[X86] Add assembler support for {disp8} and {disp32} to control the size of displacement used for memory operands.

[craig.topper]

Summary

These prefixes should override the default behavior and force a larger immediate size. I don't believe gas issues any warning if you use {disp8} when a 32-bit displacement is already required. And this patch doesn't either.

This completes the {disp8} and {disp32} support from PR46650.

[echristo]

I think I'd find the conditionals easier to "read" if they were positive rather than negative? It'd help if the comments spelled out the conditions a little more as well.

Thoughts?

[craig.topper]

Add some more variables to make the conditionals positive.
Try to improve some of the comments.

[RKSimon]

LGTM - please can you deal with the the double spaces that clang-format is warning about?

llvm/lib/Target/X86/AsmParser/X86AsmParser.cpp

bool X86AsmParser::MatchAndEmitATTInstruction(SMLoc IDLoc, unsigned &Opcode,
 
   MCInst Inst;
 
   // If VEX3 encoding is forced, we need to pass the USE_VEX3 flag to the
   // encoder.
   if (ForcedVEXEncoding == VEXEncoding_VEX3)
     Prefixes |= X86::IP_USE_VEX3;
 
+  // Set encoded flags for {disp8} and {disp32}.
+  if (ForcedDispEncoding == DispEncoding_Disp8)
+    Prefixes |= X86::IP_USE_DISP8;
+  else if (ForcedDispEncoding == DispEncoding_Disp32)
+    Prefixes |= X86::IP_USE_DISP32;
+
   if (Prefixes)
     Inst.setFlags(Prefixes);
 
   // First, try a direct match.
   FeatureBitset MissingFeatures;
   unsigned OriginalError = MatchInstruction(Operands, Inst, ErrorInfo,
                                             MissingFeatures, MatchingInlineAsm,
                                             isParsingIntelSyntax());
bool X86AsmParser::MatchAndEmitIntelInstruction(SMLoc IDLoc, unsigned &Opcode,
 
   MCInst Inst;
 
   // If VEX3 encoding is forced, we need to pass the USE_VEX3 flag to the
   // encoder.
   if (ForcedVEXEncoding == VEXEncoding_VEX3)
     Prefixes |= X86::IP_USE_VEX3;
 
+  // Set encoded flags for {disp8} and {disp32}.
+  if (ForcedDispEncoding == DispEncoding_Disp8)
+    Prefixes |= X86::IP_USE_DISP8;
+  else if (ForcedDispEncoding == DispEncoding_Disp32)
+    Prefixes |= X86::IP_USE_DISP32;
+
   if (Prefixes)
     Inst.setFlags(Prefixes);
 
   // Find one unsized memory operand, if present.
   X86Operand *UnsizedMemOp = nullptr;
   for (const auto &Op : Operands) {
     X86Operand *X86Op = static_cast<X86Operand *>(Op.get());
     if (X86Op->isMemUnsized()) {

llvm/lib/Target/X86/MCTargetDesc/X86BaseInfo.h

enum IPREFIXES {
     IP_NO_PREFIX = 0,
     IP_HAS_OP_SIZE = 1,
     IP_HAS_AD_SIZE = 2,
     IP_HAS_REPEAT_NE = 4,
     IP_HAS_REPEAT = 8,
     IP_HAS_LOCK = 16,
     IP_HAS_NOTRACK = 32,
     IP_USE_VEX3 = 64,
+    IP_USE_DISP8 = 128,
+    IP_USE_DISP32 = 256,
   };
 
   enum OperandType : unsigned {
     /// AVX512 embedded rounding control. This should only have values 0-3.
     OPERAND_ROUNDING_CONTROL = MCOI::OPERAND_FIRST_TARGET,
     OPERAND_COND_CODE,
   };
 

llvm/lib/Target/X86/MCTargetDesc/X86MCCodeEmitter.cpp

if (BaseReg) {
       emitByte(modRMByte(0, RegOpcodeField, 6), OS);
     }
 
     // Emit 16-bit displacement for plain disp16 or [REG]+disp16 cases.
     emitImmediate(Disp, MI.getLoc(), 2, FK_Data_2, StartByte, OS, Fixups);
     return;
   }
 
-  // Determine whether a SIB byte is needed.
-  // If no BaseReg, issue a RIP relative instruction only if the MCE can
-  // resolve addresses on-the-fly, otherwise use SIB (Intel Manual 2A, table
-  // 2-7) and absolute references.
+  // Check for presence of {disp8} or {disp32} pseudo prefixes.
+  bool UseDisp8 = MI.getFlags() & X86::IP_USE_DISP8;
+  bool UseDisp32 = MI.getFlags() & X86::IP_USE_DISP32;
+
+  // We only allow no displacement if no pseudo prefix is present.
+  bool AllowNoDisp = !UseDisp8 && !UseDisp32;
+  // Disp8 is allowed unless the {disp32} prefix is present.
+  bool AllowDisp8 = !UseDisp32;
 
-  if ( // The SIB byte must be used if there is an index register.
+  // Determine whether a SIB byte is needed.
+  if (// The SIB byte must be used if there is an index register or the
+      // encoding requires a SIB byte.
       !ForceSIB && IndexReg.getReg() == 0 &&
       // The SIB byte must be used if the base is ESP/RSP/R12, all of which
       // encode to an R/M value of 4, which indicates that a SIB byte is
       // present.
       BaseRegNo != N86::ESP &&
       // If there is no base register and we're in 64-bit mode, we need a SIB
       // byte to emit an addr that is just 'disp32' (the non-RIP relative form).
       (!STI.hasFeature(X86::Mode64Bit) || BaseReg != 0)) {
 
     if (BaseReg == 0) { // [disp32]     in X86-32 mode
       emitByte(modRMByte(0, RegOpcodeField, 5), OS);
       emitImmediate(Disp, MI.getLoc(), 4, FK_Data_4, StartByte, OS, Fixups);
       return;
     }
 
-    // If the base is not EBP/ESP and there is no displacement, use simple
-    // indirect register encoding, this handles addresses like [EAX].  The
-    // encoding for [EBP] with no displacement means [disp32] so we handle it
-    // by emitting a displacement of 0 below.
+    // If the base is not EBP/ESP/R12/R13 and there is no displacement, use
+    // simple indirect register encoding, this handles addresses like [EAX].
+    // The encoding for [EBP] or[R13] with no displacement means [disp32] so we
+    // handle it by emitting a displacement of 0 later.
     if (BaseRegNo != N86::EBP) {
-      if (Disp.isImm() && Disp.getImm() == 0) {
+      if (Disp.isImm() && Disp.getImm() == 0 && AllowNoDisp) {
         emitByte(modRMByte(0, RegOpcodeField, BaseRegNo), OS);
         return;
       }
 
       // If the displacement is @tlscall, treat it as a zero.
       if (Disp.isExpr()) {
         auto *Sym = dyn_cast<MCSymbolRefExpr>(Disp.getExpr());
         if (Sym && Sym->getKind() == MCSymbolRefExpr::VK_TLSCALL) {
           // This is exclusively used by call *a@tlscall(base). The relocation
           // (R_386_TLSCALL or R_X86_64_TLSCALL) applies to the beginning.
           Fixups.push_back(MCFixup::create(0, Sym, FK_NONE, MI.getLoc()));
           emitByte(modRMByte(0, RegOpcodeField, BaseRegNo), OS);
           return;
         }
       }
     }
 
     // Otherwise, if the displacement fits in a byte, encode as [REG+disp8].
-    if (Disp.isImm()) {
+    // Including a compressed disp8 for EVEX instructions that support it.
+    // This also handles the 0 displacement for [EBP] or [R13]. We can't use
+    // disp8 if the {disp32} pseudo prefix is present.
+    if (Disp.isImm() && AllowDisp8) {
       int ImmOffset = 0;
       if (isDispOrCDisp8(TSFlags, Disp.getImm(), ImmOffset)) {
         emitByte(modRMByte(1, RegOpcodeField, BaseRegNo), OS);
         emitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, StartByte, OS, Fixups,
                       ImmOffset);
         return;
       }
     }
 
-    // Otherwise, emit the most general non-SIB encoding: [REG+disp32]
+    // Otherwise, emit the most general non-SIB encoding: [REG+disp32].
+    // Displacement may be 0 for [EBP] or [R13] case if {disp32} pseudo prefix
+    // prevented using disp8 above.
     emitByte(modRMByte(2, RegOpcodeField, BaseRegNo), OS);
     unsigned Opcode = MI.getOpcode();
     unsigned FixupKind = Opcode == X86::MOV32rm ? X86::reloc_signed_4byte_relax
                                                 : X86::reloc_signed_4byte;
     emitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(FixupKind), StartByte, OS,
                   Fixups);
     return;
   }
 
   // We need a SIB byte, so start by outputting the ModR/M byte first
   assert(IndexReg.getReg() != X86::ESP && IndexReg.getReg() != X86::RSP &&
          "Cannot use ESP as index reg!");
 
   bool ForceDisp32 = false;
   bool ForceDisp8 = false;
   int ImmOffset = 0;
   if (BaseReg == 0) {
     // If there is no base register, we emit the special case SIB byte with
     // MOD=0, BASE=5, to JUST get the index, scale, and displacement.
+    BaseRegNo = 5;
     emitByte(modRMByte(0, RegOpcodeField, 4), OS);
     ForceDisp32 = true;
-  } else if (Disp.isImm() && Disp.getImm() == 0 &&
-             // Base reg can't be anything that ends up with '5' as the base
-             // reg, it is the magic [*] nomenclature that indicates no base.
+  } else if (Disp.isImm() && Disp.getImm() == 0 && AllowNoDisp &&
+             // Base reg can't be EBP/RBP/R13 as that would end up with '5' as
+             // the base field, but that is the magic [*] nomenclature that
+             // indicates no base when mod=0. For these cases we'll emit a 0
+             // displacement instead.
              BaseRegNo != N86::EBP) {
     // Emit no displacement ModR/M byte
     emitByte(modRMByte(0, RegOpcodeField, 4), OS);
-  } else if (Disp.isImm() &&
+  } else if (Disp.isImm() && AllowDisp8 &&
              isDispOrCDisp8(TSFlags, Disp.getImm(), ImmOffset)) {
-    // Emit the disp8 encoding.
+    // Displacement fits in a byte or matches an EVEX compressed disp8, use
+    // disp8 encoding. This also handles EBP/R13 base with 0 displacement unless
+    // {disp32} pseudo prefix was used.
     emitByte(modRMByte(1, RegOpcodeField, 4), OS);
-    ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
+    ForceDisp8 = true;
   } else {
-    // Emit the normal disp32 encoding.
+    // Otherwise, emit the normal disp32 encoding.
     emitByte(modRMByte(2, RegOpcodeField, 4), OS);
     ForceDisp32 = true;
   }
 
   // Calculate what the SS field value should be...
   static const unsigned SSTable[] = {~0U, 0, 1, ~0U, 2, ~0U, ~0U, ~0U, 3};
   unsigned SS = SSTable[Scale.getImm()];
 
   unsigned IndexRegNo = IndexReg.getReg() ? getX86RegNum(IndexReg) : 4;
 
-  // Handle the SIB byte for the case where there is no base, see Intel
-  // Manual 2A, table 2-7. The displacement has already been output.
-  if (BaseReg == 0)
-    BaseRegNo = 5;
-
   emitSIBByte(SS, IndexRegNo, BaseRegNo, OS);
 
   // Do we need to output a displacement?
   if (ForceDisp8)
     emitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, StartByte, OS, Fixups,
                   ImmOffset);
   else if (ForceDisp32)
     emitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(X86::reloc_signed_4byte),

llvm/test/MC/X86/x86-64.s

 
 // CHECK: ud1q %rdx, %rdi
 // CHECK:  encoding: [0x48,0x0f,0xb9,0xfa]
 ud1 %rdx, %rdi
 
 // CHECK: ud1q (%rbx), %rcx
 // CHECK:  encoding: [0x48,0x0f,0xb9,0x0b]
 ud2b (%rbx), %rcx
+
+// Requires no displacement by default
+// CHECK: movl $1, (%rax)
+// CHECK: encoding: [0xc7,0x00,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, (%rax)
+// CHECK: encoding: [0xc7,0x40,0x00,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, (%rax)
+// CHECK: encoding: [0xc7,0x80,0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x00]
+movl $1, (%rax)
+{disp8} movl $1, (%rax)
+{disp32} movl $1, (%rax)
+
+// Requires disp8 by default
+// CHECK: movl $1, (%rbp)
+// CHECK: encoding: [0xc7,0x45,0x00,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, (%rbp)
+// CHECK: encoding: [0xc7,0x45,0x00,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, (%rbp)
+// CHECK: encoding: [0xc7,0x85,0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x00]
+movl $1, (%rbp)
+{disp8} movl $1, (%rbp)
+{disp32} movl $1, (%rbp)
+
+// Requires disp8 by default
+// CHECK: movl $1, (%r13)
+// CHECK: encoding: [0x41,0xc7,0x45,0x00,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, (%r13)
+// CHECK: encoding: [0x41,0xc7,0x45,0x00,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, (%r13)
+// CHECK: encoding: [0x41,0xc7,0x85,0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x00]
+movl $1, (%r13)
+{disp8} movl $1, (%r13)
+{disp32} movl $1, (%r13)
+
+// Requires disp8 by default
+// CHECK: movl $1, 8(%rax)
+// CHECK: encoding: [0xc7,0x40,0x08,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, 8(%rax)
+// CHECK: encoding: [0xc7,0x40,0x08,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, 8(%rax)
+// CHECK: encoding: [0xc7,0x80,0x08,0x00,0x00,0x00,0x01,0x00,0x00,0x00]
+movl $1, 8(%rax)
+{disp8} movl $1, 8(%rax)
+{disp32} movl $1, 8(%rax)
+
+// Requires no displacement by default
+// CHECK: movl $1, (%rax,%rbx,4)
+// CHECK: encoding: [0xc7,0x04,0x98,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, (%rax,%rbx,4)
+// CHECK: encoding: [0xc7,0x44,0x98,0x00,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, (%rax,%rbx,4)
+// CHECK: encoding: [0xc7,0x84,0x98,0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x00]
+movl $1, (%rax,%rbx,4)
+{disp8} movl $1, (%rax,%rbx,4)
+{disp32} movl $1, (%rax,%rbx,4)
+
+// Requires disp8 by default.
+// CHECK: movl $1, 8(%rax,%rbx,4)
+// CHECK: encoding: [0xc7,0x44,0x98,0x08,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, 8(%rax,%rbx,4)
+// CHECK: encoding: [0xc7,0x44,0x98,0x08,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, 8(%rax,%rbx,4)
+// CHECK: encoding: [0xc7,0x84,0x98,0x08,0x00,0x00,0x00,0x01,0x00,0x00,0x00]
+movl $1, 8(%rax,%rbx,4)
+{disp8} movl $1, 8(%rax,%rbx,4)
+{disp32} movl $1, 8(%rax,%rbx,4)
+
+// Requires disp8 by default.
+// CHECK: movl $1, (%rbp,%rbx,4)
+// CHECK: encoding: [0xc7,0x44,0x9d,0x00,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, (%rbp,%rbx,4)
+// CHECK: encoding: [0xc7,0x44,0x9d,0x00,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, (%rbp,%rbx,4)
+// CHECK: encoding: [0xc7,0x84,0x9d,0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x00]
+movl $1, (%rbp,%rbx,4)
+{disp8} movl $1, (%rbp,%rbx,4)
+{disp32} movl $1, (%rbp,%rbx,4)
+
+// Requires disp8 by default.
+// CHECK: movl $1, (%r13,%rbx,4)
+// CHECK: encoding: [0x41,0xc7,0x44,0x9d,0x00,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, (%r13,%rbx,4)
+// CHECK: encoding: [0x41,0xc7,0x44,0x9d,0x00,0x01,0x00,0x00,0x00]
+// CHECK: movl $1, (%r13,%rbx,4)
+// CHECK: encoding: [0x41,0xc7,0x84,0x9d,0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x00]
+movl $1, (%r13,%rbx,4)
+{disp8} movl $1, (%r13,%rbx,4)
+{disp32} movl $1, (%r13,%rbx,4)