From 50f5d29e491750664c559f140dfe358339cff146 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Mon, 30 Sep 2024 17:27:26 +0800
Subject: [PATCH] zuc: eia, ppc64x optimization #253

---
 internal/cpuid/cpuid_amd64.go        | 1 +
 internal/cpuid/cpuid_arm64.go        | 1 +
 internal/cpuid/cpuid_arm64_darwin.go | 1 +
 internal/cpuid/cpuid_ppc64x.go       | 1 +
 zuc/eia_asm.go                       | 2 +-
 zuc/eia_asm_ppc64x.s                 | 6 ++++++
 6 files changed, 11 insertions(+), 1 deletion(-)

diff --git a/internal/cpuid/cpuid_amd64.go b/internal/cpuid/cpuid_amd64.go
index 7903e62..abac744 100644
--- a/internal/cpuid/cpuid_amd64.go
+++ b/internal/cpuid/cpuid_amd64.go
@@ -4,3 +4,4 @@ import "golang.org/x/sys/cpu"
 
 var HasAES = cpu.X86.HasAES
 var HasGFMUL = cpu.X86.HasPCLMULQDQ
+var HasVPMSUMD = false
diff --git a/internal/cpuid/cpuid_arm64.go b/internal/cpuid/cpuid_arm64.go
index 608f44d..c259f43 100644
--- a/internal/cpuid/cpuid_arm64.go
+++ b/internal/cpuid/cpuid_arm64.go
@@ -6,3 +6,4 @@ import "golang.org/x/sys/cpu"
 
 var HasAES = cpu.ARM64.HasAES
 var HasGFMUL = cpu.ARM64.HasPMULL
+var HasVPMSUMD = false
diff --git a/internal/cpuid/cpuid_arm64_darwin.go b/internal/cpuid/cpuid_arm64_darwin.go
index 9b051d1..aac260b 100644
--- a/internal/cpuid/cpuid_arm64_darwin.go
+++ b/internal/cpuid/cpuid_arm64_darwin.go
@@ -8,3 +8,4 @@ package cpuid
 // to all Go programs running on darwin/arm64.
 var HasAES = true
 var HasGFMUL = true
+var HasVPMSUMD = false
\ No newline at end of file
diff --git a/internal/cpuid/cpuid_ppc64x.go b/internal/cpuid/cpuid_ppc64x.go
index a15273f..0955c9b 100644
--- a/internal/cpuid/cpuid_ppc64x.go
+++ b/internal/cpuid/cpuid_ppc64x.go
@@ -4,3 +4,4 @@ package cpuid
 
 var HasAES = true
 var HasGFMUL = false
+var HasVPMSUMD = true
diff --git a/zuc/eia_asm.go b/zuc/eia_asm.go
index bf8f614..04733d1 100644
--- a/zuc/eia_asm.go
+++ b/zuc/eia_asm.go
@@ -8,7 +8,7 @@ import (
 )
 
 var supportsAES = cpuid.HasAES
-var supportsGFMUL = cpuid.HasGFMUL
+var supportsGFMUL = cpuid.HasGFMUL || cpuid.HasVPMSUMD
 var useAVX = cpu.X86.HasAVX
 
 //go:noescape
diff --git a/zuc/eia_asm_ppc64x.s b/zuc/eia_asm_ppc64x.s
index c7f320f..122b6ee 100644
--- a/zuc/eia_asm_ppc64x.s
+++ b/zuc/eia_asm_ppc64x.s
@@ -93,9 +93,15 @@ TEXT ·eia3Round16B(SB),NOSPLIT,$0
 	VXOR XTMP3, XTMP4, XTMP3
 	VSPLTW $2, XTMP3, XDIGEST
 
+	// Update tag
 	MFVSRWZ XDIGEST, PTR
 	MOVWZ (R3), R6
 	XOR R6, PTR, R6 
 	MOVW R6, (R3)
 
+	// Copy last 16 bytes of KS to the front
+	MOVD $16, PTR
+	LXVD2X (PTR)(R4), XTMP1
+	STXVD2X XTMP1, (R4)(R0)
+
 	RET