[X86] Add support for inline assembly 'x' constraint for i128.

[craig.topper]

Summary

Limiting to x86-64 since that's when __int128 is legal in clang.

[RKSimon]

LGTM

llvm/test/CodeGen/X86/inline-asm-x-i128.ll

2

test i686/x32 as well?

llvm/lib/Target/X86/X86ISelLowering.cpp

         if (VConstraint && Subtarget.hasVLX())
           return std::make_pair(0U, &X86::FR32XRegClass);
         return std::make_pair(0U, &X86::FR32RegClass);
       case MVT::f64:
       case MVT::i64:
         if (VConstraint && Subtarget.hasVLX())
           return std::make_pair(0U, &X86::FR64XRegClass);
         return std::make_pair(0U, &X86::FR64RegClass);
-      // TODO: Handle i128 in FR128RegClass after it is tested well.
+      case MVT::i128:
+        if (Subtarget.is64Bit()) {
+          if (VConstraint && Subtarget.hasVLX())
+            return std::make_pair(0U, &X86::VR128XRegClass);
+          return std::make_pair(0U, &X86::VR128RegClass);
+        }
+        break;
       // Vector types and fp128.
       case MVT::f128:
       case MVT::v16i8:
       case MVT::v8i16:
       case MVT::v4i32:
       case MVT::v2i64:
       case MVT::v4f32:
       case MVT::v2f64:

llvm/test/CodeGen/X86/inline-asm-x-i128.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu | FileCheck %s
+
+define { i64, i64 } @foo(i64 %0, i64 %1) {
+; CHECK-LABEL: foo:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movq %rsi, %xmm0
+; CHECK-NEXT:    movq %rdi, %xmm1
+; CHECK-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
+; CHECK-NEXT:    #APP
+; CHECK-NEXT:    movdqa %xmm1, %xmm0
+; CHECK-NEXT:    #NO_APP
+; CHECK-NEXT:    movq %xmm0, %rax
+; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
+; CHECK-NEXT:    movq %xmm0, %rdx
+; CHECK-NEXT:    retq
+  %3 = zext i64 %1 to i128
+  %4 = shl nuw i128 %3, 64
+  %5 = zext i64 %0 to i128
+  %6 = or i128 %4, %5
+  %7 = tail call i128 asm sideeffect "movdqa  $1, $0", "=x,x,~{dirflag},~{fpsr},~{flags}"(i128 %6)
+  %8 = trunc i128 %7 to i64
+  %9 = lshr i128 %7, 64
+  %10 = trunc i128 %9 to i64
+  %11 = insertvalue { i64, i64 } undef, i64 %8, 0
+  %12 = insertvalue { i64, i64 } %11, i64 %10, 1
+  ret { i64, i64 } %12
+}