[Spill2Reg] Adds code generation for 8/16bit spill/reloads in x86

Vasileios Porpodas · Vasileios Porpodas · commit a765fe31f106 · 2024-12-20T04:56:20.000-08:00
This patch adds support for 8/16 bit values in x86. Original review: https://reviews.llvm.org/D118305
diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp
@@ -10954,6 +10954,8 @@ bool X86InstrInfo::isLegalToSpill2Reg(Register Reg,
   switch (unsigned Bits = TRI->getRegSizeInBits(Reg, *MRI)) {
   case 64:
   case 32:
+  case 16:
+  case 8:
     return true;
   }
   return false;
@@ -11049,6 +11051,8 @@ bool X86InstrInfo::isSpill2RegProfitable(const MachineInstr *MI,
 
 static unsigned getInsertOrExtractOpcode(unsigned Bits, bool Insert) {
   switch (Bits) {
+  case 8:
+  case 16:
   case 32:
     return Insert ? X86::MOVDI2PDIrr : X86::MOVPDI2DIrr;
   case 64:
@@ -11058,6 +11062,36 @@ static unsigned getInsertOrExtractOpcode(unsigned Bits, bool Insert) {
   }
 }
 
+/// \Returns the subreg index for a getting a subregister of \p SubregBits from
+/// a register of \p RegBits.
+static unsigned spill2RegGetSubregIdx(unsigned RegBits, unsigned SubregBits) {
+  assert(RegBits > SubregBits && "From expected to cover To");
+  switch (SubregBits) {
+  case 32:
+    return X86::sub_32bit;
+  case 16:
+    return X86::sub_16bit;
+  case 8:
+    return X86::sub_8bit;
+  default:
+    llvm_unreachable("FIXME");
+  }
+}
+
+std::optional<MCRegister>
+X86InstrInfo::getMovdCompatibleReg(MCRegister OldReg, uint32_t OldRegBits,
+                                   const TargetRegisterInfo *TRI) const {
+  if (OldRegBits != 8 && OldRegBits != 16)
+    return std::nullopt;
+  // The register class of the register that movd can handle.
+  const TargetRegisterClass *NewRegClass =
+      TRI->getRegClass(X86::GR32RegClassID);
+  unsigned NewRegBits = TRI->getRegSizeInBits(*NewRegClass);
+  unsigned SubIdx = spill2RegGetSubregIdx(NewRegBits, OldRegBits);
+  MCRegister NewReg = TRI->getMatchingSuperReg(OldReg, SubIdx, NewRegClass);
+  return NewReg;
+}
+
 MachineInstr *X86InstrInfo::spill2RegInsertToVectorReg(
     Register DstReg, Register SrcReg, int OperationBits, MachineBasicBlock *MBB,
     MachineBasicBlock::iterator InsertBeforeIt,
@@ -11066,6 +11100,12 @@ MachineInstr *X86InstrInfo::spill2RegInsertToVectorReg(
   unsigned InsertOpcode =
       getInsertOrExtractOpcode(OperationBits, true /*insert*/);
   const MCInstrDesc &InsertMCID = get(InsertOpcode);
+  // `movd` does not support 8/16 bit operands. Instead, we use a 32-bit
+  // register. For example:
+  //   $al = ...
+  //   $xmm0 = MOVPDI2DIrr $eax
+  if (auto NewReg = getMovdCompatibleReg(SrcReg, OperationBits, TRI))
+    SrcReg = *NewReg;
   MachineInstr *InsertMI =
       BuildMI(*MBB, InsertBeforeIt, DL, InsertMCID, DstReg).addReg(SrcReg);
   return InsertMI;
@@ -11079,6 +11119,12 @@ MachineInstr *X86InstrInfo::spill2RegExtractFromVectorReg(
   unsigned ExtractOpcode =
       getInsertOrExtractOpcode(OperationBits, false /*extract*/);
   const MCInstrDesc &ExtractMCID = get(ExtractOpcode);
+  // `movd` does not support 8/16 bit operands. Instead, we use a 32-bit
+  // register. For example:
+  //   $eax = MOVPDI2DIrr $xmm0
+  //   ... = $al
+  if (auto NewReg = getMovdCompatibleReg(DstReg, OperationBits, TRI))
+    DstReg = *NewReg;
   MachineInstr *ExtractMI =
       BuildMI(*InsertMBB, InsertBeforeIt, DL, ExtractMCID, DstReg)
           .addReg(SrcReg);
diff --git a/llvm/lib/Target/X86/X86InstrInfo.h b/llvm/lib/Target/X86/X86InstrInfo.h
@@ -756,6 +756,10 @@ class X86InstrInfo final : public X86GenInstrInfo {
                              const TargetRegisterInfo *TRI,
                              const MachineRegisterInfo *MRI) const override;
 
+  std::optional<MCRegister>
+  getMovdCompatibleReg(MCRegister OldReg, uint32_t OldRegBits,
+                       const TargetRegisterInfo *TRI) const;
+
   MachineInstr *
   spill2RegInsertToVectorReg(Register DstReg, Register SrcReg,
                              int OperationBits, MachineBasicBlock *MBB,
diff --git a/llvm/test/CodeGen/X86/spill2reg_end_to_end_16bit.ll b/llvm/test/CodeGen/X86/spill2reg_end_to_end_16bit.ll
@@ -67,7 +67,7 @@ define dso_local void @_Z5spillv() local_unnamed_addr #0 {
 ; CHECK-NEXT:    .cfi_offset %r14, -32
 ; CHECK-NEXT:    .cfi_offset %r15, -24
 ; CHECK-NEXT:    .cfi_offset %rbp, -16
-; CHECK-NEXT:    movzwl D0(%rip), %eax
+; CHECK-NEXT:    movw D0(%rip), %ax
 ; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
 ; CHECK-NEXT:    movzwl D1(%rip), %ecx
 ; CHECK-NEXT:    movzwl D2(%rip), %edx
@@ -83,12 +83,12 @@ define dso_local void @_Z5spillv() local_unnamed_addr #0 {
 ; CHECK-NEXT:    movzwl D12(%rip), %r15d
 ; CHECK-NEXT:    movzwl D13(%rip), %r12d
 ; CHECK-NEXT:    movzwl D14(%rip), %r13d
-; CHECK-NEXT:    movzwl D15(%rip), %eax
-; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
-; CHECK-NEXT:    movzwl D16(%rip), %eax
-; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
-; CHECK-NEXT:    movzwl D17(%rip), %eax
+; CHECK-NEXT:    movw D15(%rip), %ax
+; CHECK-NEXT:    movd %eax, %xmm0
+; CHECK-NEXT:    movw D16(%rip), %ax
 ; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
+; CHECK-NEXT:    movw D17(%rip), %ax
+; CHECK-NEXT:    movd %eax, %xmm1
 ; CHECK-NEXT:    movzwl D18(%rip), %eax
 ; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
 ; CHECK-NEXT:    #APP
@@ -109,11 +109,11 @@ define dso_local void @_Z5spillv() local_unnamed_addr #0 {
 ; CHECK-NEXT:    movw %r15w, U12(%rip)
 ; CHECK-NEXT:    movw %r12w, U13(%rip)
 ; CHECK-NEXT:    movw %r13w, U14(%rip)
-; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
+; CHECK-NEXT:    movd %xmm0, %eax
 ; CHECK-NEXT:    movw %ax, U15(%rip)
 ; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
 ; CHECK-NEXT:    movw %ax, U16(%rip)
-; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
+; CHECK-NEXT:    movd %xmm1, %eax
 ; CHECK-NEXT:    movw %ax, U17(%rip)
 ; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
 ; CHECK-NEXT:    movw %ax, U18(%rip)
diff --git a/llvm/test/CodeGen/X86/spill2reg_end_to_end_8bit.ll b/llvm/test/CodeGen/X86/spill2reg_end_to_end_8bit.ll
@@ -67,7 +67,7 @@ define dso_local void @_Z5spillv() local_unnamed_addr #0 {
 ; CHECK-NEXT:    .cfi_offset %r14, -32
 ; CHECK-NEXT:    .cfi_offset %r15, -24
 ; CHECK-NEXT:    .cfi_offset %rbp, -16
-; CHECK-NEXT:    movzbl D0(%rip), %eax
+; CHECK-NEXT:    movb D0(%rip), %al
 ; CHECK-NEXT:    movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
 ; CHECK-NEXT:    movzbl D1(%rip), %ecx
 ; CHECK-NEXT:    movzbl D2(%rip), %edx
@@ -83,12 +83,12 @@ define dso_local void @_Z5spillv() local_unnamed_addr #0 {
 ; CHECK-NEXT:    movzbl D12(%rip), %r15d
 ; CHECK-NEXT:    movzbl D13(%rip), %r12d
 ; CHECK-NEXT:    movzbl D14(%rip), %r13d
-; CHECK-NEXT:    movzbl D15(%rip), %eax
-; CHECK-NEXT:    movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
-; CHECK-NEXT:    movzbl D16(%rip), %eax
-; CHECK-NEXT:    movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
-; CHECK-NEXT:    movzbl D17(%rip), %eax
+; CHECK-NEXT:    movb D15(%rip), %al
+; CHECK-NEXT:    movd %eax, %xmm0
+; CHECK-NEXT:    movb D16(%rip), %al
 ; CHECK-NEXT:    movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
+; CHECK-NEXT:    movb D17(%rip), %al
+; CHECK-NEXT:    movd %eax, %xmm1
 ; CHECK-NEXT:    movzbl D18(%rip), %eax
 ; CHECK-NEXT:    movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
 ; CHECK-NEXT:    #APP
@@ -109,11 +109,11 @@ define dso_local void @_Z5spillv() local_unnamed_addr #0 {
 ; CHECK-NEXT:    movb %r15b, U12(%rip)
 ; CHECK-NEXT:    movb %r12b, U13(%rip)
 ; CHECK-NEXT:    movb %r13b, U14(%rip)
-; CHECK-NEXT:    movzbl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 1-byte Folded Reload
+; CHECK-NEXT:    movd %xmm0, %eax
 ; CHECK-NEXT:    movb %al, U15(%rip)
 ; CHECK-NEXT:    movzbl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 1-byte Folded Reload
 ; CHECK-NEXT:    movb %al, U16(%rip)
-; CHECK-NEXT:    movzbl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 1-byte Folded Reload
+; CHECK-NEXT:    movd %xmm1, %eax
 ; CHECK-NEXT:    movb %al, U17(%rip)
 ; CHECK-NEXT:    movzbl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 1-byte Folded Reload
 ; CHECK-NEXT:    movb %al, U18(%rip)
diff --git a/llvm/test/CodeGen/X86/spill2reg_simple_1_16bit.mir b/llvm/test/CodeGen/X86/spill2reg_simple_1_16bit.mir
@@ -0,0 +1,39 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc %s -o - -mtriple=x86_64-unknown-linux -enable-spill2reg -mattr=+sse4.1 --run-pass=spill2reg -simplify-mir -spill2reg-mem-instrs=0 -spill2reg-vec-instrs=99999 | FileCheck %s
+
+# Simple test with a single 16-bit spill-reload pair:
+#   spill stack.0
+#   reload stack.0
+
+--- |
+  @D0 = dso_local local_unnamed_addr global i32 0, align 4
+  @U0 = dso_local local_unnamed_addr global i32 0, align 4
+  define void @func() { ret void }
+...
+---
+name: func
+alignment:       16
+tracksRegLiveness: true
+tracksDebugUserValues: true
+frameInfo:
+  maxAlignment:    4
+stack:
+  - { id: 0, type: spill-slot, size: 4, alignment: 4 }
+machineFunctionInfo: {}
+body:             |
+
+  bb.0:
+    ; spill
+    ; CHECK-LABEL: name: func
+    ; CHECK: $ax = MOV16rm $rip, 1, $noreg, @D0, $noreg :: (dereferenceable load (s16) from @D0)
+    ; CHECK-NEXT: $xmm0 = MOVDI2PDIrr $eax
+    ; CHECK-NEXT: $eax = MOVPDI2DIrr $xmm0
+    ; CHECK-NEXT: MOV16mr $rip, 1, $noreg, @U0, $noreg, killed renamable $ax :: (store (s16) into @U0)
+    ; CHECK-NEXT: RET 0
+    $ax = MOV16rm $rip, 1, $noreg, @D0, $noreg :: (dereferenceable load (s16) from @D0)
+    MOV16mr %stack.0, 1, $noreg, 0, $noreg, killed renamable $ax :: (store (s16) into %stack.0)
+    ; reload
+    $ax = MOV16rm %stack.0, 1, $noreg, 0, $noreg :: (load (s16) from %stack.0)
+    MOV16mr $rip, 1, $noreg, @U0, $noreg, killed renamable $ax :: (store (s16) into @U0)
+    RET 0
+...
diff --git a/llvm/test/CodeGen/X86/spill2reg_simple_1_8bit.mir b/llvm/test/CodeGen/X86/spill2reg_simple_1_8bit.mir
@@ -0,0 +1,39 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc %s -o - -mtriple=x86_64-unknown-linux -enable-spill2reg -mattr=+sse4.1 --run-pass=spill2reg -simplify-mir -spill2reg-mem-instrs=0 -spill2reg-vec-instrs=99999 | FileCheck %s
+
+# Simple test with a single 8-bit spill-reload pair:
+#   spill stack.0
+#   reload stack.0
+
+--- |
+  @D0 = dso_local local_unnamed_addr global i32 0, align 4
+  @U0 = dso_local local_unnamed_addr global i32 0, align 4
+  define void @func() { ret void }
+...
+---
+name: func
+alignment:       16
+tracksRegLiveness: true
+tracksDebugUserValues: true
+frameInfo:
+  maxAlignment:    4
+stack:
+  - { id: 0, type: spill-slot, size: 4, alignment: 4 }
+machineFunctionInfo: {}
+body:             |
+
+  bb.0:
+    ; spill
+    ; CHECK-LABEL: name: func
+    ; CHECK: $al = MOV8rm $rip, 1, $noreg, @D0, $noreg :: (dereferenceable load (s8) from @D0)
+    ; CHECK-NEXT: $xmm0 = MOVDI2PDIrr $eax
+    ; CHECK-NEXT: $eax = MOVPDI2DIrr $xmm0
+    ; CHECK-NEXT: MOV8mr $rip, 1, $noreg, @U0, $noreg, killed renamable $al :: (store (s8) into @U0)
+    ; CHECK-NEXT: RET 0
+    $al = MOV8rm $rip, 1, $noreg, @D0, $noreg :: (dereferenceable load (s8) from @D0)
+    MOV8mr %stack.0, 1, $noreg, 0, $noreg, killed renamable $al :: (store (s8) into %stack.0)
+    ; reload
+    $al = MOV8rm %stack.0, 1, $noreg, 0, $noreg :: (load (s8) from %stack.0)
+    MOV8mr $rip, 1, $noreg, @U0, $noreg, killed renamable $al :: (store (s8) into @U0)
+    RET 0
+...
