fix PR7518 - terrible codegen of <2 x float>, by only marking

v2f32 as legal in 32-bit mode.  It is just as terrible there,
but I just care about x86-64 and noone claims it is valuable
in 64-bit mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107600 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 239c42b796e5a572a92f6f8495e14350bef67396..f2054c9b244a02a37355b6ec70574ae5453921bc 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -617,7 +617,12 @@ X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
     addRegisterClass(MVT::v8i8,  X86::VR64RegisterClass, false);
     addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
     addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
-    addRegisterClass(MVT::v2f32, X86::VR64RegisterClass, false);
+    
+    // FIXME: v2f32 isn't an MMX type.  We currently claim that it is legal
+    // because of some ABI issue, but this isn't the right fix.
+    bool IsV2F32Legal = !Subtarget->is64Bit();
+    if (IsV2F32Legal)
+      addRegisterClass(MVT::v2f32, X86::VR64RegisterClass, false);
     addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
 
     setOperationAction(ISD::ADD,                MVT::v8i8,  Legal);
@@ -663,14 +668,17 @@ X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
     AddPromotedToType (ISD::LOAD,               MVT::v4i16, MVT::v1i64);
     setOperationAction(ISD::LOAD,               MVT::v2i32, Promote);
     AddPromotedToType (ISD::LOAD,               MVT::v2i32, MVT::v1i64);
-    setOperationAction(ISD::LOAD,               MVT::v2f32, Promote);
-    AddPromotedToType (ISD::LOAD,               MVT::v2f32, MVT::v1i64);
+    if (IsV2F32Legal) {
+      setOperationAction(ISD::LOAD,             MVT::v2f32, Promote);
+      AddPromotedToType (ISD::LOAD,             MVT::v2f32, MVT::v1i64);
+    }
     setOperationAction(ISD::LOAD,               MVT::v1i64, Legal);
 
     setOperationAction(ISD::BUILD_VECTOR,       MVT::v8i8,  Custom);
     setOperationAction(ISD::BUILD_VECTOR,       MVT::v4i16, Custom);
     setOperationAction(ISD::BUILD_VECTOR,       MVT::v2i32, Custom);
-    setOperationAction(ISD::BUILD_VECTOR,       MVT::v2f32, Custom);
+    if (IsV2F32Legal)
+      setOperationAction(ISD::BUILD_VECTOR,     MVT::v2f32, Custom);
     setOperationAction(ISD::BUILD_VECTOR,       MVT::v1i64, Custom);
 
     setOperationAction(ISD::VECTOR_SHUFFLE,     MVT::v8i8,  Custom);
@@ -678,7 +686,8 @@ X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
     setOperationAction(ISD::VECTOR_SHUFFLE,     MVT::v2i32, Custom);
     setOperationAction(ISD::VECTOR_SHUFFLE,     MVT::v1i64, Custom);
 
-    setOperationAction(ISD::SCALAR_TO_VECTOR,   MVT::v2f32, Custom);
+    if (IsV2F32Legal)
+      setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
     setOperationAction(ISD::SCALAR_TO_VECTOR,   MVT::v8i8,  Custom);
     setOperationAction(ISD::SCALAR_TO_VECTOR,   MVT::v4i16, Custom);
     setOperationAction(ISD::SCALAR_TO_VECTOR,   MVT::v1i64, Custom);
@@ -697,7 +706,8 @@ X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
       setOperationAction(ISD::BIT_CONVERT,        MVT::v8i8,  Custom);
       setOperationAction(ISD::BIT_CONVERT,        MVT::v4i16, Custom);
       setOperationAction(ISD::BIT_CONVERT,        MVT::v2i32, Custom);
-      setOperationAction(ISD::BIT_CONVERT,        MVT::v2f32, Custom);
+      if (IsV2F32Legal)
+        setOperationAction(ISD::BIT_CONVERT,      MVT::v2f32, Custom);
       setOperationAction(ISD::BIT_CONVERT,        MVT::v1i64, Custom);
     }
   }

diff --git a/test/CodeGen/X86/v2f32.ll b/test/CodeGen/X86/v2f32.ll
new file mode 100644 (file)
index 0000000..509d528
--- /dev/null
+++ b/test/CodeGen/X86/v2f32.ll
@@ -0,0 +1,16 @@
+; RUN: llc < %s -march=x86-64 -asm-verbose=0 -o - | FileCheck %s
+
+; PR7518
+define void @test1(<2 x float> %Q, float *%P2) nounwind {
+  %a = extractelement <2 x float> %Q, i32 0
+  %b = extractelement <2 x float> %Q, i32 1
+  %c = fadd float %a, %b
+
+  store float %c, float* %P2
+  ret void
+; CHECK: test1:
+; CHECK-NEXT: addss    %xmm1, %xmm0
+; CHECK-NEXT: movss    %xmm0, (%rdi)
+; CHECK-NEXT: ret
+}
+