[llvm] [AArch64] Apply generational deltas for Apple tuning features (PR #196792)
Tomer Shafir via llvm-commits
llvm-commits at lists.llvm.org
Sun May 10 04:11:31 PDT 2026
https://github.com/tomershafir created https://github.com/llvm/llvm-project/pull/196792
This patch refactors how we manage tuning features for AArch64 Apple CPUs. Instead of duplicating feature lists for each generation, we now form a chain such that every generation only adds or removes features from the immediate predecessor.
This creates a much more compact representation, enables to view generational changes at a glance, and can reduce potential copy-paste errors by establishing a more structured dataflow.
However, not without pitfalls: to gather a feature list for a specific gen we need to traverse a chain, and feature addition and removal becomes more subtle as it propagates across the chain. To alleviate these issues, this patch introduces a new TableGen test for complete feature lists per generation.
>From d86822e2f23b9e3de0528ad5972fd78bb76d81b0 Mon Sep 17 00:00:00 2001
From: tomershafir <tomer.shafir8 at gmail.com>
Date: Sun, 10 May 2026 13:16:48 +0300
Subject: [PATCH] [AArch64] Apply generational deltas for Apple tuning features
This patch refactors how we manage tuning features for AArch64 Apple CPUs. Instead of duplicating feature lists for each generation, we now form a chain such that every generation only adds or removes features from the immediate predecessor.
This creates a much more compact representation, enables to view generational changes at a glance, and can reduce potential copy-paste errors by establishing a more structured dataflow.
However, not without pitfalls: to gather a feature list for a specific gen we need to traverse a chain, and feature addition and removal becomes more subtle as it propagates across the chain. To alleviate these issues, this patch introduces a new TableGen test for complete feature lists per generation.
---
llvm/lib/Target/AArch64/AArch64Processors.td | 211 +++---------------
.../TableGen/aarch64-apple-tuning-features.td | 28 +++
2 files changed, 55 insertions(+), 184 deletions(-)
create mode 100644 llvm/test/TableGen/aarch64-apple-tuning-features.td
diff --git a/llvm/lib/Target/AArch64/AArch64Processors.td b/llvm/lib/Target/AArch64/AArch64Processors.td
index 74b2700d08fd1..6d2b15fb81891 100644
--- a/llvm/lib/Target/AArch64/AArch64Processors.td
+++ b/llvm/lib/Target/AArch64/AArch64Processors.td
@@ -387,27 +387,9 @@ def TuneOlympus : SubtargetFeature<"olympus", "ARMProcFamily", "Olympus",
FeatureUseFixedOverScalableIfEqualCost,
FeatureMaxInterleaveFactor4]>;
-// Note that cyclone does not fuse AES instructions, but newer apple chips do
-// perform the fusion and cyclone is used by default when targeting apple OSes.
def TuneAppleA7 : SubtargetFeature<"apple-a7", "ARMProcFamily", "AppleA7",
- "Apple A7 (the CPU formerly known as Cyclone)", [
- FeatureAlternateSExtLoadCVTF32Pattern,
- FeatureArithmeticBccFusion,
- FeatureArithmeticCbzFusion,
- FeatureDisableLatencySchedHeuristic,
- FeatureFuseAES, FeatureFuseCryptoEOR,
- FeatureStorePairSuppress,
- FeatureZCRegMoveGPR64,
- FeatureZCRegMoveFPR128,
- FeatureZCZeroingGPR32,
- FeatureZCZeroingGPR64,
- FeatureNoZCZeroingFPR64,
- FeatureZCZeroingFPR128,
- FeatureZCZeroingFPWorkaround]>;
-
-def TuneAppleA10 : SubtargetFeature<"apple-a10", "ARMProcFamily", "AppleA10",
- "Apple A10", [
- FeatureAlternateSExtLoadCVTF32Pattern,
+ "Apple A7 (the CPU formerly known as Cyclone)",
+ [FeatureAlternateSExtLoadCVTF32Pattern,
FeatureArithmeticBccFusion,
FeatureArithmeticCbzFusion,
FeatureDisableLatencySchedHeuristic,
@@ -416,198 +398,59 @@ def TuneAppleA10 : SubtargetFeature<"apple-a10", "ARMProcFamily", "AppleA10",
FeatureStorePairSuppress,
FeatureZCRegMoveGPR64,
FeatureZCRegMoveFPR128,
+ FeatureZCZeroingFPWorkaround,
FeatureZCZeroingGPR32,
FeatureZCZeroingGPR64,
FeatureNoZCZeroingFPR64,
FeatureZCZeroingFPR128]>;
+def TuneAppleA10 : SubtargetFeature<"apple-a10", "ARMProcFamily", "AppleA10",
+ "Apple A10",
+ !listremove(TuneAppleA7.Implies,
+ [FeatureZCZeroingFPWorkaround])>;
+
def TuneAppleA11 : SubtargetFeature<"apple-a11", "ARMProcFamily", "AppleA11",
- "Apple A11", [
- FeatureAlternateSExtLoadCVTF32Pattern,
- FeatureArithmeticBccFusion,
- FeatureArithmeticCbzFusion,
- FeatureDisableLatencySchedHeuristic,
- FeatureFuseAES,
- FeatureFuseCryptoEOR,
- FeatureStorePairSuppress,
- FeatureZCRegMoveGPR64,
- FeatureZCRegMoveFPR128,
- FeatureZCZeroingGPR32,
- FeatureZCZeroingGPR64,
- FeatureNoZCZeroingFPR64,
- FeatureZCZeroingFPR128]>;
+ "Apple A11", TuneAppleA10.Implies>;
def TuneAppleA12 : SubtargetFeature<"apple-a12", "ARMProcFamily", "AppleA12",
- "Apple A12", [
- FeatureAlternateSExtLoadCVTF32Pattern,
- FeatureArithmeticBccFusion,
- FeatureArithmeticCbzFusion,
- FeatureDisableLatencySchedHeuristic,
- FeatureFuseAES,
- FeatureFuseCryptoEOR,
- FeatureStorePairSuppress,
- FeatureZCRegMoveGPR64,
- FeatureZCRegMoveFPR128,
- FeatureZCZeroingGPR32,
- FeatureZCZeroingGPR64,
- FeatureNoZCZeroingFPR64,
- FeatureZCZeroingFPR128]>;
+ "Apple A12", TuneAppleA11.Implies>;
def TuneAppleA13 : SubtargetFeature<"apple-a13", "ARMProcFamily", "AppleA13",
- "Apple A13", [
- FeatureAlternateSExtLoadCVTF32Pattern,
- FeatureArithmeticBccFusion,
- FeatureArithmeticCbzFusion,
- FeatureDisableLatencySchedHeuristic,
- FeatureFuseAES,
- FeatureFuseCryptoEOR,
- FeatureStorePairSuppress,
- FeatureZCRegMoveGPR64,
- FeatureZCRegMoveFPR128,
- FeatureZCZeroingGPR32,
- FeatureZCZeroingGPR64,
- FeatureNoZCZeroingFPR64,
- FeatureZCZeroingFPR128]>;
+ "Apple A13", TuneAppleA12.Implies>;
def TuneAppleA14 : SubtargetFeature<"apple-a14", "ARMProcFamily", "AppleA14",
- "Apple A14", [
+ "Apple A14",
+ !listconcat(TuneAppleA13.Implies, [
FeatureAggressiveFMA,
- FeatureAlternateSExtLoadCVTF32Pattern,
- FeatureArithmeticBccFusion,
- FeatureArithmeticCbzFusion,
- FeatureDisableLatencySchedHeuristic,
FeatureFastLD1Single,
FeatureFuseAddress,
- FeatureFuseAES,
FeatureFuseArithmeticLogic,
FeatureFuseCmpCSel,
- FeatureFuseCryptoEOR,
FeatureFuseLiterals,
- FeatureStorePairSuppress,
- FeatureZCRegMoveGPR64,
- FeatureZCRegMoveFPR128,
- FeatureZCZeroingGPR32,
- FeatureZCZeroingGPR64,
- FeatureNoZCZeroingFPR64,
- FeatureZCZeroingFPR128,
- FeatureMaxInterleaveFactor4]>;
+ FeatureMaxInterleaveFactor4])>;
def TuneAppleA15 : SubtargetFeature<"apple-a15", "ARMProcFamily", "AppleA15",
- "Apple A15", [
- FeatureAlternateSExtLoadCVTF32Pattern,
- FeatureArithmeticBccFusion,
- FeatureArithmeticCbzFusion,
- FeatureDisableLatencySchedHeuristic,
- FeatureFastLD1Single,
- FeatureFuseAddress,
- FeatureFuseAdrpAdd,
- FeatureFuseAES,
- FeatureFuseArithmeticLogic,
- FeatureFuseCmpCSel,
- FeatureFuseCryptoEOR,
- FeatureFuseLiterals,
- FeatureStorePairSuppress,
- FeatureZCRegMoveGPR64,
- FeatureZCRegMoveFPR128,
- FeatureZCZeroingGPR32,
- FeatureZCZeroingGPR64,
- FeatureNoZCZeroingFPR64,
- FeatureZCZeroingFPR128,
- FeatureMaxInterleaveFactor4]>;
+ "Apple A15",
+ !listconcat(
+ !listremove(TuneAppleA14.Implies,
+ [FeatureAggressiveFMA]),
+ [FeatureFuseAdrpAdd])>;
def TuneAppleA16 : SubtargetFeature<"apple-a16", "ARMProcFamily", "AppleA16",
- "Apple A16", [
- FeatureAlternateSExtLoadCVTF32Pattern,
- FeatureArithmeticBccFusion,
- FeatureArithmeticCbzFusion,
- FeatureDisableLatencySchedHeuristic,
- FeatureFastLD1Single,
- FeatureFuseAddress,
- FeatureFuseAdrpAdd,
- FeatureFuseAES,
- FeatureFuseArithmeticLogic,
- FeatureFuseCmpCSel,
- FeatureFuseFCmpFCSel,
- FeatureFuseCryptoEOR,
- FeatureFuseLiterals,
- FeatureStorePairSuppress,
- FeatureZCRegMoveGPR64,
- FeatureZCRegMoveFPR128,
- FeatureZCZeroingGPR32,
- FeatureZCZeroingGPR64,
- FeatureNoZCZeroingFPR64,
- FeatureZCZeroingFPR128,
- FeatureMaxInterleaveFactor4]>;
+ "Apple A16",
+ !listconcat(TuneAppleA15.Implies,
+ [FeatureFuseFCmpFCSel])>;
def TuneAppleA17 : SubtargetFeature<"apple-a17", "ARMProcFamily", "AppleA17",
- "Apple A17", [
- FeatureAlternateSExtLoadCVTF32Pattern,
- FeatureArithmeticBccFusion,
- FeatureArithmeticCbzFusion,
- FeatureDisableLatencySchedHeuristic,
- FeatureFastLD1Single,
- FeatureFuseAddress,
- FeatureFuseAdrpAdd,
- FeatureFuseAES,
- FeatureFuseArithmeticLogic,
- FeatureFuseCmpCSel,
- FeatureFuseFCmpFCSel,
- FeatureFuseCryptoEOR,
- FeatureFuseLiterals,
- FeatureStorePairSuppress,
- FeatureZCRegMoveGPR64,
- FeatureZCRegMoveFPR128,
- FeatureZCZeroingGPR32,
- FeatureZCZeroingGPR64,
- FeatureNoZCZeroingFPR64,
- FeatureZCZeroingFPR128,
- FeatureMaxInterleaveFactor4]>;
+ "Apple A17", TuneAppleA16.Implies>;
def TuneAppleM4 : SubtargetFeature<"apple-m4", "ARMProcFamily", "AppleM4",
- "Apple M4", [
- FeatureAlternateSExtLoadCVTF32Pattern,
- FeatureArithmeticBccFusion,
- FeatureArithmeticCbzFusion,
- FeatureDisableLatencySchedHeuristic,
- FeatureFastLD1Single,
- FeatureFuseAddress,
- FeatureFuseAdrpAdd,
- FeatureFuseAES,
- FeatureFuseArithmeticLogic,
- FeatureFuseCmpCSel,
- FeatureFuseFCmpFCSel,
- FeatureFuseCryptoEOR,
- FeatureFuseLiterals,
- FeatureZCRegMoveGPR64,
- FeatureZCRegMoveFPR128,
- FeatureZCZeroingGPR32,
- FeatureZCZeroingGPR64,
- FeatureNoZCZeroingFPR64,
- FeatureZCZeroingFPR128,
- FeatureMaxInterleaveFactor4]>;
+ "Apple M4",
+ !listremove(TuneAppleA17.Implies,
+ [FeatureStorePairSuppress])>;
def TuneAppleM5 : SubtargetFeature<"apple-m5", "ARMProcFamily", "AppleM5",
- "Apple M5", [
- FeatureAlternateSExtLoadCVTF32Pattern,
- FeatureArithmeticBccFusion,
- FeatureArithmeticCbzFusion,
- FeatureDisableLatencySchedHeuristic,
- FeatureFastLD1Single,
- FeatureFuseAddress,
- FeatureFuseAdrpAdd,
- FeatureFuseAES,
- FeatureFuseArithmeticLogic,
- FeatureFuseCmpCSel,
- FeatureFuseFCmpFCSel,
- FeatureFuseCryptoEOR,
- FeatureFuseLiterals,
- FeatureZCRegMoveGPR64,
- FeatureZCRegMoveFPR128,
- FeatureZCZeroingGPR32,
- FeatureZCZeroingGPR64,
- FeatureNoZCZeroingFPR64,
- FeatureZCZeroingFPR128,
- FeatureMaxInterleaveFactor4]>;
+ "Apple M5", TuneAppleM4.Implies>;
def TuneExynosM3 : SubtargetFeature<"exynosm3", "ARMProcFamily", "ExynosM3",
"Samsung Exynos-M3 processors",
diff --git a/llvm/test/TableGen/aarch64-apple-tuning-features.td b/llvm/test/TableGen/aarch64-apple-tuning-features.td
new file mode 100644
index 0000000000000..66d4f3be5dcd4
--- /dev/null
+++ b/llvm/test/TableGen/aarch64-apple-tuning-features.td
@@ -0,0 +1,28 @@
+// RUN: llvm-tblgen -print-records %p/../../lib/Target/AArch64/AArch64.td -I %p/../../lib/Target/AArch64 -I %p/../../include | grep '^def TuneApple' -A5 | grep -E '(^def TuneApple|Implies)' | FileCheck %s
+
+// Verify the resolved tuning feature lists for Apple CPUs.
+// Each generation inherits from the previous and applies add/remove deltas.
+// When adding a tuning feature to an Apple core, update the expected list below.
+
+// CHECK-LABEL: def TuneAppleA10 {
+// CHECK-NEXT: list<SubtargetFeature> Implies = [FeatureAlternateSExtLoadCVTF32Pattern, FeatureArithmeticBccFusion, FeatureArithmeticCbzFusion, FeatureDisableLatencySchedHeuristic, FeatureFuseAES, FeatureFuseCryptoEOR, FeatureStorePairSuppress, FeatureZCRegMoveGPR64, FeatureZCRegMoveFPR128, FeatureZCZeroingGPR32, FeatureZCZeroingGPR64, FeatureNoZCZeroingFPR64, FeatureZCZeroingFPR128];
+// CHECK-LABEL: def TuneAppleA11 {
+// CHECK-NEXT: list<SubtargetFeature> Implies = [FeatureAlternateSExtLoadCVTF32Pattern, FeatureArithmeticBccFusion, FeatureArithmeticCbzFusion, FeatureDisableLatencySchedHeuristic, FeatureFuseAES, FeatureFuseCryptoEOR, FeatureStorePairSuppress, FeatureZCRegMoveGPR64, FeatureZCRegMoveFPR128, FeatureZCZeroingGPR32, FeatureZCZeroingGPR64, FeatureNoZCZeroingFPR64, FeatureZCZeroingFPR128];
+// CHECK-LABEL: def TuneAppleA12 {
+// CHECK-NEXT: list<SubtargetFeature> Implies = [FeatureAlternateSExtLoadCVTF32Pattern, FeatureArithmeticBccFusion, FeatureArithmeticCbzFusion, FeatureDisableLatencySchedHeuristic, FeatureFuseAES, FeatureFuseCryptoEOR, FeatureStorePairSuppress, FeatureZCRegMoveGPR64, FeatureZCRegMoveFPR128, FeatureZCZeroingGPR32, FeatureZCZeroingGPR64, FeatureNoZCZeroingFPR64, FeatureZCZeroingFPR128];
+// CHECK-LABEL: def TuneAppleA13 {
+// CHECK-NEXT: list<SubtargetFeature> Implies = [FeatureAlternateSExtLoadCVTF32Pattern, FeatureArithmeticBccFusion, FeatureArithmeticCbzFusion, FeatureDisableLatencySchedHeuristic, FeatureFuseAES, FeatureFuseCryptoEOR, FeatureStorePairSuppress, FeatureZCRegMoveGPR64, FeatureZCRegMoveFPR128, FeatureZCZeroingGPR32, FeatureZCZeroingGPR64, FeatureNoZCZeroingFPR64, FeatureZCZeroingFPR128];
+// CHECK-LABEL: def TuneAppleA14 {
+// CHECK-NEXT: list<SubtargetFeature> Implies = [FeatureAlternateSExtLoadCVTF32Pattern, FeatureArithmeticBccFusion, FeatureArithmeticCbzFusion, FeatureDisableLatencySchedHeuristic, FeatureFuseAES, FeatureFuseCryptoEOR, FeatureStorePairSuppress, FeatureZCRegMoveGPR64, FeatureZCRegMoveFPR128, FeatureZCZeroingGPR32, FeatureZCZeroingGPR64, FeatureNoZCZeroingFPR64, FeatureZCZeroingFPR128, FeatureAggressiveFMA, FeatureFastLD1Single, FeatureFuseAddress, FeatureFuseArithmeticLogic, FeatureFuseCmpCSel, FeatureFuseLiterals, FeatureMaxInterleaveFactor4];
+// CHECK-LABEL: def TuneAppleA15 {
+// CHECK-NEXT: list<SubtargetFeature> Implies = [FeatureAlternateSExtLoadCVTF32Pattern, FeatureArithmeticBccFusion, FeatureArithmeticCbzFusion, FeatureDisableLatencySchedHeuristic, FeatureFuseAES, FeatureFuseCryptoEOR, FeatureStorePairSuppress, FeatureZCRegMoveGPR64, FeatureZCRegMoveFPR128, FeatureZCZeroingGPR32, FeatureZCZeroingGPR64, FeatureNoZCZeroingFPR64, FeatureZCZeroingFPR128, FeatureFastLD1Single, FeatureFuseAddress, FeatureFuseArithmeticLogic, FeatureFuseCmpCSel, FeatureFuseLiterals, FeatureMaxInterleaveFactor4, FeatureFuseAdrpAdd];
+// CHECK-LABEL: def TuneAppleA16 {
+// CHECK-NEXT: list<SubtargetFeature> Implies = [FeatureAlternateSExtLoadCVTF32Pattern, FeatureArithmeticBccFusion, FeatureArithmeticCbzFusion, FeatureDisableLatencySchedHeuristic, FeatureFuseAES, FeatureFuseCryptoEOR, FeatureStorePairSuppress, FeatureZCRegMoveGPR64, FeatureZCRegMoveFPR128, FeatureZCZeroingGPR32, FeatureZCZeroingGPR64, FeatureNoZCZeroingFPR64, FeatureZCZeroingFPR128, FeatureFastLD1Single, FeatureFuseAddress, FeatureFuseArithmeticLogic, FeatureFuseCmpCSel, FeatureFuseLiterals, FeatureMaxInterleaveFactor4, FeatureFuseAdrpAdd, FeatureFuseFCmpFCSel];
+// CHECK-LABEL: def TuneAppleA17 {
+// CHECK-NEXT: list<SubtargetFeature> Implies = [FeatureAlternateSExtLoadCVTF32Pattern, FeatureArithmeticBccFusion, FeatureArithmeticCbzFusion, FeatureDisableLatencySchedHeuristic, FeatureFuseAES, FeatureFuseCryptoEOR, FeatureStorePairSuppress, FeatureZCRegMoveGPR64, FeatureZCRegMoveFPR128, FeatureZCZeroingGPR32, FeatureZCZeroingGPR64, FeatureNoZCZeroingFPR64, FeatureZCZeroingFPR128, FeatureFastLD1Single, FeatureFuseAddress, FeatureFuseArithmeticLogic, FeatureFuseCmpCSel, FeatureFuseLiterals, FeatureMaxInterleaveFactor4, FeatureFuseAdrpAdd, FeatureFuseFCmpFCSel];
+// CHECK-LABEL: def TuneAppleA7 {
+// CHECK-NEXT: list<SubtargetFeature> Implies = [FeatureAlternateSExtLoadCVTF32Pattern, FeatureArithmeticBccFusion, FeatureArithmeticCbzFusion, FeatureDisableLatencySchedHeuristic, FeatureFuseAES, FeatureFuseCryptoEOR, FeatureStorePairSuppress, FeatureZCRegMoveGPR64, FeatureZCRegMoveFPR128, FeatureZCZeroingFPWorkaround, FeatureZCZeroingGPR32, FeatureZCZeroingGPR64, FeatureNoZCZeroingFPR64, FeatureZCZeroingFPR128];
+// CHECK-LABEL: def TuneAppleM4 {
+// CHECK-NEXT: list<SubtargetFeature> Implies = [FeatureAlternateSExtLoadCVTF32Pattern, FeatureArithmeticBccFusion, FeatureArithmeticCbzFusion, FeatureDisableLatencySchedHeuristic, FeatureFuseAES, FeatureFuseCryptoEOR, FeatureZCRegMoveGPR64, FeatureZCRegMoveFPR128, FeatureZCZeroingGPR32, FeatureZCZeroingGPR64, FeatureNoZCZeroingFPR64, FeatureZCZeroingFPR128, FeatureFastLD1Single, FeatureFuseAddress, FeatureFuseArithmeticLogic, FeatureFuseCmpCSel, FeatureFuseLiterals, FeatureMaxInterleaveFactor4, FeatureFuseAdrpAdd, FeatureFuseFCmpFCSel];
+// CHECK-LABEL: def TuneAppleM5 {
+// CHECK-NEXT: list<SubtargetFeature> Implies = [FeatureAlternateSExtLoadCVTF32Pattern, FeatureArithmeticBccFusion, FeatureArithmeticCbzFusion, FeatureDisableLatencySchedHeuristic, FeatureFuseAES, FeatureFuseCryptoEOR, FeatureZCRegMoveGPR64, FeatureZCRegMoveFPR128, FeatureZCZeroingGPR32, FeatureZCZeroingGPR64, FeatureNoZCZeroingFPR64, FeatureZCZeroingFPR128, FeatureFastLD1Single, FeatureFuseAddress, FeatureFuseArithmeticLogic, FeatureFuseCmpCSel, FeatureFuseLiterals, FeatureMaxInterleaveFactor4, FeatureFuseAdrpAdd, FeatureFuseFCmpFCSel];
More information about the llvm-commits
mailing list