[X86] Remove extra MOV after widening atomic load

This change adds patterns to optimize out an extra MOV
present after widening the atomic load. It also casts floats
to ints in an atomic load during AtomicExpand to support 128 bit
vectors in SSE/AVX.

commit-id:45989503

Author: jofrn

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -32070,6 +32070,13 @@ X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
   }
 }
 
+TargetLowering::AtomicExpansionKind
+X86TargetLowering::shouldCastAtomicLoadInIR(LoadInst *LI) const {
+  if (LI->getType()->getScalarType()->isFloatingPointTy())
+    return AtomicExpansionKind::CastToInteger;
+  return AtomicExpansionKind::None;
+}
+
 LoadInst *
 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
   unsigned NativeWidth = Subtarget.is64Bit() ? 64 : 32;

llvm/lib/Target/X86/X86ISelLowering.h

@@ -1838,6 +1838,8 @@ namespace llvm {
     shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
     TargetLoweringBase::AtomicExpansionKind
     shouldExpandLogicAtomicRMWInIR(AtomicRMWInst *AI) const;
+    TargetLoweringBase::AtomicExpansionKind
+    shouldCastAtomicLoadInIR(LoadInst *LI) const override;
     void emitBitTestAtomicRMWIntrinsic(AtomicRMWInst *AI) const override;
     void emitCmpArithAtomicRMWIntrinsic(AtomicRMWInst *AI) const override;
 

llvm/lib/Target/X86/X86InstrCompiler.td

@@ -1204,6 +1204,13 @@ def : Pat<(i16 (atomic_load_nonext_16 addr:$src)), (MOV16rm addr:$src)>;
 def : Pat<(i32 (atomic_load_nonext_32 addr:$src)), (MOV32rm addr:$src)>;
 def : Pat<(i64 (atomic_load_nonext_64 addr:$src)), (MOV64rm addr:$src)>;
 
+def : Pat<(v4i32 (scalar_to_vector (i32 (zext (i16 (atomic_load_16 addr:$src)))))),
+           (MOVDI2PDIrm addr:$src)>;   // load atomic <2 x i8>
+def : Pat<(v4i32 (scalar_to_vector (i32 (atomic_load_32 addr:$src)))),
+           (MOVDI2PDIrm addr:$src)>;   // load atomic <2 x i16>
+def : Pat<(v2i64 (scalar_to_vector (i64 (atomic_load_64 addr:$src)))),
+           (MOV64toPQIrm  addr:$src)>; // load atomic <2 x i32,float>
+
 // Floating point loads/stores.
 def : Pat<(atomic_store_32 (i32 (bitconvert (f32 FR32:$src))), addr:$dst),
           (MOVSSmr addr:$dst, FR32:$src)>, Requires<[UseSSE1]>;