@@ -59,18 +59,18 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
5959 "shufps $0, %%xmm6 , %%xmm6 \n\t"
6060
6161
62- ".align 16 \n\t"
62+ ".p2align 4 \n\t"
6363 "1: \n\t"
6464 "xorps %%xmm4 , %%xmm4 \n\t"
6565 "xorps %%xmm5 , %%xmm5 \n\t"
6666 "movups (%3,%0,4), %%xmm7 \n\t" // 4 * y
6767
68- ".align 2 \n\t"
68+ ".p2align 1 \n\t"
6969 "movups (%4,%0,4), %%xmm8 \n\t"
7070 "movups (%5,%0,4), %%xmm9 \n\t"
7171 "movups (%6,%0,4), %%xmm10 \n\t"
7272 "movups (%7,%0,4), %%xmm11 \n\t"
73- ".align 2 \n\t"
73+ ".p2align 1 \n\t"
7474 "mulps %%xmm12, %%xmm8 \n\t"
7575 "mulps %%xmm13, %%xmm9 \n\t"
7676 "mulps %%xmm14, %%xmm10 \n\t"
@@ -84,7 +84,7 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
8484 "movups (%5,%8,4), %%xmm9 \n\t"
8585 "movups (%6,%8,4), %%xmm10 \n\t"
8686 "movups (%7,%8,4), %%xmm11 \n\t"
87- ".align 2 \n\t"
87+ ".p2align 1 \n\t"
8888 "mulps %%xmm0 , %%xmm8 \n\t"
8989 "mulps %%xmm1 , %%xmm9 \n\t"
9090 "mulps %%xmm2 , %%xmm10 \n\t"
@@ -154,7 +154,7 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
154154 "movss (%8), %%xmm6 \n\t" // alpha
155155 "shufps $0, %%xmm6 , %%xmm6 \n\t"
156156
157- ".align 16 \n\t"
157+ ".p2align 4 \n\t"
158158 "1: \n\t"
159159 "xorps %%xmm4 , %%xmm4 \n\t"
160160 "movups (%3,%0,4), %%xmm7 \n\t" // 4 * y
0 commit comments