From a3fa174e7184c480b2656c9e013f16c3c32268bf Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Wed, 4 Sep 2024 07:58:07 +0800
Subject: [PATCH] sm3: s390x copyResultsBy4

---
 sm3/sm3blocks_s390x.go      |  3 +++
 sm3/sm3blocks_s390x.s       | 13 +++++++++++++
 sm3/sm3blocks_s390x_test.go | 23 +++++++++++++++++++++++
 3 files changed, 39 insertions(+)

diff --git a/sm3/sm3blocks_s390x.go b/sm3/sm3blocks_s390x.go
index 01db70b..1a61e36 100644
--- a/sm3/sm3blocks_s390x.go
+++ b/sm3/sm3blocks_s390x.go
@@ -8,3 +8,6 @@ package sm3
 
 //go:noescape
 func transposeMatrix(dig **[8]uint32)
+
+//go:noescape
+func copyResultsBy4(dig *uint32, p *byte)
diff --git a/sm3/sm3blocks_s390x.s b/sm3/sm3blocks_s390x.s
index 73604be..e560774 100644
--- a/sm3/sm3blocks_s390x.s
+++ b/sm3/sm3blocks_s390x.s
@@ -56,3 +56,16 @@ TEXT ·transposeMatrix(SB),NOSPLIT,$0
 	VSTM V6, V7, (R2)
 
 	RET
+
+// func copyResultsBy4(dig *uint32, dst *byte)
+TEXT ·copyResultsBy4(SB),NOSPLIT,$0
+#define digPtr R1
+#define dstPtr R2
+	MOVD	dig+0(FP), digPtr
+	MOVD	dst+8(FP), dstPtr
+
+	// load state
+	VLM (digPtr), V0, V7
+	VSTM V0, V7, (dstPtr)
+
+	RET
diff --git a/sm3/sm3blocks_s390x_test.go b/sm3/sm3blocks_s390x_test.go
index 94021ef..a6f2a52 100644
--- a/sm3/sm3blocks_s390x_test.go
+++ b/sm3/sm3blocks_s390x_test.go
@@ -29,3 +29,26 @@ func TestTransposeMatrix(t *testing.T) {
 		fmt.Println()
 	}
 }
+
+func TestCopyResultsBy4(t *testing.T) {
+	var m [4][8]uint32
+	var k uint32 = 0
+	for i := 0; i < 4; i++ {
+		for j := 0; j < 8; j++ {
+			m[i][j] = k << 24
+			k++
+			fmt.Printf("%04x ", m[i][j])
+		}
+		fmt.Println()
+	}
+	var p [32]byte
+	copyResultsBy4(&m[0][0], &p[0])
+	fmt.Println()
+	fmt.Println()
+	for i := 0; i < 32; i++ {
+		fmt.Printf("%04x ", p[i])
+		if i%8 == 7 {
+			fmt.Println()
+		}
+	}
+}