From 8058764b4f92d9e43157433d637c22a93f51db3f Mon Sep 17 00:00:00 2001
From: MITSUNARI Shigeo <herumi@nifty.com>
Date: Sun, 29 May 2016 16:35:32 +0900
Subject: [PATCH] fpDbl_sqrPre has a bug

---
 src/gen.cpp |  2 +-
 src/mul.txt | 27 +++++++++++++++++++++++++++
 2 files changed, 28 insertions(+), 1 deletion(-)

diff --git a/src/gen.cpp b/src/gen.cpp
index 83728a6..d3f3d70 100644
--- a/src/gen.cpp
+++ b/src/gen.cpp
@@ -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 98151c7..37a3f93 100644
--- a/src/mul.txt
+++ b/src/mul.txt
@@ -1,3 +1,30 @@
+@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