From 397edeff50395b15d3c33df9d243d736b05d8a6b Mon Sep 17 00:00:00 2001
From: Evan Cheng <evan.cheng@apple.com>
Date: Tue, 11 Apr 2006 22:28:25 +0000
Subject: [PATCH] Added __builtin_ia32_storelv4si, __builtin_ia32_movqv4si,
 __builtin_ia32_loadlv4si, __builtin_ia32_loaddqu, __builtin_ia32_storedqu.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27599 91177308-0d34-0410-b5e6-96231b3b80d8
---
 include/llvm/IntrinsicsX86.td | 11 +++++++++++
 lib/Target/X86/X86InstrSSE.td | 23 +++++++++++++++++++++--
 2 files changed, 32 insertions(+), 2 deletions(-)

diff --git a/include/llvm/IntrinsicsX86.td b/include/llvm/IntrinsicsX86.td
index 48f79d25a24..77f787c63bb 100644
--- a/include/llvm/IntrinsicsX86.td
+++ b/include/llvm/IntrinsicsX86.td
@@ -269,6 +269,8 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse2_loadu_pd : GCCBuiltin<"__builtin_ia32_loadupd">,
               Intrinsic<[llvm_v2f64_ty, llvm_ptr_ty], [IntrReadMem]>;
+  def int_x86_sse2_loadu_dq : GCCBuiltin<"__builtin_ia32_loaddqu">,
+              Intrinsic<[llvm_v16i8_ty, llvm_ptr_ty], [IntrReadMem]>;
 }
 
 // SIMD store ops
@@ -276,6 +278,12 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse2_storeu_pd : GCCBuiltin<"__builtin_ia32_storeupd">,
               Intrinsic<[llvm_void_ty, llvm_ptr_ty,
                          llvm_v2f64_ty], [IntrWriteMem]>;
+  def int_x86_sse2_storeu_dq : GCCBuiltin<"__builtin_ia32_storedqu">,
+              Intrinsic<[llvm_void_ty, llvm_ptr_ty,
+                         llvm_v16i8_ty], [IntrWriteMem]>;
+  def int_x86_sse2_storel_dq : GCCBuiltin<"__builtin_ia32_storelv4si">,
+              Intrinsic<[llvm_void_ty, llvm_ptr_ty,
+                         llvm_v4i32_ty], [IntrWriteMem]>;
 }
 
 // Cacheability support ops
@@ -302,6 +310,9 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse2_packuswb_128 : GCCBuiltin<"__builtin_ia32_packuswb128">,
               Intrinsic<[llvm_v8i16_ty, llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
+  // FIXME: Temporary workaround since 2-wide shuffle is broken.
+  def int_x86_sse2_movl_dq : GCCBuiltin<"__builtin_ia32_movqv4si">,
+              Intrinsic<[llvm_v4i32_ty, llvm_v4i32_ty], [IntrNoMem]>;
   def int_x86_sse2_movmskpd : GCCBuiltin<"__builtin_ia32_movmskpd">,
               Intrinsic<[llvm_int_ty, llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_pmovmskb_128 : GCCBuiltin<"__builtin_ia32_pmovmskb128">,
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index 10f4d562766..65e8aaeffa5 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -724,6 +724,14 @@ def MOVUPDrm : PDI<0x10, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
 def MOVUPDmr : PDI<0x11, MRMDestMem, (ops f128mem:$dst, VR128:$src),
                    "movupd {$src, $dst|$dst, $src}",
                    [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
+def MOVDQUrm :   I<0x6F, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
+                   "movdqu {$src, $dst|$dst, $src}",
+                   [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
+                 XS, Requires<[HasSSE2]>;
+def MOVDQUmr :   I<0x7F, MRMDestMem, (ops i128mem:$dst, VR128:$src),
+                   "movdqu {$src, $dst|$dst, $src}",
+                   [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
+                 XS, Requires<[HasSSE2]>;
 
 let isTwoAddress = 1 in {
 def MOVLPSrm : PSI<0x12, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
@@ -1657,6 +1665,16 @@ def MOVLPDrr : SDI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
                              MOVS_shuffle_mask)))]>;
 }
 
+// Store / copy lower 64-bits of a XMM register.
+def MOVLQ128mr : PDI<0xD6, MRMDestMem, (ops i64mem:$dst, VR128:$src),
+                     "movq {$src, $dst|$dst, $src}",
+                     [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
+
+// FIXME: Temporary workaround since 2-wide shuffle is broken.
+def MOVLQ128rr : PDI<0xD6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
+                     "movq {$src, $dst|$dst, $src}",
+                     [(set VR128:$dst, (int_x86_sse2_movl_dq VR128:$src))]>;
+
 // Move to lower bits of a VR128 and zeroing upper bits.
 // Loading from memory automatically zeroing upper bits.
 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
@@ -1672,9 +1690,10 @@ def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (ops VR128:$dst, i32mem:$src),
                        [(set VR128:$dst,
                          (v4i32 (X86zexts2vec (loadi32 addr:$src))))]>;
 def MOVZQI2PQIrm : PDI<0x7E, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
-                       "movd {$src, $dst|$dst, $src}",
+                       "movq {$src, $dst|$dst, $src}",
                        [(set VR128:$dst,
-                         (v2i64 (X86zexts2vec (loadi64 addr:$src))))]>;
+                         (bc_v2i64 (v2f64 (X86zexts2vec
+                                           (loadf64 addr:$src)))))]>;
 
 //===----------------------------------------------------------------------===//
 // Non-Instruction Patterns