On Thu, May 30, 2024 at 08:38:16PM -0700, Eric Biggers wrote:

On Tue, May 28, 2024 at 02:19:54PM +0200, Jason A. Donenfeld wrote:

quoted

diff --git a/arch/x86/entry/vdso/vgetrandom-chacha.S b/arch/x86/entry/vdso/vgetrandom-chacha.S
new file mode 100644
index 000000000000..d79e2bd97598
--- /dev/null
+++ b/arch/x86/entry/vdso/vgetrandom-chacha.S

@@ -0,0 +1,178 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) 2022 Jason A. Donenfeld <Jason@zx2c4.com>. All Rights Reserved.
+ */
+
+#include <linux/linkage.h>
+#include <asm/frame.h>
+
+.section	.rodata, "a"
+.align 16
+CONSTANTS:	.octa 0x6b20657479622d323320646e61707865
+.text
+
+/*
+ * Very basic SSE2 implementation of ChaCha20. Produces a given positive number
+ * of blocks of output with a nonce of 0, taking an input key and 8-byte
+ * counter. Importantly does not spill to the stack. Its arguments are:
+ *
+ *	rdi: output bytes
+ *	rsi: 32-byte key input
+ *	rdx: 8-byte counter input/output
+ *	rcx: number of 64-byte blocks to write to output
+ */
+SYM_FUNC_START(__arch_chacha20_blocks_nostack)
+
+.set	output,		%rdi
+.set	key,		%rsi
+.set	counter,	%rdx
+.set	nblocks,	%rcx
+.set	i,		%al
+/* xmm registers are *not* callee-save. */
+.set	state0,		%xmm0
+.set	state1,		%xmm1
+.set	state2,		%xmm2
+.set	state3,		%xmm3
+.set	copy0,		%xmm4
+.set	copy1,		%xmm5
+.set	copy2,		%xmm6
+.set	copy3,		%xmm7
+.set	temp,		%xmm8
+.set	one,		%xmm9

An "interesting" x86_64 quirk: in SSE instructions, registers xmm0-xmm7 take
fewer bytes to encode than xmm8-xmm15.

Since 'temp' is used frequently, moving it into the lower range (and moving one
of the 'copy' registers, which isn't used as frequently, into the higher range)
decreases the code size of __arch_chacha20_blocks_nostack() by 5%.

That's a nice trick. Thank you very much for it.

Jason