[llvm][aarch64] Apple A16 & A17 had adrp-add fusion, but A14 did not

[jroelofs]

No description provided.

[llvmbot]

@llvm/pr-subscribers-backend-aarch64

Author: Jon Roelofs (jroelofs)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/81325.diff

1 Files Affected:

• (modified) llvm/lib/Target/AArch64/AArch64.td (+2-1)

diff --git a/llvm/lib/Target/AArch64/AArch64.td b/llvm/lib/Target/AArch64/AArch64.td
index 02fb01caf7e80..bc5cb852e13f1 100644
--- a/llvm/lib/Target/AArch64/AArch64.td
+++ b/llvm/lib/Target/AArch64/AArch64.td
@@ -1120,7 +1120,6 @@ def TuneAppleA14 : SubtargetFeature<"apple-a14", "ARMProcFamily", "AppleA14",
                                     FeatureFuseArithmeticLogic,
                                     FeatureFuseCCSelect,
                                     FeatureFuseCryptoEOR,
-                                    FeatureFuseAdrpAdd,
                                     FeatureFuseLiterals,
                                     FeatureStorePairSuppress,
                                     FeatureZCRegMove,
@@ -1149,6 +1148,7 @@ def TuneAppleA16 : SubtargetFeature<"apple-a16", "ARMProcFamily", "AppleA16",
                                     FeatureArithmeticCbzFusion,
                                     FeatureDisableLatencySchedHeuristic,
                                     FeatureFuseAddress,
+                                    FeatureFuseAdrpAdd,
                                     FeatureFuseAES,
                                     FeatureFuseArithmeticLogic,
                                     FeatureFuseCCSelect,
@@ -1165,6 +1165,7 @@ def TuneAppleA17 : SubtargetFeature<"apple-a17", "ARMProcFamily", "AppleA17",
                                     FeatureArithmeticCbzFusion,
                                     FeatureDisableLatencySchedHeuristic,
                                     FeatureFuseAddress,
+                                    FeatureFuseAdrpAdd,
                                     FeatureFuseAES,
                                     FeatureFuseArithmeticLogic,
                                     FeatureFuseCCSelect,

[jthackray]

LGTM.

[alexander-shaposhnikov]

LG

[davemgreen]

FuseAdrpAdd is also used to enable linker relaxation in lld and the gnu linkers. I guess you are assuming this will make use of linker optimization hints instead and use the Apple linker? (That sounds OK if so).

Changes like this should ideally be accompanied by a test. Hopefully that is fairly simple, like the test in llvm/test/CodeGen/AArch64/misched-fusion-addadrp.ll.

[jroelofs]

I was assuming Apple ld/ld-prime, and only thinking about the effect of this on instruction scheduling / macro-fusion + knowledge of the respective uarch's.

Can you tell me more about how that linker relaxation works? Is it an optimization that keys off of FeatureAdrpAdd itself, or is it merely enabled by placing the adrp next to the add?

Do you think A14 deserves:

+ // apple-a14 doesn't actually fuse these, but scheduling them 
+ // adjacent allows linker relaxation.
FeatureFuseAdrpAdd,

instead of:

- FeatureFuseAdrpAdd,

Changes like this should ideally be accompanied by a test.

I'll add one.

[jroelofs]

Changes like this should ideally be accompanied by a test.

I'll add one.

ffab5a0

[davemgreen]

I was assuming Apple ld/ld-prime, and only thinking about the effect of this on instruction scheduling / macro-fusion + knowledge of the respective uarch's.

Can you tell me more about how that linker relaxation works? Is it an optimization that keys off of FeatureAdrpAdd itself, or is it merely enabled by placing the adrp next to the add?

As far as I understand inside the linker it is done via adjacent instructions. It doesn't have much other information, other than the relocations. I think so long as the linker used will most likely be apple ld using LoH's, then it's fine to be more accurate with the fusion and not "mis-use" the feature to get adjacent instructions.

Changes like this should ideally be accompanied by a test.

I'll add one.

ffab5a0

Thanks!