Use umulh for Aarch64 multiply

This saves about 30% on the Integer multiply operations. For example, RSA 1024 encryption goes from 0.068 to 0.044 ms/op. RSA 2048 goes from 0.170 to 0.100 ms/op. We should have taken this sooner. Ugh...

1 files changed, 3 insertions, 1 deletions

diff --git a/integer.cpp b/integer.cpp
index f5abf83e..a74eb730 100644
--- a/integer.cpp
+++ b/integer.cpp
@@ -200,7 +200,7 @@ static word AtomicInverseModPower2(word A)
 

 // ********************************************************

 

-#if !defined(CRYPTOPP_NATIVE_DWORD_AVAILABLE) || (defined(__x86_64__) && defined(CRYPTOPP_WORD128_AVAILABLE))

+#if !defined(CRYPTOPP_NATIVE_DWORD_AVAILABLE) || ((defined(__aarch64__) || defined(__x86_64__)) && defined(CRYPTOPP_WORD128_AVAILABLE))

 	#define TWO_64_BIT_WORDS 1

 	#define Declare2Words(x)			word x##0, x##1;

 	#define AssignWord(a, b)			a##0 = b; a##1 = 0;

@@ -212,6 +212,8 @@ static word AtomicInverseModPower2(word A)
 		#ifndef __INTEL_COMPILER

 			#define Double3Words(c, d)		d##1 = __shiftleft128(d##0, d##1, 1); d##0 = __shiftleft128(c, d##0, 1); c *= 2;

 		#endif

+	#elif defined(__aarch32__) || defined(__aarch64__)

+		#define MultiplyWordsLoHi(p0, p1, a, b)		p0 = a*b; asm ("umulh %0,%1,%2" : "=r"(p1) : "r"(a), "r"(b));

 	#elif defined(__DECCXX)

 		#define MultiplyWordsLoHi(p0, p1, a, b)		p0 = a*b; p1 = asm("umulh %a0, %a1, %v0", a, b);

 	#elif defined(__x86_64__)