From c734252ab9d76d6f6c54dee35fd149db8484dc9d Mon Sep 17 00:00:00 2001
From: lloda <lloda@sarc.name>
Date: Tue, 28 Nov 2023 12:12:26 +0100
Subject: [PATCH] Another workaround for 'forwarding' issue

Doesn't seem to be a forwarding issue at all, but rather gcc thinking that rank
can be 0 in places where it can't.

gcc11 doesn't flag these, but also fails to optimize as bench-at:Small/Small as
much.

* ra/expr.hh (Match): Revert to no fwd.
* ra/big.hh (CellBig), ra/small.hh (CellSmall): Remove diagnostic guards around
  len().
* ra/ply.hh (ply_ravel): Move guards here, where the confusion happens.
* ra/expr.hh (Match::check): Idem.
---
 ra/big.hh     | 3 ---
 ra/expr.hh    | 5 ++++-
 ra/ply.hh     | 3 +++
 ra/small.hh   | 3 ---
 test/bug83.cc | 2 +-
 5 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/ra/big.hh b/ra/big.hh
index 213c28ff..69b138d9 100644
--- a/ra/big.hh
+++ b/ra/big.hh
@@ -42,10 +42,7 @@ struct CellBig
 
     consteval static rank_t rank() requires (ANY!=framer) { return framer; }
     constexpr rank_t rank() const requires (ANY==framer) { return rank_frame(std::ssize(dimv), dspec); }
-#pragma GCC diagnostic push // gcc 12.2 and 13.2 with RA_DO_CHECK=0 and -fno-sanitize=all
-#pragma GCC diagnostic warning "-Warray-bounds"
     constexpr dim_t len(int k) const { return dimv[k].len; } // len(0<=k<rank) or step(0<=k)
-#pragma GCC diagnostic pop
     constexpr static dim_t len_s(int k) { return ANY; }
     constexpr dim_t step(int k) const { return k<rank() ? dimv[k].step : 0; }
     constexpr void adv(rank_t k, dim_t d) { c.cp += step(k)*d; }
diff --git a/ra/expr.hh b/ra/expr.hh
index 0d10477b..8d9a4ff7 100644
--- a/ra/expr.hh
+++ b/ra/expr.hh
@@ -346,7 +346,10 @@ struct Match<checkp, std::tuple<P ...>, mp::int_list<I ...>>
             return false;
         } else if constexpr (1==c) {
             for (int k=0; k<rank(); ++k) {
+#pragma GCC diagnostic push // gcc 12.2 and 13.2 with RA_DO_CHECK=0 and -fno-sanitize=all.
+#pragma GCC diagnostic warning "-Warray-bounds"
                 dim_t ls = len(k);
+#pragma GCC diagnostic pop
                 if (((k<ra::rank(std::get<I>(t)) && ls!=choose_len(std::get<I>(t).len(k), ls)) || ...)) {
                     RA_CHECK(!checkp, "Mismatch on axis ", k, " [", (std::array { std::get<I>(t).len(k) ... }), "].");
                     return false;
@@ -357,7 +360,7 @@ struct Match<checkp, std::tuple<P ...>, mp::int_list<I ...>>
     }
 
     constexpr
-    Match(P ... p_): t(RA_FWD(p_) ...) // [ra1]
+    Match(P ... p_): t(p_ ...) // [ra1]
     {
 // TODO Maybe on ply, would make checkp unnecessary, make agree_xxx() unnecessary.
         if constexpr (checkp && !(has_len<P> || ...)) {
diff --git a/ra/ply.hh b/ra/ply.hh
index 313ab5e8..f3081315 100644
--- a/ra/ply.hh
+++ b/ra/ply.hh
@@ -160,6 +160,8 @@ ply_ravel(A && a, Early && early = Nop {})
 // find outermost compact dim.
     rank_t * ocd = order;
     dim_t ss = a.len(*ocd);
+#pragma GCC diagnostic push // gcc 12.2 and 13.2 with RA_DO_CHECK=0 and -fno-sanitize=all
+#pragma GCC diagnostic warning "-Warray-bounds"
     for (--rank, ++ocd; rank>0 && a.keep_step(ss, order[0], *ocd); --rank, ++ocd) {
         ss *= a.len(*ocd);
     }
@@ -203,6 +205,7 @@ ply_ravel(A && a, Early && early = Nop {})
             }
         }
     }
+#pragma GCC diagnostic pop
 }
 
 
diff --git a/ra/small.hh b/ra/small.hh
index 7ac84dde..2fa36109 100644
--- a/ra/small.hh
+++ b/ra/small.hh
@@ -265,10 +265,7 @@ struct CellSmall
     ctype c;
 
     consteval static rank_t rank() { return framer; }
-#pragma GCC diagnostic push // gcc 13.2
-#pragma GCC diagnostic warning "-Warray-bounds"
     constexpr static dim_t len(int k) { return dimv[k].len; } // len(0<=k<rank) or step(0<=k)
-#pragma GCC diagnostic pop
     constexpr static dim_t len_s(int k) { return len(k); }
     constexpr static dim_t step(int k) { return k<rank() ? dimv[k].step : 0; }
     constexpr void adv(rank_t k, dim_t d) { c.cp += step(k)*d; }
diff --git a/test/bug83.cc b/test/bug83.cc
index 68bca69e..7bfc0d54 100644
--- a/test/bug83.cc
+++ b/test/bug83.cc
@@ -20,7 +20,7 @@ main()
 {
     ra::TestRecorder tr(std::cout);
     ra::Big<double, 2> A({4, 4}, ra::_1*10 + ra::_0 + 3);
-    A = ra::expr([](auto && a) { return a(0, 0); }, ra::iter<2>(A));
+    A = ra::expr([](auto && a) { return a(0, 0); }, ra::iter<2>(A)); // [ra1]
     tr.info("A ", A).test_eq(3, A);
     return tr.summary();
 }