Add intrinsics, code gen, assembler and disassembler support for the SSE4a extrq and insertq instructions.

This required light surgery on the assembler and disassembler
because the instructions use an uncommon encoding. They are
the only two instructions in x86 that use register operands
and two immediates.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@157634 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/IntrinsicsX86.td

@@ -1008,6 +1008,17 @@ let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.".
 // SSE4A
 
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_sse4a_extrqi : GCCBuiltin<"__builtin_ia32_extrqi">,
+    Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_i8_ty, llvm_i8_ty], []>;
+  def int_x86_sse4a_extrq  : GCCBuiltin<"__builtin_ia32_extrq">,
+    Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v16i8_ty], []>;
+
+  def int_x86_sse4a_insertqi : GCCBuiltin<"__builtin_ia32_insertqi">,
+    Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty,
+                                llvm_i8_ty, llvm_i8_ty], []>;
+  def int_x86_sse4a_insertq  : GCCBuiltin<"__builtin_ia32_insertq">,
+    Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty], []>;
+
   def int_x86_sse4a_movnt_ss : GCCBuiltin<"__builtin_ia32_movntss">,
     Intrinsic<[], [llvm_ptr_ty, llvm_v4f32_ty], []>;
   def int_x86_sse4a_movnt_sd : GCCBuiltin<"__builtin_ia32_movntsd">,

lib/Target/X86/MCTargetDesc/X86MCCodeEmitter.cpp

@@ -1150,8 +1150,9 @@ EncodeInstruction(const MCInst &MI, raw_ostream &OS,
   }
 
   // If there is a remaining operand, it must be a trailing immediate.  Emit it
-  // according to the right size for the instruction.
-  if (CurOp != NumOps) {
+  // according to the right size for the instruction. Some instructions
+  // (SSE4a extrq and insertq) have two trailing immediates.
+  while (CurOp != NumOps && NumOps - CurOp <= 2) {
     // The last source register of a 4 operand instruction in AVX is encoded
     // in bits[7:4] of a immediate byte.
     if ((TSFlags >> X86II::VEXShift) & X86II::VEX_I8IMM) {

lib/Target/X86/X86InstrSSE.td

@@ -7268,6 +7268,31 @@ defm : pclmul_alias<"lqlq", 0x00>;
 //===----------------------------------------------------------------------===//
 
 let Predicates = [HasSSE4A] in {
+
+let Constraints = "$src = $dst" in {
+def EXTRQI : Ii8<0x78, MRM0r, (outs VR128:$dst),
+                 (ins VR128:$src, i8imm:$len, i8imm:$idx),
+                 "extrq\t{$idx, $len, $src|$src, $len, $idx}",
+                 [(set VR128:$dst, (int_x86_sse4a_extrqi VR128:$src, imm:$len,
+                                    imm:$idx))]>, TB, OpSize;
+def EXTRQ  : I<0x79, MRMSrcReg, (outs VR128:$dst),
+              (ins VR128:$src, VR128:$mask),
+              "extrq\t{$mask, $src|$src, $mask}",
+              [(set VR128:$dst, (int_x86_sse4a_extrq VR128:$src,
+                                 VR128:$mask))]>, TB, OpSize;
+
+def INSERTQI : Ii8<0x78, MRMSrcReg, (outs VR128:$dst),
+                   (ins VR128:$src, VR128:$src2, i8imm:$len, i8imm:$idx),
+                   "insertq\t{$idx, $len, $src2, $src|$src, $src2, $len, $idx}",
+                   [(set VR128:$dst, (int_x86_sse4a_insertqi VR128:$src,
+                                      VR128:$src2, imm:$len, imm:$idx))]>, XD;
+def INSERTQ  : I<0x79, MRMSrcReg, (outs VR128:$dst),
+                 (ins VR128:$src, VR128:$mask),
+                 "insertq\t{$mask, $src|$src, $mask}",
+                 [(set VR128:$dst, (int_x86_sse4a_insertq VR128:$src,
+                                    VR128:$mask))]>, XD;
+}
+
 def MOVNTSS : I<0x2B, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
                 "movntss\t{$src, $dst|$dst, $src}",
                 [(int_x86_sse4a_movnt_ss addr:$dst, VR128:$src)]>, XS;

test/CodeGen/X86/sse4a.ll

@@ -1,8 +1,8 @@
 ; RUN: llc < %s -mtriple=i686-apple-darwin9 -mattr=sse4a | FileCheck %s
 
 define void @test1(i8* %p, <4 x float> %a) nounwind optsize ssp {
+; CHECK: test1:
 ; CHECK: movntss
-entry:
   tail call void @llvm.x86.sse4a.movnt.ss(i8* %p, <4 x float> %a) nounwind
   ret void
 }
@@ -10,10 +10,47 @@ entry:
 declare void @llvm.x86.sse4a.movnt.ss(i8*, <4 x float>)
 
 define void @test2(i8* %p, <2 x double> %a) nounwind optsize ssp {
+; CHECK: test2:
 ; CHECK: movntsd
-entry:
   tail call void @llvm.x86.sse4a.movnt.sd(i8* %p, <2 x double> %a) nounwind
   ret void
 }
 
 declare void @llvm.x86.sse4a.movnt.sd(i8*, <2 x double>)
+
+define <2 x i64> @test3(<2 x i64> %x) nounwind uwtable ssp {
+; CHECK: test3:
+; CHECK: extrq
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.extrqi(<2 x i64> %x, i8 3, i8 2)
+  ret <2 x i64> %1
+}
+
+declare <2 x i64> @llvm.x86.sse4a.extrqi(<2 x i64>, i8, i8) nounwind
+
+define <2 x i64> @test4(<2 x i64> %x, <2 x i64> %y) nounwind uwtable ssp {
+; CHECK: test4:
+; CHECK: extrq
+  %1 = bitcast <2 x i64> %y to <16 x i8>
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.extrq(<2 x i64> %x, <16 x i8> %1) nounwind
+  ret <2 x i64> %2
+}
+
+declare <2 x i64> @llvm.x86.sse4a.extrq(<2 x i64>, <16 x i8>) nounwind
+
+define <2 x i64> @test5(<2 x i64> %x, <2 x i64> %y) nounwind uwtable ssp {
+; CHECK: test5:
+; CHECK: insertq
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %x, <2 x i64> %y, i8 5, i8 6)
+  ret <2 x i64> %1
+}
+
+declare <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64>, <2 x i64>, i8, i8) nounwind
+
+define <2 x i64> @test6(<2 x i64> %x, <2 x i64> %y) nounwind uwtable ssp {
+; CHECK: test6:
+; CHECK: insertq
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertq(<2 x i64> %x, <2 x i64> %y) nounwind
+  ret <2 x i64> %1
+}
+
+declare <2 x i64> @llvm.x86.sse4a.insertq(<2 x i64>, <2 x i64>) nounwind

test/MC/Disassembler/X86/x86-32.txt

@@ -612,3 +612,21 @@
 
 # CHECK: shrxl %esi, %ebx, %edx
 0xc4 0xe2 0x0b 0xf7 0xd3
+
+# CHECK: extrq  $2, $3, %xmm0
+0x66 0x0f 0x78 0xc0 0x03 0x02
+
+# CHECK: extrq  %xmm1, %xmm0
+0x66 0x0f 0x79 0xc1
+
+# CHECK: insertq $6, $5, %xmm1, %xmm0
+0xf2 0x0f 0x78 0xc1 0x05 0x06
+
+# CHECK: insertq %xmm1, %xmm0
+0xf2 0x0f 0x79 0xc1
+
+# CHECK: movntsd %xmm0, (%edi)
+0xf2 0x0f 0x2b 0x07
+
+# CHECK: movntss %xmm0, (%edi)
+0xf3 0x0f 0x2b 0x07

test/MC/Disassembler/X86/x86-64.txt

@@ -61,3 +61,21 @@
 
 # CHECK: cmpordsd 
 0xf2 0x0f 0xc2 0xc7 0x07
+
+# CHECK: extrq  $2, $3, %xmm0
+0x66 0x0f 0x78 0xc0 0x03 0x02
+
+# CHECK: extrq  %xmm1, %xmm0
+0x66 0x0f 0x79 0xc1
+
+# CHECK: insertq $6, $5, %xmm1, %xmm0
+0xf2 0x0f 0x78 0xc1 0x05 0x06
+
+# CHECK: insertq %xmm1, %xmm0
+0xf2 0x0f 0x79 0xc1
+
+# CHECK: movntsd %xmm0, (%rdi)
+0xf2 0x0f 0x2b 0x07
+
+# CHECK: movntss %xmm0, (%rdi)
+0xf3 0x0f 0x2b 0x07

test/MC/X86/x86_64-sse4a.s

@@ -0,0 +1,25 @@
+# RUN: llvm-mc -triple x86_64-unknown-unknown --show-encoding %s | FileCheck %s
+
+extrq  $2, $3, %xmm0
+# CHECK: extrq  $2, $3, %xmm0
+# CHECK: encoding: [0x66,0x0f,0x78,0xc0,0x03,0x02]
+
+extrq  %xmm1, %xmm0
+# CHECK: extrq  %xmm1, %xmm0
+# CHECK: encoding: [0x66,0x0f,0x79,0xc1]
+
+insertq $6, $5, %xmm1, %xmm0
+# CHECK: insertq $6, $5, %xmm1, %xmm0
+# CHECK: encoding: [0xf2,0x0f,0x78,0xc1,0x05,0x06]
+
+insertq %xmm1, %xmm0
+# CHECK: insertq %xmm1, %xmm0
+# CHECK: encoding: [0xf2,0x0f,0x79,0xc1]
+
+movntsd %xmm0, (%rdi)
+# CHECK: movntsd %xmm0, (%rdi)
+# CHECK: encoding: [0xf2,0x0f,0x2b,0x07]
+
+movntss %xmm0, (%rdi)
+# CHECK: movntss %xmm0, (%rdi)
+# CHECK: encoding: [0xf3,0x0f,0x2b,0x07]

utils/TableGen/X86RecognizableInstr.cpp

@@ -690,12 +690,13 @@ void RecognizableInstr::emitInstructionSpecifier(DisassemblerTables &tables) {
     // Operand 2 is a register operand in the R/M field.
     // - In AVX, there is a register operand in the VEX.vvvv field here -
     // Operand 3 (optional) is an immediate.
+    // Operand 4 (optional) is an immediate.
 
     if (HasVEX_4VPrefix || HasVEX_4VOp3Prefix)
       assert(numPhysicalOperands >= 3 && numPhysicalOperands <= 5 &&
              "Unexpected number of operands for MRMSrcRegFrm with VEX_4V"); 
     else
-      assert(numPhysicalOperands >= 2 && numPhysicalOperands <= 3 &&
+      assert(numPhysicalOperands >= 2 && numPhysicalOperands <= 4 &&
              "Unexpected number of operands for MRMSrcRegFrm");
   
     HANDLE_OPERAND(roRegister)
@@ -716,6 +717,7 @@ void RecognizableInstr::emitInstructionSpecifier(DisassemblerTables &tables) {
     if (!HasMemOp4Prefix)
       HANDLE_OPTIONAL(immediate)
     HANDLE_OPTIONAL(immediate) // above might be a register in 7:4
+    HANDLE_OPTIONAL(immediate)
     break;
   case X86Local::MRMSrcMem:
     // Operand 1 is a register operand in the Reg/Opcode field.
@@ -759,16 +761,18 @@ void RecognizableInstr::emitInstructionSpecifier(DisassemblerTables &tables) {
   case X86Local::MRM7r:
     // Operand 1 is a register operand in the R/M field.
     // Operand 2 (optional) is an immediate or relocation.
+    // Operand 3 (optional) is an immediate.
     if (HasVEX_4VPrefix)
       assert(numPhysicalOperands <= 3 &&
              "Unexpected number of operands for MRMnRFrm with VEX_4V");
     else
-      assert(numPhysicalOperands <= 2 &&
+      assert(numPhysicalOperands <= 3 &&
              "Unexpected number of operands for MRMnRFrm");
     if (HasVEX_4VPrefix)
       HANDLE_OPERAND(vvvvRegister)
     HANDLE_OPTIONAL(rmRegister)
     HANDLE_OPTIONAL(relocation)
+    HANDLE_OPTIONAL(immediate)
     break;
   case X86Local::MRM0m:
   case X86Local::MRM1m: