[PATCH] D108725: [AArch64][GlobalISel] Implement custom legalization for s32/s64 G_FCOPYSIGN

[Jessica Paquette via Phabricator via llvm-commits]

paquette added inline comments.


================
Comment at: llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp:1393
+  auto Zero = MIRBuilder.buildConstant(DstTy, 0);
+  auto Ins1 = MIRBuilder.buildInsertVectorElement(
+      VecTy, Undef, MI.getOperand(1).getReg(), Zero);
----------------
aemerson wrote:
> What if we use G_MERGE instead? Do we get an INSERT_SUBREG?
Not quite.

We get the following for s32 with G_MERGE_VALUES:

```
legalize_s32:
	adrp	x8, .LCPI0_0
	mov	v0.s[1], v0.s[0]
	mov	v1.s[1], v0.s[0]
	ldr	q2, [x8, :lo12:.LCPI0_0]
	mov	v0.s[2], v0.s[0]
	mov	v0.s[3], v0.s[0]
	mov	v1.s[2], v0.s[0]
	mov	v1.s[3], v0.s[0]
	bit	v0.16b, v1.16b, v2.16b
	ret
```

Meanwhile, with SDAG we get

```
	movi	v2.4s, #128, lsl #24 ; We should emit the constant like this, but we don't have that optimization
	bit	v0.16b, v1.16b, v2.16b
	ret
```

We can probably change the selector code to recognize the pattern though. Using G_INSERT_VECTOR_ELT is only slightly better.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D108725/new/

https://reviews.llvm.org/D108725