[llvm-branch-commits] [llvm] [LoongArch] Fix misaligned frame-index offset for vldrepl.{h, w, d} (PR #239516)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Fri Nov 6 00:11:45 PST 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-loongarch
Author: hev (heiher)
<details>
<summary>Changes</summary>
eliminateFrameIndex did not check that the final offset is a multiple of the access size for [X]VLDREPL_{H,W,D}, so a byte-aligned stack object could produce an unencodable immediate. Materialize the address in a scratch register in that case, as is already done for [X]VSTELM_*.
Assisted-by: LLM
Fixes #<!-- -->239508
---
Full diff: https://github.com/llvm/llvm-project/pull/239516.diff
3 Files Affected:
- (modified) llvm/lib/Target/LoongArch/LoongArchRegisterInfo.cpp (+12)
- (modified) llvm/test/CodeGen/LoongArch/lasx/xvldrepl-frame-index.ll (+6-3)
- (modified) llvm/test/CodeGen/LoongArch/lsx/vldrepl-frame-index.ll (+6-3)
``````````diff
diff --git a/llvm/lib/Target/LoongArch/LoongArchRegisterInfo.cpp b/llvm/lib/Target/LoongArch/LoongArchRegisterInfo.cpp
index 682a147d94678..392cbd0d0f0ab 100644
--- a/llvm/lib/Target/LoongArch/LoongArchRegisterInfo.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchRegisterInfo.cpp
@@ -202,6 +202,18 @@ bool LoongArchRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
// Handle offsets that exceed the immediate range of the instruction.
switch (MIOpc) {
+ case LoongArch::VLDREPL_H:
+ case LoongArch::XVLDREPL_H:
+ OffsetLegal = isShiftedInt<11, 1>(FixedOffset);
+ break;
+ case LoongArch::VLDREPL_W:
+ case LoongArch::XVLDREPL_W:
+ OffsetLegal = isShiftedInt<10, 2>(FixedOffset);
+ break;
+ case LoongArch::VLDREPL_D:
+ case LoongArch::XVLDREPL_D:
+ OffsetLegal = isShiftedInt<9, 3>(FixedOffset);
+ break;
case LoongArch::VSTELM_B:
case LoongArch::XVSTELM_B:
OffsetLegal = isInt<8>(FixedOffset);
diff --git a/llvm/test/CodeGen/LoongArch/lasx/xvldrepl-frame-index.ll b/llvm/test/CodeGen/LoongArch/lasx/xvldrepl-frame-index.ll
index c8bbb89d8cb5b..efbc7a495708b 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/xvldrepl-frame-index.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/xvldrepl-frame-index.ll
@@ -13,7 +13,8 @@ define <16 x i16> @repl_h() noinline nounwind "frame-pointer"="all" {
; CHECK-NEXT: addi.d $a0, $fp, -89
; CHECK-NEXT: pcaddu18i $ra, %call36(use)
; CHECK-NEXT: jirl $ra, $ra, 0
-; CHECK-NEXT: xvldrepl.h $xr0, $fp, -73
+; CHECK-NEXT: addi.d $a0, $fp, -73
+; CHECK-NEXT: xvldrepl.h $xr0, $a0, 0
; CHECK-NEXT: ld.d $fp, $sp, 80 # 8-byte Folded Reload
; CHECK-NEXT: ld.d $ra, $sp, 88 # 8-byte Folded Reload
; CHECK-NEXT: addi.d $sp, $sp, 96
@@ -37,7 +38,8 @@ define <8 x i32> @repl_w() noinline nounwind "frame-pointer"="all" {
; CHECK-NEXT: addi.d $a0, $fp, -90
; CHECK-NEXT: pcaddu18i $ra, %call36(use)
; CHECK-NEXT: jirl $ra, $ra, 0
-; CHECK-NEXT: xvldrepl.w $xr0, $fp, -74
+; CHECK-NEXT: addi.d $a0, $fp, -74
+; CHECK-NEXT: xvldrepl.w $xr0, $a0, 0
; CHECK-NEXT: ld.d $fp, $sp, 80 # 8-byte Folded Reload
; CHECK-NEXT: ld.d $ra, $sp, 88 # 8-byte Folded Reload
; CHECK-NEXT: addi.d $sp, $sp, 96
@@ -61,7 +63,8 @@ define <4 x i64> @repl_d() noinline nounwind "frame-pointer"="all" {
; CHECK-NEXT: addi.d $a0, $fp, -91
; CHECK-NEXT: pcaddu18i $ra, %call36(use)
; CHECK-NEXT: jirl $ra, $ra, 0
-; CHECK-NEXT: xvldrepl.d $xr0, $fp, -75
+; CHECK-NEXT: addi.d $a0, $fp, -75
+; CHECK-NEXT: xvldrepl.d $xr0, $a0, 0
; CHECK-NEXT: ld.d $fp, $sp, 80 # 8-byte Folded Reload
; CHECK-NEXT: ld.d $ra, $sp, 88 # 8-byte Folded Reload
; CHECK-NEXT: addi.d $sp, $sp, 96
diff --git a/llvm/test/CodeGen/LoongArch/lsx/vldrepl-frame-index.ll b/llvm/test/CodeGen/LoongArch/lsx/vldrepl-frame-index.ll
index cc49fb96550be..d4aecb0ffc95d 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/vldrepl-frame-index.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/vldrepl-frame-index.ll
@@ -13,7 +13,8 @@ define <8 x i16> @repl_h() noinline nounwind "frame-pointer"="all" {
; CHECK-NEXT: addi.d $a0, $fp, -89
; CHECK-NEXT: pcaddu18i $ra, %call36(use)
; CHECK-NEXT: jirl $ra, $ra, 0
-; CHECK-NEXT: vldrepl.h $vr0, $fp, -73
+; CHECK-NEXT: addi.d $a0, $fp, -73
+; CHECK-NEXT: vldrepl.h $vr0, $a0, 0
; CHECK-NEXT: ld.d $fp, $sp, 80 # 8-byte Folded Reload
; CHECK-NEXT: ld.d $ra, $sp, 88 # 8-byte Folded Reload
; CHECK-NEXT: addi.d $sp, $sp, 96
@@ -37,7 +38,8 @@ define <4 x i32> @repl_w() noinline nounwind "frame-pointer"="all" {
; CHECK-NEXT: addi.d $a0, $fp, -90
; CHECK-NEXT: pcaddu18i $ra, %call36(use)
; CHECK-NEXT: jirl $ra, $ra, 0
-; CHECK-NEXT: vldrepl.w $vr0, $fp, -74
+; CHECK-NEXT: addi.d $a0, $fp, -74
+; CHECK-NEXT: vldrepl.w $vr0, $a0, 0
; CHECK-NEXT: ld.d $fp, $sp, 80 # 8-byte Folded Reload
; CHECK-NEXT: ld.d $ra, $sp, 88 # 8-byte Folded Reload
; CHECK-NEXT: addi.d $sp, $sp, 96
@@ -61,7 +63,8 @@ define <2 x i64> @repl_d() noinline nounwind "frame-pointer"="all" {
; CHECK-NEXT: addi.d $a0, $fp, -91
; CHECK-NEXT: pcaddu18i $ra, %call36(use)
; CHECK-NEXT: jirl $ra, $ra, 0
-; CHECK-NEXT: vldrepl.d $vr0, $fp, -75
+; CHECK-NEXT: addi.d $a0, $fp, -75
+; CHECK-NEXT: vldrepl.d $vr0, $a0, 0
; CHECK-NEXT: ld.d $fp, $sp, 80 # 8-byte Folded Reload
; CHECK-NEXT: ld.d $ra, $sp, 88 # 8-byte Folded Reload
; CHECK-NEXT: addi.d $sp, $sp, 96
``````````
</details>
https://github.com/llvm/llvm-project/pull/239516
More information about the llvm-branch-commits
mailing list