From: jsing <jsing@openbsd.org>
Date: Thu, 16 Feb 2023 10:51:58 +0000 (+0000)
Subject: Bring in word_clz.S from s2n-bignum for amd64.
X-Git-Url: http://artulab.com/gitweb/?a=commitdiff_plain;h=5dab377ea1088a20cfd9c4bfac2a26c6727540c6;p=openbsd

Bring in word_clz.S from s2n-bignum for amd64.
---

diff --git a/lib/libcrypto/bn/arch/amd64/word_clz.S b/lib/libcrypto/bn/arch/amd64/word_clz.S
new file mode 100644
index 00000000000..6388b75d530
--- /dev/null
+++ b/lib/libcrypto/bn/arch/amd64/word_clz.S
@@ -0,0 +1,48 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: Apache-2.0 OR ISC
+
+// ----------------------------------------------------------------------------
+// Count leading zero bits in a single word
+// Input a; output function return
+//
+//    extern uint64_t word_clz (uint64_t a);
+//
+// Standard x86-64 ABI: RDI = a, returns RAX
+// Microsoft x64 ABI:   RCX = a, returns RAX
+// ----------------------------------------------------------------------------
+
+#include "_internal_s2n_bignum.h"
+
+        .intel_syntax noprefix
+        S2N_BN_SYM_VISIBILITY_DIRECTIVE(word_clz)
+        S2N_BN_SYM_PRIVACY_DIRECTIVE(word_clz)
+        .text
+
+S2N_BN_SYMBOL(word_clz):
+
+#if WINDOWS_ABI
+        push    rdi
+        push    rsi
+        mov     rdi, rcx
+#endif
+
+// First do rax = 63 - bsr(a), which is right except (maybe) for zero inputs
+
+        bsr     rax, rdi
+        xor     rax, 63
+
+// Force return of 64 in the zero-input case
+
+        mov     edx, 64
+        test    rdi, rdi
+        cmove   rax, rdx
+
+#if WINDOWS_ABI
+        pop    rsi
+        pop    rdi
+#endif
+        ret
+
+#if defined(__linux__) && defined(__ELF__)
+.section .note.GNU-stack,"",%progbits
+#endif