[PATCH] D28681: [AArch64] Enable Rounding Double Multiply Add/Subtract instructions on Falkor.

[Ahmed Bougacha via Phabricator via llvm-commits]

ab added inline comments.


================
Comment at: test/CodeGen/AArch64/arm64-neon-v8.1a.ll:2
 ; RUN: llc < %s -verify-machineinstrs -mtriple=arm64-eabi -aarch64-neon-syntax=generic | FileCheck %s --check-prefix=CHECK-V8a
+; RUN: llc < %s -verify-machineinstrs -mtriple=arm64-eabi -mattr=+rdm -aarch64-neon-syntax=generic | FileCheck %s --check-prefix=CHECK-RDM
 ; RUN: llc < %s -verify-machineinstrs -mtriple=arm64-eabi -mattr=+v8.1a -aarch64-neon-syntax=generic | FileCheck %s --check-prefix=CHECK-V81a
----------------
mcrosier wrote:
> ab wrote:
> > Maybe add --check-prefix=CHECK-RDM to the v8.1a line, and remove the rdm CHECK-V81a lines?
> I'm not sure I follow, Ahmed.  Are you suggesting that I not add an additional RUN line and replace
> 
>   ; RUN: llc < %s -verify-machineinstrs -mtriple=arm64-eabi -mattr=+v8.1a -aarch64-neon-syntax=generic | FileCheck %s --check-prefix=CHECK-V81a
> 
> with
> 
>   ; RUN: llc < %s -verify-machineinstrs -mtriple=arm64-eabi -mattr=+rdm -aarch64-neon-syntax=generic | FileCheck %s --check-prefix=CHECK-RDM
> 
> and update the CHECK-V81a directives to CHECK-RDM?
> 
> I'm happy to do that, I just want to confirm that is what you were suggesting.
Oh sorry: I was thinking you'd add the RUN line:


```
; RUN: llc < %s -verify-machineinstrs -mtriple=arm64-eabi -mattr=+rdm -aarch64-neon-syntax=generic | FileCheck %s --check-prefix=CHECK-RDM
; RUN: llc < %s -verify-machineinstrs -mtriple=arm64-eabi -mattr=+v8.1a -aarch64-neon-syntax=generic | FileCheck %s --check-prefix=CHECK-V81a --check-prefix=CHECK-RDM

```

Which lets you avoid duplicating the CHECK-RDM lines; e.g., instead of:


```
    %prod = call <4 x i16> @llvm.aarch64.neon.sqrdmulh.v4i16(<4 x i16> %mhs,  <4 x i16> %rhs)
    %retval =  call <4 x i16> @llvm.aarch64.neon.sqadd.v4i16(<4 x i16> %acc,  <4 x i16> %prod)
 ; CHECK-V8a:        sqrdmulh    v1.4h, v1.4h, v2.4h
+; CHECK-RDM:        sqrdmlah    v0.4h, v1.4h, v2.4h
 ; CHECK-V81a:       sqrdmlah    v0.4h, v1.4h, v2.4h
 ; CHECK-V81a-apple: sqrdmlah.4h v0,    v1,    v2
    ret <4 x i16> %retval

```
you'd do:



```
    %prod = call <4 x i16> @llvm.aarch64.neon.sqrdmulh.v4i16(<4 x i16> %mhs,  <4 x i16> %rhs)
    %retval =  call <4 x i16> @llvm.aarch64.neon.sqadd.v4i16(<4 x i16> %acc,  <4 x i16> %prod)
 ; CHECK-V8a:        sqrdmulh    v1.4h, v1.4h, v2.4h
-; CHECK-V81a:       sqrdmlah    v0.4h, v1.4h, v2.4h
+; CHECK-RDM:        sqrdmlah    v0.4h, v1.4h, v2.4h
 ; CHECK-V81a-apple: sqrdmlah.4h v0,    v1,    v2
    ret <4 x i16> %retval

```

Not a big deal either way, really.


https://reviews.llvm.org/D28681