agx: Mask shifts in the backend

This gives our shifts SM5 behaviour at the cost of a little extra ALU. That way,
we match NIR's shifts.

This fixes unsoundness of GLSL expressions like "a << (b & 31)", where the &
would mistakenly get optimized away.

Closes: #8181
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reported-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21673>

diff --git a/src/asahi/compiler/agx_compile.c b/src/asahi/compiler/agx_compile.c
index 2d69bac..d9782ea 100644 (file)
--- a/src/asahi/compiler/agx_compile.c
+++ b/src/asahi/compiler/agx_compile.c
@@ -31,6 +31,7 @@
 #include "agx_builder.h"
 #include "agx_compiler.h"
 #include "agx_internal_formats.h"
+#include "agx_nir.h"
 
 /* Alignment for shader programs. I'm not sure what the optimal value is. */
 #define AGX_CODE_ALIGN 0x100
@@ -1918,6 +1919,7 @@ agx_optimize_nir(nir_shader *nir, unsigned *preamble_size)
    NIR_PASS_V(nir, nir_opt_peephole_select, 64, false, true);
 
    NIR_PASS_V(nir, nir_opt_algebraic_late);
+   NIR_PASS_V(nir, agx_nir_lower_algebraic_late);
    NIR_PASS_V(nir, nir_opt_constant_folding);
 
    /* Must run after uses are fixed but before a last round of copyprop + DCE */

diff --git a/src/asahi/compiler/agx_nir.h b/src/asahi/compiler/agx_nir.h
new file mode 100644 (file)
index 0000000..6829116
--- /dev/null
+++ b/src/asahi/compiler/agx_nir.h
@@ -0,0 +1,14 @@
+/*
+ * Copyright 2023 Alyssa Rosenzweig
+ * SPDX-License-Identifier: MIT
+ */
+#ifndef AGX_NIR_H
+#define AGX_NIR_H
+
+#include <stdbool.h>
+
+struct nir_shader;
+
+bool agx_nir_lower_algebraic_late(struct nir_shader *shader);
+
+#endif

diff --git a/src/asahi/compiler/agx_nir_algebraic.py b/src/asahi/compiler/agx_nir_algebraic.py
new file mode 100644 (file)
index 0000000..2195f43
--- /dev/null
+++ b/src/asahi/compiler/agx_nir_algebraic.py
@@ -0,0 +1,41 @@
+# Copyright 2022 Alyssa Rosenzweig
+# Copyright 2021 Collabora, Ltd.
+# Copyright 2016 Intel Corporation
+# SPDX-License-Identifier: MIT
+
+import argparse
+import sys
+import math
+
+a = 'a'
+b = 'b'
+c = 'c'
+
+lower_sm5_shift = []
+
+# Our shifts differ from SM5 for the upper bits. Mask to match the NIR
+# behaviour. Because this happens as a late lowering, NIR won't optimize the
+# masking back out (that happens in the main nir_opt_algebraic).
+for s in [8, 16, 32, 64]:
+    for shift in ["ishl", "ishr", "ushr"]:
+        lower_sm5_shift += [((shift, f'a@{s}', b),
+                             (shift, a, ('iand', b, s - 1)))]
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-p', '--import-path', required=True)
+    args = parser.parse_args()
+    sys.path.insert(0, args.import_path)
+    run()
+
+def run():
+    import nir_algebraic  # pylint: disable=import-error
+
+    print('#include "agx_nir.h"')
+
+    print(nir_algebraic.AlgebraicPass("agx_nir_lower_algebraic_late",
+                                      lower_sm5_shift).render())
+
+
+if __name__ == '__main__':
+    main()

diff --git a/src/asahi/compiler/meson.build b/src/asahi/compiler/meson.build
index 61f87ca..0404ed3 100644 (file)
--- a/src/asahi/compiler/meson.build
+++ b/src/asahi/compiler/meson.build
@@ -43,6 +43,17 @@ libasahi_agx_files = files(
   'agx_validate.c',
 )
 
+agx_nir_algebraic_c = custom_target(
+  'agx_nir_algebraic.c',
+  input : 'agx_nir_algebraic.py',
+  output : 'agx_nir_algebraic.c',
+  command : [
+    prog_python, '@INPUT@', '-p', dir_compiler_nir,
+  ],
+  capture : true,
+  depend_files : nir_algebraic_depends,
+)
+
 agx_opcodes_h = custom_target(
   'agx_opcodes.h',
   input : ['agx_opcodes.h.py'],
@@ -82,7 +93,7 @@ idep_agx_builder_h = declare_dependency(
 
 libasahi_compiler = static_library(
   'asahi_compiler',
-  [libasahi_agx_files, agx_opcodes_c],
+  [libasahi_agx_files, agx_opcodes_c, agx_nir_algebraic_c],
   include_directories : [inc_include, inc_src, inc_mesa, inc_gallium, inc_gallium_aux, inc_mapi],
   dependencies: [idep_nir, idep_agx_opcodes_h, idep_agx_builder_h, idep_agx_pack],
   c_args : [no_override_init_args],