Backends: Namespace assembly local labels with mlk_ prefix

Previously, local labels in the assembly backends were prefixed only by
the name of the enclosing function, e.g. Lntt_layer123_start in
ntt_aarch64_asm.S. Global entry symbols are namespaced per project via
MLK_ASM_NAMESPACE(), but local labels are emitted as literal text and
survive verbatim into the autogenerated mlkem/*.S files.

mldsa-native derives its assembly from mlkem-native and kept the same
function and label names, so its local labels are byte-for-byte identical
(Lntt_layer123_start, Lrej_uniform_loop48, ...). This is harmless as long
as each backend is assembled into its own object file, but breaks when
both backends are merged into a single link unit: AWS-LC's FIPS build
concatenates all sources and runs them through the delocate tool, which
flattens them into one symbol namespace and panics on the duplicate
labels.

This commit adds a fixed mlk_ project prefix to every local label across
all backends (AArch64 arithmetic and FIPS202, x86_64 AVX2, Armv8.1-M MVE,
ppc64le), not just the ones that clash today, so mlkem-native emits
Lmlk_ntt_layer123_start and a future mld_-prefixed mldsa-native drop can
no longer collide, independent of the integrator or platform.

The change is made in the dev/ assembly sources and propagated to
mlkem/*.S and the proofs/hol_light/* drops via scripts/simpasm. Local
labels emit no object-code bytes, and branch targets encode PC-relative
offsets rather than names, so the assembled .text is byte-identical; the
byte-locked HOL Light and CBMC proofs are unaffected.

check_asm_loop_labels in scripts/autogen is updated to require the
mlk_-prefixed function-derived prefix, so a newly added un-namespaced
label fails CI.

Signed-off-by: Hanno Becker <beckphan@amazon.co.uk>

Author: hanno-becker

dev/aarch64_clean/src/intt_aarch64_asm.S

@@ -239,7 +239,7 @@ MLK_ASM_FN_SYMBOL(intt_aarch64_asm)
         mov count, #8
 
         .p2align 2
-intt_layer4567_start:
+mlk_intt_layer4567_start:
 
         ldr q_data0, [inp, #(16*0)]
         ldr q_data1, [inp, #(16*1)]
@@ -301,7 +301,7 @@ intt_layer4567_start:
         str q_data3, [inp, #(-64 + 16*3)]
 
         sub count, count, #1
-        cbnz count, intt_layer4567_start
+        cbnz count, mlk_intt_layer4567_start
 
         // ---------------------------------------------------------------------
 
@@ -310,7 +310,7 @@ intt_layer4567_start:
 
         .p2align 2
 
-intt_layer123_start:
+mlk_intt_layer123_start:
 
         ldr q_data0, [in, #0]
         ldr q_data1, [in, #(1*(512/8))]
@@ -349,7 +349,7 @@ intt_layer123_start:
         str q_data3, [in, #(-16 + 3*(512/8))]
 
         sub count, count, #1
-        cbnz count, intt_layer123_start
+        cbnz count, mlk_intt_layer123_start
 
         pop_stack
         ret

dev/aarch64_clean/src/ntt_aarch64_asm.S

@@ -221,7 +221,7 @@ MLK_ASM_FN_SYMBOL(ntt_aarch64_asm)
         //
         // See test/test_bounds.py for more details.
 
-ntt_layer123_start:
+mlk_ntt_layer123_start:
 
         ldr q_data0, [in, #0]
         ldr q_data1, [in, #(1*(512/8))]
@@ -257,13 +257,13 @@ ntt_layer123_start:
         str q_data7, [in, #(-16 + 7*(512/8))]
 
         sub count, count, #1
-        cbnz count, ntt_layer123_start
+        cbnz count, mlk_ntt_layer123_start
 
         mov in, inp
         mov count, #8
 
         .p2align 2
-ntt_layer4567_start:
+mlk_ntt_layer4567_start:
 
         ldr q_data0, [in, #(16*0)]
         ldr q_data1, [in, #(16*1)]
@@ -293,7 +293,7 @@ ntt_layer4567_start:
         str q_data3, [in, #(-16*1)]
 
         sub count, count, #1
-        cbnz count, ntt_layer4567_start
+        cbnz count, mlk_ntt_layer4567_start
 
         pop_stack
         ret

dev/aarch64_clean/src/poly_mulcache_compute_aarch64_asm.S

@@ -86,7 +86,7 @@ MLK_ASM_FN_SYMBOL(poly_mulcache_compute_aarch64_asm)
         dup modulus_twisted.8h, wtmp
 
         mov count, #16
-poly_mulcache_compute_loop_start:
+mlk_poly_mulcache_compute_loop_start:
         ldr q_tmp0, [data_ptr], #32
         ldr q_tmp1, [data_ptr, #-16]
         ldr q_zeta, [zeta_ptr], #16
@@ -101,7 +101,7 @@ poly_mulcache_compute_loop_start:
         str q_dst, [cache_ptr], #16
 
         subs count, count, #1
-        cbnz count, poly_mulcache_compute_loop_start
+        cbnz count, mlk_poly_mulcache_compute_loop_start
 
         ret
 

dev/aarch64_clean/src/poly_reduce_aarch64_asm.S

@@ -67,7 +67,7 @@ MLK_ASM_FN_SYMBOL(poly_reduce_aarch64_asm)
         dup modulus_twisted.8h, wtmp
 
         mov count, #8
-poly_reduce_loop_start:
+mlk_poly_reduce_loop_start:
         ldr q_data, [ptr], #64
         barrett_reduce data
         scalar_signed_to_unsigned data
@@ -89,7 +89,7 @@ poly_reduce_loop_start:
         str q_data, [ptr, #-16]
 
         subs count, count, #1
-        cbnz count, poly_reduce_loop_start
+        cbnz count, mlk_poly_reduce_loop_start
 
         ret
 

dev/aarch64_clean/src/poly_tobytes_aarch64_asm.S

@@ -59,7 +59,7 @@
 MLK_ASM_FN_SYMBOL(poly_tobytes_aarch64_asm)
 
         mov count, #16
-poly_tobytes_loop_start:
+mlk_poly_tobytes_loop_start:
         ld2_wrap data, src
 
         // r[3 * i + 0] = (t0 >> 0);
@@ -77,7 +77,7 @@ poly_tobytes_loop_start:
         st3 {out0.8b, out1.8b, out2.8b}, [dst], #24
 
         subs count, count, #1
-        cbnz count, poly_tobytes_loop_start
+        cbnz count, mlk_poly_tobytes_loop_start
         ret
 
         .unreq data0

dev/aarch64_clean/src/poly_tomont_aarch64_asm.S

@@ -64,7 +64,7 @@ MLK_ASM_FN_SYMBOL(poly_tomont_aarch64_asm)
         dup factor_t.8h, wtmp
 
         mov count, #8
-poly_tomont_loop:
+mlk_poly_tomont_loop:
 
         ldr q_data, [src], #64
         mulmod res, data, factor, factor_t
@@ -83,7 +83,7 @@ poly_tomont_loop:
         str q_res, [src, #-16]
 
         sub count, count, #1
-        cbnz count, poly_tomont_loop
+        cbnz count, mlk_poly_tomont_loop
 
         ret
 

dev/aarch64_clean/src/polyvec_basemul_acc_montgomery_cached_k2_aarch64_asm.S

@@ -201,7 +201,7 @@ MLK_ASM_FN_SYMBOL(polyvec_basemul_acc_montgomery_cached_k2_aarch64_asm)
         add b1_cache_ptr, b0_cache_ptr, #(1 * 512/2)
 
         mov count, #(MLKEM_N / 16)
-polyvec_basemul_acc_montgomery_cached_k2_loop_start:
+mlk_polyvec_basemul_acc_montgomery_cached_k2_loop_start:
 
         load_polys aa, bb, a0_ptr, b0_ptr, b0_cache_ptr
         pmull res, aa, bb
@@ -214,7 +214,7 @@ polyvec_basemul_acc_montgomery_cached_k2_loop_start:
         st2_wrap out, out
 
         subs count, count, #1
-        cbnz count, polyvec_basemul_acc_montgomery_cached_k2_loop_start
+        cbnz count, mlk_polyvec_basemul_acc_montgomery_cached_k2_loop_start
 
         pop_stack
         ret

dev/aarch64_clean/src/polyvec_basemul_acc_montgomery_cached_k3_aarch64_asm.S

@@ -203,7 +203,7 @@ MLK_ASM_FN_SYMBOL(polyvec_basemul_acc_montgomery_cached_k3_aarch64_asm)
         add b2_cache_ptr, b0_cache_ptr, #(2 * 512/2)
 
         mov count, #(MLKEM_N / 16)
-polyvec_basemul_acc_montgomery_cached_k3_loop_start:
+mlk_polyvec_basemul_acc_montgomery_cached_k3_loop_start:
 
         load_polys aa, bb, a0_ptr, b0_ptr, b0_cache_ptr
         pmull res, aa, bb
@@ -218,7 +218,7 @@ polyvec_basemul_acc_montgomery_cached_k3_loop_start:
         st2_wrap out, out
 
         subs count, count, #1
-        cbnz count, polyvec_basemul_acc_montgomery_cached_k3_loop_start
+        cbnz count, mlk_polyvec_basemul_acc_montgomery_cached_k3_loop_start
 
         pop_stack
         ret

dev/aarch64_clean/src/polyvec_basemul_acc_montgomery_cached_k4_aarch64_asm.S

@@ -211,7 +211,7 @@ MLK_ASM_FN_SYMBOL(polyvec_basemul_acc_montgomery_cached_k4_aarch64_asm)
         // before Montgomery reduction is bound by 2^30.
 
         mov count, #(MLKEM_N / 16)
-polyvec_basemul_acc_montgomery_cached_k4_loop_start:
+mlk_polyvec_basemul_acc_montgomery_cached_k4_loop_start:
 
         load_polys aa, bb, a0_ptr, b0_ptr, b0_cache_ptr
         pmull res, aa, bb
@@ -228,7 +228,7 @@ polyvec_basemul_acc_montgomery_cached_k4_loop_start:
         st2_wrap out, out
 
         subs count, count, #1
-        cbnz count, polyvec_basemul_acc_montgomery_cached_k4_loop_start
+        cbnz count, mlk_polyvec_basemul_acc_montgomery_cached_k4_loop_start
 
         pop_stack
         ret

dev/aarch64_clean/src/rej_uniform_aarch64_asm.S

@@ -174,27 +174,27 @@ MLK_ASM_FN_SYMBOL(rej_uniform_aarch64_asm)
     // in case not enough bytes have been sampled, zero the temporary buffer.
     mov initial_zero_count, #0
     eor val0.16b, val0.16b, val0.16b
-rej_uniform_initial_zero:
+mlk_rej_uniform_initial_zero:
         str val0q, [output_tmp], #64
         str val0q, [output_tmp, #-48]
         str val0q, [output_tmp, #-32]
         str val0q, [output_tmp, #-16]
         add initial_zero_count, initial_zero_count, #32
         cmp initial_zero_count, #MLKEM_N
-        b.lt rej_uniform_initial_zero
+        b.lt mlk_rej_uniform_initial_zero
 
     mov output_tmp, output_tmp_base
 
     mov count, #0
     mov len, #MLKEM_N
 
     cmp buflen, #48
-    b.lo rej_uniform_loop48_end
+    b.lo mlk_rej_uniform_loop48_end
 
-rej_uniform_loop48:
+mlk_rej_uniform_loop48:
         // Finish once we've generated sufficiently many coefficients
         cmp     count, len
-        b.hs    rej_uniform_memory_copy
+        b.hs    mlk_rej_uniform_memory_copy
 
         // First, we unpack the byte stream into a stream of signed
         // coefficients, interpreting each consecutive 3 bytes as two
@@ -304,15 +304,15 @@ rej_uniform_loop48:
         add     count, count, ctr23
 
         cmp buflen, #48
-        b.hs rej_uniform_loop48
-rej_uniform_loop48_end:
+        b.hs mlk_rej_uniform_loop48
+mlk_rej_uniform_loop48_end:
 
     // Finish once we've generated sufficiently many coefficients
     cmp     count, len
-    b.hs    rej_uniform_memory_copy
+    b.hs    mlk_rej_uniform_memory_copy
 
     cmp buflen, #24
-    b.lo rej_uniform_memory_copy
+    b.lo mlk_rej_uniform_memory_copy
 
         ld3     {buf0.8b, buf1.8b, buf2.8b}, [buf], #24
 
@@ -360,7 +360,7 @@ rej_uniform_loop48_end:
         add     count, count, ctr0
         add     count, count, ctr1
 
-rej_uniform_memory_copy:
+mlk_rej_uniform_memory_copy:
     // min = min(count,len)
     cmp  count, len
     csel count, count, len, lo
@@ -370,7 +370,7 @@ rej_uniform_memory_copy:
     // allows us to stick to vectorized code.
     mov final_copy_count, #0
     mov output_tmp, output_tmp_base
-rej_uniform_final_copy:
+mlk_rej_uniform_final_copy:
         ldr val0q, [output_tmp], #64
         ldr val1q, [output_tmp, #-48]
         ldr val2q, [output_tmp, #-32]
@@ -381,11 +381,11 @@ rej_uniform_final_copy:
         str val3q, [output, #-16]
         add final_copy_count, final_copy_count, #32
         cmp final_copy_count, #MLKEM_N
-        b.lt rej_uniform_final_copy
+        b.lt mlk_rej_uniform_final_copy
 
     mov x0, count
 
-rej_uniform_return:
+mlk_rej_uniform_return:
     pop_stack
     ret