powerpc64le build error with OpenBLAS 0.3.34

[eschnett]

I am cross-building OpenBLAS 0.3.34 for powerpc64le-linux-gnu using GCC 12. I see this build error:

[14:18:39] cc -O2 -DSMALL_MATRIX_OPT -DGEMM_GEMV_FORWARD -DSBGEMM_GEMV_FORWARD -DMAX_STACK_ALLOC=2048 -fopenmp -DHAVE_P10_SUPPORT -Wall -m64 -DF_INTERFACE_GFORT -fPIC -DDYNAMIC_ARCH -DSMP_SERVER -DUSE_OPENMP -DNO_WARMUP -DMAX_CPU_NUMBER=512 -DMAX_PARALLEL_NUMBER=1 -DBUILD_BFLOAT16 -DBUILD_SINGLE=1 -DBUILD_DOUBLE=1 -DBUILD_COMPLEX=1 -DBUILD_COMPLEX16=1 -DVERSION=\"0.3.34\" -O3 -mcpu=power10 -mtune=power10 -mvsx -fno-fast-math -DUSE_OPENMP -fopenmp -UASMNAME -UASMFNAME -UNAME -UCNAME -UCHAR_NAME -UCHAR_CNAME -DASMNAME=sbgemm_otcopy_POWER10 -DASMFNAME=sbgemm_otcopy_POWER10_ -DNAME=sbgemm_otcopy_POWER10_ -DCNAME=sbgemm_otcopy_POWER10 -DCHAR_NAME=\"sbgemm_otcopy_POWER10_\" -DCHAR_CNAME=\"sbgemm_otcopy_POWER10\" -DNO_AFFINITY -DTS=_POWER10 -I.. -DBUILD_KERNEL -DTABLE_NAME=gotoblas_POWER10 -DBFLOAT16 -UDOUBLE  -UCOMPLEX -c -DBFLOAT16 -UDOUBLE -UCOMPLEX ../kernel/power/sbgemm_tcopy_8_power10.c -o sbgemm_otcopy_POWER10.o
2026-07-21 14:18:39 UTC
[14:18:39] In file included from ../kernel/power/sbgemv_common.c:30,
2026-07-21 14:18:39 UTC
[14:18:39]                  from ../kernel/power/sbgemv_common_power10.c:30,
2026-07-21 14:18:39 UTC
[14:18:39]                  from ../kernel/power/sbgemv_n_power10.c:34:
2026-07-21 14:18:39 UTC
[14:18:39] ../kernel/power/gemm_common.c: In function ‘vec_load_pair’:
2026-07-21 14:18:39 UTC
[14:18:39] ../kernel/power/gemm_common.c:50:3: error: invalid conversion to type ‘* __vector_pair’
2026-07-21 14:18:39 UTC
[14:18:39]    50 |   vy0p = __builtin_vsx_lxvp(0L, (const __vector_pair *)(src));
2026-07-21 14:18:39 UTC
[14:18:39]       |   ^~~~
2026-07-21 14:18:39 UTC

[eschnett]

I find that this patch remedies the problem. This mechanically replaces __builtin_vsx_lxvp by explicit pointer casts.

--- a/kernel/power/gemm_common.c
+++ b/kernel/power/gemm_common.c
@@ -46,11 +46,11 @@
 FORCEINLINE void vec_load_pair(vec_f32 *dst, vec_f32 *src)
 {
 #ifdef USE_VECTOR_PAIRS
   __vector_pair vy0p;
-//#ifdef __clang__
+#ifdef __clang__
   vy0p = __builtin_vsx_lxvp(0L, (const __vector_pair *)(src));
-//#else
-//  vy0p = *(__vector_pair *)((void *)src);
-//#endif
+#else
+  vy0p = *(__vector_pair *)((void *)src);
+#endif
   __builtin_vsx_disassemble_pair((void *)(dst), &vy0p);
 #else



diff --git a/kernel/power/dgemm_kernel_power10.c b/kernel/power/dgemm_kernel_power10.c
index ea9462b..778232a 100644
--- a/kernel/power/dgemm_kernel_power10.c
+++ b/kernel/power/dgemm_kernel_power10.c
@@ -94,8 +94,8 @@ typedef FLOAT v4sf_t __attribute__ ((vector_size (16)));
 #endif
 #define KERNEL(i) \
     rowA = (vec_t *)&AO[(i)<< 3];\
-    rowB = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[(i) << 3])); \
-    rowB1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[((i) << 3) + 4])); \
+    rowB = (*(__vector_pair *)((void *)(&BO[(i) << 3]))); \
+    rowB1 = (*(__vector_pair *)((void *)(&BO[((i) << 3) + 4]))); \
     __builtin_mma_xvf64gerpp(&acc0, rowB, rowA[0]);\
     __builtin_mma_xvf64gerpp(&acc1, rowB1, rowA[0]);\
     __builtin_mma_xvf64gerpp(&acc2, rowB, rowA[1]);\
@@ -200,8 +200,8 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
             BLASLONG l = 1;
             vec_t *rowA = (vec_t *) & AO[0];
             __vector_pair rowB, rowB1;
-            rowB = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[0]));
-            rowB1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[4]));
+            rowB = (*(__vector_pair *)((void *)(&BO[0])));
+            rowB1 = (*(__vector_pair *)((void *)(&BO[4])));
             __builtin_mma_xvf64ger (&acc0, rowB, rowA[0]);
             __builtin_mma_xvf64ger (&acc1, rowB1, rowA[0]);
             __builtin_mma_xvf64ger (&acc2, rowB, rowA[1]);
@@ -283,16 +283,16 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
             BLASLONG l = 0;
             vec_t *rowA = (vec_t *) & AO[0];
             __vector_pair rowB, rowB1;
-            rowB = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[0]));
-            rowB1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[4]));
+            rowB = (*(__vector_pair *)((void *)(&BO[0])));
+            rowB1 = (*(__vector_pair *)((void *)(&BO[4])));
             __builtin_mma_xvf64ger (&acc0, rowB, rowA[0]);
             __builtin_mma_xvf64ger (&acc1, rowB1, rowA[0]);
             __builtin_mma_xvf64ger (&acc2, rowB, rowA[1]);
             __builtin_mma_xvf64ger (&acc3, rowB1, rowA[1]);
             for (l = 1; l < temp; l++) {
                 rowA = (vec_t *) & AO[l << 2];
-                rowB = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[l << 3]));
-                rowB1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[(l << 3) + 4]));
+                rowB = (*(__vector_pair *)((void *)(&BO[l << 3])));
+                rowB1 = (*(__vector_pair *)((void *)(&BO[(l << 3) + 4])));
                 __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
                 __builtin_mma_xvf64gerpp (&acc1, rowB1, rowA[0]);
                 __builtin_mma_xvf64gerpp (&acc2, rowB, rowA[1]);
@@ -323,14 +323,14 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
             BLASLONG l = 0;
             vec_t *rowA = (vec_t *) & AO[0];
             __vector_pair rowB, rowB1;
-            rowB = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[0]));
-            rowB1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[4]));
+            rowB = (*(__vector_pair *)((void *)(&BO[0])));
+            rowB1 = (*(__vector_pair *)((void *)(&BO[4])));
             __builtin_mma_xvf64ger (&acc0, rowB, rowA[0]);
             __builtin_mma_xvf64ger (&acc1, rowB1, rowA[0]);
             for (l = 1; l < temp; l++) {
                 rowA = (vec_t *) & AO[l << 1];
-                rowB = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[l << 3]));
-                rowB1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[(l << 3) + 4]));
+                rowB = (*(__vector_pair *)((void *)(&BO[l << 3])));
+                rowB1 = (*(__vector_pair *)((void *)(&BO[(l << 3) + 4])));
                 __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
                 __builtin_mma_xvf64gerpp (&acc1, rowB1, rowA[0]);
             }
@@ -428,14 +428,14 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
             BLASLONG l = 0;
             vec_t *rowA = (vec_t *) & AO[0];
             __vector_pair rowB;
-            rowB = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[0]));
+            rowB = (*(__vector_pair *)((void *)(&BO[0])));
             __builtin_mma_xvf64ger (&acc0, rowB, rowA[0]);
             __builtin_mma_xvf64ger (&acc1, rowB, rowA[1]);
             __builtin_mma_xvf64ger (&acc2, rowB, rowA[2]);
             __builtin_mma_xvf64ger (&acc3, rowB, rowA[3]);
             for (l = 1; l < temp; l++) {
                 rowA = (vec_t *) & AO[l << 3];
-                rowB = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[l << 2]));
+                rowB = (*(__vector_pair *)((void *)(&BO[l << 2])));
                 __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
                 __builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
                 __builtin_mma_xvf64gerpp (&acc2, rowB, rowA[2]);
@@ -466,12 +466,12 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
             BLASLONG l = 0;
             vec_t *rowA = (vec_t *) & AO[0];
             __vector_pair rowB;
-            rowB = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[0]));
+            rowB = (*(__vector_pair *)((void *)(&BO[0])));
             __builtin_mma_xvf64ger (&acc0, rowB, rowA[0]);
             __builtin_mma_xvf64ger (&acc1, rowB, rowA[1]);
             for (l = 1; l < temp; l++) {
                 rowA = (vec_t *) & AO[l << 2];
-                rowB = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[l << 2]));
+                rowB = (*(__vector_pair *)((void *)(&BO[l << 2])));
                 __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
                 __builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
             }
@@ -498,11 +498,11 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
             BLASLONG l = 0;
             vec_t *rowA = (vec_t *) & AO[0];
             __vector_pair rowB;
-            rowB = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[0]));
+            rowB = (*(__vector_pair *)((void *)(&BO[0])));
             __builtin_mma_xvf64ger (&acc0, rowB, rowA[0]);
             for (l = 1; l < temp; l++) {
                 rowA = (vec_t *) & AO[l << 1];
-                rowB = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&BO[l << 2]));
+                rowB = (*(__vector_pair *)((void *)(&BO[l << 2])));
                 __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
             }
             SAVE_ACC (&acc0, 0);
diff --git a/kernel/power/dgemm_small_kernel_nn_power10.c b/kernel/power/dgemm_small_kernel_nn_power10.c
index e3ce979..8ce0b05 100644
--- a/kernel/power/dgemm_small_kernel_nn_power10.c
+++ b/kernel/power/dgemm_small_kernel_nn_power10.c
@@ -314,8 +314,8 @@ typedef __vector unsigned char vec_t;
   *((__vector_pair *)(void *)(packB+(k*8)+4+offset)) = pb1;
 
 #define LOAD_PACKED_B(pb0, pb1, offset)                        \
-  pb0 = __builtin_vsx_lxvp(0L, (const __vector_pair *)((packB+(k*8)+0+offset)));  \
-  pb1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)((packB+(k*8)+4+offset)));
+  pb0 = (*(__vector_pair *)((void *)((packB+(k*8)+0+offset))));  \
+  pb1 = (*(__vector_pair *)((void *)((packB+(k*8)+4+offset))));
 
 #ifdef B0
 int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, IFLOAT * A, BLASLONG lda, FLOAT alpha, IFLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
diff --git a/kernel/power/dgemm_small_kernel_nt_power10.c b/kernel/power/dgemm_small_kernel_nt_power10.c
index a587bdb..fd90dd7 100644
--- a/kernel/power/dgemm_small_kernel_nt_power10.c
+++ b/kernel/power/dgemm_small_kernel_nt_power10.c
@@ -144,11 +144,11 @@ typedef __vector unsigned char vec_t;
 #define LOAD_A_1x1(K, M) ra0 = vec_splats(A[K*lda+M]);
 
 #define LOAD_BP_1x8(K, N)                                 \
-  pb0 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&B[((K)*ldb)+N+0]));  \
-  pb1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&B[((K)*ldb)+N+4]));
+  pb0 = (*(__vector_pair *)((void *)(&B[((K)*ldb)+N+0])));  \
+  pb1 = (*(__vector_pair *)((void *)(&B[((K)*ldb)+N+4])));
 
 #define LOAD_BP_1x4(K, N)                                \
-  pb0 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&B[((K)*ldb)+N+0]));
+  pb0 = (*(__vector_pair *)((void *)(&B[((K)*ldb)+N+0])));
 
 #define LOAD_BP_1x2(K, N)                                  \
   t0 = vec_xl(0, B+(K*ldb)+N);                             \
diff --git a/kernel/power/dgemm_small_kernel_tt_power10.c b/kernel/power/dgemm_small_kernel_tt_power10.c
index 92a67e1..bc3a84f 100644
--- a/kernel/power/dgemm_small_kernel_tt_power10.c
+++ b/kernel/power/dgemm_small_kernel_tt_power10.c
@@ -207,11 +207,11 @@ typedef __vector unsigned char vec_t;
 #define LOAD_A_1x1(M, K) ra0 = vec_splats(A[(M)*lda+K]);
 
 #define LOAD_BP_1x8(K, N)                                 \
-  pb0 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&B[((K)*ldb)+N+0]));  \
-  pb1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&B[((K)*ldb)+N+4]));
+  pb0 = (*(__vector_pair *)((void *)(&B[((K)*ldb)+N+0])));  \
+  pb1 = (*(__vector_pair *)((void *)(&B[((K)*ldb)+N+4])));
 
 #define LOAD_BP_1x4(K, N)                                 \
-  pb0 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&B[((K)*ldb)+N+0]));
+  pb0 = (*(__vector_pair *)((void *)(&B[((K)*ldb)+N+0])));
 
 #define LOAD_BP_1x2(K, N)                                  \
   t0 = vec_xl(0, B+((K)*ldb)+N);                           \
diff --git a/kernel/power/dgemv_t_power10.c b/kernel/power/dgemv_t_power10.c
index 429d7e9..f7f69f4 100644
--- a/kernel/power/dgemv_t_power10.c
+++ b/kernel/power/dgemv_t_power10.c
@@ -61,37 +61,37 @@ static void dgemv_kernel_4x8(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
     a6 = a5 + lda;
     a7 = a6 + lda;
     for (i = 0; i < n/2; i += 2) {
-        vp = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&a0[i*2]));
-        vx = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&x[i*2]));
+        vp = (*(__vector_pair *)((void *)(&a0[i*2])));
+        vx = (*(__vector_pair *)((void *)(&x[i*2])));
         __builtin_vsx_disassemble_pair (res, &vx);
         __builtin_vsx_disassemble_pair (res1, &vp);
         temp0 = vec_madd ((__vector double)res[0], (__vector double)res1[0], temp0);
         temp0 = vec_madd ((__vector double)res[1], (__vector double)res1[1], temp0);
-        vp = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&a1[i*2]));
+        vp = (*(__vector_pair *)((void *)(&a1[i*2])));
         __builtin_vsx_disassemble_pair (res1, &vp);
         temp1 = vec_madd ((__vector double)res[0], (__vector double)res1[0], temp1);
         temp1 = vec_madd ((__vector double)res[1], (__vector double)res1[1], temp1);
-        vp = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&a2[i*2]));
+        vp = (*(__vector_pair *)((void *)(&a2[i*2])));
         __builtin_vsx_disassemble_pair (res1, &vp);
         temp2 = vec_madd ((__vector double)res[0], (__vector double)res1[0], temp2);
         temp2 = vec_madd ((__vector double)res[1], (__vector double)res1[1], temp2);
-        vp = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&a3[i*2]));
+        vp = (*(__vector_pair *)((void *)(&a3[i*2])));
         __builtin_vsx_disassemble_pair (res1, &vp);
         temp3 = vec_madd ((__vector double)res[0], (__vector double)res1[0], temp3);
         temp3 = vec_madd ((__vector double)res[1], (__vector double)res1[1], temp3);
-        vp = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&a4[i*2]));
+        vp = (*(__vector_pair *)((void *)(&a4[i*2])));
         __builtin_vsx_disassemble_pair (res1, &vp);
         temp4 = vec_madd ((__vector double)res[0], (__vector double)res1[0], temp4);
         temp4 = vec_madd ((__vector double)res[1], (__vector double)res1[1], temp4);
-        vp = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&a5[i*2]));
+        vp = (*(__vector_pair *)((void *)(&a5[i*2])));
         __builtin_vsx_disassemble_pair (res1, &vp);
         temp5 = vec_madd ((__vector double)res[0], (__vector double)res1[0], temp5);
         temp5 = vec_madd ((__vector double)res[1], (__vector double)res1[1], temp5);
-        vp = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&a6[i*2]));
+        vp = (*(__vector_pair *)((void *)(&a6[i*2])));
         __builtin_vsx_disassemble_pair (res1, &vp);
         temp6 = vec_madd ((__vector double)res[0], (__vector double)res1[0], temp6);
         temp6 = vec_madd ((__vector double)res[1], (__vector double)res1[1], temp6);
-        vp = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&a7[i*2]));
+        vp = (*(__vector_pair *)((void *)(&a7[i*2])));
         __builtin_vsx_disassemble_pair (res1, &vp);
         temp7 = vec_madd ((__vector double)res[0], (__vector double)res1[0], temp7);
         temp7 = vec_madd ((__vector double)res[1], (__vector double)res1[1], temp7);
diff --git a/kernel/power/zgemm_kernel_power10.c b/kernel/power/zgemm_kernel_power10.c
index 70e73d1..3f9793e 100644
--- a/kernel/power/zgemm_kernel_power10.c
+++ b/kernel/power/zgemm_kernel_power10.c
@@ -316,10 +316,10 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
           SET_ACC_ZERO()
 	  for (l = 0; l < temp; ++l)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<4]));
-              __vector_pair rowA2 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<4)+4]));
-              __vector_pair rowA3 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<4)+8]));
-              __vector_pair rowA4 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<4)+12]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<4])));
+              __vector_pair rowA2 = (*(__vector_pair *)((void *)(&AO[(l<<4)+4])));
+              __vector_pair rowA3 = (*(__vector_pair *)((void *)(&AO[(l<<4)+8])));
+              __vector_pair rowA4 = (*(__vector_pair *)((void *)(&AO[(l<<4)+12])));
               vec_t rowB1 = *(vec_t *) & BO[l<<2];
               vec_t rowB2 = *(vec_t *) & BO[(l<<2)+2];
               __builtin_mma_xvf64gerpp(&acc0, rowA1, rowB1);
@@ -406,10 +406,10 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
           SET_ACC_ZERO()
 	  for (l = 0; l < (temp & (~1)); l+=2)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<3]));
-              __vector_pair rowA2 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<3)+4]));
-              __vector_pair rowA3 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<3)+8]));
-              __vector_pair rowA4 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<3)+12]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<3])));
+              __vector_pair rowA2 = (*(__vector_pair *)((void *)(&AO[(l<<3)+4])));
+              __vector_pair rowA3 = (*(__vector_pair *)((void *)(&AO[(l<<3)+8])));
+              __vector_pair rowA4 = (*(__vector_pair *)((void *)(&AO[(l<<3)+12])));
               vec_t rowB1 = *(vec_t *) & BO[l<<2];
               vec_t rowB2 = *(vec_t *) & BO[(l<<2)+2];
               vec_t rowB3 = *(vec_t *) & BO[(l<<2)+4];
@@ -425,8 +425,8 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
 	    }
 	  for (l = (temp & (~1)); l < temp; ++l)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<3]));
-              __vector_pair rowA2 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<3)+4]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<3])));
+              __vector_pair rowA2 = (*(__vector_pair *)((void *)(&AO[(l<<3)+4])));
               vec_t rowB1 = *(vec_t *) & BO[l<<2];
               vec_t rowB2 = *(vec_t *) & BO[(l<<2)+2];
               __builtin_mma_xvf64gerpp(&acc0, rowA1, rowB1);
@@ -454,10 +454,10 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
           SET_ACC_ZERO()
 	  for (l = 0; l < (temp & (~3)); l+=4)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<2]));
-              __vector_pair rowA2 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<2)+4]));
-              __vector_pair rowA3 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<2)+8]));
-              __vector_pair rowA4 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<2)+12]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<2])));
+              __vector_pair rowA2 = (*(__vector_pair *)((void *)(&AO[(l<<2)+4])));
+              __vector_pair rowA3 = (*(__vector_pair *)((void *)(&AO[(l<<2)+8])));
+              __vector_pair rowA4 = (*(__vector_pair *)((void *)(&AO[(l<<2)+12])));
               vec_t rowB1 = *(vec_t *) & BO[l<<2];
               vec_t rowB2 = *(vec_t *) & BO[(l<<2)+2];
               vec_t rowB3 = *(vec_t *) & BO[(l<<2)+4];
@@ -477,7 +477,7 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
 	    }
 	  for (l = (temp & (~3)); l < temp; ++l)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<2]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<2])));
               vec_t rowB1 = *(vec_t *) & BO[l<<2];
               vec_t rowB2 = *(vec_t *) & BO[(l<<2)+2];
               __builtin_mma_xvf64gerpp(&acc0, rowA1, rowB1);
@@ -503,10 +503,10 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
           SET_ACC_ZERO()
 	  for (l = 0; l < (temp & (~3)); l+=4)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<1]));
-              __vector_pair rowA2 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<1)+2]));
-              __vector_pair rowA3 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<1)+4]));
-              __vector_pair rowA4 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<1)+6]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<1])));
+              __vector_pair rowA2 = (*(__vector_pair *)((void *)(&AO[(l<<1)+2])));
+              __vector_pair rowA3 = (*(__vector_pair *)((void *)(&AO[(l<<1)+4])));
+              __vector_pair rowA4 = (*(__vector_pair *)((void *)(&AO[(l<<1)+6])));
               vec_t rowB1 = *(vec_t *) & BO[l<<2];
               vec_t rowB2 = *(vec_t *) & BO[(l<<2)+2];
               vec_t rowB3 = *(vec_t *) & BO[(l<<2)+4];
@@ -526,7 +526,7 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
 	    }
 	  for (l = (temp & (~3)); l < temp; ++l)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<1]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<1])));
               vec_t rowB1 = *(vec_t *) & BO[l<<2];
               vec_t rowB2 = *(vec_t *) & BO[(l<<2)+2];
               __builtin_mma_xvf64gerpp(&acc0, rowA1, rowB1);
@@ -564,14 +564,14 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
           SET_ACC_ZERO()
 	  for (l = 0; l < (temp & (~1)); l+=2)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<4]));
-              __vector_pair rowA2 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<4)+4]));
-              __vector_pair rowA3 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<4)+8]));
-              __vector_pair rowA4 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<4)+12]));
-              __vector_pair rowA5 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<4)+16]));
-              __vector_pair rowA6 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<4)+20]));
-              __vector_pair rowA7 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<4)+24]));
-              __vector_pair rowA8 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<4)+28]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<4])));
+              __vector_pair rowA2 = (*(__vector_pair *)((void *)(&AO[(l<<4)+4])));
+              __vector_pair rowA3 = (*(__vector_pair *)((void *)(&AO[(l<<4)+8])));
+              __vector_pair rowA4 = (*(__vector_pair *)((void *)(&AO[(l<<4)+12])));
+              __vector_pair rowA5 = (*(__vector_pair *)((void *)(&AO[(l<<4)+16])));
+              __vector_pair rowA6 = (*(__vector_pair *)((void *)(&AO[(l<<4)+20])));
+              __vector_pair rowA7 = (*(__vector_pair *)((void *)(&AO[(l<<4)+24])));
+              __vector_pair rowA8 = (*(__vector_pair *)((void *)(&AO[(l<<4)+28])));
               vec_t rowB1 = *(vec_t *) & BO[l<<1];
               vec_t rowB2 = *(vec_t *) & BO[(l<<1)+2];
               __builtin_mma_xvf64gerpp(&acc0, rowA1, rowB1);
@@ -585,10 +585,10 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
 	    }
 	  for (l = (temp & (~1)); l < temp; ++l)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<4]));
-              __vector_pair rowA2 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<4)+4]));
-              __vector_pair rowA3 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<4)+8]));
-              __vector_pair rowA4 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<4)+12]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<4])));
+              __vector_pair rowA2 = (*(__vector_pair *)((void *)(&AO[(l<<4)+4])));
+              __vector_pair rowA3 = (*(__vector_pair *)((void *)(&AO[(l<<4)+8])));
+              __vector_pair rowA4 = (*(__vector_pair *)((void *)(&AO[(l<<4)+12])));
               vec_t rowB1 = *(vec_t *) & BO[l<<1];
               __builtin_mma_xvf64gerpp(&acc0, rowA1, rowB1);
               __builtin_mma_xvf64gerpp(&acc1, rowA2, rowB1);
@@ -615,14 +615,14 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
           SET_ACC_ZERO()
 	  for (l = 0; l < (temp & (~3)); l+=4)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<3]));
-              __vector_pair rowA2 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<3)+4]));
-              __vector_pair rowA3 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<3)+8]));
-              __vector_pair rowA4 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<3)+12]));
-              __vector_pair rowA5 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<3)+16]));
-              __vector_pair rowA6 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<3)+20]));
-              __vector_pair rowA7 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<3)+24]));
-              __vector_pair rowA8 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<3)+28]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<3])));
+              __vector_pair rowA2 = (*(__vector_pair *)((void *)(&AO[(l<<3)+4])));
+              __vector_pair rowA3 = (*(__vector_pair *)((void *)(&AO[(l<<3)+8])));
+              __vector_pair rowA4 = (*(__vector_pair *)((void *)(&AO[(l<<3)+12])));
+              __vector_pair rowA5 = (*(__vector_pair *)((void *)(&AO[(l<<3)+16])));
+              __vector_pair rowA6 = (*(__vector_pair *)((void *)(&AO[(l<<3)+20])));
+              __vector_pair rowA7 = (*(__vector_pair *)((void *)(&AO[(l<<3)+24])));
+              __vector_pair rowA8 = (*(__vector_pair *)((void *)(&AO[(l<<3)+28])));
               vec_t rowB1 = *(vec_t *) & BO[l<<1];
               vec_t rowB2 = *(vec_t *) & BO[(l<<1)+2];
               vec_t rowB3 = *(vec_t *) & BO[(l<<1)+4];
@@ -638,8 +638,8 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
 	    }
 	  for (l = (temp & (~3)); l < temp; ++l)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<3]));
-              __vector_pair rowA2 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<3)+4]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<3])));
+              __vector_pair rowA2 = (*(__vector_pair *)((void *)(&AO[(l<<3)+4])));
               vec_t rowB1 = *(vec_t *) & BO[l<<1];
               __builtin_mma_xvf64gerpp(&acc0, rowA1, rowB1);
               __builtin_mma_xvf64gerpp(&acc1, rowA2, rowB1);
@@ -662,14 +662,14 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
           SET_ACC_ZERO()
 	  for (l = 0; l < (temp & (~7)); l+=8)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<2]));
-              __vector_pair rowA2 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<2)+4]));
-              __vector_pair rowA3 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<2)+8]));
-              __vector_pair rowA4 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<2)+12]));
-              __vector_pair rowA5 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<2)+16]));
-              __vector_pair rowA6 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<2)+20]));
-              __vector_pair rowA7 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<2)+24]));
-              __vector_pair rowA8 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<2)+28]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<2])));
+              __vector_pair rowA2 = (*(__vector_pair *)((void *)(&AO[(l<<2)+4])));
+              __vector_pair rowA3 = (*(__vector_pair *)((void *)(&AO[(l<<2)+8])));
+              __vector_pair rowA4 = (*(__vector_pair *)((void *)(&AO[(l<<2)+12])));
+              __vector_pair rowA5 = (*(__vector_pair *)((void *)(&AO[(l<<2)+16])));
+              __vector_pair rowA6 = (*(__vector_pair *)((void *)(&AO[(l<<2)+20])));
+              __vector_pair rowA7 = (*(__vector_pair *)((void *)(&AO[(l<<2)+24])));
+              __vector_pair rowA8 = (*(__vector_pair *)((void *)(&AO[(l<<2)+28])));
               vec_t rowB1 = *(vec_t *) & BO[l<<1];
               vec_t rowB2 = *(vec_t *) & BO[(l<<1)+2];
               vec_t rowB3 = *(vec_t *) & BO[(l<<1)+4];
@@ -689,7 +689,7 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
 	    }
 	  for (l = (temp & (~7)); l < temp; ++l)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<2]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<2])));
               vec_t rowB1 = *(vec_t *) & BO[l<<1];
               __builtin_mma_xvf64gerpp(&acc0, rowA1, rowB1);
 	    }
@@ -713,14 +713,14 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
           SET_ACC_ZERO()
 	  for (l = 0; l < (temp & (~7)); l+=8)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<1]));
-              __vector_pair rowA2 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<1)+2]));
-              __vector_pair rowA3 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<1)+4]));
-              __vector_pair rowA4 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<1)+6]));
-              __vector_pair rowA5 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<1)+8]));
-              __vector_pair rowA6 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<1)+10]));
-              __vector_pair rowA7 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<1)+12]));
-              __vector_pair rowA8 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[(l<<1)+14]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<1])));
+              __vector_pair rowA2 = (*(__vector_pair *)((void *)(&AO[(l<<1)+2])));
+              __vector_pair rowA3 = (*(__vector_pair *)((void *)(&AO[(l<<1)+4])));
+              __vector_pair rowA4 = (*(__vector_pair *)((void *)(&AO[(l<<1)+6])));
+              __vector_pair rowA5 = (*(__vector_pair *)((void *)(&AO[(l<<1)+8])));
+              __vector_pair rowA6 = (*(__vector_pair *)((void *)(&AO[(l<<1)+10])));
+              __vector_pair rowA7 = (*(__vector_pair *)((void *)(&AO[(l<<1)+12])));
+              __vector_pair rowA8 = (*(__vector_pair *)((void *)(&AO[(l<<1)+14])));
               vec_t rowB1 = *(vec_t *) & BO[l<<1];
               vec_t rowB2 = *(vec_t *) & BO[(l<<1)+2];
               vec_t rowB3 = *(vec_t *) & BO[(l<<1)+4];
@@ -740,7 +740,7 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i, FLOAT *
 	    }
 	  for (l = (temp & (~7)); l < temp; ++l)
 	    {
-              __vector_pair rowA1 = __builtin_vsx_lxvp(0L, (const __vector_pair *)(&AO[l<<1]));
+              __vector_pair rowA1 = (*(__vector_pair *)((void *)(&AO[l<<1])));
               vec_t rowB1 = *(vec_t *) & BO[l<<1];
               __builtin_mma_xvf64gerpp(&acc0, rowA1, rowB1);
 	    }

[martin-frbg]

This appears to have been brought in by #5833 (to make it compile with IBM's version of clang) @ayappanec - I wonder if restoring the (void*) casts of the A0/B0 argument would make both happy - but if OpenXL accepts the direct cast too(?) we can certainly go with that

[ayappanec]

GCC 13 is a requirement to compile OpenBLAS power10 code because of the availability of new Power vsx builtins.
Please try with GCC 13.