Fix for top-of-tree LLVM (#7194)

halide · Dec 2, 2022 · a9ea9b5 · a9ea9b5
1 parent 43911f4
commit a9ea9b5
Showing 1 changed file with 12 additions and 2 deletions.
diff --git a/test/correctness/simd_op_check_hvx.cpp b/test/correctness/simd_op_check_hvx.cpp
@@ -290,8 +290,18 @@ class SimdOpCheckHVX : public SimdOpCheckTest {
         check("vdelta(v*,v*)", hvx_width / 2, in_u32(3 * x / 2));
         check("vdelta(v*,v*)", hvx_width * 3, in_u16(x * 3));
         check("vdelta(v*,v*)", hvx_width * 3, in_u8(x * 3));
-        check("vdelta(v*,v*)", hvx_width * 4, in_u16(x * 4));
-        check("vdelta(v*,v*)", hvx_width * 4, in_u8(x * 4));
+
+        if (Halide::Internal::get_llvm_version() >= 160) {
+            // As of commit 073d5e5945c4, this pattern doesn't generate vdelta,
+            // but does emit vrdelta, which the author claims is superior codegen
+            // for this case.
+            check("vrdelta(v*,v*)", hvx_width * 4, in_u16(x * 4));
+            check("vrdelta(v*,v*)", hvx_width * 4, in_u8(x * 4));
+        } else {
+            abort();
+            check("vdelta(v*,v*)", hvx_width * 4, in_u16(x * 4));
+            check("vdelta(v*,v*)", hvx_width * 4, in_u8(x * 4));
+        }
 
         check("vlut32(v*.b,v*.b,r*)", hvx_width / 1, in_u8(u8_1));
         check("vlut32(v*.b,v*.b,r*)", hvx_width / 1, in_u8(clamp(u16_1, 0, 63)));