.td
index d2f03bac88917..cd94a2590c6d2 100644
--- a/llvm/lib/Target/AArch64/AArch64RegisterInfo.td
+++ b/llvm/lib/Target/AArch64/AArch64RegisterInfo.td
@@ -202,13 +202,11 @@ def GPR64common : RegisterClass<"AArch64", [i64], 64,
 def GPR32 : RegisterClass<"AArch64", [i32], 32, (add GPR32common, WZR)> {
   let AltOrders = [(rotl GPR32, 8)];
   let AltOrderSelect = [{ return 1; }];
-  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::GPR32RegClassID, 0, 32>";
 }
 def GPR64 : RegisterClass<"AArch64", [i64], 64, (add GPR64common, XZR)> {
   let AltOrders = [(rotl GPR64, 8)];
   let AltOrderSelect = [{ return 1; }];
-  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::GPR64RegClassID, 0, 32>";
 }
 
@@ -216,13 +214,11 @@ def GPR64 : RegisterClass<"AArch64", [i64], 64, (add GPR64common, XZR)> {
 def GPR32sp : RegisterClass<"AArch64", [i32], 32, (add GPR32common, WSP)> {
   let AltOrders = [(rotl GPR32sp, 8)];
   let AltOrderSelect = [{ return 1; }];
-  let BaseClassOrder = 1;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::GPR32spRegClassID, 0, 32>";
 }
 def GPR64sp : RegisterClass<"AArch64", [i64], 64, (add GPR64common, SP)> {
   let AltOrders = [(rotl GPR64sp, 8)];
   let AltOrderSelect = [{ return 1; }];
-  let BaseClassOrder = 1;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::GPR64spRegClassID,0, 32>";
 }
 
@@ -504,12 +500,10 @@ def Q31   : AArch64Reg<31, "q31", [D31, D31_HI], ["v31", ""]>, DwarfRegAlias<B31
 
 def FPR8  : RegisterClass<"AArch64", [i8, aarch64mfp8], 8, (sequence "B%u", 0, 31)> {
   let Size = 8;
-  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::FPR8RegClassID, 0, 32>";
 }
 def FPR16 : RegisterClass<"AArch64", [f16, bf16, i16], 16, (sequence "H%u", 0, 31)> {
   let Size = 16;
-  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::FPR16RegClassID, 0, 32>";
 }
 
@@ -517,13 +511,11 @@ def FPR16_lo : RegisterClass<"AArch64", [f16], 16, (trunc FPR16, 16)> {
   let Size = 16;
 }
 def FPR32 : RegisterClass<"AArch64", [f32, i32], 32,(sequence "S%u", 0, 31)> {
-  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::FPR32RegClassID, 0, 32>";
 }
 def FPR64 : RegisterClass<"AArch64", [f64, i64, v2f32, v1f64, v8i8, v4i16, v2i32,
                                       v1i64, v4f16, v4bf16],
                                      64, (sequence "D%u", 0, 31)> {
-  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::FPR64RegClassID, 0, 32>";
 }
 def FPR64_lo : RegisterClass<"AArch64",
@@ -538,7 +530,6 @@ def FPR128 : RegisterClass<"AArch64",
                            [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64, f128,
                             v8f16, v8bf16],
                            128, (sequence "Q%u", 0, 31)> {
-  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::FPR128RegClassID, 0, 32>";
 }
 
@@ -996,7 +987,6 @@ class PPRClass<int firstreg, int lastreg, int step = 1> : RegisterClass<"AArch64
 }
 
 def PPR    : PPRClass<0, 15> {
-  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::PPRRegClassID, 0, 16>";
 }
 def PPR_3b : PPRClass<0, 7> { // Restricted 3 bit SVE predicate register class.
@@ -1200,7 +1190,6 @@ def ZPRMul2_Lo : ZPRClass<0,  14, 2>;
 def ZPRMul2_Hi : ZPRClass<16, 30, 2>;
 
 def ZPR    : ZPRClass<0, 31> {
-  let BaseClassOrder = 0;
   let DecoderMethod = "DecodeSimpleRegisterClass<AArch64::ZPRRegClassID, 0, 32>";
 }
 def ZPR_4b : ZPRClass<0, 15> { // Restricted 4 bit SVE vector register class.
@@ -2021,7 +2010,5 @@ let isAllocatable = 0, GeneratePressureSet = 0 in {
 
 def GeneralPurposeRegisters : RegisterCategory<[GPR64, GPR32]>;
 
-def FIXED_REGS : RegisterClass<"AArch64", [i64], 64, (add FP, SP, VG, FFR)> {
-  let BaseClassOrder = 2;
-}
+def FIXED_REGS : RegisterClass<"AArch64", [i64], 64, (add FP, SP, VG, FFR)>;
 def FixedRegisters : RegisterCategory<[CCR, FIXED_REGS]>;
diff --git a/llvm/utils/TableGen/RegisterInfoEmitter.cpp b/llvm/utils/TableGen/RegisterInfoEmitter.cpp
index 0338d0b588352..577c434e2e014 100644
--- a/llvm/utils/TableGen/RegisterInfoEmitter.cpp
+++ b/llvm/utils/TableGen/RegisterInfoEmitter.cpp
@@ -1263,6 +1263,10 @@ void RegisterInfoEmitter::runTargetHeader(raw_ostream &OS, raw_ostream &MainOS,
     OS << "  const TargetRegisterClass *getPhysRegBaseClass(MCRegister Reg) "
           "const override;\n";
   }
+  if (!RegisterClasses.empty()) {
+    OS << "  const TargetRegisterClass *getDefaultMinimalPhysRegClass("
+          "MCRegister Reg) const override;\n";
+  }
 
   OS << "};\n\n";
 
@@ -1718,6 +1722,38 @@ void RegisterInfoEmitter::runTargetDesc(raw_ostream &OS, raw_ostream &MainOS,
     }
   }
 
+  if (!RegisterClasses.empty()) {
+    assert(RegisterClasses.size() < UINT16_MAX &&
+           "Too many minimal register classes");
+
+    OS << "\n// Register to minimal register class mapping\n\n";
+    OS << "const TargetRegisterClass *" << ClassName
+       << "::getDefaultMinimalPhysRegClass(MCRegister Reg)" << " const {\n";
+    OS << "  static const uint16_t InvalidRegClassID = UINT16_MAX;\n\n";
+    OS << "  static const uint16_t Mapping[" << Regs.size() + 1 << "] = {\n";
+    OS << "    InvalidRegClassID,  // NoRegister\n";
+    for (const CodeGenRegister &Reg : Regs) {
+      const CodeGenRegisterClass *MinimalRC = nullptr;
+      for (const auto &RC : RegisterClasses) {
+        if (RC.contains(&Reg) && (!MinimalRC || MinimalRC->hasSubClass(&RC)))
+          MinimalRC = &RC;
+      }
+
+      OS << "    "
+         << (MinimalRC ? MinimalRC->getQualifiedIdName() : "InvalidRegClassID")
+         << ",  // " << Reg.getName() << "\n";
+    }
+    OS << "  };\n\n"
+          "  assert(Reg < ArrayRef(Mapping).size());\n"
+          "  unsigned RCID = Mapping[Reg.id()];\n"
+          "  if (RCID == InvalidRegClassID)\n"
+          "    return nullptr;\n"
+          "  return "
+       << TargetName
+       << "RegisterClasses[RCID];\n"
+          "}\n";
+  }
+
   // Emit the constructor of the class...
   OS << "extern const MCRegisterDesc " << TargetName << "RegDesc[];\n";
   OS << "extern const int16_t " << TargetName << "RegDiffLists[];\n";

>From 2f8233bba9bbe48e3496ab479eefb9cd0a2974ae Mon Sep 17 00:00:00 2001
From: Cullen Rhodes <cullen.rhodes at arm.com>
Date: Tue, 5 May 2026 06:32:50 +0000
Subject: [PATCH 4/5] address comments

---
 llvm/lib/CodeGen/TargetRegisterInfo.cpp | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/llvm/lib/CodeGen/TargetRegisterInfo.cpp b/llvm/lib/CodeGen/TargetRegisterInfo.cpp
index 2bd92582910ee..98cf65d465f63 100644
--- a/llvm/lib/CodeGen/TargetRegisterInfo.cpp
+++ b/llvm/lib/CodeGen/TargetRegisterInfo.cpp
@@ -218,6 +218,11 @@ getMinimalPhysRegClass(const TargetRegisterInfo *TRI, MCRegister Reg,
       return !Ty.isValid();
   }();
 
+  if (IsDefault) {
+    if (const TargetRegisterClass *RC = TRI->getDefaultMinimalPhysRegClass(Reg))
+      return RC;
+  }
+
   // Pick the most sub register class of the right type that contains
   // this physreg.
   const TargetRegisterClass *BestRC = nullptr;
@@ -263,10 +268,6 @@ getCommonMinimalPhysRegClass(const TargetRegisterInfo *TRI, MCRegister Reg1,
 
 const TargetRegisterClass *
 TargetRegisterInfo::getMinimalPhysRegClass(MCRegister Reg, MVT VT) const {
-  if (VT == MVT::Other) {
-    if (const TargetRegisterClass *RC = getDefaultMinimalPhysRegClass(Reg))
-      return RC;
-  }
   return ::getMinimalPhysRegClass(this, Reg, VT);
 }
 

>From a7cdfd67e921b3c7458934b605ba2cc0166a0f1a Mon Sep 17 00:00:00 2001
From: Cullen Rhodes <cullen.rhodes at arm.com>
Date: Tue, 5 May 2026 08:46:00 +0000
Subject: [PATCH 5/5] address comments

---
 llvm/include/llvm/CodeGen/TargetRegisterInfo.h | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/llvm/include/llvm/CodeGen/TargetRegisterInfo.h b/llvm/include/llvm/CodeGen/TargetRegisterInfo.h
index 06d812ae2ef52..8148d11417128 100644
--- a/llvm/include/llvm/CodeGen/TargetRegisterInfo.h
+++ b/llvm/include/llvm/CodeGen/TargetRegisterInfo.h
@@ -829,9 +829,8 @@ class LLVM_ABI TargetRegisterInfo : public MCRegisterInfo {
     return nullptr;
   }
 
-  /// Return the target-defined minimal register class for an untyped physical
-  /// register query.
-  /// Will be nullptr if the register is not in any register class.
+  /// Returns the target-defined minimal register class for an untyped physical
+  /// register query or nullptr if the register is not in any register class.
   virtual const TargetRegisterClass *
   getDefaultMinimalPhysRegClass(MCRegister Reg) const {
     return nullptr;



More information about the llvm-commits mailing list