From da9c0e77d9b70a42caea0bbdeea435a021ee27e2 Mon Sep 17 00:00:00 2001
From: Violet <c01368481@gmail.com>
Date: Tue, 23 Sep 2025 18:16:03 +0000
Subject: [PATCH] Improve constant folding for sign extension

---
 ptx/lib/zluda_ptx_impl.bc  | Bin 24796 -> 24896 bytes
 ptx/lib/zluda_ptx_impl.cpp |  15 ++++++++++++++-
 2 files changed, 14 insertions(+), 1 deletion(-)

diff --git a/ptx/lib/zluda_ptx_impl.bc b/ptx/lib/zluda_ptx_impl.bc
index b89e0d120aa69ba530b9a6a739cdbc6d1dcf250e..afc9c2c502e413f473ecd473e5f781656d4a9e4c 100644
GIT binary patch
delta 303
zcmca}knzAF#tABn=QgV53Ny)QZJr{0$F$ysfq_Ao#ZgCqfq{XSfq}t9qf<zuxk;nl
zNu%|UMsJhm1SXLc4h{|u9b7CdEgB9E4lPV9N9r{k9GaM(*lYY}WME)$_qDns5RgzH
z)+!!Q#kh5Ppw(RkmZQ0%aT{J1I3H*{SScQN;3eywc?VyXGz*BBv&qL9uvy&^n7}A8
zr(SXjW1It96{F-E$&>|`d7T38NcAwrZ2-w{V3e36J%usu09zHK^c%?>uyn!%HuDBA
zsXGTa4=Sm$n-@535U;RKDqv9LU=!AUaZrxsFmGA{dqTk>=@%u=1{{ZZ6DP1I7~EO#
zl2L-k+^3p>?J{q%0{g*|W|<=g3!F9xU*;_mU_ZurP)_)TX7K`bmfegDn`OQBiva**
CU1fRz

delta 202
zcmX?bi1E%r#tABnTQ{oa3N!U+Y@Q-~$FyFXfq_Ao#ZgCqfq{XKfq}t9qy3PEgM-74
zCJhINIZhf5eS!^4A{q`8m{?R+I5;?TaIvtkv}iau{Ap)mVR6xL_}BbNo`Hcu-L)#X
zMBJq+6ht_0X!k5{Tw=~r?YK#qja{HA!C<fBbJpvrR(Av<5(=7^UgmXCh>=LJx}ztM
zkx-C4;bnpIfyRSV8<!pCZDeP=%sX4m$2>!_&}qYE-dpW+1l~9{#vOjic`$Etqt|{h
E0L~Fb<p2Nx

diff --git a/ptx/lib/zluda_ptx_impl.cpp b/ptx/lib/zluda_ptx_impl.cpp
index f985ebc..f247f45 100644
--- a/ptx/lib/zluda_ptx_impl.cpp
+++ b/ptx/lib/zluda_ptx_impl.cpp
@@ -812,6 +812,14 @@ typedef uint32_t ShflSyncResult __attribute__((ext_vector_type(2)));
         {
             uint8_t sel = static_cast<uint8_t>(s >> (i * 4));
             uint8_t addr = sel & 0x7;
+            if (sel & 0x8)
+            {
+                if (addr % 2 == 1)
+                {
+                    v_perm_selector.u8x4[i] = 0x8 + addr / 2;
+                    continue;
+                }
+            }
             v_perm_selector.u8x4[i] = addr;
         }
 
@@ -821,9 +829,14 @@ typedef uint32_t ShflSyncResult __attribute__((ext_vector_type(2)));
         for (size_t i = 0; i < 4; i++)
         {
             uint8_t sel = static_cast<uint8_t>(s >> (i * 4));
+            uint8_t addr = sel & 0x7;
             if (sel & 0x8)
             {
-                output.u8x4[i] = (output.u8x4[i] & 0x80) * 0xff;
+                if (addr % 2 != 1)
+                {
+                    output.u8x4[i] = (output.u8x4[i] & 0x80) * 0xff;
+                    continue;
+                }
             }
         }