[llvm-branch-commits] [llvm] [AArch64][PAC] Eliminate excessive MOVs when computing blend (PR #115185)

[Anatoly Trosinenko via llvm-branch-commits]

================
@@ -163,8 +163,15 @@ class AArch64AsmPrinter : public AsmPrinter {
   // Emit the sequence for AUT or AUTPAC.
   void emitPtrauthAuthResign(const MachineInstr *MI);
 
-  // Emit the sequence to compute a discriminator into x17, or reuse AddrDisc.
-  unsigned emitPtrauthDiscriminator(uint16_t Disc, unsigned AddrDisc);
+  // Emit the sequence to compute the discriminator.
+  // ScratchReg should be x16/x17.
+  // The returned register is either unmodified AddrDisc or x16/x17.
+  // If the expanded pseudo is allowed to clobber AddrDisc register, setting
+  // MayUseAddrAsScratch may save one MOV instruction, provided the address
----------------
atrosinenko wrote:

I agree the comment is not very clear - hopefully the updated version is more readable. The reason that the commit message mentions two instructions and this comment mentions only one is that the second `mov` is saved due to relaxing the register class.

When the particular instruction clobbers both x16 and x17 and AddrDisc is not used after the discriminator is computed, by relaxing the register class this
```
	ldr	x9, [x16]
	mov	x8, x16
	mov	x17, x8
	movk	x17, #34646, lsl #48
	blraa	x9, x17
```
can be simplified to
```
	ldr	x8, [x16]
	mov	x17, x16
	movk	x17, #34646, lsl #48
	blraa	x8, x17
```
which can be further simplified to
```
	ldr	x8, [x16]
	movk	x16, #34646, lsl #48
	blraa	x8, x16
```
by reusing AddrDisc in-place.

https://github.com/llvm/llvm-project/pull/115185