From ab201ce8cb84c780742dac71fc2d2996f2b9c775 Mon Sep 17 00:00:00 2001
From: Davit Grigoryan <ggc.language@gmail.com>
Date: Sat, 18 Apr 2026 18:28:57 +0000
Subject: impl xor lane shuffling; add delay option for operand reads

---
 src/abstract_hardware_model.cc | 46 ++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 46 insertions(+)

(limited to 'src/abstract_hardware_model.cc')

diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc
index 9262a92..86427ad 100644
--- a/src/abstract_hardware_model.cc
+++ b/src/abstract_hardware_model.cc
@@ -129,6 +129,52 @@ void warp_inst_t::compute_simd_sets_compacted(unsigned num_sets,
   }
 }
 
+void warp_inst_t::compute_simd_sets_xor_static(unsigned num_sets,
+                                                unsigned set_width,
+                                                unsigned start_set) {
+  // start_set is unused for XOR-static: the lane permutation is fully
+  // determined by warp_id. Accepted for signature compatibility with
+  // compute_simd_sets_compacted.
+  (void)start_set;
+
+  m_simd_sets.resize(num_sets);
+  for (unsigned s = 0; s < num_sets; s++) {
+    simd_set_info &info = m_simd_sets[s];
+    info.set_id = s;
+    info.warp_id = m_warp_id;
+    info.set_active_mask.reset();
+    info.active_mask_in_warp.reset();
+    info.num_active_threads = 0;
+  }
+
+  // xor_key = reverse of the low log2(warp_size) bits of m_warp_id
+  unsigned warp_size = m_config->warp_size;
+  unsigned bits = 0;
+  while ((1u << bits) < warp_size) bits++;
+  unsigned xor_key = 0;
+  for (unsigned i = 0; i < bits; i++) {
+    if (m_warp_id & (1u << i)) xor_key |= (1u << (bits - 1 - i));
+  }
+
+  // Physical lane P runs logical thread T = P ^ xor_key.
+  for (unsigned P = 0; P < warp_size; P++) {
+    unsigned T = P ^ xor_key;
+    unsigned s = P / set_width;
+    unsigned lane = P % set_width;
+    simd_set_info &info = m_simd_sets[s];
+    info.thread_ids[lane] = T;
+    if (m_warp_active_mask.test(T)) {
+      info.set_active_mask.set(lane);
+      info.active_mask_in_warp.set(T);
+      info.num_active_threads++;
+    }
+  }
+
+  for (unsigned s = 0; s < num_sets; s++) {
+    m_simd_sets[s].valid = (m_simd_sets[s].num_active_threads > 0);
+  }
+}
+
 unsigned warp_inst_t::num_active_simd_sets() const {
   unsigned count = 0;
   for (unsigned i = 0; i < m_simd_sets.size(); i++) {
-- 
cgit v1.3