[llvm] AMDGPU: Fix packed 16-bit inline constants (PR #76522)

[Joe Nash via llvm-commits]

================
@@ -2508,53 +2508,95 @@ bool isInlinableLiteral16(int16_t Literal, bool HasInv2Pi) {
          Val == 0x3118;   // 1/2pi
 }
 
-bool isInlinableLiteralV216(int32_t Literal, bool HasInv2Pi) {
-  assert(HasInv2Pi);
-
-  if (isInt<16>(Literal) || isUInt<16>(Literal)) {
-    int16_t Trunc = static_cast<int16_t>(Literal);
-    return AMDGPU::isInlinableLiteral16(Trunc, HasInv2Pi);
+std::optional<unsigned> getInlineEncodingV216(bool IsFloat, uint32_t Literal) {
+  // Unfortunately, the Instruction Set Architecture Reference Guide is
+  // misleading about how the inline operands work for (packed) 16-bit
+  // instructions. In a nutshell, the actual HW behavior is:
+  //
+  //  - integer encodings (-16 .. 64) are always produced as sign-extended
+  //    32-bit values
+  //  - float encodings are produces as:
----------------
Sisyph wrote:

```suggestion
  //  - float encodings are produced as:
```

https://github.com/llvm/llvm-project/pull/76522