From 14fe2e6948c394b0565787eddf59feec29a765da Mon Sep 17 00:00:00 2001
From: Ahmed Bougacha <ahmed.bougacha@gmail.com>
Date: Mon, 1 Dec 2014 20:59:00 +0000
Subject: [PATCH] [AArch64] Don't combine "select (setcc i1 LHS, RHS), vL, vR".

r208210 introduced an optimization that improves the vector select
codegen by doing the setcc on vectors directly.
This is a problem they the setcc operands are i1s, because the
optimization would create vectors of i1, which aren't legal.

Part of PR21549.

Differential Revision: http://reviews.llvm.org/D6308


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223075 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/AArch64/AArch64ISelLowering.cpp   |  6 ++++++
 test/CodeGen/AArch64/arm64-neon-select_cc.ll | 15 +++++++++++++++
 2 files changed, 21 insertions(+)

diff --git a/lib/Target/AArch64/AArch64ISelLowering.cpp b/lib/Target/AArch64/AArch64ISelLowering.cpp
index 16ad2f6e3b5..622b0e1d73d 100644
--- a/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -8479,6 +8479,12 @@ static SDValue performSelectCombine(SDNode *N, SelectionDAG &DAG) {
   // largest real NEON comparison is 64-bits per lane, which means the result is
   // at most 32-bits and an illegal vector. Just bail out for now.
   EVT SrcVT = N0.getOperand(0).getValueType();
+
+  // Don't try to do this optimization when the setcc itself has i1 operands.
+  // There are no legal vectors of i1, so this would be pointless.
+  if (SrcVT == MVT::i1)
+    return SDValue();
+
   int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits();
   if (!ResVT.isVector() || NumMaskElts == 0)
     return SDValue();
diff --git a/test/CodeGen/AArch64/arm64-neon-select_cc.ll b/test/CodeGen/AArch64/arm64-neon-select_cc.ll
index 95c582a5348..d334c0846ac 100644
--- a/test/CodeGen/AArch64/arm64-neon-select_cc.ll
+++ b/test/CodeGen/AArch64/arm64-neon-select_cc.ll
@@ -204,3 +204,18 @@ define <2 x double> @test_select_cc_v2f64(double %a, double %b, <2 x double> %c,
   %e = select i1 %cmp31, <2 x double> %c, <2 x double> %d
   ret <2 x double> %e
 }
+
+; Special case: when the select condition is an icmp with i1 operands, don't
+; do the comparison on vectors.
+; Part of PR21549.
+define <2 x i32> @test_select_cc_v2i32_icmpi1(i1 %cc, <2 x i32> %a, <2 x i32> %b) {
+; CHECK-LABEL: test_select_cc_v2i32_icmpi1:
+; CHECK: tst   w0, #0x1
+; CHECK: csetm [[MASK:w[0-9]+]], ne
+; CHECK: dup   [[DUPMASK:v[0-9]+]].2s, [[MASK]]
+; CHECK: bsl   [[DUPMASK]].8b, v0.8b, v1.8b
+; CHECK: mov   v0.16b, [[DUPMASK]].16b
+  %cmp = icmp ne i1 %cc, 0
+  %e = select i1 %cmp, <2 x i32> %a, <2 x i32> %b
+  ret <2 x i32> %e
+}