[x86] Fix crashes in lowering bitcast instructions with the widening

mode.

This also runs the test in that mode which would reproduce the crash.
What I love is that *every single FIXME* in the test is addressed by
switching to widening.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@212254 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index ede29662e14262b159abac96aa6b4cfda94b9f9d..09052eddc930ec0af72c25eb6b2cd63ceedb6334 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -16410,6 +16410,13 @@ void X86TargetLowering::ReplaceNodeResults(SDNode *N,
                                    MVT::v2f64, N->getOperand(0));
     SDValue ToVecInt = DAG.getNode(ISD::BITCAST, dl, WiderVT, Expanded);
 
+    if (ExperimentalVectorWideningLegalization) {
+      // If we are legalizing vectors by widening, we already have the desired
+      // legal vector type, just return it.
+      Results.push_back(ToVecInt);
+      return;
+    }
+
     SmallVector<SDValue, 8> Elts;
     for (unsigned i = 0, e = NumElts; i != e; ++i)
       Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,

diff --git a/test/CodeGen/X86/lower-bitcast.ll b/test/CodeGen/X86/lower-bitcast.ll
index 769831ee818551b4f5e6185478066f3d8051bcb3..1c6327bbc232c20de43de39c965689849bc15d87 100644 (file)
--- a/test/CodeGen/X86/lower-bitcast.ll
+++ b/test/CodeGen/X86/lower-bitcast.ll
@@ -1,4 +1,5 @@
 ; RUN: llc < %s -march=x86-64 -mcpu=core2 -mattr=+sse2 | FileCheck %s
+; RUN: llc < %s -march=x86-64 -mcpu=core2 -mattr=+sse2 -x86-experimental-vector-widening-legalization | FileCheck %s --check-prefix=CHECK-WIDE
 
 
 define double @test1(double %A) {
@@ -9,14 +10,19 @@ define double @test1(double %A) {
 }
 ; FIXME: Ideally we should be able to fold the entire body of @test1 into a
 ; single paddd instruction. At the moment we produce the sequence 
-; pshufd+paddq+pshufd.
-
+; pshufd+paddq+pshufd. This is fixed with the widening legalization.
+;
 ; CHECK-LABEL: test1
 ; CHECK-NOT: movsd
 ; CHECK: pshufd
 ; CHECK-NEXT: paddd
 ; CHECK-NEXT: pshufd
 ; CHECK-NEXT: ret
+;
+; CHECK-WIDE-LABEL: test1
+; CHECK-WIDE-NOT: movsd
+; CHECK-WIDE: paddd
+; CHECK-WIDE-NEXT: ret
 
 
 define double @test2(double %A, double %B) {
@@ -30,6 +36,11 @@ define double @test2(double %A, double %B) {
 ; CHECK-NOT: movsd
 ; CHECK: paddd
 ; CHECK-NEXT: ret
+;
+; CHECK-WIDE-LABEL: test2
+; CHECK-WIDE-NOT: movsd
+; CHECK-WIDE: paddd
+; CHECK-WIDE-NEXT: ret
 
 
 define i64 @test3(i64 %A) {
@@ -43,6 +54,12 @@ define i64 @test3(i64 %A) {
 ; CHECK: addps
 ; CHECK-NOT: pshufd
 ; CHECK: ret
+;
+; CHECK-WIDE-LABEL: test3
+; CHECK-WIDE-NOT: pshufd
+; CHECK-WIDE: addps
+; CHECK-WIDE-NOT: pshufd
+; CHECK-WIDE: ret
 
 
 define i64 @test4(i64 %A) {
@@ -52,13 +69,20 @@ define i64 @test4(i64 %A) {
   ret i64 %2
 }
 ; FIXME: At the moment we still produce the sequence pshufd+paddq+pshufd.
-; Ideally, we should fold that sequence into a single paddd.
-
+; Ideally, we should fold that sequence into a single paddd. This is fixed with
+; the widening legalization.
+;
 ; CHECK-LABEL: test4
 ; CHECK: pshufd
 ; CHECK-NEXT: paddq
 ; CHECK-NEXT: pshufd
 ; CHECK: ret
+;
+; CHECK-WIDE-LABEL: test4
+; CHECK-WIDE: movd %rdi,
+; CHECK-WIDE-NEXT: paddd
+; CHECK-WIDE-NEXT: movd {{.*}}, %rax
+; CHECK-WIDE: ret
 
 
 define double @test5(double %A) {
@@ -70,6 +94,10 @@ define double @test5(double %A) {
 ; CHECK-LABEL: test5
 ; CHECK: addps
 ; CHECK-NEXT: ret
+;
+; CHECK-WIDE-LABEL: test5
+; CHECK-WIDE: addps
+; CHECK-WIDE-NEXT: ret
 
 
 define double @test6(double %A) {
@@ -79,14 +107,20 @@ define double @test6(double %A) {
   ret double %2
 }
 ; FIXME: Ideally we should be able to fold the entire body of @test6 into a
-; single paddw instruction.
-
+; single paddw instruction. This is fixed with the widening legalization.
+;
 ; CHECK-LABEL: test6
 ; CHECK-NOT: movsd
 ; CHECK: punpcklwd
 ; CHECK-NEXT: paddw
 ; CHECK-NEXT: pshufb
 ; CHECK-NEXT: ret
+;
+; CHECK-WIDE-LABEL: test6
+; CHECK-WIDE-NOT: mov
+; CHECK-WIDE-NOT: punpcklwd
+; CHECK-WIDE: paddw
+; CHECK-WIDE-NEXT: ret
 
 
 define double @test7(double %A, double %B) {
@@ -101,6 +135,12 @@ define double @test7(double %A, double %B) {
 ; CHECK-NOT: punpcklwd
 ; CHECK: paddw
 ; CHECK-NEXT: ret
+;
+; CHECK-WIDE-LABEL: test7
+; CHECK-WIDE-NOT: movsd
+; CHECK-WIDE-NOT: punpcklwd
+; CHECK-WIDE: paddw
+; CHECK-WIDE-NEXT: ret
 
 
 define double @test8(double %A) {
@@ -111,14 +151,20 @@ define double @test8(double %A) {
 }
 ; FIXME: Ideally we should be able to fold the entire body of @test8 into a
 ; single paddb instruction. At the moment we produce the sequence 
-; pshufd+paddw+pshufd.
-
+; pshufd+paddw+pshufd. This is fixed with the widening legalization.
+;
 ; CHECK-LABEL: test8
 ; CHECK-NOT: movsd
 ; CHECK: punpcklbw
 ; CHECK-NEXT: paddb
 ; CHECK-NEXT: pshufb
 ; CHECK-NEXT: ret
+;
+; CHECK-WIDE-LABEL: test8
+; CHECK-WIDE-NOT: movsd
+; CHECK-WIDE-NOT: punpcklbw
+; CHECK-WIDE: paddb
+; CHECK-WIDE-NEXT: ret
 
 
 define double @test9(double %A, double %B) {
@@ -133,4 +179,10 @@ define double @test9(double %A, double %B) {
 ; CHECK-NOT: punpcklbw
 ; CHECK: paddb
 ; CHECK-NEXT: ret
+;
+; CHECK-WIDE-LABEL: test9
+; CHECK-WIDE-NOT: movsd
+; CHECK-WIDE-NOT: punpcklbw
+; CHECK-WIDE: paddb
+; CHECK-WIDE-NEXT: ret