[LoongArch] Insert nops and emit align reloc when handle alignment directive (#72962)

Refer to RISCV, we will fix up the alignment if linker relaxation
changes code size and breaks alignment. Insert enough Nops and emit
R_LARCH_ALIGN relocation type so that linker could satisfy the alignment
by removing Nops.
It does so only in sections with the SHF_EXECINSTR flag.

In LoongArch psABI v2.30, R_LARCH_ALIGN requires symbol index. The
lowest 8 bits of addend represent alignment and the other bits of addend
represent the maximum number of bytes to emit.

Author: MQ-mengqing

llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchAsmBackend.cpp

@@ -17,9 +17,12 @@
 #include "llvm/MC/MCAssembler.h"
 #include "llvm/MC/MCContext.h"
 #include "llvm/MC/MCELFObjectWriter.h"
+#include "llvm/MC/MCExpr.h"
+#include "llvm/MC/MCSection.h"
 #include "llvm/MC/MCValue.h"
 #include "llvm/Support/EndianStream.h"
 #include "llvm/Support/LEB128.h"
+#include "llvm/Support/MathExtras.h"
 
 #define DEBUG_TYPE "loongarch-asmbackend"
 
@@ -176,6 +179,70 @@ void LoongArchAsmBackend::applyFixup(const MCAssembler &Asm,
   }
 }
 
+// Linker relaxation may change code size. We have to insert Nops
+// for .align directive when linker relaxation enabled. So then Linker
+// could satisfy alignment by removing Nops.
+// The function returns the total Nops Size we need to insert.
+bool LoongArchAsmBackend::shouldInsertExtraNopBytesForCodeAlign(
+    const MCAlignFragment &AF, unsigned &Size) {
+  // Calculate Nops Size only when linker relaxation enabled.
+  if (!AF.getSubtargetInfo()->hasFeature(LoongArch::FeatureRelax))
+    return false;
+
+  // Ignore alignment if MaxBytesToEmit is less than the minimum Nop size.
+  const unsigned MinNopLen = 4;
+  if (AF.getMaxBytesToEmit() < MinNopLen)
+    return false;
+  Size = AF.getAlignment().value() - MinNopLen;
+  return AF.getAlignment() > MinNopLen;
+}
+
+// We need to insert R_LARCH_ALIGN relocation type to indicate the
+// position of Nops and the total bytes of the Nops have been inserted
+// when linker relaxation enabled.
+// The function inserts fixup_loongarch_align fixup which eventually will
+// transfer to R_LARCH_ALIGN relocation type.
+// The improved R_LARCH_ALIGN requires symbol index. The lowest 8 bits of
+// addend represent alignment and the other bits of addend represent the
+// maximum number of bytes to emit. The maximum number of bytes is zero
+// means ignore the emit limit.
+bool LoongArchAsmBackend::shouldInsertFixupForCodeAlign(
+    MCAssembler &Asm, const MCAsmLayout &Layout, MCAlignFragment &AF) {
+  // Insert the fixup only when linker relaxation enabled.
+  if (!AF.getSubtargetInfo()->hasFeature(LoongArch::FeatureRelax))
+    return false;
+
+  // Calculate total Nops we need to insert. If there are none to insert
+  // then simply return.
+  unsigned Count;
+  if (!shouldInsertExtraNopBytesForCodeAlign(AF, Count))
+    return false;
+
+  MCSection *Sec = AF.getParent();
+  MCContext &Ctx = Asm.getContext();
+  const MCExpr *Dummy = MCConstantExpr::create(0, Ctx);
+  // Create fixup_loongarch_align fixup.
+  MCFixup Fixup =
+      MCFixup::create(0, Dummy, MCFixupKind(LoongArch::fixup_loongarch_align));
+  const MCSymbolRefExpr *MCSym = getSecToAlignSym()[Sec];
+  if (MCSym == nullptr) {
+    // Create a symbol and make the value of symbol is zero.
+    MCSymbol *Sym = Ctx.createNamedTempSymbol("la-relax-align");
+    Sym->setFragment(&*Sec->getBeginSymbol()->getFragment());
+    Asm.registerSymbol(*Sym);
+    MCSym = MCSymbolRefExpr::create(Sym, Ctx);
+    getSecToAlignSym()[Sec] = MCSym;
+  }
+
+  uint64_t FixedValue = 0;
+  unsigned Lo = Log2_64(Count) + 1;
+  unsigned Hi = AF.getMaxBytesToEmit() >= Count ? 0 : AF.getMaxBytesToEmit();
+  MCValue Value = MCValue::get(MCSym, nullptr, Hi << 8 | Lo);
+  Asm.getWriter().recordRelocation(Asm, Layout, &AF, Fixup, Value, FixedValue);
+
+  return true;
+}
+
 bool LoongArchAsmBackend::shouldForceRelocation(const MCAssembler &Asm,
                                                 const MCFixup &Fixup,
                                                 const MCValue &Target,

llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchAsmBackend.h

@@ -17,7 +17,9 @@
 #include "MCTargetDesc/LoongArchFixupKinds.h"
 #include "MCTargetDesc/LoongArchMCTargetDesc.h"
 #include "llvm/MC/MCAsmBackend.h"
+#include "llvm/MC/MCExpr.h"
 #include "llvm/MC/MCFixupKindInfo.h"
+#include "llvm/MC/MCSection.h"
 #include "llvm/MC/MCSubtargetInfo.h"
 
 namespace llvm {
@@ -27,6 +29,7 @@ class LoongArchAsmBackend : public MCAsmBackend {
   uint8_t OSABI;
   bool Is64Bit;
   const MCTargetOptions &TargetOptions;
+  DenseMap<MCSection *, const MCSymbolRefExpr *> SecToAlignSym;
 
 public:
   LoongArchAsmBackend(const MCSubtargetInfo &STI, uint8_t OSABI, bool Is64Bit,
@@ -45,6 +48,15 @@ class LoongArchAsmBackend : public MCAsmBackend {
                   uint64_t Value, bool IsResolved,
                   const MCSubtargetInfo *STI) const override;
 
+  // Return Size with extra Nop Bytes for alignment directive in code section.
+  bool shouldInsertExtraNopBytesForCodeAlign(const MCAlignFragment &AF,
+                                             unsigned &Size) override;
+
+  // Insert target specific fixup type for alignment directive in code section.
+  bool shouldInsertFixupForCodeAlign(MCAssembler &Asm,
+                                     const MCAsmLayout &Layout,
+                                     MCAlignFragment &AF) override;
+
   bool shouldForceRelocation(const MCAssembler &Asm, const MCFixup &Fixup,
                              const MCValue &Target,
                              const MCSubtargetInfo *STI) override;
@@ -80,6 +92,9 @@ class LoongArchAsmBackend : public MCAsmBackend {
   std::unique_ptr<MCObjectTargetWriter>
   createObjectTargetWriter() const override;
   const MCTargetOptions &getTargetOptions() const { return TargetOptions; }
+  DenseMap<MCSection *, const MCSymbolRefExpr *> &getSecToAlignSym() {
+    return SecToAlignSym;
+  }
 };
 } // end namespace llvm
 

llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchFixupKinds.h

@@ -109,6 +109,8 @@ enum Fixups {
   fixup_loongarch_tls_gd_hi20,
   // Generate an R_LARCH_RELAX which indicates the linker may relax here.
   fixup_loongarch_relax = FirstLiteralRelocationKind + ELF::R_LARCH_RELAX,
+  // Generate an R_LARCH_ALIGN which indicates the linker may fixup align here.
+  fixup_loongarch_align = FirstLiteralRelocationKind + ELF::R_LARCH_ALIGN,
   // 36-bit fixup corresponding to %call36(foo) for a pair instructions:
   // pcaddu18i+jirl.
   fixup_loongarch_call36 = FirstLiteralRelocationKind + ELF::R_LARCH_CALL36,

llvm/test/MC/LoongArch/Relocations/align-non-executable.s

@@ -0,0 +1,27 @@
+## A label difference separated by an alignment directive, when the
+## referenced symbols are in a non-executable section with instructions,
+## should generate ADD/SUB relocations.
+## https://github.com/llvm/llvm-project/pull/76552
+
+# RUN: llvm-mc --filetype=obj --triple=loongarch64 --mattr=+relax %s \
+# RUN:     | llvm-readobj -r - | FileCheck --check-prefixes=CHECK,RELAX %s
+# RUN: llvm-mc --filetype=obj --triple=loongarch64 --mattr=-relax %s \
+# RUN:     | llvm-readobj -r - | FileCheck %s
+
+.section ".dummy", "a"
+.L1:
+  la.pcrel $t0, sym
+.p2align 3
+.L2:
+.dword .L2 - .L1
+
+# CHECK:       Relocations [
+# CHECK-NEXT:    Section ({{.*}}) .rela.dummy {
+# CHECK-NEXT:      0x0 R_LARCH_PCALA_HI20 sym 0x0
+# RELAX-NEXT:      0x0 R_LARCH_RELAX - 0x0
+# CHECK-NEXT:      0x4 R_LARCH_PCALA_LO12 sym 0x0
+# RELAX-NEXT:      0x4 R_LARCH_RELAX - 0x0
+# RELAX-NEXT:      0x8 R_LARCH_ADD64 .L2 0x0
+# RELAX-NEXT:      0x8 R_LARCH_SUB64 .L1 0x0
+# CHECK-NEXT:    }
+# CHECK-NEXT:  ]

llvm/test/MC/LoongArch/Relocations/relax-addsub.s

@@ -28,12 +28,23 @@
 
 # RELAX:       Relocations [
 # RELAX-NEXT:    Section ({{.*}}) .rela.text {
+# RELAX-NEXT:      0x4 R_LARCH_ALIGN {{.*}} 0x4
 # RELAX-NEXT:      0x10 R_LARCH_PCALA_HI20 .L1 0x0
 # RELAX-NEXT:      0x10 R_LARCH_RELAX - 0x0
 # RELAX-NEXT:      0x14 R_LARCH_PCALA_LO12 .L1 0x0
 # RELAX-NEXT:      0x14 R_LARCH_RELAX - 0x0
 # RELAX-NEXT:    }
 # RELAX-NEXT:    Section ({{.*}}) .rela.data {
+# RELAX-NEXT:      0x10 R_LARCH_ADD8 .L3 0x0
+# RELAX-NEXT:      0x10 R_LARCH_SUB8 .L2 0x0
+# RELAX-NEXT:      0x11 R_LARCH_ADD16 .L3 0x0
+# RELAX-NEXT:      0x11 R_LARCH_SUB16 .L2 0x0
+# RELAX-NEXT:      0x13 R_LARCH_ADD32 .L3 0x0
+# RELAX-NEXT:      0x13 R_LARCH_SUB32 .L2 0x0
+# RELAX-NEXT:      0x17 R_LARCH_ADD64 .L3 0x0
+# RELAX-NEXT:      0x17 R_LARCH_SUB64 .L2 0x0
+# RELAX-NEXT:      0x1F R_LARCH_ADD_ULEB128 .L3 0x0
+# RELAX-NEXT:      0x1F R_LARCH_SUB_ULEB128 .L2 0x0
 # RELAX-NEXT:      0x20 R_LARCH_ADD8 .L4 0x0
 # RELAX-NEXT:      0x20 R_LARCH_SUB8 .L3 0x0
 # RELAX-NEXT:      0x21 R_LARCH_ADD16 .L4 0x0
@@ -57,7 +68,7 @@
 
 # RELAX:      Hex dump of section '.data':
 # RELAX-NEXT: 0x00000000 04040004 00000004 00000000 00000004
-# RELAX-NEXT: 0x00000010 0c0c000c 0000000c 00000000 0000000c
+# RELAX-NEXT: 0x00000010 00000000 00000000 00000000 00000000
 # RELAX-NEXT: 0x00000020 00000000 00000000 00000000 00000000
 # RELAX-NEXT: 0x00000030 00000000 00000000 00000000 000000
 
@@ -78,7 +89,7 @@
 .word  .L2 - .L1
 .dword .L2 - .L1
 .uleb128 .L2 - .L1
-## TODO Handle alignment directive.
+## With relaxation, emit relocs because the .align makes the diff variable.
 .byte  .L3 - .L2
 .short .L3 - .L2
 .word  .L3 - .L2

llvm/test/MC/LoongArch/Relocations/relax-align.s

@@ -0,0 +1,79 @@
+## The file testing Nop insertion with R_LARCH_ALIGN for relaxation.
+
+# RUN: llvm-mc --filetype=obj --triple=loongarch64 --mattr=-relax %s -o %t
+# RUN: llvm-objdump -d %t | FileCheck %s --check-prefix=INSTR
+# RUN: llvm-readobj -r %t | FileCheck %s --check-prefix=RELOC
+# RUN: llvm-mc --filetype=obj --triple=loongarch64 --mattr=+relax %s -o %t.r
+# RUN: llvm-objdump -d %t.r | FileCheck %s --check-prefixes=INSTR,RELAX-INSTR
+# RUN: llvm-readobj -r %t.r | FileCheck %s --check-prefixes=RELOC,RELAX-RELOC
+
+.text
+break 0
+# INSTR: break 0
+
+## Not emit R_LARCH_ALIGN if alignment directive is less than or equal to
+## minimum code alignment(a.k.a 4).
+.p2align 2
+.p2align 1
+.p2align 0
+
+## Not emit instructions if max emit bytes less than min nop size.
+.p2align 4, , 2
+
+## Not emit R_LARCH_ALIGN if alignment directive with specific padding value.
+## The behavior is the same as GNU assembler.
+break 1
+.p2align 4, 1
+# INSTR-NEXT:    break 1
+# INSTR-COUNT-2: 01 01 01 01
+
+break 2
+.p2align 4, 1, 12
+# INSTR-NEXT:    break 2
+# INSTR-COUNT-3: 01 01 01 01
+
+break 3
+.p2align 4
+# INSTR-NEXT:    break 3
+# INSTR-COUNT-3: nop
+
+break 4
+.p2align 5
+.p2align 4
+# INSTR-NEXT:          break 4
+# INSTR-COUNT-3:       nop
+# RELAX-INSTR-COUNT-7: nop
+
+break 5
+.p2align 4, , 11
+# INSTR-NEXT: break 5
+# RELAX-INSTR-COUNT-3: nop
+
+break 6
+## Not emit the third parameter.
+.p2align 4, , 12
+# INSTR-NEXT:       break 6
+# INSTR-NEXT:       nop
+# INSTR-NEXT:       nop
+# RELAX-INSTR-NEXT: nop
+
+ret
+# INSNR-NEXT: ret
+
+## Test the symbol index is different from .text.
+.section .text2, "ax"
+.p2align 4
+break 7
+
+# RELOC:            Relocations [
+# RELAX-RELOC-NEXT:   Section ({{.*}}) .rela.text {
+# RELAX-RELOC-NEXT:     0x24 R_LARCH_ALIGN .Lla-relax-align0 0x4
+# RELAX-RELOC-NEXT:     0x34 R_LARCH_ALIGN .Lla-relax-align0 0x5
+# RELAX-RELOC-NEXT:     0x50 R_LARCH_ALIGN .Lla-relax-align0 0x4
+# RELAX-RELOC-NEXT:     0x60 R_LARCH_ALIGN .Lla-relax-align0 0xB04
+# RELAX-RELOC-NEXT:     0x70 R_LARCH_ALIGN .Lla-relax-align0 0x4
+# RELAX-RELOC-NEXT:   }
+# RELAX-RELOC-NEXT:   Section ({{.*}}) .rela.text2 {
+# RELAX-RELOC-NEXT:     0x0 R_LARCH_ALIGN .Lla-relax-align1 0x4
+# RELAX-RELOC-NEXT:   }
+# RELOC-NEXT:       ]