[llvm] r190147 - SelectionDAG: create correct BooleanContent constants

[Tim Northover]

Author: tnorthover
Date: Fri Sep  6 07:38:12 2013
New Revision: 190147

URL: http://llvm.org/viewvc/llvm-project?rev=190147&view=rev
Log:
SelectionDAG: create correct BooleanContent constants

Occasionally DAGCombiner can spot that a SETCC operation is completely
redundant and reduce it to "all true" or "all false". If this happens to a
vector, the value produced has to take account of what a normal comparison
would have produced, which may be an all-1s bitmask.

The fix in SelectionDAG.cpp is tested, however, as far as I can see the code in
TargetLowering.cpp is possibly unreachable and almost certainly irrelevant when
triggered so there are no tests. However, I believe it's still clearly the
right change and may save someone else some hassle if it suddenly becomes
reachable. So I'm doing it anyway.

Modified:
    llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
    llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp
    llvm/trunk/test/CodeGen/X86/vec_setcc.ll

Modified: llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp?rev=190147&r1=190146&r2=190147&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp (original)
+++ llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp Fri Sep  6 07:38:12 2013
@@ -1554,7 +1554,12 @@ SDValue SelectionDAG::FoldSetCC(EVT VT,
   case ISD::SETFALSE:
   case ISD::SETFALSE2: return getConstant(0, VT);
   case ISD::SETTRUE:
-  case ISD::SETTRUE2:  return getConstant(1, VT);
+  case ISD::SETTRUE2: {
+    const TargetLowering *TLI = TM.getTargetLowering();
+    TargetLowering::BooleanContent Cnt = TLI->getBooleanContents(VT.isVector());
+    return getConstant(
+        Cnt == TargetLowering::ZeroOrNegativeOneBooleanContent ? -1ULL : 1, VT);
+  }
 
   case ISD::SETOEQ:
   case ISD::SETOGT:

Modified: llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp?rev=190147&r1=190146&r2=190147&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp (original)
+++ llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp Fri Sep  6 07:38:12 2013
@@ -1080,7 +1080,11 @@ TargetLowering::SimplifySetCC(EVT VT, SD
   case ISD::SETFALSE:
   case ISD::SETFALSE2: return DAG.getConstant(0, VT);
   case ISD::SETTRUE:
-  case ISD::SETTRUE2:  return DAG.getConstant(1, VT);
+  case ISD::SETTRUE2: {
+    TargetLowering::BooleanContent Cnt = getBooleanContents(VT.isVector());
+    return DAG.getConstant(
+        Cnt == TargetLowering::ZeroOrNegativeOneBooleanContent ? -1ULL : 1, VT);
+  }
   }
 
   // Ensure that the constant occurs on the RHS, and fold constant

Modified: llvm/trunk/test/CodeGen/X86/vec_setcc.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/vec_setcc.ll?rev=190147&r1=190146&r2=190147&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/vec_setcc.ll (original)
+++ llvm/trunk/test/CodeGen/X86/vec_setcc.ll Fri Sep  6 07:38:12 2013
@@ -124,3 +124,64 @@ define <4 x i32> @v4i32_icmp_ule(<4 x i3
 ; AVX: pcmpeqd %xmm1, %xmm0, %xmm0
 }
 
+; At one point we were incorrectly constant-folding a setcc to 0x1 instead of
+; 0xff, leading to a constpool load. The instruction doesn't matter here, but it
+; should set all bits to 1.
+define <16 x i8> @test_setcc_constfold_vi8(<16 x i8> %l, <16 x i8> %r) {
+  %test1 = icmp eq <16 x i8> %l, %r
+  %mask1 = sext <16 x i1> %test1 to <16 x i8>
+
+  %test2 = icmp ne <16 x i8> %l, %r
+  %mask2 = sext <16 x i1> %test2 to <16 x i8>
+
+  %res = or <16 x i8> %mask1, %mask2
+  ret <16 x i8> %res
+; SSE2-LABEL: test_setcc_constfold_vi8:
+; SSE2: pcmpeqd %xmm0, %xmm0
+
+; SSE41-LABEL: test_setcc_constfold_vi8:
+; SSE41: pcmpeqd %xmm0, %xmm0
+
+; AVX-LABEL: test_setcc_constfold_vi8:
+; AVX: vpcmpeqd %xmm0, %xmm0, %xmm0
+}
+
+; Make sure sensible results come from doing extension afterwards
+define <16 x i8> @test_setcc_constfold_vi1(<16 x i8> %l, <16 x i8> %r) {
+  %test1 = icmp eq <16 x i8> %l, %r
+  %test2 = icmp ne <16 x i8> %l, %r
+
+  %res = or <16 x i1> %test1, %test2
+  %mask = sext <16 x i1> %res to <16 x i8>
+  ret <16 x i8> %mask
+; SSE2-LABEL: test_setcc_constfold_vi1:
+; SSE2: pcmpeqd %xmm0, %xmm0
+
+; SSE41-LABEL: test_setcc_constfold_vi1:
+; SSE41: pcmpeqd %xmm0, %xmm0
+
+; AVX-LABEL: test_setcc_constfold_vi1:
+; AVX: vpcmpeqd %xmm0, %xmm0, %xmm0
+}
+
+
+; 64-bit case is also particularly important, as the constant "-1" is probably
+; just 32-bits wide.
+define <2 x i64> @test_setcc_constfold_vi64(<2 x i64> %l, <2 x i64> %r) {
+  %test1 = icmp eq <2 x i64> %l, %r
+  %mask1 = sext <2 x i1> %test1 to <2 x i64>
+
+  %test2 = icmp ne <2 x i64> %l, %r
+  %mask2 = sext <2 x i1> %test2 to <2 x i64>
+
+  %res = or <2 x i64> %mask1, %mask2
+  ret <2 x i64> %res
+; SSE2-LABEL: test_setcc_constfold_vi64:
+; SSE2: pcmpeqd %xmm0, %xmm0
+
+; SSE41-LABEL: test_setcc_constfold_vi64:
+; SSE41: pcmpeqd %xmm0, %xmm0
+
+; AVX-LABEL: test_setcc_constfold_vi64:
+; AVX: vpcmpeqd %xmm0, %xmm0, %xmm0
+}