[llvm] [BOLT][AArch64] Reduce the number of ADR relaxations (PR #111577)
via llvm-commits
llvm-commits at lists.llvm.org
Tue Oct 8 12:16:14 PDT 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-bolt
Author: Maksim Panchenko (maksfb)
<details>
<summary>Changes</summary>
If ADR instruction references the same function, we can skip relaxation even if the function is split, but ADR is in the main fragment.
---
Full diff: https://github.com/llvm/llvm-project/pull/111577.diff
2 Files Affected:
- (modified) bolt/lib/Passes/ADRRelaxationPass.cpp (+4-3)
- (added) bolt/test/AArch64/adr-relaxation.s (+31)
``````````diff
diff --git a/bolt/lib/Passes/ADRRelaxationPass.cpp b/bolt/lib/Passes/ADRRelaxationPass.cpp
index 256034a841c706..52811edcb82731 100644
--- a/bolt/lib/Passes/ADRRelaxationPass.cpp
+++ b/bolt/lib/Passes/ADRRelaxationPass.cpp
@@ -56,13 +56,14 @@ void ADRRelaxationPass::runOnFunction(BinaryFunction &BF) {
continue;
}
- // Don't relax adr if it points to the same function and it is not split
- // and BF initial size is < 1MB.
+ // Don't relax ADR if it points to the same function and is in the main
+ // fragment and BF initial size is < 1MB.
const unsigned OneMB = 0x100000;
if (BF.getSize() < OneMB) {
BinaryFunction *TargetBF = BC.getFunctionForSymbol(Symbol);
- if (TargetBF == &BF && !BF.isSplit())
+ if (TargetBF == &BF && !BB.isSplit())
continue;
+
// No relaxation needed if ADR references a basic block in the same
// fragment.
if (BinaryBasicBlock *TargetBB = BF.getBasicBlockForLabel(Symbol))
diff --git a/bolt/test/AArch64/adr-relaxation.s b/bolt/test/AArch64/adr-relaxation.s
new file mode 100644
index 00000000000000..55b7d80edaab49
--- /dev/null
+++ b/bolt/test/AArch64/adr-relaxation.s
@@ -0,0 +1,31 @@
+## Check that llvm-bolt will unnecessarily relax ADR instruction.
+## ADR below references containing function that is split. But ADR is always
+## in the main fragment, thus there is no need to relax it.
+
+# RUN: llvm-mc -filetype=obj -triple aarch64-unknown-unknown %s -o %t.o
+# RUN: %clang %cflags %t.o -o %t.exe -Wl,-q -static
+# RUN: llvm-bolt %t.exe -o %t.bolt --split-functions --split-strategy=randomN \
+# RUN: 2>&1 | FileCheck %s
+# RUN: llvm-objdump -d --disassemble-symbols=_start %t.bolt | FileCheck %s
+
+# CHECK-NOT: adrp
+
+ .text
+ .globl _start
+ .type _start, %function
+_start:
+ .cfi_startproc
+ adr x1, _start
+ cmp x1, x11
+ b.hi .L1
+
+ mov x0, #0x0
+
+.L1:
+ ret x30
+
+ .cfi_endproc
+.size _start, .-_start
+
+## Force relocation mode.
+ .reloc 0, R_AARCH64_NONE
``````````
</details>
https://github.com/llvm/llvm-project/pull/111577
More information about the llvm-commits
mailing list