From bc92b2ca377145f4d39fa0283c41c69e3b42527b Mon Sep 17 00:00:00 2001
From: Ahmed Bougacha <ahmed.bougacha@gmail.com>
Date: Mon, 27 Apr 2015 21:01:20 +0000
Subject: [PATCH] [AArch64] Don't assert when combining (v3f32 select (setcc
 f64)).

When the setcc has f64 operands, we can't build a vector setcc mask
to feed a vselect, because f64 doesn't divide v3f32 evenly.
Just bail out when that happens.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@235917 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/AArch64/AArch64ISelLowering.cpp   |  6 ++++
 test/CodeGen/AArch64/arm64-neon-select_cc.ll | 32 +++++++++++++++++++-
 2 files changed, 37 insertions(+), 1 deletion(-)

diff --git a/lib/Target/AArch64/AArch64ISelLowering.cpp b/lib/Target/AArch64/AArch64ISelLowering.cpp
index 0607c4a3b91..e66b07ea2f4 100644
--- a/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -8668,6 +8668,12 @@ static SDValue performSelectCombine(SDNode *N, SelectionDAG &DAG) {
   SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts);
   EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
 
+  // Also bail out if the vector CCVT isn't the same size as ResVT.
+  // This can happen if the SETCC operand size doesn't divide the ResVT size
+  // (e.g., f64 vs v3f32).
+  if (CCVT.getSizeInBits() != ResVT.getSizeInBits())
+    return SDValue();
+
   // First perform a vector comparison, where lane 0 is the one we're interested
   // in.
   SDLoc DL(N0);
diff --git a/test/CodeGen/AArch64/arm64-neon-select_cc.ll b/test/CodeGen/AArch64/arm64-neon-select_cc.ll
index d334c0846ac..b98d2d9219b 100644
--- a/test/CodeGen/AArch64/arm64-neon-select_cc.ll
+++ b/test/CodeGen/AArch64/arm64-neon-select_cc.ll
@@ -1,4 +1,5 @@
-; RUN: llc < %s -verify-machineinstrs -mtriple=arm64-none-linux-gnu -mattr=+neon -fp-contract=fast | FileCheck %s
+; RUN: llc  -mtriple=arm64-none-linux-gnu -mattr=+neon -fp-contract=fast \
+; RUN:        < %s -verify-machineinstrs -asm-verbose=false | FileCheck %s
 
 define <8x i8> @test_select_cc_v8i8_i8(i8 %a, i8 %b, <8x i8> %c, <8x i8> %d ) {
 ; CHECK-LABEL: test_select_cc_v8i8_i8:
@@ -219,3 +220,32 @@ define <2 x i32> @test_select_cc_v2i32_icmpi1(i1 %cc, <2 x i32> %a, <2 x i32> %b
   %e = select i1 %cmp, <2 x i32> %a, <2 x i32> %b
   ret <2 x i32> %e
 }
+
+; Also make sure we support irregular/non-power-of-2 types such as v3f32.
+define <3 x float> @test_select_cc_v3f32_fcmp_f32(<3 x float> %a, <3 x float> %b, float %c1, float %c2) #0 {
+; CHECK-LABEL: test_select_cc_v3f32_fcmp_f32:
+; CHECK-NEXT: fcmeq [[MASK:v[0-9]+]].4s, v2.4s, v3.4s
+; CHECK-NEXT: dup [[VMASK:v[0-9]+]].4s, [[MASK]].s[0]
+; CHECK-NEXT: bsl [[RES:v[0-9]+]].16b, v0.16b, v1.16b
+; CHECK-NEXT: mov v0.16b, [[RES]].16b
+; CHECK-NEXT: ret
+  %cc = fcmp oeq float %c1, %c2
+  %r = select i1 %cc, <3 x float> %a, <3 x float> %b
+  ret <3 x float> %r
+}
+
+define <3 x float> @test_select_cc_v3f32_fcmp_f64(<3 x float> %a, <3 x float> %b, double %c1, double %c2) #0 {
+; CHECK-LABEL: test_select_cc_v3f32_fcmp_f64:
+; CHECK-NEXT: fcmp d2, d3
+; CHECK-NEXT: movn [[N0:w[0-9]+]], #0
+; CHECK-NEXT: csel [[MASK:w[0-9]+]], [[N0]], wzr, eq
+; CHECK-NEXT: dup [[VMASK:v[0-9]+]].4s, [[MASK]]
+; CHECK-NEXT: bsl [[RES:v[0-9]+]].16b, v0.16b, v1.16b
+; CHECK-NEXT: mov v0.16b, [[RES]].16b
+; CHECK-NEXT: ret
+  %cc = fcmp oeq double %c1, %c2
+  %r = select i1 %cc, <3 x float> %a, <3 x float> %b
+  ret <3 x float> %r
+}
+
+attributes #0 = { nounwind}