Inline quintic extension (#197)

* perf(quintic): force-inline quintic_mul + packed Mul impls

  LLVM was not force-inlining quintic_mul despite #[inline] — the monomorphized
  body is large enough that LLVM's cost heuristic declined. Each call-site paid
  ~5 cycles of function-call overhead. With quintic_mul called millions of times
  per proof, this accumulated to ~2.4% of total runtime.

  Zen 4 (c7a.2xlarge): -2.38% on xmss_leaf_1400sigs, p=0.0, revert-A/B confirmed.

* perf(quintic): force-inline quintic_square, quintic_mul_packed, MulAssign

  Extends the previous commit's inlining pattern to additional multiplication-
  related functions: quintic_square, all platform-specific quintic_mul_packed
  variants (AVX-512/AVX2/NEON/fallback), and MulAssign<Self>/MulAssign<QEF>.

  Testing established the I-cache budget boundary for forced inlining on Zen 4:
  these 9 functions are the optimal set. Inlining more (e.g. Add/Sub/Neg) causes
  regression from expanded code size.

  Zen 4 (c7a.2xlarge): additional -1.25% on xmss_leaf_1400sigs, p=0.0,
  revert-A/B confirmed. Combined with previous commit: ~-3.6% total.

Author: Barnadrot

crates/backend/koala-bear/src/quintic_extension/extension.rs

@@ -527,7 +527,7 @@ impl<F: TwoAdicField + QuinticExtendable> TwoAdicField for QuinticExtensionField
 }
 
 /// Quintic extension field multiplication in F[X]/(X^5 + X^2 - 1).
-#[inline]
+#[inline(always)]
 pub fn quintic_mul<T: Copy + Sub<Output = T>>(
     a: &[T; 5],
     b: &[T; 5],
@@ -546,7 +546,7 @@ pub fn quintic_mul<T: Copy + Sub<Output = T>>(
     ]
 }
 
-#[inline]
+#[inline(always)]
 pub(crate) fn quintic_square<F, R>(a: &[R; 5], res: &mut [R; 5])
 where
     F: Field,

crates/backend/koala-bear/src/quintic_extension/packed_extension.rs

@@ -461,7 +461,7 @@ where
 {
     type Output = Self;
 
-    #[inline]
+    #[inline(always)]
     fn mul(self, rhs: Self) -> Self {
         Self {
             value: super::extension::quintic_mul(&self.value, &rhs.value, PF::dot_product::<5>),
@@ -476,7 +476,7 @@ where
 {
     type Output = Self;
 
-    #[inline]
+    #[inline(always)]
     fn mul(self, rhs: QuinticExtensionField<F>) -> Self {
         let b: [PF; 5] = rhs.value.map(|x| x.into());
         Self {
@@ -516,7 +516,7 @@ where
     F: QuinticExtendable,
     PF: PackedField<Scalar = F>,
 {
-    #[inline]
+    #[inline(always)]
     fn mul_assign(&mut self, rhs: Self) {
         *self = *self * rhs;
     }
@@ -527,7 +527,7 @@ where
     F: QuinticExtendable,
     PF: PackedField<Scalar = F>,
 {
-    #[inline]
+    #[inline(always)]
     fn mul_assign(&mut self, rhs: QuinticExtensionField<F>) {
         *self = *self * rhs;
     }

crates/backend/koala-bear/src/quintic_extension/packing.rs

@@ -8,15 +8,15 @@ use crate::KoalaBear;
     all(target_arch = "aarch64", target_feature = "neon"),
     all(target_arch = "x86_64", target_feature = "avx2",)
 )))]
-#[inline]
+#[inline(always)]
 pub(crate) fn quintic_mul_packed(a: &[KoalaBear; 5], b: &[KoalaBear; 5], res: &mut [KoalaBear; 5]) {
     use field::PrimeCharacteristicRing;
     *res = super::extension::quintic_mul(a, b, KoalaBear::dot_product::<5>);
 }
 
 #[cfg(all(target_arch = "x86_64", target_feature = "avx2", not(target_feature = "avx512f")))]
 /// Multiplication in a quintic binomial extension field.
-#[inline]
+#[inline(always)]
 pub(crate) fn quintic_mul_packed(a: &[KoalaBear; 5], b: &[KoalaBear; 5], res: &mut [KoalaBear; 5]) {
     // TODO: This could likely be optimised further with more effort.
     // in particular it would benefit from a custom AVX2 implementation.
@@ -73,7 +73,7 @@ pub(crate) fn quintic_mul_packed(a: &[KoalaBear; 5], b: &[KoalaBear; 5], res: &m
 
 #[cfg(all(target_arch = "x86_64", target_feature = "avx512f"))]
 /// Multiplication in a quintic binomial extension field.
-#[inline]
+#[inline(always)]
 pub(crate) fn quintic_mul_packed(a: &[KoalaBear; 5], b: &[KoalaBear; 5], res: &mut [KoalaBear; 5]) {
     use crate::{PackedMontyField31AVX512, dot_product_2};
     use field::PrimeCharacteristicRing;
@@ -157,7 +157,7 @@ pub(crate) fn quintic_mul_packed(a: &[KoalaBear; 5], b: &[KoalaBear; 5], res: &m
 
 #[cfg(all(target_arch = "aarch64", target_feature = "neon"))]
 /// Multiplication in a quintic binomial extension field.
-#[inline]
+#[inline(always)]
 pub(crate) fn quintic_mul_packed(a: &[KoalaBear; 5], b: &[KoalaBear; 5], res: &mut [KoalaBear; 5]) {
     // TODO: This could be optimised further with a custom NEON implementation.