[X86] Also create+widen FMIN/FMAX nodes for v2f32.

This happens in the HINT benchmark, where the SLP-vectorizer created
v2f32 fcmp/select code.  The "correct" solution would have been to
teach the vectorizer cost model that v2f32 isn't legal (because really,
it isn't), but if we can vectorize we might as well do so.

We legalize these v2f32 FMIN/FMAX nodes by widening to v4f32 later on.
v3f32 were already widened to v4f32 by the generic unroll-and-build-vector
legalization.

rdar://15763436
Differential Revision: http://reviews.llvm.org/D6557

llvm-svn: 225691

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 65440a0..f455700 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -19658,6 +19658,22 @@ void X86TargetLowering::ReplaceNodeResults(SDNode *N,
   switch (N->getOpcode()) {
   default:
     llvm_unreachable("Do not know how to custom type legalize this operation!");
+  // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
+  case X86ISD::FMINC:
+  case X86ISD::FMIN:
+  case X86ISD::FMAXC:
+  case X86ISD::FMAX: {
+    EVT VT = N->getValueType(0);
+    if (VT != MVT::v2f32)
+      llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
+    SDValue UNDEF = DAG.getUNDEF(VT);
+    SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
+                              N->getOperand(0), UNDEF);
+    SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
+                              N->getOperand(1), UNDEF);
+    Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
+    return;
+  }
   case ISD::SIGN_EXTEND_INREG:
   case ISD::ADDC:
   case ISD::ADDE:
@@ -23077,8 +23093,9 @@ static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
   // instructions match the semantics of the common C idiom x<y?x:y but not
   // x<=y?x:y, because of how they handle negative zero (which can be
   // ignored in unsafe-math mode).
+  // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
   if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
-      VT != MVT::f80 && TLI.isTypeLegal(VT) &&
+      VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
       (Subtarget->hasSSE2() ||
        (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
     ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();

diff --git a/llvm/test/CodeGen/X86/sse-minmax.ll b/llvm/test/CodeGen/X86/sse-minmax.ll
index 46f2b56..4dcb54c 100644 (file)
--- a/llvm/test/CodeGen/X86/sse-minmax.ll
+++ b/llvm/test/CodeGen/X86/sse-minmax.ll
@@ -989,3 +989,39 @@ define <4 x float> @test_minps(<4 x float> %x, <4 x float> %y) nounwind {
   %min = select <4 x i1> %min_is_x, <4 x float> %x, <4 x float> %y
   ret <4 x float> %min
 }
+
+; UNSAFE-LABEL: test_maxps_illegal_v2f32:
+; UNSAFE-NEXT: maxps %xmm1, %xmm0
+; UNSAFE-NEXT: ret
+define <2 x float> @test_maxps_illegal_v2f32(<2 x float> %x, <2 x float> %y) nounwind {
+  %max_is_x = fcmp oge <2 x float> %x, %y
+  %max = select <2 x i1> %max_is_x, <2 x float> %x, <2 x float> %y
+  ret <2 x float> %max
+}
+
+; UNSAFE-LABEL: test_minps_illegal_v2f32:
+; UNSAFE-NEXT: minps %xmm1, %xmm0
+; UNSAFE-NEXT: ret
+define <2 x float> @test_minps_illegal_v2f32(<2 x float> %x, <2 x float> %y) nounwind {
+  %min_is_x = fcmp ole <2 x float> %x, %y
+  %min = select <2 x i1> %min_is_x, <2 x float> %x, <2 x float> %y
+  ret <2 x float> %min
+}
+
+; UNSAFE-LABEL: test_maxps_illegal_v3f32:
+; UNSAFE-NEXT: maxps %xmm1, %xmm0
+; UNSAFE-NEXT: ret
+define <3 x float> @test_maxps_illegal_v3f32(<3 x float> %x, <3 x float> %y) nounwind {
+  %max_is_x = fcmp oge <3 x float> %x, %y
+  %max = select <3 x i1> %max_is_x, <3 x float> %x, <3 x float> %y
+  ret <3 x float> %max
+}
+
+; UNSAFE-LABEL: test_minps_illegal_v3f32:
+; UNSAFE-NEXT: minps %xmm1, %xmm0
+; UNSAFE-NEXT: ret
+define <3 x float> @test_minps_illegal_v3f32(<3 x float> %x, <3 x float> %y) nounwind {
+  %min_is_x = fcmp ole <3 x float> %x, %y
+  %min = select <3 x i1> %min_is_x, <3 x float> %x, <3 x float> %y
+  ret <3 x float> %min
+}