Revert changes from PR#1419

at least one of these changes apparently is an oversimplification, leading to TRMM breakage on some platforms as observed in #1563

kernel/generic/trmm_ltcopy_2.c

@@ -116,22 +116,34 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
       if (m & 1) {
 
 	if (X > posY) {
-	  /* ao1 += 1;
-	  ao2 += 1; */
+	  ao1 += 1;
+	  ao2 += 1;
 	  b += 2;
 	} else
-#ifdef UNIT
 	  if (X < posY) {
-#endif 
-	     b[ 0] = *(ao1 + 0);
-#ifdef UNIT
+	    data01 = *(ao1 + 0);
+	    data02 = *(ao1 + 1);
+
+	    b[ 0] = data01;
+	    b[ 1] = data02;
+	    ao1 += lda;
+	    b += 2;
 	  } else {
+#ifdef UNIT
+	    data02 = *(ao1 + 1);
 
 	    b[ 0] = ONE;
+	    b[ 1] = data02;
+#else
+	    data01 = *(ao1 + 0);
+	    data02 = *(ao1 + 1);
+
+	    b[ 0] = data01;
+	    b[ 1] = data02;
+#endif
+	    ao1 += 2;
+	    b += 2;
 	  }
-#endif 
-	  b[ 1] = *(ao1 + 1);
-	  b += 2;
       }
 
       posY += 2;
@@ -178,7 +190,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
 	} while (i > 0);
       }
 
-      // posY += 1;
+      posY += 1;
   }
 
   return 0;

kernel/generic/trmm_utcopy_16.c

@@ -518,7 +518,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
       i = (m & 15);
       if (i > 0) {
 	if (X < posY) {
-	  /* a01 += i;
+	  a01 += i;
 	  a02 += i;
 	  a03 += i;
 	  a04 += i;
@@ -533,7 +533,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
 	  a13 += i;
 	  a14 += i;
 	  a15 += i;
-	  a16 += i; */
+	  a16 += i;
 	  b += 16 * i;
 	} else
 	  if (X > posY) {
@@ -1130,14 +1130,14 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
       i = (m & 7);
       if (i > 0) {
 	if (X < posY) {
-	  /* a01 += i;
+	  a01 += i;
 	  a02 += i;
 	  a03 += i;
 	  a04 += i;
 	  a05 += i;
 	  a06 += i;
 	  a07 += i;
-	  a08 += i; */
+	  a08 += i;
 	  b += 8 * i;
 	} else
 	  if (X > posY) {
@@ -1156,13 +1156,13 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
 	      b += 8;
 	    }
 
-	    /* a02 += i * lda;
+	    a02 += i * lda;
 	    a03 += i * lda;
 	    a04 += i * lda;
 	    a05 += i * lda;
 	    a06 += i * lda;
 	    a07 += i * lda;
-	    a08 += i * lda; */
+	    a08 += i * lda;
 	  } else {
 #ifdef UNIT
 	    b[ 0] = ONE;
@@ -1371,10 +1371,10 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
       i = (m & 3);
       if (i > 0) {
 	if (X < posY) {
-	  /* a01 += i;
+	  a01 += i;
 	  a02 += i;
 	  a03 += i;
-	  a04 += i; */
+	  a04 += i;
 	  b += 4 * i;
 	} else
 	  if (X > posY) {
@@ -1387,9 +1387,9 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
 	      a01 += lda;
 	      b += 4;
 	    }
-	    /* a02 += lda;
+	    a02 += lda;
 	    a03 += lda;
-	    a04 += lda; */
+	    a04 += lda;
 	  } else {
 
 #ifdef UNIT
@@ -1487,19 +1487,23 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
       if (X < posY) {
 	a01 ++;
 	a02 ++;
-      } else {
-#ifdef UNIT
+	b += 2;
+      } else
 	if (X > posY) {
-#endif
 	  b[  0] = *(a01 +  0);
-#ifdef UNIT
+	  b[  1] = *(a01 +  1);
+	  a01 += lda;
+	  b += 2;
 	} else {
+#ifdef UNIT
 	  b[  0] = ONE;
-	}
+	  b[  1] = *(a01 +  1);
+#else
+	  b[  0] = *(a01 +  0);
+	  b[  1] = *(a01 +  1);
 #endif
-	b[  1] = *(a01 +  1);
-      }
-      b += 2;
+	  b += 2;
+	}
     }
     posY += 2;
   }
@@ -1518,25 +1522,28 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
     if (i > 0) {
       do {
 	if (X < posY) {
-	  a01 ++;
-	} else {
-#ifdef UNIT
+	  a01 += 1;
+	  b ++;
+	} else
 	  if (X > posY) {
-#endif
 	    b[  0] = *(a01 +  0);
-#ifdef UNIT
+	    a01 += lda;
+	    b ++;
 	  } else {
+#ifdef UNIT
 	    b[ 0] = ONE;
-	  }
+#else
+	    b[ 0] = *(a01 +  0);
 #endif
-	  a01 += lda;
-	}
-	b ++;
-	X ++;
-	i --;
+	    a01 += lda;
+	    b ++;
+	  }
+
+	  X += 1;
+	  i --;
       } while (i > 0);
     }
-    // posY += 1;
+    posY += 1;
   }
 
   return 0;

kernel/generic/trmm_utcopy_2.c

@@ -117,8 +117,8 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
       if (m & 1) {
 
 	if (X < posY) {
-	  /* ao1 += 1;
-	  ao2 += 1; */
+	  ao1 += 1;
+	  ao2 += 1;
 	  b += 2;
 	} else
 	  if (X > posY) {
@@ -127,7 +127,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
 
 	    b[ 0] = data01;
 	    b[ 1] = data02;
-	    // ao1 += lda;
+	    ao1 += lda;
 	    b += 2;
 	  } else {
 #ifdef UNIT
@@ -139,7 +139,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
 	    b[ 0] = data01;
 	    b[ 1] = ZERO;
 #endif
-	    // ao1 += lda;
+	    ao1 += lda;
 	    b += 2;
 	  }
       }
@@ -161,18 +161,27 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
     i = m;
     if (m > 0) {
       do {
+	if (X < posY) {
+	  b += 1;
+  	  ao1 += 1;
+	} else
+	  if (X > posY) {
+	    data01 = *(ao1 + 0);
+	    b[ 0] = data01;
+	    b += 1;
+	    ao1 += lda;
+	  } else {
 #ifdef UNIT
-        if (X > posY) {
-#endif
-	  b[ 0] = *(ao1 + 0);
-#ifdef UNIT
-	} else {
-	  b[ 0] = ONE;
-	}
+	    b[ 0] = ONE;
+#else
+	    data01 = *(ao1 + 0);
+	    b[ 0] = data01;
 #endif
-        b ++;
-        ao1 += lda;
-	X ++;
+	    b += 1;
+	    ao1 += lda;
+	  }
+
+	X += 1;
 	i --;
       } while (i > 0);
     }

kernel/generic/trmm_utcopy_4.c

@@ -201,18 +201,18 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
 	if (X < posY) {
 
 	  if (m & 2) {
-	    /* ao1 += 2;
+	    ao1 += 2;
 	    ao2 += 2;
 	    ao3 += 2;
-	    ao4 += 2; */
+	    ao4 += 2;
 	    b += 8;
 	  }
 
 	  if (m & 1) {
-	    /* ao1 += 1;
+	    ao1 += 1;
 	    ao2 += 1;
 	    ao3 += 1;
-	    ao4 += 1; */
+	    ao4 += 1;
 	    b += 4;
 	  }
 
@@ -238,7 +238,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
 	      b[ 7] = data08;
 
 	      ao1 += 2 * lda;
-	      // ao2 += 2 * lda;
+	      ao2 += 2 * lda;
 	      b   += 8;
 	    }
 
@@ -253,7 +253,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
 	      b[ 2] = data03;
 	      b[ 3] = data04;
 
-	      // ao1 += lda;
+	      ao1 += lda;
 	      b += 4;
 	    }
 
@@ -401,7 +401,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
       if (i) {
 
 	if (X < posY) {
-	  // ao1 += 2;
+	  ao1 += 2;
 	  b += 2;
 	} else
 	  if (X > posY) {
@@ -411,7 +411,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
 	    b[ 0] = data01;
 	    b[ 1] = data02;
 
-	    // ao1 += lda;
+	    ao1 += lda;
 	    b += 2;
 	  } else {
 #ifdef UNIT
@@ -443,21 +443,26 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
 	do {
 
 	  if (X < posY) {
+	    b += 1;
 	    ao1 += 1;
-	  } else {
-#ifdef UNIT
+	  } else
 	    if (X > posY) {
-#endif
-	      b[ 0] = *(ao1 + 0);
-#ifdef UNIT
+	      data01 = *(ao1 + 0);
+	      b[ 0] = data01;
+	      ao1 += lda;
+	      b += 1;
 	    } else {
+#ifdef UNIT
 	      b[ 0] = ONE;
-	    }
+#else
+	      data01 = *(ao1 + 0);
+	      b[ 0] = data01;
 #endif
-	    ao1 += lda;
-	  }
-	  b ++;
-	  X ++;
+	      ao1 += lda;
+	      b += 1;
+	    }
+
+	  X += 1;
 	  i --;
 	} while (i > 0);
       }

kernel/generic/trsm_ltcopy_4.c

@@ -206,7 +206,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
       }
 
       a1 += 2 * lda;
-      // a2 += 2 * lda;
+      a2 += 2 * lda;
       b += 8;
 
       ii += 2;

kernel/generic/ztrmm_ltcopy_2.c

@@ -139,18 +139,48 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
       }
 
       if (m & 1) {
-#ifdef UNIT
+
+	if (X > posY) {
+	  ao1 += 2;
+	  ao2 += 2;
+	  b += 4;
+
+	} else
 	  if (X < posY) {
-#endif
-	    b[ 0] = *(ao1 + 0);
-	    b[ 1] = *(ao1 + 1);
-#ifdef UNIT
+	    data1 = *(ao1 + 0);
+	    data2 = *(ao1 + 1);
+	    data3 = *(ao1 + 2);
+	    data4 = *(ao1 + 3);
+
+	    b[ 0] = data1;
+	    b[ 1] = data2;
+	    b[ 2] = data3;
+	    b[ 3] = data4;
+
+	    ao1 += lda;
+	    b += 4;
 	  } else {
+#ifdef UNIT
+	    data3 = *(ao1 + 2);
+	    data4 = *(ao1 + 3);
+
 	    b[ 0] = ONE;
 	    b[ 1] = ZERO;
-	  }
+	    b[ 2] = data3;
+	    b[ 3] = data4;
+#else
+	    data1 = *(ao1 + 0);
+	    data2 = *(ao1 + 1);
+	    data3 = *(ao1 + 2);
+	    data4 = *(ao1 + 3);
+
+	    b[ 0] = data1;
+	    b[ 1] = data2;
+	    b[ 2] = data3;
+	    b[ 3] = data4;
 #endif
-	  b += 4;
+	    b += 4;
+	  }
       }
 
       posY += 2;
@@ -203,7 +233,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
 	} while (i > 0);
       }
 
-      // posY += 1;
+      posY += 1;
   }
 
   return 0;

kernel/generic/ztrsm_utcopy_1.c

@@ -43,7 +43,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
 
   BLASLONG i, ii, j, jj;
 
-  FLOAT data01 = 0.0, data02 = 0.0;
+  FLOAT data01, data02;
   FLOAT *a1;
 
   lda *= 2;

kernel/generic/ztrsm_utcopy_2.c

@@ -43,8 +43,8 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
 
   BLASLONG i, ii, j, jj;
 
-  FLOAT data01 = 0.0, data02 = 0.0, data03, data04;
-  FLOAT data05, data06, data07 = 0.0, data08 = 0.0;
+  FLOAT data01, data02, data03, data04;
+  FLOAT data05, data06, data07, data08;
   FLOAT *a1, *a2;
 
   lda *= 2;