[llvm] [AMDGPU] Fix decoder for BF16 inline constants (PR #82276)

[Stanislav Mekhanoshin via llvm-commits]

================
@@ -12,3 +12,32 @@ v_dot2_bf16_bf16 v2, 1.0, v0, v2
 
 v_dot2_bf16_bf16 v5, v1, v2, 1.0
 // CHECK: v_dot2_bf16_bf16 v5, v1, v2, 1.0 ; encoding: [0x05,0x00,0x67,0xd6,0x01,0x05,0xca,0x03]
+
+v_dot2_bf16_bf16 v2, v0, -1.0, v2
+// CHECK: v_dot2_bf16_bf16 v2, v0, -1.0, v2       ; encoding: [0x02,0x00,0x67,0xd6,0x00,0xe7,0x09,0x04]
+
+v_dot2_bf16_bf16 v2, v0, 0.5, v2
+// CHECK: v_dot2_bf16_bf16 v2, v0, 0.5, v2        ; encoding: [0x02,0x00,0x67,0xd6,0x00,0xe1,0x09,0x04]
+
+v_dot2_bf16_bf16 v2, v0, -0.5, v2
+// CHECK: v_dot2_bf16_bf16 v2, v0, -0.5, v2       ; encoding: [0x02,0x00,0x67,0xd6,0x00,0xe3,0x09,0x04]
+
+v_dot2_bf16_bf16 v2, v0, 2.0, v2
+// CHECK: v_dot2_bf16_bf16 v2, v0, 2.0, v2        ; encoding: [0x02,0x00,0x67,0xd6,0x00,0xe9,0x09,0x04]
+
+v_dot2_bf16_bf16 v2, v0, -2.0, v2
+// CHECK: v_dot2_bf16_bf16 v2, v0, -2.0, v2       ; encoding: [0x02,0x00,0x67,0xd6,0x00,0xeb,0x09,0x04]
+
+v_dot2_bf16_bf16 v2, v0, 4.0, v2
+// CHECK: v_dot2_bf16_bf16 v2, v0, 4.0, v2        ; encoding: [0x02,0x00,0x67,0xd6,0x00,0xed,0x09,0x04]
+
+v_dot2_bf16_bf16 v2, v0, -4.0, v2
+// CHECK: v_dot2_bf16_bf16 v2, v0, -4.0, v2       ; encoding: [0x02,0x00,0x67,0xd6,0x00,0xef,0x09,0x04]
+
+// FIXME: pi/2 rounded value is incorrect in the inst printer.
----------------
rampitec wrote:

WRT this value: 0.15915494 cannot be represented as bf16. The actual hex value is 0x3e22 which is 0.158203125. The next representable value is 0x3e23 which is 0.1591796875.

We can accept non-representable FP32 value of 0.15915494 out of courtesy, but that is a separate patch.

https://github.com/llvm/llvm-project/pull/82276