From fde2c7f3c0d8455d158b8edbe6a8adc58db73b61 Mon Sep 17 00:00:00 2001
From: Brian Smith <brian@briansmith.org>
Date: Sun, 23 Jun 2024 13:46:36 -0700
Subject: [PATCH] aes_gcm: Prefer integrated implementations, avoid inlining
 fallbacks.

The `#[inline(always)]` on `open_strided` was just a mistake.

But, even correcting that, the compiler won't structure `open()` and
`seal()` in the expected way. With these changes, the integrated
implementations will get inlined (as far as possible) and the
compiler will be biased towards assuming they will be used.
---
 src/aead/aes_gcm.rs | 41 ++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 40 insertions(+), 1 deletion(-)

diff --git a/src/aead/aes_gcm.rs b/src/aead/aes_gcm.rs
index ca44ac3403..8472c93c37 100644
--- a/src/aead/aes_gcm.rs
+++ b/src/aead/aes_gcm.rs
@@ -218,6 +218,26 @@ pub(super) fn seal(
     }
 }
 
+#[cfg_attr(
+    any(
+        all(
+            any(target_arch = "aarch64", target_arch = "arm"),
+            target_feature = "neon"
+        ),
+        all(
+            any(target_arch = "x86", target_arch = "x86_64"),
+            target_feature = "sse"
+        )
+    ),
+    inline(never)
+)]
+#[cfg_attr(
+    any(
+        all(target_arch = "aarch64", target_feature = "neon"),
+        all(target_arch = "x86_64", target_feature = "sse")
+    ),
+    cold
+)]
 fn seal_strided<A: aes::EncryptBlock + aes::EncryptCtr32, G: gcm::UpdateBlocks + gcm::Gmult>(
     Combo { aes_key, gcm_key }: &Combo<A, G>,
     aad: Aad<&[u8]>,
@@ -392,7 +412,26 @@ pub(super) fn open(
     }
 }
 
-#[inline(always)]
+#[cfg_attr(
+    any(
+        all(
+            any(target_arch = "aarch64", target_arch = "arm"),
+            target_feature = "neon"
+        ),
+        all(
+            any(target_arch = "x86", target_arch = "x86_64"),
+            target_feature = "sse"
+        )
+    ),
+    inline(never)
+)]
+#[cfg_attr(
+    any(
+        all(target_arch = "aarch64", target_feature = "neon"),
+        all(target_arch = "x86_64", target_feature = "sse")
+    ),
+    cold
+)]
 fn open_strided<A: aes::EncryptBlock + aes::EncryptCtr32, G: gcm::UpdateBlocks + gcm::Gmult>(
     Combo { aes_key, gcm_key }: &Combo<A, G>,
     aad: Aad<&[u8]>,