From e35d49b7964691831617faf54a9c3d5757d5367b Mon Sep 17 00:00:00 2001
From: MITSUNARI Shigeo <herumi@nifty.com>
Date: Sun, 29 May 2016 17:20:32 +0900
Subject: [PATCH] fix generic_fpDbl_mul

---
 src/gen.cpp |  4 ++--
 src/mul.txt | 26 --------------------------
 2 files changed, 2 insertions(+), 28 deletions(-)

diff --git a/src/gen.cpp b/src/gen.cpp
index d3f3d70..23ecae8 100644
--- a/src/gen.cpp
+++ b/src/gen.cpp
@@ -545,7 +545,7 @@ struct Code : public mcl::Generator {
 		ret(Void);
 		endFunc();
 	}
-	void generic_fpDbl_mul(Operand& pz, Operand& px, Operand& py)
+	void generic_fpDbl_mul(Operand pz, const Operand& px, Operand py)
 	{
 		const int bu = bit + unit;
 		Operand y = load(py);
@@ -614,7 +614,7 @@ struct Code : public mcl::Generator {
 		gen_mulPv();
 		gen_mcl_fp_mul_UnitPre();
 		gen_mcl_fpDbl_mulPre();
-//		gen_mcl_fpDbl_sqrPre();
+		gen_mcl_fpDbl_sqrPre();
 	}
 	void setBit(uint32_t bit)
 	{
diff --git a/src/mul.txt b/src/mul.txt
index 37a3f93..cbe2133 100644
--- a/src/mul.txt
+++ b/src/mul.txt
@@ -1,31 +1,5 @@
 @define bu = bit + unit
 
-define void @mcl_fpDbl_sqrPre$(bit)(i$(unit)* %py, i$(unit)* %px) {
-  %x0 = load i$(unit)* %px
-  %xx0 = call i$(bu) @mulPv$(bit)x$(unit)(i$(unit) *%px, i$(unit) %x0)
-  %s0 = trunc i$(bu) %xx0 to i$(unit)
-  store i$(unit) %s0, i$(unit)* %py
-  %t0 = lshr i$(bu) %xx0, $(unit)
-
-@for i, 1, N
-  %px$(i) = getelementptr i$(unit)* %px, i32 $(i)
-  %x$(i) = load i$(unit)* %px$(i)
-  %xx$(i) = call i$(bu) @mulPv$(bit)x$(unit)(i$(unit)* %px, i$(unit) %x$(i))
-  %a$(i) = add i$(bu) %t$(i-1), %xx$(i)
-  %s$(i) = trunc i$(bu) %a$(i) to i$(unit)
-  %py$(i) = getelementptr i$(unit)* %py, i32 $(i)
-  @if i < N - 1
-    store i$(unit) %s$(i), i$(unit)* %py$(i)
-    %t$(i) = lshr i$(bu) %a$(i), $(unit)
-  @endif
-@endfor
-
-  %py$(N-1)e = bitcast i$(unit)* %py$(N-1) to i$(bu)*
-  store i$(bu) %a$(N-1), i$(bu)* %py$(N-1)e
-
-  ret void
-}
-
 @define bu = bit + unit
 @define bu2 = bit + unit * 2
 define void @mcl_fp_mont$(bit)(i$(bit)* %pz, i$(unit)* %px, i$(unit)* %py, i$(unit)* %pp, i$(unit) %r) {