def FeatureFP16 : SubtargetFeature<"fp16", "HasFP16", "true",
"Enable half-precision floating point">;
+// Some processors have multiply-accumulate instructions that don't
+// play nicely with other VFP instructions, and it's generally better
+// to just not use them.
+// FIXME: Currently, this is only flagged for Cortex-A8. It may be true for
+// others as well. We should do more benchmarking and confirm one way or
+// the other.
+def HasSlowVMLx : SubtargetFeature<"vmlx", "SlowVMLx", "true",
+ "Disable VFP MAC instructions">;
+
//===----------------------------------------------------------------------===//
// ARM Processors supported.
//
// V7 Processors.
def : Processor<"cortex-a8", CortexA8Itineraries,
- [ArchV7A, FeatureThumb2, FeatureNEON]>;
+ [ArchV7A, FeatureThumb2, FeatureNEON, HasSlowVMLx]>;
def : ProcNoItin<"cortex-a9", [ArchV7A, FeatureThumb2, FeatureNEON]>;
//===----------------------------------------------------------------------===//
UseNEONFP("arm-use-neon-fp",
cl::desc("Use NEON for single-precision FP"),
cl::init(false), cl::Hidden);
-static cl::opt<bool>
-UseVMLxInstructions("arm-use-vmlx",
- cl::desc("Use VFP vmla and vmls instructions"),
- cl::init(true), cl::Hidden);
static cl::opt<bool>
UseMOVT("arm-use-movt",
: ARMArchVersion(V4)
, ARMFPUType(None)
, UseNEONForSinglePrecisionFP(UseNEONFP)
- , UseVMLx(UseVMLxInstructions)
+ , SlowVMLx(false)
, IsThumb(isT)
, ThumbMode(Thumb1)
, PostRAScheduler(false)
// operations with NEON instructions.
if (UseNEONFP.getPosition() == 0)
UseNEONForSinglePrecisionFP = true;
- // The VFP vlma and vlms instructions don't play nicely with others;
- // disable them.
- // FIXME: This may be true for other variants as well. Get benchmark
- // numbers and add them if determined that's the case.
- if (UseVMLxInstructions.getPosition() == 0)
- UseVMLx = false;
}
}
/// determine if NEON should actually be used.
bool UseNEONForSinglePrecisionFP;
- /// UseVMLx - If the VFP2 instructions are available, indicates whether
- /// the VML[AS] instructions should be used.
- bool UseVMLx;
+ /// SlowVMLx - If the VFP2 instructions are available, indicates whether
+ /// the VML[AS] instructions are slow (if so, don't use them).
+ bool SlowVMLx;
/// IsThumb - True if we are in thumb mode, false if in ARM mode.
bool IsThumb;
bool hasNEON() const { return ARMFPUType >= NEON; }
bool useNEONForSinglePrecisionFP() const {
return hasNEON() && UseNEONForSinglePrecisionFP; }
- bool useVMLx() const {return hasVFP2() && UseVMLx; }
+ bool useVMLx() const {return hasVFP2() && !SlowVMLx; }
bool hasFP16() const { return HasFP16; }