vmovd instead of movd

dkostic · hanno-becker · commit ca2ff7dc2e81 · 2025-10-03T19:02:58.000+01:00
Signed-off-by: Dusan Kostic &lt;dkostic@protonmail.com&gt;
diff --git a/dev/x86_64/src/mulcache_compute.S b/dev/x86_64/src/mulcache_compute.S
@@ -38,7 +38,7 @@ MLK_ASM_FN_SYMBOL(poly_mulcache_compute_avx2)
 
 // Broadcast 3329 (0x0D01) to all elements of ymm0
 movl $0x0D010D01, %eax
-movd %eax, %xmm0
+vmovd %eax, %xmm0
 vpbroadcastd %xmm0, %ymm0
 
 mulcache_compute_iter 0
diff --git a/dev/x86_64/src/ntt.S b/dev/x86_64/src/ntt.S
@@ -230,7 +230,7 @@ MLK_ASM_FN_SYMBOL(ntt_avx2)
 
 // Broadcast 3329 (0x0D01) to all elements of ymm0
 movl $0x0D010D01, %eax
-movd %eax, %xmm0
+vmovd %eax, %xmm0
 vpbroadcastd %xmm0, %ymm0
 
 level0		0
diff --git a/dev/x86_64/src/nttfrombytes.S b/dev/x86_64/src/nttfrombytes.S
@@ -99,7 +99,7 @@ MLK_ASM_FN_SYMBOL(nttfrombytes_avx2)
 
 // Broadcast 0xFFF to all elements of ymm0
 movl $0x0FFF0FFF, %eax
-movd %eax, %xmm0
+vmovd %eax, %xmm0
 vpbroadcastd %xmm0, %ymm0
 
 nttfrombytes_128_coefficients 0 0
diff --git a/dev/x86_64/src/ntttobytes.S b/dev/x86_64/src/ntttobytes.S
@@ -94,7 +94,7 @@ MLK_ASM_FN_SYMBOL(ntttobytes_avx2)
 
 // Broadcast 3329 (0x0D01) to all elements of ymm0
 movl $0x0D010D01, %eax
-movd %eax, %xmm0
+vmovd %eax, %xmm0
 vpbroadcastd %xmm0, %ymm0
 
 ntttobytes_128_coefficients 0 0
diff --git a/dev/x86_64/src/poly_basemul.inc b/dev/x86_64/src/poly_basemul.inc
@@ -127,12 +127,12 @@ schoolbook 7 \k
 .macro polyvec_basemul k
   // Broadcast 3329 (0x0D01) to all elements of ymm0
   movl $0x0D010D01, %eax
-  movd %eax, _qx
+  vmovd %eax, _qx
   vpbroadcastd _qx, _q
 
   // Broadcast -3327 (0xF301) to all elements of ymm0
   movl $0xF301F301, %eax
-  movd %eax, _qinvx
+  vmovd %eax, _qinvx
   vpbroadcastd _qinvx, _qinv
 
   .if \k > 0
diff --git a/dev/x86_64/src/reduce.S b/dev/x86_64/src/reduce.S
@@ -77,12 +77,12 @@ MLK_ASM_FN_SYMBOL(reduce_avx2)
 
 // Broadcast 3329 (0x0D01) to all elements of ymm0
 movl $0x0D010D01, %eax
-movd %eax, %xmm0
+vmovd %eax, %xmm0
 vpbroadcastd %xmm0, %ymm0
 
 // Broadcast 20159 (0x4EBF) to all elements of ymm1
 movl $0x4ebf4ebf, %eax
-movd %eax, %xmm1
+vmovd %eax, %xmm1
 vpbroadcastd %xmm1, %ymm1
 
 reduce_128_coefficients 0
diff --git a/dev/x86_64/src/tomont.S b/dev/x86_64/src/tomont.S
@@ -68,17 +68,17 @@ MLK_ASM_FN_SYMBOL(tomont_avx2)
 
 // Broadcast 3329 (0x0D01) to all elements of ymm0
 movl $0x0D010D01, %eax
-movd %eax, %xmm0
+vmovd %eax, %xmm0
 vpbroadcastd %xmm0, %ymm0
 
 // Broadcast 20553 (0x5049) to all elements of ymm1
 movl $0x50495049, %eax
-movd %eax, %xmm1
+vmovd %eax, %xmm1
 vpbroadcastd %xmm1, %ymm1
 
 // Broadcast 1353 (0x0549) to all elements of ymm2
 movl $0x05490549, %eax
-movd %eax, %xmm2
+vmovd %eax, %xmm2
 vpbroadcastd %xmm2, %ymm2
 
 tomont_128_coefficients 0
diff --git a/mlkem/src/native/x86_64/src/mulcache_compute.S b/mlkem/src/native/x86_64/src/mulcache_compute.S
@@ -20,7 +20,7 @@ MLK_ASM_FN_SYMBOL(poly_mulcache_compute_avx2)
 
         .cfi_startproc
         movl	$0xd010d01, %eax        # imm = 0xD010D01
-        movd	%eax, %xmm0
+        vmovd	%eax, %xmm0
         vpbroadcastd	%xmm0, %ymm0
         vmovdqa	0x20(%rsi), %ymm2
         vmovdqa	0x60(%rsi), %ymm3
diff --git a/mlkem/src/native/x86_64/src/ntt.S b/mlkem/src/native/x86_64/src/ntt.S
@@ -41,7 +41,7 @@ MLK_ASM_FN_SYMBOL(ntt_avx2)
 
         .cfi_startproc
         movl	$0xd010d01, %eax        # imm = 0xD010D01
-        movd	%eax, %xmm0
+        vmovd	%eax, %xmm0
         vpbroadcastd	%xmm0, %ymm0
         vpbroadcastq	0x140(%rsi), %ymm15
         vmovdqa	0x100(%rdi), %ymm8
diff --git a/mlkem/src/native/x86_64/src/nttfrombytes.S b/mlkem/src/native/x86_64/src/nttfrombytes.S
@@ -35,7 +35,7 @@ MLK_ASM_FN_SYMBOL(nttfrombytes_avx2)
 
         .cfi_startproc
         movl	$0xfff0fff, %eax        # imm = 0xFFF0FFF
-        movd	%eax, %xmm0
+        vmovd	%eax, %xmm0
         vpbroadcastd	%xmm0, %ymm0
         vmovdqu	(%rsi), %ymm4
         vmovdqu	0x20(%rsi), %ymm5
diff --git a/mlkem/src/native/x86_64/src/ntttobytes.S b/mlkem/src/native/x86_64/src/ntttobytes.S
@@ -35,7 +35,7 @@ MLK_ASM_FN_SYMBOL(ntttobytes_avx2)
 
         .cfi_startproc
         movl	$0xd010d01, %eax        # imm = 0xD010D01
-        movd	%eax, %xmm0
+        vmovd	%eax, %xmm0
         vpbroadcastd	%xmm0, %ymm0
         vmovdqa	(%rsi), %ymm5
         vmovdqa	0x20(%rsi), %ymm6
diff --git a/mlkem/src/native/x86_64/src/poly_basemul.inc b/mlkem/src/native/x86_64/src/poly_basemul.inc
@@ -127,12 +127,12 @@ schoolbook 7 \k
 .macro polyvec_basemul k
   // Broadcast 3329 (0x0D01) to all elements of ymm0
   movl $0x0D010D01, %eax
-  movd %eax, _qx
+  vmovd %eax, _qx
   vpbroadcastd _qx, _q
 
   // Broadcast -3327 (0xF301) to all elements of ymm0
   movl $0xF301F301, %eax
-  movd %eax, _qinvx
+  vmovd %eax, _qinvx
   vpbroadcastd _qinvx, _qinv
 
   .if \k > 0
diff --git a/mlkem/src/native/x86_64/src/polyvec_basemul_acc_montgomery_cached_asm_k2.S b/mlkem/src/native/x86_64/src/polyvec_basemul_acc_montgomery_cached_asm_k2.S
@@ -21,10 +21,10 @@ MLK_ASM_FN_SYMBOL(polyvec_basemul_acc_montgomery_cached_asm_k2)
 
         .cfi_startproc
         movl	$0xd010d01, %eax        # imm = 0xD010D01
-        movd	%eax, %xmm0
+        vmovd	%eax, %xmm0
         vpbroadcastd	%xmm0, %ymm0
         movl	$0xf301f301, %eax       # imm = 0xF301F301
-        movd	%eax, %xmm1
+        vmovd	%eax, %xmm1
         vpbroadcastd	%xmm1, %ymm1
         vmovdqa	(%rsi), %ymm2
         vmovdqa	0x20(%rsi), %ymm3
diff --git a/mlkem/src/native/x86_64/src/polyvec_basemul_acc_montgomery_cached_asm_k3.S b/mlkem/src/native/x86_64/src/polyvec_basemul_acc_montgomery_cached_asm_k3.S
@@ -21,10 +21,10 @@ MLK_ASM_FN_SYMBOL(polyvec_basemul_acc_montgomery_cached_asm_k3)
 
         .cfi_startproc
         movl	$0xd010d01, %eax        # imm = 0xD010D01
-        movd	%eax, %xmm0
+        vmovd	%eax, %xmm0
         vpbroadcastd	%xmm0, %ymm0
         movl	$0xf301f301, %eax       # imm = 0xF301F301
-        movd	%eax, %xmm1
+        vmovd	%eax, %xmm1
         vpbroadcastd	%xmm1, %ymm1
         vmovdqa	(%rsi), %ymm2
         vmovdqa	0x20(%rsi), %ymm3
diff --git a/mlkem/src/native/x86_64/src/polyvec_basemul_acc_montgomery_cached_asm_k4.S b/mlkem/src/native/x86_64/src/polyvec_basemul_acc_montgomery_cached_asm_k4.S
@@ -21,10 +21,10 @@ MLK_ASM_FN_SYMBOL(polyvec_basemul_acc_montgomery_cached_asm_k4)
 
         .cfi_startproc
         movl	$0xd010d01, %eax        # imm = 0xD010D01
-        movd	%eax, %xmm0
+        vmovd	%eax, %xmm0
         vpbroadcastd	%xmm0, %ymm0
         movl	$0xf301f301, %eax       # imm = 0xF301F301
-        movd	%eax, %xmm1
+        vmovd	%eax, %xmm1
         vpbroadcastd	%xmm1, %ymm1
         vmovdqa	(%rsi), %ymm2
         vmovdqa	0x20(%rsi), %ymm3
diff --git a/mlkem/src/native/x86_64/src/reduce.S b/mlkem/src/native/x86_64/src/reduce.S
@@ -41,10 +41,10 @@ MLK_ASM_FN_SYMBOL(reduce_avx2)
 
         .cfi_startproc
         movl	$0xd010d01, %eax        # imm = 0xD010D01
-        movd	%eax, %xmm0
+        vmovd	%eax, %xmm0
         vpbroadcastd	%xmm0, %ymm0
         movl	$0x4ebf4ebf, %eax       # imm = 0x4EBF4EBF
-        movd	%eax, %xmm1
+        vmovd	%eax, %xmm1
         vpbroadcastd	%xmm1, %ymm1
         vmovdqa	(%rdi), %ymm2
         vmovdqa	0x20(%rdi), %ymm3
diff --git a/mlkem/src/native/x86_64/src/tomont.S b/mlkem/src/native/x86_64/src/tomont.S
@@ -39,13 +39,13 @@ MLK_ASM_FN_SYMBOL(tomont_avx2)
 
         .cfi_startproc
         movl	$0xd010d01, %eax        # imm = 0xD010D01
-        movd	%eax, %xmm0
+        vmovd	%eax, %xmm0
         vpbroadcastd	%xmm0, %ymm0
         movl	$0x50495049, %eax       # imm = 0x50495049
-        movd	%eax, %xmm1
+        vmovd	%eax, %xmm1
         vpbroadcastd	%xmm1, %ymm1
         movl	$0x5490549, %eax        # imm = 0x5490549
-        movd	%eax, %xmm2
+        vmovd	%eax, %xmm2
         vpbroadcastd	%xmm2, %ymm2
         vmovdqa	(%rdi), %ymm3
         vmovdqa	0x20(%rdi), %ymm4
