[llvm] r311258 - [ARM] Check the right order for halves of VZIP/VUZP if both parts are used

[Martin Storsjo via llvm-commits]

Author: mstorsjo
Date: Sat Aug 19 12:47:48 2017
New Revision: 311258

URL: http://llvm.org/viewvc/llvm-project?rev=311258&view=rev
Log:
[ARM] Check the right order for halves of VZIP/VUZP if both parts are used

This is the exact same fix as in SVN r247254. In that commit, the fix was
applied only for isVTRNMask and isVTRN_v_undef_Mask, but the same issue
is present for VZIP/VUZP as well.

This fixes PR33921.

Differential Revision: https://reviews.llvm.org/D36899

Modified:
    llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp
    llvm/trunk/test/CodeGen/ARM/vzip.ll

Modified: llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp?rev=311258&r1=311257&r2=311258&view=diff
==============================================================================
--- llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp Sat Aug 19 12:47:48 2017
@@ -5901,7 +5901,10 @@ static bool isVUZPMask(ArrayRef<int> M,
     return false;
 
   for (unsigned i = 0; i < M.size(); i += NumElts) {
-    WhichResult = M[i] == 0 ? 0 : 1;
+    if (M.size() == NumElts * 2)
+      WhichResult = i / NumElts;
+    else
+      WhichResult = M[i] == 0 ? 0 : 1;
     for (unsigned j = 0; j < NumElts; ++j) {
       if (M[i+j] >= 0 && (unsigned) M[i+j] != 2 * j + WhichResult)
         return false;
@@ -5932,7 +5935,10 @@ static bool isVUZP_v_undef_Mask(ArrayRef
 
   unsigned Half = NumElts / 2;
   for (unsigned i = 0; i < M.size(); i += NumElts) {
-    WhichResult = M[i] == 0 ? 0 : 1;
+    if (M.size() == NumElts * 2)
+      WhichResult = i / NumElts;
+    else
+      WhichResult = M[i] == 0 ? 0 : 1;
     for (unsigned j = 0; j < NumElts; j += Half) {
       unsigned Idx = WhichResult;
       for (unsigned k = 0; k < Half; ++k) {
@@ -5972,7 +5978,10 @@ static bool isVZIPMask(ArrayRef<int> M,
     return false;
 
   for (unsigned i = 0; i < M.size(); i += NumElts) {
-    WhichResult = M[i] == 0 ? 0 : 1;
+    if (M.size() == NumElts * 2)
+      WhichResult = i / NumElts;
+    else
+      WhichResult = M[i] == 0 ? 0 : 1;
     unsigned Idx = WhichResult * NumElts / 2;
     for (unsigned j = 0; j < NumElts; j += 2) {
       if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
@@ -6005,7 +6014,10 @@ static bool isVZIP_v_undef_Mask(ArrayRef
     return false;
 
   for (unsigned i = 0; i < M.size(); i += NumElts) {
-    WhichResult = M[i] == 0 ? 0 : 1;
+    if (M.size() == NumElts * 2)
+      WhichResult = i / NumElts;
+    else
+      WhichResult = M[i] == 0 ? 0 : 1;
     unsigned Idx = WhichResult * NumElts / 2;
     for (unsigned j = 0; j < NumElts; j += 2) {
       if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||

Modified: llvm/trunk/test/CodeGen/ARM/vzip.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/ARM/vzip.ll?rev=311258&r1=311257&r2=311258&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/ARM/vzip.ll (original)
+++ llvm/trunk/test/CodeGen/ARM/vzip.ll Sat Aug 19 12:47:48 2017
@@ -282,6 +282,25 @@ entry:
   ret <8 x i16> %0
 }
 
+; NOTE: The mask here looks like something that could be done with a vzip,
+; but which the current handling of two-result vzip can't do - thus ending up
+; as a vtrn.
+define <8 x i16> @vzip_lower_shufflemask_undef_rev(<4 x i16>* %A, <4 x i16>* %B) {
+; CHECK-LABEL: vzip_lower_shufflemask_undef_rev:
+; CHECK:       @ BB#0: @ %entry
+; CHECK-NEXT:    vldr d16, [r1]
+; CHECK-NEXT:    vldr d19, [r0]
+; CHECK-NEXT:    vtrn.16 d19, d16
+; CHECK-NEXT:    vmov r0, r1, d18
+; CHECK-NEXT:    vmov r2, r3, d19
+; CHECK-NEXT:    mov pc, lr
+entry:
+  %tmp1 = load <4 x i16>, <4 x i16>* %A
+  %tmp2 = load <4 x i16>, <4 x i16>* %B
+  %0 = shufflevector <4 x i16> %tmp1, <4 x i16> %tmp2, <8 x i32> <i32 undef, i32 undef, i32 undef, i32 undef, i32 0, i32 4, i32 undef, i32 undef>
+  ret <8 x i16> %0
+}
+
 define <4 x i32> @vzip_lower_shufflemask_zeroed(<2 x i32>* %A) {
 ; CHECK-LABEL: vzip_lower_shufflemask_zeroed:
 ; CHECK:       @ BB#0: @ %entry