Accept 'inreg' attribute on x86 functions as

meaning sse_regparm (i.e. float/double values go
in XMM0 instead of ST0).  Update documentation
to reflect reality.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56619 91177308-0d34-0410-b5e6-96231b3b80d8

docs/LangRef.html

@@ -845,9 +845,11 @@ declare i32 @atoi(i8*) nounwind readonly
     a call to this function.</dd>
 
     <dt><tt>inreg</tt></dt>
-    <dd>This indicates that the parameter should be placed in register (if
-    possible) during assembling function call. Support for this attribute is
-    target-specific</dd>
+    <dd>This indicates that this parameter or return value should be treated
+    in a special target-dependent fashion during while emitting code for a
+    function call or return (usually, by putting it in a register as opposed 
+    to memory; in some places it is used to distinguish between two different
+    kinds of registers).  Use of this attribute is target-specific</dd>
 
     <dt><tt>byval</tt></dt>
     <dd>This indicates that the pointer parameter should really be passed by

include/llvm/CodeGen/SelectionDAGNodes.h

@@ -480,8 +480,8 @@ namespace ISD {
     
     // RET - Return from function.  The first operand is the chain,
     // and any subsequent operands are pairs of return value and return value
-    // signness for the function.  This operation can have variable number of
-    // operands.
+    // attributes (see CALL for description of attributes) for the function.
+    // This operation can have variable number of operands.
     RET,
 
     // INLINEASM - Represents an inline asm block.  This node always has two

lib/CodeGen/SelectionDAG/SelectionDAGBuild.cpp

@@ -913,7 +913,8 @@ void SelectionDAGLowering::visitRet(ReturnInst &I) {
       MVT VT = ValueVTs[j];
 
       // FIXME: C calling convention requires the return type to be promoted to
-      // at least 32-bit. But this is not necessary for non-C calling conventions.
+      // at least 32-bit. But this is not necessary for non-C calling
+      // conventions.
       if (VT.isInteger()) {
         MVT MinVT = TLI.getRegisterType(MVT::i32);
         if (VT.bitsLT(MinVT))
@@ -934,9 +935,13 @@ void SelectionDAGLowering::visitRet(ReturnInst &I) {
       getCopyToParts(DAG, SDValue(RetOp.getNode(), RetOp.getResNo() + j),
                      &Parts[0], NumParts, PartVT, ExtendKind);
 
+      // 'inreg' on function refers to return value
+      ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
+      if (F->paramHasAttr(0, ParamAttr::InReg))
+        Flags.setInReg();
       for (unsigned i = 0; i < NumParts; ++i) {
         NewValues.push_back(Parts[i]);
-        NewValues.push_back(DAG.getArgFlags(ISD::ArgFlagsTy()));
+        NewValues.push_back(DAG.getArgFlags(Flags));
       }
     }
   }

lib/Target/X86/X86CallingConv.td

@@ -43,10 +43,14 @@ def RetCC_X86Common : CallingConv<[
 
 // X86-32 C return-value convention.
 def RetCC_X86_32_C : CallingConv<[
-  // The X86-32 calling convention returns FP values in ST0, otherwise it is the
-  // same as the common X86 calling conv.
-  CCIfType<[f32], CCAssignToReg<[ST0, ST1]>>,
-  CCIfType<[f64], CCAssignToReg<[ST0, ST1]>>,
+  // The X86-32 calling convention returns FP values in ST0, unless marked
+  // with "inreg" (used here to distinguish one kind of reg from another,
+  // weirdly; this is really the sse-regparm calling convention) in which
+  // case they use XMM0, otherwise it is the same as the common X86 calling
+  // conv.
+  CCIfInReg<CCIfSubtarget<"hasSSE2()", 
+    CCIfType<[f32, f64], CCAssignToReg<[XMM0,XMM1,XMM2]>>>>,
+  CCIfType<[f32,f64], CCAssignToReg<[ST0, ST1]>>,
   CCDelegateTo<RetCC_X86Common>
 ]>;
 

test/CodeGen/X86/2008-09-25-sseregparm-1.ll

@@ -0,0 +1,19 @@
+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2 | grep movs | count 2
+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2 | grep fld | count 2
+; check 'inreg' attribute for sse_regparm
+
+define double @foo1() inreg nounwind {
+  ret double 1.0
+}
+
+define float @foo2() inreg nounwind {
+  ret float 1.0
+}
+
+define double @bar() nounwind {
+  ret double 1.0
+}
+
+define float @bar2() nounwind {
+  ret float 1.0
+}