[X86][MC] Support Enc/Dec for NF BMI instructions (#76709)

Promoted BMI instructions were supported in #73899

Author: XinWang10

llvm/lib/Target/X86/Disassembler/X86Disassembler.cpp

@@ -1134,6 +1134,27 @@ static int getInstructionIDWithAttrMask(uint16_t *instructionID,
   return 0;
 }
 
+static bool isNF(InternalInstruction *insn) {
+  if (!nfFromEVEX4of4(insn->vectorExtensionPrefix[3]))
+    return false;
+  if (insn->opcodeType == MAP4)
+    return true;
+  // Below NF instructions are not in map4.
+  if (insn->opcodeType == THREEBYTE_38 &&
+      ppFromEVEX3of4(insn->vectorExtensionPrefix[2]) == VEX_PREFIX_NONE) {
+    switch (insn->opcode) {
+    case 0xf2: // ANDN
+    case 0xf3: // BLSI, BLSR, BLSMSK
+    case 0xf5: // BZHI
+    case 0xf7: // BEXTR
+      return true;
+    default:
+      break;
+    }
+  }
+  return false;
+}
+
 // Determine the ID of an instruction, consuming the ModR/M byte as appropriate
 // for extended and escape opcodes. Determines the attributes and context for
 // the instruction before doing so.
@@ -1169,9 +1190,7 @@ static int getInstructionID(struct InternalInstruction *insn,
         attrMask |= ATTR_EVEXKZ;
       if (bFromEVEX4of4(insn->vectorExtensionPrefix[3]))
         attrMask |= ATTR_EVEXB;
-      // nf bit is the MSB of aaa
-      if (nfFromEVEX4of4(insn->vectorExtensionPrefix[3]) &&
-          insn->opcodeType == MAP4)
+      if (isNF(insn)) // NF bit is the MSB of aaa.
         attrMask |= ATTR_EVEXNF;
       else if (aaaFromEVEX4of4(insn->vectorExtensionPrefix[3]))
         attrMask |= ATTR_EVEXK;

llvm/lib/Target/X86/X86InstrArithmetic.td

@@ -1312,30 +1312,30 @@ def : Pat<(X86testpat (loadi64 addr:$src1), i64relocImmSExt32_su:$src2),
 //===----------------------------------------------------------------------===//
 // ANDN Instruction
 //
-multiclass AndN<X86TypeInfo t, string suffix> {
+multiclass AndN<X86TypeInfo t, SDPatternOperator node, string suffix = ""> {
   defvar andn_rr_p =
-    [(set t.RegClass:$dst, EFLAGS, (X86and_flag (not t.RegClass:$src1),
+    [(set t.RegClass:$dst, EFLAGS, (node (not t.RegClass:$src1),
      t.RegClass:$src2))];
   defvar andn_rm_p =
-    [(set t.RegClass:$dst, EFLAGS, (X86and_flag (not t.RegClass:$src1),
+    [(set t.RegClass:$dst, EFLAGS, (node (not t.RegClass:$src1),
      (t.LoadNode addr:$src2)))];
   def rr#suffix : ITy<0xF2, MRMSrcReg, t, (outs t.RegClass:$dst),
                       (ins t.RegClass:$src1, t.RegClass:$src2), "andn",
-                      binop_ndd_args, andn_rr_p>, VVVV, Sched<[WriteALU]>,
-                     T8, DefEFLAGS;
+                      binop_ndd_args, andn_rr_p>, VVVV, Sched<[WriteALU]>, T8;
   def rm#suffix : ITy<0xF2, MRMSrcMem, t, (outs t.RegClass:$dst),
                        (ins t.RegClass:$src1, t.MemOperand:$src2), "andn",
                        binop_ndd_args, andn_rm_p>, VVVV,
-                       Sched<[WriteALU.Folded, WriteALU.ReadAfterFold]>,
-                       T8, DefEFLAGS;
+                       Sched<[WriteALU.Folded, WriteALU.ReadAfterFold]>, T8;
 }
 
 // Complexity is reduced to give and with immediate a chance to match first.
 let AddedComplexity = -6 in {
-defm ANDN32 : AndN<Xi32, "">, VEX, Requires<[HasBMI, NoEGPR]>;
-defm ANDN64 : AndN<Xi64, "">, VEX, REX_W, Requires<[HasBMI, NoEGPR]>;
-defm ANDN32 : AndN<Xi32, "_EVEX">, EVEX, Requires<[HasBMI, HasEGPR, In64BitMode]>;
-defm ANDN64 : AndN<Xi64, "_EVEX">, EVEX, REX_W, Requires<[HasBMI, HasEGPR, In64BitMode]>;
+defm ANDN32 : AndN<Xi32, X86and_flag>, VEX, Requires<[HasBMI, NoEGPR]>, DefEFLAGS;
+defm ANDN64 : AndN<Xi64, X86and_flag>, VEX, Requires<[HasBMI, NoEGPR]>, DefEFLAGS;
+defm ANDN32 : AndN<Xi32, X86and_flag, "_EVEX">, EVEX, Requires<[HasBMI, HasEGPR, In64BitMode]>, DefEFLAGS;
+defm ANDN64 : AndN<Xi64, X86and_flag, "_EVEX">, EVEX, Requires<[HasBMI, HasEGPR, In64BitMode]>, DefEFLAGS;
+defm ANDN32 : AndN<Xi32, null_frag, "_NF">, EVEX, EVEX_NF, Requires<[In64BitMode]>;
+defm ANDN64 : AndN<Xi64, null_frag, "_NF">, EVEX, EVEX_NF, Requires<[In64BitMode]>;
 }
 
 multiclass Andn_Pats<string suffix> {

llvm/lib/Target/X86/X86InstrMisc.td

@@ -1223,7 +1223,7 @@ multiclass Bls<string m, Format RegMRM, Format MemMRM, X86TypeInfo t, string Suf
                              (outs t.RegClass:$dst), []>, T8, VVVV;
 }
 
-let Predicates = [HasBMI], Defs = [EFLAGS] in {
+let Defs = [EFLAGS] in {
   defm BLSR32 : Bls<"blsr", MRM1r, MRM1m, Xi32>, VEX;
   defm BLSR64 : Bls<"blsr", MRM1r, MRM1m, Xi64>, VEX;
   defm BLSMSK32 : Bls<"blsmsk", MRM2r, MRM2m, Xi32>, VEX;
@@ -1232,7 +1232,7 @@ let Predicates = [HasBMI], Defs = [EFLAGS] in {
   defm BLSI64 : Bls<"blsi", MRM3r, MRM3m, Xi64>, VEX;
 }
 
-let Predicates = [HasBMI, In64BitMode], Defs = [EFLAGS] in {
+let Predicates = [In64BitMode], Defs = [EFLAGS] in {
   defm BLSR32 : Bls<"blsr", MRM1r, MRM1m, Xi32, "_EVEX">, EVEX;
   defm BLSR64 : Bls<"blsr", MRM1r, MRM1m, Xi64, "_EVEX">, EVEX;
   defm BLSMSK32 : Bls<"blsmsk", MRM2r, MRM2m, Xi32, "_EVEX">, EVEX;
@@ -1241,6 +1241,15 @@ let Predicates = [HasBMI, In64BitMode], Defs = [EFLAGS] in {
   defm BLSI64 : Bls<"blsi", MRM3r, MRM3m, Xi64, "_EVEX">, EVEX;
 }
 
+let Predicates = [In64BitMode] in {
+  defm BLSR32 : Bls<"blsr", MRM1r, MRM1m, Xi32, "_NF">, EVEX, EVEX_NF;
+  defm BLSR64 : Bls<"blsr", MRM1r, MRM1m, Xi64, "_NF">, EVEX, EVEX_NF;
+  defm BLSMSK32 : Bls<"blsmsk", MRM2r, MRM2m, Xi32, "_NF">, EVEX, EVEX_NF;
+  defm BLSMSK64 : Bls<"blsmsk", MRM2r, MRM2m, Xi64, "_NF">, EVEX, EVEX_NF;
+  defm BLSI32 : Bls<"blsi", MRM3r, MRM3m, Xi32, "_NF">, EVEX, EVEX_NF;
+  defm BLSI64 : Bls<"blsi", MRM3r, MRM3m, Xi64, "_NF">, EVEX, EVEX_NF;
+}
+
 multiclass Bls_Pats<string suffix> {
   // FIXME(1): patterns for the load versions are not implemented
   // FIXME(2): By only matching `add_su` and `ineg_su` we may emit
@@ -1315,6 +1324,13 @@ let Predicates = [HasBMI2, HasEGPR, In64BitMode], Defs = [EFLAGS] in {
   defm BZHI64 : Bmi4VOp3<0xF5, "bzhi", Xi64, X86bzhi, WriteBZHI, "_EVEX">, EVEX;
 }
 
+let Predicates = [In64BitMode] in {
+  defm BEXTR32 : Bmi4VOp3<0xF7, "bextr", Xi32, null_frag, WriteBEXTR, "_NF">, EVEX, EVEX_NF;
+  defm BEXTR64 : Bmi4VOp3<0xF7, "bextr", Xi64, null_frag, WriteBEXTR, "_NF">, EVEX, EVEX_NF;
+  defm BZHI32 : Bmi4VOp3<0xF5, "bzhi", Xi32, null_frag, WriteBZHI, "_NF">, EVEX, EVEX_NF;
+  defm BZHI64 : Bmi4VOp3<0xF5, "bzhi", Xi64, null_frag, WriteBZHI, "_NF">, EVEX, EVEX_NF;
+}
+
 def CountTrailingOnes : SDNodeXForm<imm, [{
   // Count the trailing ones in the immediate.
   return getI8Imm(llvm::countr_one(N->getZExtValue()), SDLoc(N));

llvm/test/MC/Disassembler/X86/apx/andn.txt

@@ -1,6 +1,38 @@
 # RUN: llvm-mc --disassemble %s -triple=x86_64 | FileCheck %s --check-prefixes=ATT
 # RUN: llvm-mc --disassemble %s -triple=x86_64 -x86-asm-syntax=intel --output-asm-variant=1 | FileCheck %s --check-prefixes=INTEL
 
+# ATT:   {nf}	andnl	%ecx, %edx, %r10d
+# INTEL: {nf}	andn	r10d, edx, ecx
+0x62,0x72,0x6c,0x0c,0xf2,0xd1
+
+# ATT:   andnl	%ecx, %edx, %r10d
+# INTEL: andn	r10d, edx, ecx
+0x62,0x72,0x6c,0x08,0xf2,0xd1
+
+# ATT:   {nf}	andnq	%r9, %r15, %r11
+# INTEL: {nf}	andn	r11, r15, r9
+0x62,0x52,0x84,0x0c,0xf2,0xd9
+
+# ATT:   andnq	%r9, %r15, %r11
+# INTEL: andn	r11, r15, r9
+0x62,0x52,0x84,0x08,0xf2,0xd9
+
+# ATT:   {nf}	andnl	123(%rax,%rbx,4), %ecx, %edx
+# INTEL: {nf}	andn	edx, ecx, dword ptr [rax + 4*rbx + 123]
+0x62,0xf2,0x74,0x0c,0xf2,0x54,0x98,0x7b
+
+# ATT:   andnl	123(%rax,%rbx,4), %ecx, %edx
+# INTEL: andn	edx, ecx, dword ptr [rax + 4*rbx + 123]
+0x62,0xf2,0x74,0x08,0xf2,0x54,0x98,0x7b
+
+# ATT:   {nf}	andnq	123(%rax,%rbx,4), %r9, %r15
+# INTEL: {nf}	andn	r15, r9, qword ptr [rax + 4*rbx + 123]
+0x62,0x72,0xb4,0x0c,0xf2,0x7c,0x98,0x7b
+
+# ATT:   andnq	123(%rax,%rbx,4), %r9, %r15
+# INTEL: andn	r15, r9, qword ptr [rax + 4*rbx + 123]
+0x62,0x72,0xb4,0x08,0xf2,0x7c,0x98,0x7b
+
 # ATT:   andnl	%r18d, %r22d, %r26d
 # INTEL: andn	r26d, r22d, r18d
 0x62,0x6a,0x4c,0x00,0xf2,0xd2

llvm/test/MC/Disassembler/X86/apx/bextr.txt

@@ -1,6 +1,38 @@
 # RUN: llvm-mc --disassemble %s -triple=x86_64 | FileCheck %s --check-prefixes=ATT
 # RUN: llvm-mc --disassemble %s -triple=x86_64 -x86-asm-syntax=intel --output-asm-variant=1 | FileCheck %s --check-prefixes=INTEL
 
+# ATT:   {nf}	bextrl	%ecx, %edx, %r10d
+# INTEL: {nf}	bextr	r10d, edx, ecx
+0x62,0x72,0x74,0x0c,0xf7,0xd2
+
+# ATT:   bextrl	%ecx, %edx, %r10d
+# INTEL: bextr	r10d, edx, ecx
+0x62,0x72,0x74,0x08,0xf7,0xd2
+
+# ATT:   {nf}	bextrl	%ecx, 123(%rax,%rbx,4), %edx
+# INTEL: {nf}	bextr	edx, dword ptr [rax + 4*rbx + 123], ecx
+0x62,0xf2,0x74,0x0c,0xf7,0x54,0x98,0x7b
+
+# ATT:   bextrl	%ecx, 123(%rax,%rbx,4), %edx
+# INTEL: bextr	edx, dword ptr [rax + 4*rbx + 123], ecx
+0x62,0xf2,0x74,0x08,0xf7,0x54,0x98,0x7b
+
+# ATT:   {nf}	bextrq	%r9, %r15, %r11
+# INTEL: {nf}	bextr	r11, r15, r9
+0x62,0x52,0xb4,0x0c,0xf7,0xdf
+
+# ATT:   bextrq	%r9, %r15, %r11
+# INTEL: bextr	r11, r15, r9
+0x62,0x52,0xb4,0x08,0xf7,0xdf
+
+# ATT:   {nf}	bextrq	%r9, 123(%rax,%rbx,4), %r15
+# INTEL: {nf}	bextr	r15, qword ptr [rax + 4*rbx + 123], r9
+0x62,0x72,0xb4,0x0c,0xf7,0x7c,0x98,0x7b
+
+# ATT:   bextrq	%r9, 123(%rax,%rbx,4), %r15
+# INTEL: bextr	r15, qword ptr [rax + 4*rbx + 123], r9
+0x62,0x72,0xb4,0x08,0xf7,0x7c,0x98,0x7b
+
 # ATT:   bextrl	%r18d, %r22d, %r26d
 # INTEL: bextr	r26d, r22d, r18d
 0x62,0x6a,0x6c,0x00,0xf7,0xd6

llvm/test/MC/Disassembler/X86/apx/blsi.txt

@@ -1,6 +1,38 @@
 # RUN: llvm-mc --disassemble %s -triple=x86_64 | FileCheck %s --check-prefixes=ATT
 # RUN: llvm-mc --disassemble %s -triple=x86_64 -x86-asm-syntax=intel --output-asm-variant=1 | FileCheck %s --check-prefixes=INTEL
 
+# ATT:   {nf}	blsil	%ecx, %edx
+# INTEL: {nf}	blsi	edx, ecx
+0x62,0xf2,0x6c,0x0c,0xf3,0xd9
+
+# ATT:   blsil	%ecx, %edx
+# INTEL: blsi	edx, ecx
+0x62,0xf2,0x6c,0x08,0xf3,0xd9
+
+# ATT:   {nf}	blsiq	%r9, %r15
+# INTEL: {nf}	blsi	r15, r9
+0x62,0xd2,0x84,0x0c,0xf3,0xd9
+
+# ATT:   blsiq	%r9, %r15
+# INTEL: blsi	r15, r9
+0x62,0xd2,0x84,0x08,0xf3,0xd9
+
+# ATT:   {nf}	blsil	123(%rax,%rbx,4), %ecx
+# INTEL: {nf}	blsi	ecx, dword ptr [rax + 4*rbx + 123]
+0x62,0xf2,0x74,0x0c,0xf3,0x5c,0x98,0x7b
+
+# ATT:   blsil	123(%rax,%rbx,4), %ecx
+# INTEL: blsi	ecx, dword ptr [rax + 4*rbx + 123]
+0x62,0xf2,0x74,0x08,0xf3,0x5c,0x98,0x7b
+
+# ATT:   {nf}	blsiq	123(%rax,%rbx,4), %r9
+# INTEL: {nf}	blsi	r9, qword ptr [rax + 4*rbx + 123]
+0x62,0xf2,0xb4,0x0c,0xf3,0x5c,0x98,0x7b
+
+# ATT:   blsiq	123(%rax,%rbx,4), %r9
+# INTEL: blsi	r9, qword ptr [rax + 4*rbx + 123]
+0x62,0xf2,0xb4,0x08,0xf3,0x5c,0x98,0x7b
+
 # ATT:   blsil	%r18d, %r22d
 # INTEL: blsi	r22d, r18d
 0x62,0xfa,0x4c,0x00,0xf3,0xda

llvm/test/MC/Disassembler/X86/apx/blsmsk.txt

@@ -1,6 +1,38 @@
 # RUN: llvm-mc --disassemble %s -triple=x86_64 | FileCheck %s --check-prefixes=ATT
 # RUN: llvm-mc --disassemble %s -triple=x86_64 -x86-asm-syntax=intel --output-asm-variant=1 | FileCheck %s --check-prefixes=INTEL
 
+# ATT:   {nf}	blsmskl	%ecx, %edx
+# INTEL: {nf}	blsmsk	edx, ecx
+0x62,0xf2,0x6c,0x0c,0xf3,0xd1
+
+# ATT:   blsmskl	%ecx, %edx
+# INTEL: blsmsk	edx, ecx
+0x62,0xf2,0x6c,0x08,0xf3,0xd1
+
+# ATT:   {nf}	blsmskq	%r9, %r15
+# INTEL: {nf}	blsmsk	r15, r9
+0x62,0xd2,0x84,0x0c,0xf3,0xd1
+
+# ATT:   blsmskq	%r9, %r15
+# INTEL: blsmsk	r15, r9
+0x62,0xd2,0x84,0x08,0xf3,0xd1
+
+# ATT:   {nf}	blsmskl	123(%rax,%rbx,4), %ecx
+# INTEL: {nf}	blsmsk	ecx, dword ptr [rax + 4*rbx + 123]
+0x62,0xf2,0x74,0x0c,0xf3,0x54,0x98,0x7b
+
+# ATT:   blsmskl	123(%rax,%rbx,4), %ecx
+# INTEL: blsmsk	ecx, dword ptr [rax + 4*rbx + 123]
+0x62,0xf2,0x74,0x08,0xf3,0x54,0x98,0x7b
+
+# ATT:   {nf}	blsmskq	123(%rax,%rbx,4), %r9
+# INTEL: {nf}	blsmsk	r9, qword ptr [rax + 4*rbx + 123]
+0x62,0xf2,0xb4,0x0c,0xf3,0x54,0x98,0x7b
+
+# ATT:   blsmskq	123(%rax,%rbx,4), %r9
+# INTEL: blsmsk	r9, qword ptr [rax + 4*rbx + 123]
+0x62,0xf2,0xb4,0x08,0xf3,0x54,0x98,0x7b
+
 # ATT:   blsmskl	%r18d, %r22d
 # INTEL: blsmsk	r22d, r18d
 0x62,0xfa,0x4c,0x00,0xf3,0xd2

llvm/test/MC/Disassembler/X86/apx/blsr.txt

@@ -1,6 +1,38 @@
 # RUN: llvm-mc --disassemble %s -triple=x86_64 | FileCheck %s --check-prefixes=ATT
 # RUN: llvm-mc --disassemble %s -triple=x86_64 -x86-asm-syntax=intel --output-asm-variant=1 | FileCheck %s --check-prefixes=INTEL
 
+# ATT:   {nf}	blsrl	%ecx, %edx
+# INTEL: {nf}	blsr	edx, ecx
+0x62,0xf2,0x6c,0x0c,0xf3,0xc9
+
+# ATT:   blsrl	%ecx, %edx
+# INTEL: blsr	edx, ecx
+0x62,0xf2,0x6c,0x08,0xf3,0xc9
+
+# ATT:   {nf}	blsrq	%r9, %r15
+# INTEL: {nf}	blsr	r15, r9
+0x62,0xd2,0x84,0x0c,0xf3,0xc9
+
+# ATT:   blsrq	%r9, %r15
+# INTEL: blsr	r15, r9
+0x62,0xd2,0x84,0x08,0xf3,0xc9
+
+# ATT:   {nf}	blsrl	123(%rax,%rbx,4), %ecx
+# INTEL: {nf}	blsr	ecx, dword ptr [rax + 4*rbx + 123]
+0x62,0xf2,0x74,0x0c,0xf3,0x4c,0x98,0x7b
+
+# ATT:   blsrl	123(%rax,%rbx,4), %ecx
+# INTEL: blsr	ecx, dword ptr [rax + 4*rbx + 123]
+0x62,0xf2,0x74,0x08,0xf3,0x4c,0x98,0x7b
+
+# ATT:   {nf}	blsrq	123(%rax,%rbx,4), %r9
+# INTEL: {nf}	blsr	r9, qword ptr [rax + 4*rbx + 123]
+0x62,0xf2,0xb4,0x0c,0xf3,0x4c,0x98,0x7b
+
+# ATT:   blsrq	123(%rax,%rbx,4), %r9
+# INTEL: blsr	r9, qword ptr [rax + 4*rbx + 123]
+0x62,0xf2,0xb4,0x08,0xf3,0x4c,0x98,0x7b
+
 # ATT:   blsrl	%r18d, %r22d
 # INTEL: blsr	r22d, r18d
 0x62,0xfa,0x4c,0x00,0xf3,0xca

llvm/test/MC/Disassembler/X86/apx/bzhi.txt

@@ -1,6 +1,42 @@
 # RUN: llvm-mc --disassemble %s -triple=x86_64 | FileCheck %s --check-prefixes=ATT
 # RUN: llvm-mc --disassemble %s -triple=x86_64 -x86-asm-syntax=intel --output-asm-variant=1 | FileCheck %s --check-prefixes=INTEL
 
+# ATT:   {nf}	bzhil	%ecx, %edx, %r10d
+# INTEL: {nf}	bzhi	r10d, edx, ecx
+0x62,0x72,0x74,0x0c,0xf5,0xd2
+
+# ATT:   bzhil	%ecx, %edx, %r10d
+# INTEL: bzhi	r10d, edx, ecx
+0x62,0x72,0x74,0x08,0xf5,0xd2
+
+# ATT:   {nf}	bzhil	%ecx, 123(%rax,%rbx,4), %edx
+# INTEL: {nf}	bzhi	edx, dword ptr [rax + 4*rbx + 123], ecx
+0x62,0xf2,0x74,0x0c,0xf5,0x54,0x98,0x7b
+
+# ATT:   bzhil	%ecx, 123(%rax,%rbx,4), %edx
+# INTEL: bzhi	edx, dword ptr [rax + 4*rbx + 123], ecx
+0x62,0xf2,0x74,0x08,0xf5,0x54,0x98,0x7b
+
+# ATT:   {nf}	bzhiq	%r9, %r15, %r11
+# INTEL: {nf}	bzhi	r11, r15, r9
+0x62,0x52,0xb4,0x0c,0xf5,0xdf
+
+# ATT:   bzhiq	%r9, %r15, %r11
+# INTEL: bzhi	r11, r15, r9
+0x62,0x52,0xb4,0x08,0xf5,0xdf
+
+# ATT:   {nf}	bzhiq	%r9, 123(%rax,%rbx,4), %r15
+# INTEL: {nf}	bzhi	r15, qword ptr [rax + 4*rbx + 123], r9
+0x62,0x72,0xb4,0x0c,0xf5,0x7c,0x98,0x7b
+
+# ATT:   bzhiq	%r9, 123(%rax,%rbx,4), %r15
+# INTEL: bzhi	r15, qword ptr [rax + 4*rbx + 123], r9
+0x62,0x72,0xb4,0x08,0xf5,0x7c,0x98,0x7b
+
+# ATT:   bzhiq	%r9, 123(%rax,%rbx,4), %r15
+# INTEL: bzhi	r15, qword ptr [rax + 4*rbx + 123], r9
+0x62,0x72,0xb4,0x08,0xf5,0xbc,0x98,0x7b,0x00,0x00,0x00
+
 # ATT:   bzhil	%r18d, %r22d, %r26d
 # INTEL: bzhi	r26d, r22d, r18d
 0x62,0x6a,0x6c,0x00,0xf5,0xd6