Teach FoldBitCast to be able to handle bitcasts from (e.g.) i128 -> <4 x float>.

This allows us to simplify this:
union vec2d {
    double e[2];
    double v __attribute__((vector_size(16)));
};
typedef union vec2d vec2d;

static vec2d a={{1,2}}, b={{3,4}};
    
vec2d foo () {
    return (vec2d){ .v = a.v + b.v * (vec2d){{5,5}}.v };
}

down to:

define %0 @foo() nounwind ssp {
entry:
  %mrv5 = insertvalue %0 undef, double 1.600000e+01, 0 ; <%0> [#uses=1]
  %mrv6 = insertvalue %0 %mrv5, double 2.200000e+01, 1 ; <%0> [#uses=1]
  ret %0 %mrv6
}

instead of:

define %0 @foo() nounwind ssp {
entry:
  %mrv5 = insertvalue %0 undef, double extractelement (<2 x double> fadd (<2 x double> fmul (<2 x double> bitcast (<1 x i128> <i128 85174437667405312423031577302488055808> to <2 x double>), <2 x double> <double 3.000000e+00, double 4.000000e+00>), <2 x double> <double 1.000000e+00, double 2.000000e+00>), i32 0), 0 ; <%0> [#uses=1]
  %mrv6 = insertvalue %0 %mrv5, double extractelement (<2 x double> fadd (<2 x double> fmul (<2 x double> bitcast (<1 x i128> <i128 85174437667405312423031577302488055808> to <2 x double>), <2 x double> <double 3.000000e+00, double 4.000000e+00>), <2 x double> <double 1.000000e+00, double 2.000000e+00>), i32 1), 1 ; <%0> [#uses=1]
  ret %0 %mrv6
}

llvm-svn: 85040

1 files changed, 13 insertions, 4 deletions

diff --git a/llvm/lib/Analysis/ConstantFolding.cpp b/llvm/lib/Analysis/ConstantFolding.cpp
index 8bb3673195d..33a5792796f 100644
--- a/llvm/lib/Analysis/ConstantFolding.cpp
+++ b/llvm/lib/Analysis/ConstantFolding.cpp
@@ -44,15 +44,24 @@ using namespace llvm;
 /// ConstantExpr if unfoldable.
 static Constant *FoldBitCast(Constant *C, const Type *DestTy,
                              const TargetData &TD) {
-  // If this is a bitcast from constant vector -> vector, fold it.
-  ConstantVector *CV = dyn_cast<ConstantVector>(C);
-  if (CV == 0)
-    return ConstantExpr::getBitCast(C, DestTy);
   
+  // This only handles casts to vectors currently.
   const VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
   if (DestVTy == 0)
     return ConstantExpr::getBitCast(C, DestTy);
   
+  // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
+  // vector so the code below can handle it uniformly.
+  if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
+    Constant *Ops = C; // don't take the address of C!
+    return FoldBitCast(ConstantVector::get(&Ops, 1), DestTy, TD);
+  }
+  
+  // If this is a bitcast from constant vector -> vector, fold it.
+  ConstantVector *CV = dyn_cast<ConstantVector>(C);
+  if (CV == 0)
+    return ConstantExpr::getBitCast(C, DestTy);
+  
   // If the element types match, VMCore can fold it.
   unsigned NumDstElt = DestVTy->getNumElements();
   unsigned NumSrcElt = CV->getNumOperands();