arith: Only strided unary int ops

This should have negligible overhead and will save a whole bunch of C
code duplication when the FUnops are also converted to strided form.

1 files changed, 10 insertions, 16 deletions

diff --git a/cbits/arith.c b/cbits/arith.c
index f08e456..6ea197d 100644
--- a/cbits/arith.c
+++ b/cbits/arith.c
@@ -113,7 +113,7 @@ static void print_shape(FILE *stream, i64 rank, const i64 *shape) {
 // Provides idx, outlinidx, arrlinidx.
 #define TARRAY_WALK_NOINNER(again_label_name, rank, shape, strides, body) \
   do { \
-    i64 idx[(rank) - 1]; \
+    i64 idx[(rank) /* - 1 */]; /* Note: [zero-length VLA] */ \
     memset(idx, 0, ((rank) - 1) * sizeof(idx[0])); \
     i64 arrlinidx = 0; \
     i64 outlinidx = 0; \
@@ -138,7 +138,7 @@ static void print_shape(FILE *stream, i64 rank, const i64 *shape) {
 // Provides idx, outlinidx, arrlinidx1, arrlinidx2.
 #define TARRAY_WALK2_NOINNER(again_label_name, rank, shape, strides1, strides2, body) \
   do { \
-    i64 idx[(rank) - 1]; \
+    i64 idx[(rank) /* - 1 */]; /* Note: [zero-length VLA] */ \
     memset(idx, 0, ((rank) - 1) * sizeof(idx[0])); \
     i64 arrlinidx1 = 0, arrlinidx2 = 0; \
     i64 outlinidx = 0; \
@@ -433,16 +433,6 @@ enum unop_tag_t {
 #define LIST_UNOP(name, id, _)
 };
 
-#define ENTRY_UNARY_OPS(typ) \
-  void oxarop_unary_ ## typ(enum unop_tag_t tag, i64 n, typ *out, const typ *x) { \
-    switch (tag) { \
-      case UO_NEG: oxarop_op_neg_ ## typ(n, out, x); break; \
-      case UO_ABS: oxarop_op_abs_ ## typ(n, out, x); break; \
-      case UO_SIGNUM: oxarop_op_signum_ ## typ(n, out, x); break; \
-      default: wrong_op("unary", tag); \
-    } \
-  }
-
 #define ENTRY_UNARY_STRIDED_OPS(typ) \
   void oxarop_unary_ ## typ ## _strided(enum unop_tag_t tag, i64 rank, typ *out, const i64 *shape, const i64 *strides, const typ *x) { \
     switch (tag) { \
@@ -526,9 +516,6 @@ enum redop_tag_t {
   COMM_OP(add, +, typ) \
   NONCOMM_OP(sub, -, typ) \
   COMM_OP(mul, *, typ) \
-  UNARY_OP(neg, -, typ) \
-  UNARY_OP(abs, GEN_ABS, typ) \
-  UNARY_OP(signum, GEN_SIGNUM, typ) \
   UNARY_OP_STRIDED(neg, -, typ) \
   UNARY_OP_STRIDED(abs, GEN_ABS, typ) \
   UNARY_OP_STRIDED(signum, GEN_SIGNUM, typ) \
@@ -537,7 +524,6 @@ enum redop_tag_t {
   REDUCEFULL_OP(sumfull, +, typ) \
   REDUCEFULL_OP(productfull, *, typ) \
   ENTRY_BINARY_OPS(typ) \
-  ENTRY_UNARY_OPS(typ) \
   ENTRY_UNARY_STRIDED_OPS(typ) \
   ENTRY_REDUCE1_OPS(typ) \
   ENTRY_REDUCEFULL_OPS(typ) \
@@ -552,6 +538,7 @@ NUM_TYPES_XLIST
   NONCOMM_OP(fdiv, /, typ) \
   PREFIX_BINOP(pow, GEN_POW, typ) \
   PREFIX_BINOP(logbase, GEN_LOGBASE, typ) \
+  /* TODO: when replaced with UNARY_OP_STRIDED, remove UNARY_OP entirely */ \
   UNARY_OP(recip, 1.0/, typ) \
   UNARY_OP(exp, GEN_EXP, typ) \
   UNARY_OP(log, GEN_LOG, typ) \
@@ -576,3 +563,10 @@ NUM_TYPES_XLIST
   ENTRY_FUNARY_OPS(typ)
 FLOAT_TYPES_XLIST
 #undef X
+
+// Note: [zero-length VLA]
+//
+// Zero-length variable-length arrays are not allowed in C(99). Thus whenever we
+// have a VLA that could sometimes suffice to be empty (e.g. `idx` in the
+// TARRAY_WALK_NOINNER macros), we tweak the length formula (typically by just
+// adding 1) so that it never ends up empty.