nir/algebraic: add patterns for a >> #b << #b

Fixes compilation of a Battlefront 2 shader with ACO by removing VGPR
spilling. The reassociation makes it worse on LLVM though.

pipeline-db (ACO):
Totals from affected shaders:
SGPRS: 10704 -> 10688 (-0.15 %)
VGPRS: 18736 -> 18528 (-1.11 %)
Spilled SGPRs: 70 -> 70 (0.00 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 0 -> 0 (0.00 %) dwords per thread
Code Size: 909696 -> 885796 (-2.63 %) bytes
LDS: 225 -> 225 (0.00 %) blocks
Max Waves: 1115 -> 1129 (1.26 %)

pipeline-db (LLVM):
Totals from affected shaders:
SGPRS: 8472 -> 8424 (-0.57 %)
VGPRS: 14284 -> 14368 (0.59 %)
Spilled SGPRs: 0 -> 0 (0.00 %)
Spilled VGPRs: 442 -> 503 (13.80 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 268 -> 396 (47.76 %) dwords per thread
Code Size: 862568 -> 853028 (-1.11 %) bytes
LDS: 0 -> 0 (0.00 %) blocks
Max Waves: 971 -> 964 (-0.72 %)

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/2271>

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index 7b9a6a8e45d3ee3915f3a97372337d928a2f8c35..fd0007bb54d1bf8bad15cb9a6b2fee32136c7316 100644 (file)
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -255,6 +255,22 @@ for s in [8, 16, 32, 64]:
         (ishr, a, ('imin', ('iadd', ('iand', b, mask), ('iand', c, mask)), s - 1))),
    ])
 
+# Optimize a pattern of address calculation created by DXVK where the offset is
+# divided by 4 and then multipled by 4. This can be turned into an iand and the
+# additions before can be reassociated to CSE the iand instruction.
+for log2 in range(1, 7): # powers of two from 2 to 64
+   v = 1 << log2
+   mask = 0xffffffff & ~(v - 1)
+   b_is_multiple = '#b(is_unsigned_multiple_of_{})'.format(v)
+
+   optimizations.extend([
+       # 'a >> #b << #b' -> 'a & ~((1 << #b) - 1)'
+       (('ishl@32', ('ushr@32', a, log2), log2), ('iand', a, mask)),
+
+       # Reassociate for improved CSE
+       (('iand@32', ('iadd@32', a, b_is_multiple), mask), ('iadd', ('iand', a, mask), b)),
+   ])
+
 optimizations.extend([
    # This is common for address calculations.  Reassociating may enable the
    # 'a<<c' to be CSE'd.  It also helps architectures that have an ISHLADD

diff --git a/src/compiler/nir/nir_search_helpers.h b/src/compiler/nir/nir_search_helpers.h
index 0c2f631ac39029026a45127801bd907af610cd8b..abfd660ae792597475df99956a411b6a1eab9398 100644 (file)
--- a/src/compiler/nir/nir_search_helpers.h
+++ b/src/compiler/nir/nir_search_helpers.h
@@ -90,6 +90,32 @@ is_neg_power_of_two(UNUSED struct hash_table *ht, nir_alu_instr *instr,
    return true;
 }
 
+#define MULTIPLE(test)                                                  \
+static inline bool                                                      \
+is_unsigned_multiple_of_ ## test(UNUSED struct hash_table *ht, nir_alu_instr *instr, \
+                                 unsigned src, unsigned num_components, \
+                                 const uint8_t *swizzle)                \
+{                                                                       \
+   /* only constant srcs: */                                            \
+   if (!nir_src_is_const(instr->src[src].src))                          \
+      return false;                                                     \
+                                                                        \
+   for (unsigned i = 0; i < num_components; i++) {                      \
+      uint64_t val = nir_src_comp_as_uint(instr->src[src].src, swizzle[i]); \
+      if (val % test != 0)                                              \
+         return false;                                                  \
+   }                                                                    \
+                                                                        \
+   return true;                                                         \
+}
+
+MULTIPLE(2)
+MULTIPLE(4)
+MULTIPLE(8)
+MULTIPLE(16)
+MULTIPLE(32)
+MULTIPLE(64)
+
 static inline bool
 is_zero_to_one(UNUSED struct hash_table *ht, nir_alu_instr *instr, unsigned src,
                unsigned num_components,