From 13fd88322f7a9debc32bdcdb86de7a0d11062feb Mon Sep 17 00:00:00 2001
From: Chris Duncan <chris@codecow.com>
Date: Thu, 1 Oct 2026 13:51:02 -0700
Subject: [PATCH] Interleave carry assignments to improve cpu scheduling.

---
 src/assembly/ed25519/fe.ts | 45 +++++++++++++++++++-------------------
 1 file changed, 23 insertions(+), 22 deletions(-)

diff --git a/src/assembly/ed25519/fe.ts b/src/assembly/ed25519/fe.ts
index fba24bd..e157b1d 100644
--- a/src/assembly/ed25519/fe.ts
+++ b/src/assembly/ed25519/fe.ts
@@ -461,36 +461,37 @@ export function fe_mul (h: FieldElement, f: FieldElement, g: FieldElement): void
 	const bias25 = v128.splat<i64>((<i64>1 << 25) + (<i64>1 << 62))
 	const unbias25 = v128.splat<i64>(<i64>1 << 37)
 	const unbias26 = v128.splat<i64>(<i64>1 << 36)
-	let carry: v128
+	let carryA: v128
+	let carryB: v128
 
-	carry = v128.sub<i64>(v128.shr<u64>(v128.add<i64>(h04_sum, bias25), 26), unbias26)
-	h04_sum = v128.sub<i64>(h04_sum, v128.shl<i64>(carry, 26))
-	h15_sum = v128.add<i64>(h15_sum, carry)
+	carryA = v128.sub<i64>(v128.shr<u64>(v128.add<i64>(h04_sum, bias25), 26), unbias26)
+	h04_sum = v128.sub<i64>(h04_sum, v128.shl<i64>(carryA, 26))
+	h15_sum = v128.add<i64>(h15_sum, carryA)
 
-	carry = v128.sub<i64>(v128.shr<u64>(v128.add<i64>(h15_sum, bias24), 25), unbias25)
-	h15_sum = v128.sub<i64>(h15_sum, v128.shl<i64>(carry, 25))
-	h26_sum = v128.add<i64>(h26_sum, carry)
+	carryB = v128.sub<i64>(v128.shr<u64>(v128.add<i64>(h15_sum, bias24), 25), unbias25)
+	h15_sum = v128.sub<i64>(h15_sum, v128.shl<i64>(carryB, 25))
+	h26_sum = v128.add<i64>(h26_sum, carryB)
 
-	carry = v128.sub<i64>(v128.shr<u64>(v128.add<i64>(h26_sum, bias25), 26), unbias26)
-	h26_sum = v128.sub<i64>(h26_sum, v128.shl<i64>(carry, 26))
-	h37_sum = v128.add<i64>(h37_sum, carry)
+	carryA = v128.sub<i64>(v128.shr<u64>(v128.add<i64>(h26_sum, bias25), 26), unbias26)
+	h26_sum = v128.sub<i64>(h26_sum, v128.shl<i64>(carryA, 26))
+	h37_sum = v128.add<i64>(h37_sum, carryA)
 
-	carry = v128.sub<i64>(v128.shr<u64>(v128.add<i64>(h37_sum, bias24), 25), unbias25)
-	h37_sum = v128.sub<i64>(h37_sum, v128.shl<i64>(carry, 25))
-	let h48_sum = v128.add<i64>(v128.shuffle<i64>(h04_sum, h89_sum, 1, 2), carry)
+	carryB = v128.sub<i64>(v128.shr<u64>(v128.add<i64>(h37_sum, bias24), 25), unbias25)
+	h37_sum = v128.sub<i64>(h37_sum, v128.shl<i64>(carryB, 25))
+	let h48_sum = v128.add<i64>(v128.shuffle<i64>(h04_sum, h89_sum, 1, 2), carryB)
 
-	carry = v128.sub<i64>(v128.shr<u64>(v128.add<i64>(h48_sum, bias25), 26), unbias26)
-	h48_sum = v128.sub<i64>(h48_sum, v128.shl<i64>(carry, 26))
-	let h59_sum = v128.add<i64>(v128.shuffle<i64>(h15_sum, h89_sum, 1, 3), carry)
+	carryA = v128.sub<i64>(v128.shr<u64>(v128.add<i64>(h48_sum, bias25), 26), unbias26)
+	h48_sum = v128.sub<i64>(h48_sum, v128.shl<i64>(carryA, 26))
+	let h59_sum = v128.add<i64>(v128.shuffle<i64>(h15_sum, h89_sum, 1, 3), carryA)
 
-	carry = v128.sub<i64>(v128.shr<u64>(v128.add<i64>(h59_sum, bias24), 25), unbias25)
-	h59_sum = v128.sub<i64>(h59_sum, v128.shl<i64>(carry, 25))
+	carryB = v128.sub<i64>(v128.shr<u64>(v128.add<i64>(h59_sum, bias24), 25), unbias25)
+	h59_sum = v128.sub<i64>(h59_sum, v128.shl<i64>(carryB, 25))
 	// wrap requires multiplying carry from h9 by 19
-	let h60_sum = v128.add<i64>(v128.shuffle<i64>(h26_sum, h04_sum, 1, 2), v128.mul<i64>(carry, i64x2(1, 19)))
+	let h60_sum = v128.add<i64>(v128.shuffle<i64>(h26_sum, h04_sum, 1, 2), v128.mul<i64>(carryB, i64x2(1, 19)))
 
-	carry = v128.sub<i64>(v128.shr<u64>(v128.add<i64>(h60_sum, bias25), 26), unbias26)
-	h60_sum = v128.sub<i64>(h60_sum, v128.shl<i64>(carry, 26))
-	let h71_sum = v128.add<i64>(v128.shuffle<i64>(h37_sum, h15_sum, 1, 2), carry)
+	carryA = v128.sub<i64>(v128.shr<u64>(v128.add<i64>(h60_sum, bias25), 26), unbias26)
+	h60_sum = v128.sub<i64>(h60_sum, v128.shl<i64>(carryA, 26))
+	let h71_sum = v128.add<i64>(v128.shuffle<i64>(h37_sum, h15_sum, 1, 2), carryA)
 
 	// assign scalar results to output
 	v128.store_lane<i32>(h_ptr, h60_sum, 2, 0)
-- 
2.52.0