X86: Don't generate 64-bit movd after cmpneqsd in 32-bit mode (PR19059)

This fixes the bug where we would bitcast the 64-bit floating point result
of cmpneqsd to a 64-bit integer even on 32-bit targets.

Differential Revision: http://llvm-reviews.chandlerc.com/D3009

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203581 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index f651205bf3d2b0aa4fa9136d9449febb1ebadf4f..e22a6a5c324ad27f1630428c46898254bf34d494 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -18044,7 +18044,6 @@ static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
 
         if ((cc0 == X86::COND_E  && cc1 == X86::COND_NP) ||
             (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
-          bool is64BitFP = (CMP00.getValueType() == MVT::f64);
           // FIXME: need symbolic constants for these magic numbers.
           // See X86ATTInstPrinter.cpp:printSSECC().
           unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
@@ -18059,9 +18058,26 @@ static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
           SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
                                               CMP00.getValueType(), CMP00, CMP01,
                                               DAG.getConstant(x86cc, MVT::i8));
-          MVT IntVT = (is64BitFP ? MVT::i64 : MVT::i32); 
-          SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT,
-                                              OnesOrZeroesF);
+
+          bool is64BitFP = (CMP00.getValueType() == MVT::f64);
+          MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
+
+          if (is64BitFP && !Subtarget->is64Bit()) {
+            // On a 32-bit target, we cannot bitcast the 64-bit float to a
+            // 64-bit integer, since that's not a legal type. Since
+            // OnesOrZeroesF is all ones of all zeroes, we don't need all the
+            // bits, but can do this little dance to extract the lowest 32 bits
+            // and work with those going forward.
+            SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
+                                           OnesOrZeroesF);
+            SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32,
+                                           Vector64);
+            OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
+                                        Vector32, DAG.getIntPtrConstant(0));
+            IntVT = MVT::i32;
+          }
+
+          SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT, OnesOrZeroesF);
           SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
                                       DAG.getConstant(1, IntVT));
           SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);

diff --git a/test/CodeGen/X86/isint.ll b/test/CodeGen/X86/isint.ll
index 96c89605b6ae907e35dfb040f10440a407671f99..38d05c662bd5d346782d6caf2c8dde1041bac63d 100644 (file)
--- a/test/CodeGen/X86/isint.ll
+++ b/test/CodeGen/X86/isint.ll
@@ -1,6 +1,11 @@
-; RUN: llc < %s -march=x86 -mattr=+sse2 -mcpu=penryn| FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-pc-unknown -mattr=+sse2 -mcpu=penryn | FileCheck %s
+; RUN: llc < %s -mtriple=i686-pc-unknown -mattr=+sse2 -mcpu=penryn | FileCheck %s
+
+; PR19059
+; RUN: llc < %s -mtriple=i686-pc-unknown -mattr=+sse2 -mcpu=penryn | FileCheck -check-prefix=CHECK32 %s
 
 define i32 @isint_return(double %d) nounwind {
+; CHECK-LABEL: isint_return:
 ; CHECK-NOT: xor
 ; CHECK: cvt
   %i = fptosi double %d to i32
@@ -8,8 +13,26 @@ define i32 @isint_return(double %d) nounwind {
   %e = sitofp i32 %i to double
 ; CHECK: cmpeqsd
   %c = fcmp oeq double %d, %e
+; CHECK32-NOT: movd {{.*}}, %r{{.*}}
+; CHECK32-NOT: andq
+; CHECK-NEXT: movd
+; CHECK-NEXT: andl
+  %z = zext i1 %c to i32
+  ret i32 %z
+}
+
+define i32 @isint_float_return(float %f) nounwind {
+; CHECK-LABEL: isint_float_return:
+; CHECK-NOT: xor
+; CHECK: cvt
+  %i = fptosi float %f to i32
+; CHECK-NEXT: cvt
+  %g = sitofp i32 %i to float
+; CHECK: cmpeqss
+  %c = fcmp oeq float %f, %g
+; CHECK-NOT: movd {{.*}}, %r{{.*}}
 ; CHECK-NEXT: movd
-; CHECK-NEXT: andq
+; CHECK-NEXT: andl
   %z = zext i1 %c to i32
   ret i32 %z
 }
@@ -17,6 +40,7 @@ define i32 @isint_return(double %d) nounwind {
 declare void @foo()
 
 define void @isint_branch(double %d) nounwind {
+; CHECK-LABEL: isint_branch:
 ; CHECK: cvt
   %i = fptosi double %d to i32
 ; CHECK-NEXT: cvt