JIT: more optimizing of float ops based on known input characteristics

If the inputs are both float singles, and the top half is known to be identical
to the bottom half, we can use packed arithmetic instead of scalar to skip
the movddup.

This is slower on a few rather old CPUs, plus the Atom+Silvermont, so detect
Atom and disable it in that case.

Also avoid PPC_FP on stores if we know that the output came from a float op.

2 files changed, 7 insertions, 1 deletions

diff --git a/Source/Core/Common/CPUDetect.h b/Source/Core/Common/CPUDetect.h
index 752d26afb2..c63076ff7b 100644
--- a/Source/Core/Common/CPUDetect.h
+++ b/Source/Core/Common/CPUDetect.h
@@ -50,10 +50,10 @@ struct CPUInfo
 	bool bMOVBE;
 	// This flag indicates that the hardware supports some mode
 	// in which denormal inputs _and_ outputs are automatically set to (signed) zero.
-	// TODO: ARM
 	bool bFlushToZero;
 	bool bLAHFSAHF64;
 	bool bLongMode;
+	bool bAtom;
 
 	// ARM specific CPUInfo
 	bool bSwp;
diff --git a/Source/Core/Common/x64CPUDetect.cpp b/Source/Core/Common/x64CPUDetect.cpp
index 31409685e8..8ad8046c8b 100644
--- a/Source/Core/Common/x64CPUDetect.cpp
+++ b/Source/Core/Common/x64CPUDetect.cpp
@@ -129,6 +129,12 @@ void CPUInfo::Detect()
 	if (max_std_fn >= 1)
 	{
 		__cpuid(cpu_id, 0x00000001);
+		int family = ((cpu_id[0] >> 8) & 0xf) + ((cpu_id[0] >> 20) & 0xff);
+		int model = ((cpu_id[0] >> 4) & 0xf) + ((cpu_id[0] >> 12) & 0xf0);
+		// Detect people unfortunate enough to be running Dolphin on an Atom
+		if (family == 6 && (model == 0x1C || model == 0x26 ||model == 0x27 || model == 0x35 || model == 0x36 ||
+		                    model == 0x37 || model == 0x4A || model == 0x4D || model == 0x5A || model == 0x5D))
+			bAtom = true;
 		logical_cpu_count = (cpu_id[1] >> 16) & 0xFF;
 		ht = (cpu_id[3] >> 28) & 1;