Use fp unpack instructions to unpack int types. Until we have AVX2, this

is the best we can do for these patterns. This fix PR10554.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@137161 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 65 insertions, 8 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 71b23e2ccb..e9510af775 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -5940,7 +5940,9 @@ static inline unsigned getUNPCKLOpcode(EVT VT) {
   case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
   case MVT::v4f32: return X86ISD::UNPCKLPS;
   case MVT::v2f64: return X86ISD::UNPCKLPD;
+  case MVT::v8i32: // Use fp unit for int unpack.
   case MVT::v8f32: return X86ISD::VUNPCKLPSY;
+  case MVT::v4i64: // Use fp unit for int unpack.
   case MVT::v4f64: return X86ISD::VUNPCKLPDY;
   case MVT::v16i8: return X86ISD::PUNPCKLBW;
   case MVT::v8i16: return X86ISD::PUNPCKLWD;
@@ -5956,7 +5958,9 @@ static inline unsigned getUNPCKHOpcode(EVT VT) {
   case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
   case MVT::v4f32: return X86ISD::UNPCKHPS;
   case MVT::v2f64: return X86ISD::UNPCKHPD;
+  case MVT::v8i32: // Use fp unit for int unpack.
   case MVT::v8f32: return X86ISD::VUNPCKHPSY;
+  case MVT::v4i64: // Use fp unit for int unpack.
   case MVT::v4f64: return X86ISD::VUNPCKHPDY;
   case MVT::v16i8: return X86ISD::PUNPCKHBW;
   case MVT::v8i16: return X86ISD::PUNPCKHWD;
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index 2b8180aac0..56b05e17ac 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -5799,18 +5799,24 @@ def : Pat<(X86Movddup (bc_v2f64
 // Shuffle with UNPCKLPS
 def : Pat<(v4f32 (X86Unpcklps VR128:$src1, (memopv4f32 addr:$src2))),
           (VUNPCKLPSrm VR128:$src1, addr:$src2)>, Requires<[HasAVX]>;
-def : Pat<(v8f32 (X86Unpcklpsy VR256:$src1, (memopv8f32 addr:$src2))),
-          (VUNPCKLPSYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
 def : Pat<(v4f32 (X86Unpcklps VR128:$src1, (memopv4f32 addr:$src2))),
           (UNPCKLPSrm VR128:$src1, addr:$src2)>;
 
 def : Pat<(v4f32 (X86Unpcklps VR128:$src1, VR128:$src2)),
           (VUNPCKLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasAVX]>;
-def : Pat<(v8f32 (X86Unpcklpsy VR256:$src1, VR256:$src2)),
-          (VUNPCKLPSYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
 def : Pat<(v4f32 (X86Unpcklps VR128:$src1, VR128:$src2)),
           (UNPCKLPSrr VR128:$src1, VR128:$src2)>;
 
+// Shuffle with VUNPCKHPSY
+def : Pat<(v8f32 (X86Unpcklpsy VR256:$src1, (memopv8f32 addr:$src2))),
+          (VUNPCKLPSYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
+def : Pat<(v8f32 (X86Unpcklpsy VR256:$src1, VR256:$src2)),
+          (VUNPCKLPSYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
+def : Pat<(v8i32 (X86Unpcklpsy VR256:$src1, VR256:$src2)),
+          (VUNPCKLPSYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
+def : Pat<(v8i32 (X86Unpcklpsy VR256:$src1, (memopv8i32 addr:$src2))),
+          (VUNPCKLPSYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
+
 // Shuffle with UNPCKHPS
 def : Pat<(v4f32 (X86Unpckhps VR128:$src1, (memopv4f32 addr:$src2))),
           (VUNPCKHPSrm VR128:$src1, addr:$src2)>, Requires<[HasAVX]>;
@@ -5828,21 +5834,33 @@ def : Pat<(v8f32 (X86Unpckhpsy VR256:$src1, (memopv8f32 addr:$src2))),
 def : Pat<(v8f32 (X86Unpckhpsy VR256:$src1, VR256:$src2)),
           (VUNPCKHPSYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
 
+def : Pat<(v8i32 (X86Unpckhpsy VR256:$src1, (memopv8i32 addr:$src2))),
+          (VUNPCKHPSYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
+def : Pat<(v8i32 (X86Unpckhpsy VR256:$src1, VR256:$src2)),
+          (VUNPCKHPSYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
+
 // Shuffle with UNPCKLPD
 def : Pat<(v2f64 (X86Unpcklpd VR128:$src1, (memopv2f64 addr:$src2))),
           (VUNPCKLPDrm VR128:$src1, addr:$src2)>, Requires<[HasAVX]>;
-def : Pat<(v4f64 (X86Unpcklpdy VR256:$src1, (memopv4f64 addr:$src2))),
-          (VUNPCKLPDYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
 def : Pat<(v2f64 (X86Unpcklpd VR128:$src1, (memopv2f64 addr:$src2))),
           (UNPCKLPDrm VR128:$src1, addr:$src2)>;
 
 def : Pat<(v2f64 (X86Unpcklpd VR128:$src1, VR128:$src2)),
           (VUNPCKLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasAVX]>;
-def : Pat<(v4f64 (X86Unpcklpdy VR256:$src1, VR256:$src2)),
-          (VUNPCKLPDYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
 def : Pat<(v2f64 (X86Unpcklpd VR128:$src1, VR128:$src2)),
           (UNPCKLPDrr VR128:$src1, VR128:$src2)>;
 
+// Shuffle with VUNPCKLPDY
+def : Pat<(v4f64 (X86Unpcklpdy VR256:$src1, (memopv4f64 addr:$src2))),
+          (VUNPCKLPDYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
+def : Pat<(v4f64 (X86Unpcklpdy VR256:$src1, VR256:$src2)),
+          (VUNPCKLPDYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
+
+def : Pat<(v4i64 (X86Unpcklpdy VR256:$src1, (memopv4i64 addr:$src2))),
+          (VUNPCKLPDYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
+def : Pat<(v4i64 (X86Unpcklpdy VR256:$src1, VR256:$src2)),
+          (VUNPCKLPDYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
+
 // Shuffle with UNPCKHPD
 def : Pat<(v2f64 (X86Unpckhpd VR128:$src1, (memopv2f64 addr:$src2))),
           (VUNPCKHPDrm VR128:$src1, addr:$src2)>, Requires<[HasAVX]>;
@@ -5859,6 +5877,10 @@ def : Pat<(v4f64 (X86Unpckhpdy VR256:$src1, (memopv4f64 addr:$src2))),
           (VUNPCKHPDYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
 def : Pat<(v4f64 (X86Unpckhpdy VR256:$src1, VR256:$src2)),
           (VUNPCKHPDYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
+def : Pat<(v4i64 (X86Unpckhpdy VR256:$src1, (memopv4i64 addr:$src2))),
+          (VUNPCKHPDYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
+def : Pat<(v4i64 (X86Unpckhpdy VR256:$src1, VR256:$src2)),
+          (VUNPCKHPDYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
 
 // Shuffle with MOVLHPS
 def : Pat<(X86Movlhps VR128:$src1,
diff --git a/test/CodeGen/X86/avx-unpack.ll b/test/CodeGen/X86/avx-unpack.ll
index 4e906ee155..d420101339 100644
--- a/test/CodeGen/X86/avx-unpack.ll
+++ b/test/CodeGen/X86/avx-unpack.ll
@@ -56,3 +56,34 @@ entry:
   ret <4 x double> %shuffle.i
 }
 
+;;;;
+;;;; Unpack versions using the fp unit for int unpacking
+;;;;
+
+; CHECK: vunpckhps
+define <8 x i32> @unpackhips1(<8 x i32> %src1, <8 x i32> %src2) nounwind uwtable readnone ssp {
+entry:
+  %shuffle.i = shufflevector <8 x i32> %src1, <8 x i32> %src2, <8 x i32> <i32 2, i32 10, i32 3, i32 11, i32 6, i32 14, i32 7, i32 15>
+  ret <8 x i32> %shuffle.i
+}
+
+; CHECK: vunpckhpd
+define <4 x i64> @unpackhipd1(<4 x i64> %src1, <4 x i64> %src2) nounwind uwtable readnone ssp {
+entry:
+  %shuffle.i = shufflevector <4 x i64> %src1, <4 x i64> %src2, <4 x i32> <i32 1, i32 5, i32 3, i32 7>
+  ret <4 x i64> %shuffle.i
+}
+
+; CHECK: vunpcklps
+define <8 x i32> @unpacklops1(<8 x i32> %src1, <8 x i32> %src2) nounwind uwtable readnone ssp {
+entry:
+  %shuffle.i = shufflevector <8 x i32> %src1, <8 x i32> %src2, <8 x i32> <i32 0, i32 8, i32 1, i32 9, i32 4, i32 12, i32 5, i32 13>
+  ret <8 x i32> %shuffle.i
+}
+
+; CHECK: vunpcklpd
+define <4 x i64> @unpacklopd1(<4 x i64> %src1, <4 x i64> %src2) nounwind uwtable readnone ssp {
+entry:
+  %shuffle.i = shufflevector <4 x i64> %src1, <4 x i64> %src2, <4 x i32> <i32 0, i32 4, i32 2, i32 6>
+  ret <4 x i64> %shuffle.i
+}