Merging r203581:
------------------------------------------------------------------------
r203581 | hans | 2014-03-11 11:49:24 -0400 (Tue, 11 Mar 2014) | 7 lines
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
------------------------------------------------------------------------
llvm-svn: 206071
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 06929dd..716c146 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -17583,12 +17583,30 @@
// FIXME: need symbolic constants for these magic numbers.
// See X86ATTInstPrinter.cpp:printSSECC().
unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
- SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, MVT::f32, CMP00, CMP01,
+ SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, CMP00.getValueType(),
+ CMP00, CMP01,
DAG.getConstant(x86cc, MVT::i8));
- SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, MVT::i32,
- OnesOrZeroesF);
- SDValue ANDed = DAG.getNode(ISD::AND, DL, MVT::i32, OnesOrZeroesI,
- DAG.getConstant(1, MVT::i32));
+
+ 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);
return OneBitOfTruth;
}
diff --git a/llvm/test/CodeGen/X86/isint.ll b/llvm/test/CodeGen/X86/isint.ll
index 4a98e63..38d05c6 100644
--- a/llvm/test/CodeGen/X86/isint.ll
+++ b/llvm/test/CodeGen/X86/isint.ll
@@ -1,6 +1,11 @@
-; RUN: llc < %s -march=x86 -mattr=+sse2 | 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
@@ -8,6 +13,24 @@
%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: andl
%z = zext i1 %c to i32
@@ -17,6 +40,7 @@
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
