zuc: eia, ppc64x optimization #253

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

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

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

internal/cpuid/cpuid_ppc64x.go

@@ -4,3 +4,4 @@ package cpuid
 
 var HasAES = true
 var HasGFMUL = false
+var HasVPMSUMD = true

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

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