Fix a logic bug in x86 vector codegen: sext (zext (x) ) != sext (x) (PR20472).

Remove a block of code from LowerSIGN_EXTEND_INREG() that was added with:
http://llvm.org/viewvc/llvm-project?view=revision&revision=177421

And caused:
http://llvm.org/bugs/show_bug.cgi?id=20472 (more analysis here)
http://llvm.org/bugs/show_bug.cgi?id=18054

The testcases confirm that we (1) don't remove a zext op that is necessary and (2) generate
a pmovz instead of punpck if SSE4.1 is available. Although pmovz is 1 byte longer, it allows
folding of the load, and so saves 3 bytes overall.

Differential Revision: http://reviews.llvm.org/D4909

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@216679 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index a181a9070473db322d9e1935a80337e92bdce869..b3a02726ebfd8d4f1a8ae6763eb07e71561a09e4 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -16446,6 +16446,11 @@ static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
   return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
 }
 
+// Sign extension of the low part of vector elements. This may be used either
+// when sign extend instructions are not available or if the vector element
+// sizes already match the sign-extended size. If the vector elements are in
+// their pre-extended size and sign extend instructions are available, that will
+// be handled by LowerSIGN_EXTEND.
 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
                                                   SelectionDAG &DAG) const {
   SDLoc dl(Op);
@@ -16491,32 +16496,13 @@ SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
     case MVT::v4i32:
     case MVT::v8i16: {
       SDValue Op0 = Op.getOperand(0);
-      SDValue Op00 = Op0.getOperand(0);
-      SDValue Tmp1;
-      // Hopefully, this VECTOR_SHUFFLE is just a VZEXT.
-      if (Op0.getOpcode() == ISD::BITCAST &&
-          Op00.getOpcode() == ISD::VECTOR_SHUFFLE) {
-        // (sext (vzext x)) -> (vsext x)
-        Tmp1 = LowerVectorIntExtend(Op00, Subtarget, DAG);
-        if (Tmp1.getNode()) {
-          EVT ExtraEltVT = ExtraVT.getVectorElementType();
-          // This folding is only valid when the in-reg type is a vector of i8,
-          // i16, or i32.
-          if (ExtraEltVT == MVT::i8 || ExtraEltVT == MVT::i16 ||
-              ExtraEltVT == MVT::i32) {
-            SDValue Tmp1Op0 = Tmp1.getOperand(0);
-            assert(Tmp1Op0.getOpcode() == X86ISD::VZEXT &&
-                   "This optimization is invalid without a VZEXT.");
-            return DAG.getNode(X86ISD::VSEXT, dl, VT, Tmp1Op0.getOperand(0));
-          }
-          Op0 = Tmp1;
-        }
-      }
 
-      // If the above didn't work, then just use Shift-Left + Shift-Right.
-      Tmp1 = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Op0, BitsDiff,
-                                        DAG);
-      return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, Tmp1, BitsDiff,
+      // This is a sign extension of some low part of vector elements without
+      // changing the size of the vector elements themselves:
+      // Shift-Left + Shift-Right-Algebraic.
+      SDValue Shl = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Op0,
+                                               BitsDiff, DAG);
+      return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, Shl, BitsDiff,
                                         DAG);
     }
   }

diff --git a/test/CodeGen/X86/vec_trunc_sext.ll b/test/CodeGen/X86/vec_trunc_sext.ll
new file mode 100644 (file)
index 0000000..3c446bb
--- /dev/null
+++ b/test/CodeGen/X86/vec_trunc_sext.ll
@@ -0,0 +1,30 @@
+; RUN: llc %s -mtriple=x86_64-unknown-unknown -mattr='-sse4.1' -o - | FileCheck %s -check-prefix=NO_SSE_41
+; RUN: llc %s -mtriple=x86_64-unknown-unknown -mattr='+sse4.1' -o - | FileCheck %s -check-prefix=SSE_41
+
+; PR20472 ( http://llvm.org/bugs/show_bug.cgi?id=20472 )
+; When sexting a trunc'd vector value, we can't eliminate the zext.
+; If we don't have SSE4.1, use punpck.
+; If we have SSE4.1, use pmovzx because it combines the load op.
+; There may be a better way to do this using pshufb + pmovsx,
+; but that is beyond our current codegen capabilities.
+
+define <4 x i32> @trunc_sext(<4 x i16>* %in) {
+  %load = load <4 x i16>* %in
+  %trunc = trunc <4 x i16> %load to <4 x i8>
+  %sext = sext <4 x i8> %trunc to <4 x i32>
+  ret <4 x i32> %sext
+
+; NO_SSE_41-LABEL: trunc_sext:
+; NO_SSE_41: movq (%rdi), %xmm0
+; NO_SSE_41-NEXT: punpcklwd %xmm0, %xmm0
+; NO_SSE_41-NEXT: pslld $24, %xmm0
+; NO_SSE_41-NEXT: psrad $24, %xmm0
+; NO_SSE_41-NEXT: retq
+
+; SSE_41-LABEL: trunc_sext:
+; SSE_41: pmovzxwd (%rdi), %xmm0
+; SSE_41-NEXT: pslld $24, %xmm0
+; SSE_41-NEXT: psrad $24, %xmm0
+; SSE_41-NEXT: retq
+}
+