[llvm] [AArch64] Correct FDIV/FSQRT throughput for Neoverse V3ae. (PR #219113)
David Green via llvm-commits
llvm-commits at lists.llvm.org
Wed Aug 26 22:12:32 PDT 2026
https://github.com/davemgreen created https://github.com/llvm/llvm-project/pull/219113
This attempts to improve the throughput of fdiv / fsqrt to the correct throughputs of 1 per cycle. I've moved the SchedWriteRes with multiple ReleaseAtCycles to the correct section / naming whilst here. Which has the result of changing the fdiv and fsqrt throughput to 1.
>From 9324a261bab7b0267b78a1562d30b9918291e17b Mon Sep 17 00:00:00 2001
From: David Green <david.green at arm.com>
Date: Thu, 27 Aug 2026 05:38:48 +0100
Subject: [PATCH] [AArch64] Correct FDIV/FSQRT throughput for Neoverse V3ae.
This attempts to improve the throughput of fdiv / fsqrt to the correct
throughputs of 1 per cycle. I've moved the SchedWriteRes with multiple
ReleaseAtCycles to the correct section / naming whilst here. Which has the
result of changing the fdiv and fsqrt throughput to 1.
---
.../AArch64/AArch64SchedNeoverseV3AE.td | 43 ++++++++-----------
.../Neoverse/V3AE-basic-instructions.s | 10 ++---
2 files changed, 24 insertions(+), 29 deletions(-)
diff --git a/llvm/lib/Target/AArch64/AArch64SchedNeoverseV3AE.td b/llvm/lib/Target/AArch64/AArch64SchedNeoverseV3AE.td
index 29f8fa783467a..1f4f0206be908 100644
--- a/llvm/lib/Target/AArch64/AArch64SchedNeoverseV3AE.td
+++ b/llvm/lib/Target/AArch64/AArch64SchedNeoverseV3AE.td
@@ -106,10 +106,6 @@ def V3AEWrite_3c_1M : SchedWriteRes<[V3AEUnitM]> { let Latency = 3; }
def V3AEWrite_2c_1M0 : SchedWriteRes<[V3AEUnitM0]> { let Latency = 2; }
def V3AEWrite_3c_1M0 : SchedWriteRes<[V3AEUnitM0]> { let Latency = 3; }
def V3AEWrite_4c_1M0 : SchedWriteRes<[V3AEUnitM0]> { let Latency = 4; }
-def V3AEWrite_12c_1M0 : SchedWriteRes<[V3AEUnitM0]> { let Latency = 12;
- let ReleaseAtCycles = [12]; }
-def V3AEWrite_20c_1M0 : SchedWriteRes<[V3AEUnitM0]> { let Latency = 20;
- let ReleaseAtCycles = [20]; }
def V3AEWrite_4c_1L : SchedWriteRes<[V3AEUnitL]> { let Latency = 4; }
def V3AEWrite_6c_1L : SchedWriteRes<[V3AEUnitL]> { let Latency = 6; }
def V3AEWrite_2c_1V : SchedWriteRes<[V3AEUnitV]> { let Latency = 2; }
@@ -123,16 +119,11 @@ def V3AEWrite_3c_1V0 : SchedWriteRes<[V3AEUnitV0]> { let Latency = 3; }
def V3AEWrite_4c_1V0 : SchedWriteRes<[V3AEUnitV0]> { let Latency = 4; }
def V3AEWrite_9c_1V0 : SchedWriteRes<[V3AEUnitV0]> { let Latency = 9; }
def V3AEWrite_10c_1V0 : SchedWriteRes<[V3AEUnitV0]> { let Latency = 10; }
-def V3AEWrite_8c_1V1 : SchedWriteRes<[V3AEUnitV1]> { let Latency = 8; }
-def V3AEWrite_12c_1V0 : SchedWriteRes<[V3AEUnitV0]> { let Latency = 12;
- let ReleaseAtCycles = [11]; }
+def V3AEWrite_8c_1V1 : SchedWriteRes<[V3AEUnitV1]> { let Latency = 8; }
def V3AEWrite_13c_1V0 : SchedWriteRes<[V3AEUnitV0]> { let Latency = 13; }
def V3AEWrite_15c_1V0 : SchedWriteRes<[V3AEUnitV0]> { let Latency = 15; }
-def V3AEWrite_13c_1V1 : SchedWriteRes<[V3AEUnitV1]> { let Latency = 13;
- let ReleaseAtCycles = [8]; }
+def V3AEWrite_13c_1V1 : SchedWriteRes<[V3AEUnitV1]> { let Latency = 13; }
def V3AEWrite_16c_1V0 : SchedWriteRes<[V3AEUnitV0]> { let Latency = 16; }
-def V3AEWrite_20c_1V0 : SchedWriteRes<[V3AEUnitV0]> { let Latency = 20;
- let ReleaseAtCycles = [20]; }
def V3AEWrite_2c_1V1 : SchedWriteRes<[V3AEUnitV1]> { let Latency = 2; }
def V3AEWrite_3c_1V1 : SchedWriteRes<[V3AEUnitV1]> { let Latency = 3; }
def V3AEWrite_4c_1V1 : SchedWriteRes<[V3AEUnitV1]> { let Latency = 4; }
@@ -942,13 +933,17 @@ def V3AERd_ZBFMAL : SchedReadAdvance<3, [V3AEWr_ZBFMAL]>;
//===----------------------------------------------------------------------===//
// Define types with long resource cycles (rc)
-def V3AEWrite_6c_1V1_5rc : SchedWriteRes<[V3AEUnitV1]> { let Latency = 6; let ReleaseAtCycles = [ 5]; }
-def V3AEWrite_9c_1V1_2rc : SchedWriteRes<[V3AEUnitV1]> { let Latency = 9; let ReleaseAtCycles = [ 2]; }
-def V3AEWrite_9c_1V1_4rc : SchedWriteRes<[V3AEUnitV1]> { let Latency = 9; let ReleaseAtCycles = [ 4]; }
-def V3AEWrite_10c_1V1_9rc : SchedWriteRes<[V3AEUnitV1]> { let Latency = 10; let ReleaseAtCycles = [ 9]; }
-def V3AEWrite_11c_1V1_4rc : SchedWriteRes<[V3AEUnitV1]> { let Latency = 11; let ReleaseAtCycles = [ 4]; }
-def V3AEWrite_13c_1V1_8rc : SchedWriteRes<[V3AEUnitV1]> { let Latency = 13; let ReleaseAtCycles = [8]; }
-def V3AEWrite_14c_1V1_2rc : SchedWriteRes<[V3AEUnitV1]> { let Latency = 14; let ReleaseAtCycles = [2]; }
+def V3AEWrite_6c_1V1_5rc : SchedWriteRes<[V3AEUnitV1]> { let Latency = 6; let ReleaseAtCycles = [5]; }
+def V3AEWrite_9c_1V1_2rc : SchedWriteRes<[V3AEUnitV1]> { let Latency = 9; let ReleaseAtCycles = [2]; }
+def V3AEWrite_9c_1V1_4rc : SchedWriteRes<[V3AEUnitV1]> { let Latency = 9; let ReleaseAtCycles = [4]; }
+def V3AEWrite_10c_1V1_9rc : SchedWriteRes<[V3AEUnitV1]> { let Latency = 10; let ReleaseAtCycles = [9]; }
+def V3AEWrite_11c_1V1_4rc : SchedWriteRes<[V3AEUnitV1]> { let Latency = 11; let ReleaseAtCycles = [4]; }
+def V3AEWrite_12c_1M0_12rc : SchedWriteRes<[V3AEUnitM0]> { let Latency = 12; let ReleaseAtCycles = [12]; }
+def V3AEWrite_12c_1V0_11rc : SchedWriteRes<[V3AEUnitV0]> { let Latency = 12; let ReleaseAtCycles = [11]; }
+def V3AEWrite_13c_1V1_8rc : SchedWriteRes<[V3AEUnitV1]> { let Latency = 13; let ReleaseAtCycles = [8]; }
+def V3AEWrite_14c_1V1_2rc : SchedWriteRes<[V3AEUnitV1]> { let Latency = 14; let ReleaseAtCycles = [2]; }
+def V3AEWrite_20c_1V0_20rc : SchedWriteRes<[V3AEUnitV0]> { let Latency = 20; let ReleaseAtCycles = [20]; }
+def V3AEWrite_20c_1M0_20rc : SchedWriteRes<[V3AEUnitM0]> { let Latency = 20; let ReleaseAtCycles = [20]; }
// Miscellaneous
// -----------------------------------------------------------------------------
@@ -1029,8 +1024,8 @@ def : SchedAlias<WriteImm, V3AEWrite_1c_1I>;
// -----------------------------------------------------------------------------
// SDIV, UDIV
-def : SchedAlias<WriteID32, V3AEWrite_12c_1M0>;
-def : SchedAlias<WriteID64, V3AEWrite_20c_1M0>;
+def : SchedAlias<WriteID32, V3AEWrite_12c_1M0_12rc>;
+def : SchedAlias<WriteID64, V3AEWrite_20c_1M0_20rc>;
def : SchedAlias<WriteIM32, V3AEWrite_2c_1M>;
def : SchedAlias<WriteIM64, V3AEWrite_2c_1M>;
@@ -2103,12 +2098,12 @@ def : InstRW<[V3AEWrite_2c_1V], (instregex "^CPY_ZPm[IV]_[BHSD]",
"^CPY_ZPzI_[BHSD]")>;
// Divides, 32 bit
-def : InstRW<[V3AEWrite_12c_1V0], (instregex "^[SU]DIVR?_ZPmZ_S",
- "^[SU]DIV_ZPZZ_S")>;
+def : InstRW<[V3AEWrite_12c_1V0_11rc], (instregex "^[SU]DIVR?_ZPmZ_S",
+ "^[SU]DIV_ZPZZ_S")>;
// Divides, 64 bit
-def : InstRW<[V3AEWrite_20c_1V0], (instregex "^[SU]DIVR?_ZPmZ_D",
- "^[SU]DIV_ZPZZ_D")>;
+def : InstRW<[V3AEWrite_20c_1V0_20rc], (instregex "^[SU]DIVR?_ZPmZ_D",
+ "^[SU]DIV_ZPZZ_D")>;
// Dot product, 8 bit
def : InstRW<[V3AEWr_ZDOTB, V3AERd_ZDOTB], (instregex "^[SU]DOT_ZZZI?_BtoS")>;
diff --git a/llvm/test/tools/llvm-mca/AArch64/Neoverse/V3AE-basic-instructions.s b/llvm/test/tools/llvm-mca/AArch64/Neoverse/V3AE-basic-instructions.s
index b9c64a3ddca8f..c9eb6f6b85b1c 100644
--- a/llvm/test/tools/llvm-mca/AArch64/Neoverse/V3AE-basic-instructions.s
+++ b/llvm/test/tools/llvm-mca/AArch64/Neoverse/V3AE-basic-instructions.s
@@ -560,7 +560,7 @@
# CHECK-NEXT: 1 2 0.50 fabs d2, d3
# CHECK-NEXT: 1 2 0.50 fneg d4, d5
# CHECK-NEXT: 1 6 1.00 fsqrt h13, h24
-# CHECK-NEXT: 1 13 8.00 fsqrt d6, d7
+# CHECK-NEXT: 1 13 1.00 fsqrt d6, d7
# CHECK-NEXT: 1 3 0.50 fcvt s8, d9
# CHECK-NEXT: 1 3 0.50 fcvt h10, d11
# CHECK-NEXT: 1 3 1.00 frintn d12, d13
@@ -588,7 +588,7 @@
# CHECK-NEXT: 1 3 0.50 fnmul h3, h15, h7
# CHECK-NEXT: 1 3 0.50 fnmul s22, s23, s2
# CHECK-NEXT: 1 3 0.50 fmul d20, d19, d17
-# CHECK-NEXT: 1 13 8.00 fdiv d1, d2, d3
+# CHECK-NEXT: 1 13 1.00 fdiv d1, d2, d3
# CHECK-NEXT: 1 2 0.50 fadd d4, d5, d6
# CHECK-NEXT: 1 2 0.50 fsub d7, d8, d9
# CHECK-NEXT: 1 2 0.50 fmax d10, d11, d12
@@ -1278,7 +1278,7 @@
# CHECK: Resource pressure per iteration:
# CHECK-NEXT: [0.0] [0.1] [0.2] [1.0] [1.1] [2.0] [2.1] [2.2] [2.3] [3.0] [3.1] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15]
-# CHECK-NEXT: 8.67 8.67 8.67 34.00 34.00 40.75 40.75 40.75 40.75 99.33 99.33 170.33 290.63 163.63 79.96 79.96 79.96 79.96 79.96 79.96 71.00 219.50 96.50
+# CHECK-NEXT: 8.67 8.67 8.67 34.00 34.00 40.75 40.75 40.75 40.75 99.33 99.33 170.33 290.63 163.63 79.96 79.96 79.96 79.96 79.96 79.96 71.00 219.50 82.50
# CHECK: Resource pressure by instruction:
# CHECK-NEXT: [0.0] [0.1] [0.2] [1.0] [1.1] [2.0] [2.1] [2.2] [2.3] [3.0] [3.1] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] Instructions:
@@ -1832,7 +1832,7 @@
# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - 0.50 0.50 fabs d2, d3
# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - 0.50 0.50 fneg d4, d5
# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - - 1.00 fsqrt h13, h24
-# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - - 8.00 fsqrt d6, d7
+# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - - 1.00 fsqrt d6, d7
# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - 0.50 0.50 fcvt s8, d9
# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - 0.50 0.50 fcvt h10, d11
# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - 1.00 - frintn d12, d13
@@ -1860,7 +1860,7 @@
# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - 0.50 0.50 fnmul h3, h15, h7
# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - 0.50 0.50 fnmul s22, s23, s2
# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - 0.50 0.50 fmul d20, d19, d17
-# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - - 8.00 fdiv d1, d2, d3
+# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - - 1.00 fdiv d1, d2, d3
# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - 0.50 0.50 fadd d4, d5, d6
# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - 0.50 0.50 fsub d7, d8, d9
# CHECK-NEXT: - - - - - - - - - - - - - - - - - - - - - 0.50 0.50 fmax d10, d11, d12
More information about the llvm-commits
mailing list