From e6ea2bef6b8e4c22a51e46f676de55dd4dfdb2ba Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?M=C3=A1t=C3=A9=20Pinczel?= <mpinczel90@gmail.com>
Date: Mon, 19 Jan 2026 01:45:18 +0100
Subject: [PATCH] nak: implement uror and urol using shf

Reviewed-by: Mel Henning <mhenning@darkrefraction.com>
Reviewed-by: Mohamed Ahmed <mohamedahmedegypt2001@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39409>
---
 src/nouveau/compiler/nak/api.rs      |  1 +
 src/nouveau/compiler/nak/builder.rs  | 36 +++++++++++++++
 src/nouveau/compiler/nak/from_nir.rs |  8 ++++
 src/nouveau/compiler/nak/hw_tests.rs | 68 ++++++++++++++++++++++++++++
 4 files changed, 113 insertions(+)

diff --git a/src/nouveau/compiler/nak/api.rs b/src/nouveau/compiler/nak/api.rs
index 74c4aa2258b..81bf4dc096e 100644
--- a/src/nouveau/compiler/nak/api.rs
+++ b/src/nouveau/compiler/nak/api.rs
@@ -168,6 +168,7 @@ fn nir_options(dev: &nv_device_info) -> nir_shader_compiler_options {
         lower_scmp: true,
         lower_uadd_carry: true,
         lower_usub_borrow: true,
+        has_rotate32: dev.sm >= 32,
         has_iadd3: dev.sm >= 70,
         has_imad32: dev.sm >= 70,
         has_sdot_4x8: dev.sm >= 70,
diff --git a/src/nouveau/compiler/nak/builder.rs b/src/nouveau/compiler/nak/builder.rs
index a91f8358bf8..829898d9c9d 100644
--- a/src/nouveau/compiler/nak/builder.rs
+++ b/src/nouveau/compiler/nak/builder.rs
@@ -241,6 +241,42 @@ pub trait SSABuilder: Builder {
         dst
     }
 
+    fn urol(&mut self, x: Src, shift: Src) -> SSAValue {
+        let dst = self.alloc_ssa(RegFile::GPR);
+        assert!(self.sm() >= 32);
+
+        self.push_op(OpShf {
+            dst: dst.into(),
+            low: x.clone(),
+            high: x,
+            shift: shift,
+            right: false,
+            wrap: true,
+            data_type: IntType::U32,
+            dst_high: true,
+        });
+
+        dst
+    }
+
+    fn uror(&mut self, x: Src, shift: Src) -> SSAValue {
+        let dst = self.alloc_ssa(RegFile::GPR);
+        assert!(self.sm() >= 32);
+
+        self.push_op(OpShf {
+            dst: dst.into(),
+            low: x.clone(),
+            high: x,
+            shift: shift,
+            right: true,
+            wrap: true,
+            data_type: IntType::U32,
+            dst_high: false,
+        });
+
+        dst
+    }
+
     fn fadd(&mut self, x: Src, y: Src) -> SSAValue {
         let dst = self.alloc_ssa(RegFile::GPR);
         self.push_op(OpFAdd {
diff --git a/src/nouveau/compiler/nak/from_nir.rs b/src/nouveau/compiler/nak/from_nir.rs
index e71f5884a5d..5fc5fc156c3 100644
--- a/src/nouveau/compiler/nak/from_nir.rs
+++ b/src/nouveau/compiler/nak/from_nir.rs
@@ -1533,6 +1533,14 @@ impl<'a> ShaderFromNir<'a> {
                     b.shr(srcs(0), srcs(1), true).into()
                 }
             }
+            nir_op_urol => {
+                assert!(alu.get_src(0).bit_size() == 32);
+                b.urol(srcs(0), srcs(1)).into()
+            }
+            nir_op_uror => {
+                assert!(alu.get_src(0).bit_size() == 32);
+                b.uror(srcs(0), srcs(1)).into()
+            }
             nir_op_lea_nv => {
                 let src_a = srcs(1);
                 let src_b = srcs(0);
diff --git a/src/nouveau/compiler/nak/hw_tests.rs b/src/nouveau/compiler/nak/hw_tests.rs
index 0bb55639aa9..e6b6eb225ac 100644
--- a/src/nouveau/compiler/nak/hw_tests.rs
+++ b/src/nouveau/compiler/nak/hw_tests.rs
@@ -974,6 +974,74 @@ fn test_op_shf() {
     }
 }
 
+#[test]
+fn test_op_urol() {
+    let run = RunSingleton::get();
+    if run.sm.sm() < 32 {
+        return;
+    }
+
+    let mut b = TestShaderBuilder::new(&run.sm);
+    let invocations = 100;
+
+    let x = Src::from(b.ld_test_data(0, MemType::B32)[0]);
+    let y = Src::from(b.ld_test_data(4, MemType::B32)[0]);
+
+    let dst = b.urol(x, y);
+    b.st_test_data(8, MemType::B32, dst.into());
+
+    let bin = b.compile();
+
+    let mut a = Acorn::new();
+    let mut data = Vec::new();
+    for _ in 0..invocations {
+        data.push([a.get_u32(), a.get_u32() as u32, 0]);
+    }
+
+    run.run.run(&bin, &mut data).unwrap();
+
+    for d in &data {
+        let x = d[0];
+        let y = d[1];
+        let dst = x.rotate_left(y);
+        assert_eq!(d[2], dst as u32);
+    }
+}
+
+#[test]
+fn test_op_uror() {
+    let run = RunSingleton::get();
+    if run.sm.sm() < 32 {
+        return;
+    }
+
+    let mut b = TestShaderBuilder::new(&run.sm);
+    let invocations = 100;
+
+    let x = Src::from(b.ld_test_data(0, MemType::B32)[0]);
+    let y = Src::from(b.ld_test_data(4, MemType::B32)[0]);
+
+    let dst = b.uror(x, y);
+    b.st_test_data(8, MemType::B32, dst.into());
+
+    let bin = b.compile();
+
+    let mut a = Acorn::new();
+    let mut data = Vec::new();
+    for _ in 0..invocations {
+        data.push([a.get_u32(), a.get_u32() as u32, 0]);
+    }
+
+    run.run.run(&bin, &mut data).unwrap();
+
+    for d in &data {
+        let x = d[0];
+        let y = d[1];
+        let dst = x.rotate_right(y);
+        assert_eq!(d[2], dst as u32);
+    }
+}
+
 #[test]
 fn test_op_shr() {
     let sm = &RunSingleton::get().sm;