Merge pull request #5203 from quic/fix-sgemmdirect-sme1

Add vector registers to clobber list to prevent compiler optimization.

kernel/arm64/sgemm_direct_arm64_sme1.c

@@ -7,7 +7,6 @@
 #include <stdlib.h>
 #include <inttypes.h>
 #include <math.h>
-
 #if defined(HAVE_SME)
 
 /* Function prototypes */
@@ -44,7 +43,17 @@ void CNAME (BLASLONG M, BLASLONG N, BLASLONG K, float * __restrict A,\
         m_mod = ceil((double)M/(double)vl_elms) * vl_elms;
 
         float *A_mod = (float *) malloc(m_mod*K*sizeof(float));
-		
+	    
+	    /* Prevent compiler optimization by reading from memory instead
+	     * of reading directly from vector (z) registers.
+	     * */
+        asm volatile("" : : :"p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7",
+                         "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15",
+                         "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7",
+                         "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15",
+                         "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23",
+                         "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31");
+      
         /* Pre-process the left matrix to make it suitable for 
            matrix sum of outer-product calculation
          */
@@ -52,7 +61,13 @@ void CNAME (BLASLONG M, BLASLONG N, BLASLONG K, float * __restrict A,\
         
         /* Calculate C = A*B */
         sgemm_direct_sme1_2VLx2VL(M, K, N, A_mod, B, R);
-        
+       
+        asm volatile("" : : :"p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7",
+                         "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15",
+                         "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7",
+                         "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15",
+                         "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23",
+                         "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31");
         free(A_mod);
 }