[AVX] Fix mask predicates for 256-bit UNPCKLPS/D and implement

missing patterns for them.

      Add a SIMD test subdirectory to hold tests for SIMD instruction
      selection correctness and quality.
'


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@126845 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/X86/Utils/X86ShuffleDecode.cpp

@@ -165,12 +165,25 @@ void DecodeUNPCKLPDMask(unsigned NElts,
 /// datatypes and vector widths.
 void DecodeUNPCKLPMask(EVT VT,
                        SmallVectorImpl<unsigned> &ShuffleMask) {
+  unsigned NumElts = VT.getVectorNumElements();
 
-  unsigned NElts = VT.getVectorNumElements();
+  // Handle vector lengths > 128 bits.  Define a "section" as a set of
+  // 128 bits.  AVX defines UNPCK* to operate independently on 128-bit
+  // sections.
+  unsigned NumSections = VT.getSizeInBits() / 128;
+  if (NumSections == 0 ) NumSections = 1;  // Handle MMX
+  unsigned NumSectionElts = NumElts / NumSections;
 
-  for (unsigned i = 0; i != NElts/2; ++i) {
+  unsigned Start = 0;
-    ShuffleMask.push_back(i);        // Reads from dest
+  unsigned End = NumSectionElts / 2;
-    ShuffleMask.push_back(i+NElts);  // Reads from src
+  for (unsigned s = 0; s < NumSections; ++s) {
+    for (unsigned i = Start; i != End; ++i) {
+      ShuffleMask.push_back(i);                 // Reads from dest/src1
+      ShuffleMask.push_back(i+NumSectionElts);  // Reads from src/src2
+    }
+    // Process the next 128 bits.
+    Start += NumSectionElts;
+    End += NumSectionElts;
   }
 }
 

lib/Target/X86/X86ISelLowering.cpp

@@ -3173,7 +3173,8 @@ bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
   unsigned NumElems = N->getValueType(0).getVectorNumElements();
 
-  if (NumElems != 2 && NumElems != 4)
+  if ((NumElems != 2 && NumElems != 4)
+      || N->getValueType(0).getSizeInBits() > 128)
     return false;
 
   for (unsigned i = 0; i < NumElems/2; ++i)
@@ -3195,19 +3196,36 @@ static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
   if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
     return false;
 
-  for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
+  // Handle vector lengths > 128 bits.  Define a "section" as a set of
-    int BitI  = Mask[i];
+  // 128 bits.  AVX defines UNPCK* to operate independently on 128-bit
-    int BitI1 = Mask[i+1];
+  // sections.
-    if (!isUndefOrEqual(BitI, j))
+  unsigned NumSections = VT.getSizeInBits() / 128;
-      return false;
+  if (NumSections == 0 ) NumSections = 1;  // Handle MMX
-    if (V2IsSplat) {
+  unsigned NumSectionElts = NumElts / NumSections;
-      if (!isUndefOrEqual(BitI1, NumElts))
+
-        return false;
+  unsigned Start = 0;
-    } else {
+  unsigned End = NumSectionElts;
-      if (!isUndefOrEqual(BitI1, j + NumElts))
+  for (unsigned s = 0; s < NumSections; ++s) {
+    for (unsigned i = Start, j = s * NumSectionElts;
+         i != End;
+         i += 2, ++j) {
+      int BitI  = Mask[i];
+      int BitI1 = Mask[i+1];
+      if (!isUndefOrEqual(BitI, j))
         return false;
+      if (V2IsSplat) {
+        if (!isUndefOrEqual(BitI1, NumElts))
+          return false;
+      } else {
+        if (!isUndefOrEqual(BitI1, j + NumElts))
+          return false;
+      }
     }
+    // Process the next 128 bits.
+    Start += NumSectionElts;
+    End += NumSectionElts;
   }
+
   return true;
 }
 
@@ -3255,14 +3273,27 @@ static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
   if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
     return false;
 
-  for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
+  // Handle vector lengths > 128 bits.  Define a "section" as a set of
-    int BitI  = Mask[i];
+  // 128 bits.  AVX defines UNPCK* to operate independently on 128-bit
-    int BitI1 = Mask[i+1];
+  // sections.
-    if (!isUndefOrEqual(BitI, j))
+  unsigned NumSections = VT.getSizeInBits() / 128;
-      return false;
+  if (NumSections == 0 ) NumSections = 1;  // Handle MMX
-    if (!isUndefOrEqual(BitI1, j))
+  unsigned NumSectionElts = NumElems / NumSections;
-      return false;
+
+  for (unsigned s = 0; s < NumSections; ++s) {
+    for (unsigned i = s * NumSectionElts, j = s * NumSectionElts;
+         i != NumSectionElts * (s + 1);
+         i += 2, ++j) {
+      int BitI  = Mask[i];
+      int BitI1 = Mask[i+1];
+
+      if (!isUndefOrEqual(BitI, j))
+        return false;
+      if (!isUndefOrEqual(BitI1, j))
+        return false;
+    }
   }
+
   return true;
 }
 

lib/Target/X86/X86InstrFragmentsSIMD.td

@@ -132,6 +132,8 @@ def X86Movlpd : SDNode<"X86ISD::MOVLPD", SDTShuff2Op>;
 
 def X86Unpcklps : SDNode<"X86ISD::UNPCKLPS", SDTShuff2Op>;
 def X86Unpcklpd : SDNode<"X86ISD::UNPCKLPD", SDTShuff2Op>;
+def X86Unpcklpsy : SDNode<"X86ISD::VUNPCKLPSY", SDTShuff2Op>;
+def X86Unpcklpdy : SDNode<"X86ISD::VUNPCKLPDY", SDTShuff2Op>;
 def X86Unpckhps : SDNode<"X86ISD::UNPCKHPS", SDTShuff2Op>;
 def X86Unpckhpd : SDNode<"X86ISD::UNPCKHPD", SDTShuff2Op>;
 

lib/Target/X86/X86InstrSSE.td

@@ -5622,11 +5622,15 @@ 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)>;
 
@@ -5644,11 +5648,15 @@ def : Pat<(v4f32 (X86Unpckhps VR128:$src1, VR128:$src2)),
 // 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)>;
 

test/CodeGen/X86/SIMD/dg.exp

@@ -0,0 +1,5 @@
+load_lib llvm.exp
+
+if { [llvm_supports_target X86] } {
+  RunLLVMTests [lsort [glob -nocomplain $srcdir/$subdir/*.{ll,c,cpp}]]
+}

test/CodeGen/X86/SIMD/notvunpcklpd.ll

@@ -0,0 +1,20 @@
+; RUN: llc < %s -mattr=+avx | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @try_([2 x <4 x double>]* noalias %incarray, [2 x <4 x double>]* noalias %incarrayb ) {
+entry:
+	%incarray1 = alloca [2 x <4 x double>]*, align 8
+	%incarrayb1 = alloca [2 x <4 x double>]*, align 8
+	%carray = alloca [2 x <4 x double>], align 16
+	%r = getelementptr [2 x <4 x double>]* %incarray, i32 0, i32 0
+	%rb = getelementptr [2 x <4 x double>]* %incarrayb, i32 0, i32 0
+	%r3 = load <4 x double>* %r, align 8
+	%r4 = load <4 x double>* %rb, align 8
+	%r11 = shufflevector <4 x double> %r3, <4 x double> %r4, <4 x i32> < i32 0, i32 4, i32 1, i32 5 >		; <<4 x double>> [#uses=1]
+; CHECK-NOT: vunpcklpd
+	%r12 = getelementptr [2 x <4 x double>]* %carray, i32 0, i32 1
+	store <4 x double> %r11, <4 x double>* %r12, align 4
+	ret void
+}

test/CodeGen/X86/SIMD/notvunpcklps.ll

@@ -0,0 +1,20 @@
+; RUN: llc < %s -mattr=+avx | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @try_([2 x <8 x float>]* noalias %incarray, [2 x <8 x float>]* noalias %incarrayb ) {
+enmtry:
+	%incarray1 = alloca [2 x <8 x float>]*, align 8
+	%incarrayb1 = alloca [2 x <8 x float>]*, align 8
+	%carray = alloca [2 x <8 x float>], align 16
+	%r = getelementptr [2 x <8 x float>]* %incarray, i32 0, i32 0
+	%rb = getelementptr [2 x <8 x float>]* %incarrayb, i32 0, i32 0
+	%r3 = load <8 x float>* %r, align 8
+	%r4 = load <8 x float>* %rb, align 8
+	%r8 = shufflevector <8 x float> %r3, <8 x float> %r4, <8 x i32> < i32 0, i32 8, i32 1, i32 9, i32 2, i32 10, i32 3, i32 11 >		; <<8 x float>> [#uses=1]
+; CHECK-NOT: vunpcklps
+	%r9 = getelementptr [2 x <8 x float>]* %carray, i32 0, i32 0
+	store <8 x float> %r8, <8 x float>* %r9, align 4
+	ret void
+}

test/CodeGen/X86/SIMD/vunpcklpd.ll

@@ -0,0 +1,20 @@
+; RUN: llc < %s -mattr=+avx | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @try_([2 x <4 x double>]* noalias %incarray, [2 x <4 x double>]* noalias %incarrayb ) {
+entry:
+	%incarray1 = alloca [2 x <4 x double>]*, align 8
+	%incarrayb1 = alloca [2 x <4 x double>]*, align 8
+	%carray = alloca [2 x <4 x double>], align 16
+	%r = getelementptr [2 x <4 x double>]* %incarray, i32 0, i32 0
+	%rb = getelementptr [2 x <4 x double>]* %incarrayb, i32 0, i32 0
+	%r3 = load <4 x double>* %r, align 8
+	%r4 = load <4 x double>* %rb, align 8
+	%r11 = shufflevector <4 x double> %r3, <4 x double> %r4, <4 x i32> < i32 0, i32 4, i32 2, i32 6 >		; <<4 x double>> [#uses=1]
+; CHECK: vunpcklpd
+	%r12 = getelementptr [2 x <4 x double>]* %carray, i32 0, i32 1
+	store <4 x double> %r11, <4 x double>* %r12, align 4
+	ret void
+}

test/CodeGen/X86/SIMD/vunpcklps.ll

@@ -0,0 +1,20 @@
+; RUN: llc < %s -mattr=+avx | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @try_([2 x <8 x float>]* noalias %incarray, [2 x <8 x float>]* noalias %incarrayb ) {
+entry:
+	%incarray1 = alloca [2 x <8 x float>]*, align 8
+	%incarrayb1 = alloca [2 x <8 x float>]*, align 8
+	%carray = alloca [2 x <8 x float>], align 16
+	%r = getelementptr [2 x <8 x float>]* %incarray, i32 0, i32 0
+	%rb = getelementptr [2 x <8 x float>]* %incarrayb, i32 0, i32 0
+	%r3 = load <8 x float>* %r, align 8
+	%r4 = load <8 x float>* %rb, align 8
+	%r11 = shufflevector <8 x float> %r3, <8 x float> %r4, <8 x i32> < i32 0, i32 8, i32 1, i32 9, i32 4, i32 12, i32 5, i32 13 >		; <<8 x float>> [#uses=1]
+; CHECK: vunpcklps
+	%r12 = getelementptr [2 x <8 x float>]* %carray, i32 0, i32 1
+	store <8 x float> %r11, <8 x float>* %r12, align 4
+	ret void
+}