[llvm-branch-commits] [llvm] [LoongArch] Override `isLSRCostLess` to set `Insns` as the first priority (PR #168035)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Fri Nov 13 01:47:00 PST 2026
https://github.com/zhaoqi5 updated https://github.com/llvm/llvm-project/pull/168035
>From 1beaed1ec4a6d8209053dcb5d6df28352aabc681 Mon Sep 17 00:00:00 2001
From: Qi Zhao <zhaoqi01 at loongson.cn>
Date: Fri, 14 Nov 2025 16:33:31 +0800
Subject: [PATCH 1/2] [LoongArch] Override `isLSRCostLess` to set `Insns` as
the first priority
Similar to several other targets, this commit override
`isLSRCostLess` to set instruction number as the first priority
when LSR pass deciding the cost.
Besides, this commit also takes the extra temporary register
may be used into account in `NumRegs`. This is same as riscv,
see the reason in https://github.com/llvm/llvm-project/pull/92296.
---
.../LoongArch/LoongArchTargetTransformInfo.cpp | 14 ++++++++++++++
.../LoongArch/LoongArchTargetTransformInfo.h | 3 +++
2 files changed, 17 insertions(+)
diff --git a/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.cpp b/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.cpp
index ec1f33ae2c523..9a839eb122a8d 100644
--- a/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.cpp
@@ -506,3 +506,17 @@ InstructionCost LoongArchTTIImpl::getShuffleCost(
return BaseT::getShuffleCost(Kind, DstTy, SrcTy, CostKind, Mask, Index, SubTp,
Args, CxtI, VIC);
}
+
+bool LoongArchTTIImpl::isLSRCostLess(
+ const TargetTransformInfo::LSRCost &C1,
+ const TargetTransformInfo::LSRCost &C2) const {
+ // LoongArch specific here are "instruction number 1st priority".
+ // If we need to emit adds inside the loop to add up base registers, then
+ // we need at least one extra temporary register.
+ unsigned C1NumRegs = C1.NumRegs + (C1.NumBaseAdds != 0);
+ unsigned C2NumRegs = C2.NumRegs + (C2.NumBaseAdds != 0);
+ return std::tie(C1.Insns, C1NumRegs, C1.AddRecCost, C1.NumIVMuls,
+ C1.NumBaseAdds, C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
+ std::tie(C2.Insns, C2NumRegs, C2.AddRecCost, C2.NumIVMuls,
+ C2.NumBaseAdds, C2.ScaleCost, C2.ImmCost, C2.SetupCost);
+}
diff --git a/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.h b/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.h
index a61a339a1fe7d..ee963c0d57716 100644
--- a/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.h
+++ b/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.h
@@ -93,6 +93,9 @@ class LoongArchTTIImpl : public BasicTTIImplBase<LoongArchTTIImpl> {
std::optional<FastMathFlags> FMF) const override {
return InstructionCost::getInvalid();
}
+
+ bool isLSRCostLess(const TargetTransformInfo::LSRCost &C1,
+ const TargetTransformInfo::LSRCost &C2) const override;
};
} // end namespace llvm
>From 528282eadd1baec27f540c6723a7ea29f7f83a80 Mon Sep 17 00:00:00 2001
From: Qi Zhao <zhaoqi01 at loongson.cn>
Date: Fri, 13 Nov 2026 16:57:32 +0800
Subject: [PATCH 2/2] update after rebase
---
.../LoopStrengthReduce/LoongArch/lsr-insns.ll | 65 +++++++++----------
1 file changed, 30 insertions(+), 35 deletions(-)
diff --git a/llvm/test/Transforms/LoopStrengthReduce/LoongArch/lsr-insns.ll b/llvm/test/Transforms/LoopStrengthReduce/LoongArch/lsr-insns.ll
index 1cb47afb9103a..a09a9a37034e1 100644
--- a/llvm/test/Transforms/LoopStrengthReduce/LoongArch/lsr-insns.ll
+++ b/llvm/test/Transforms/LoopStrengthReduce/LoongArch/lsr-insns.ll
@@ -5,19 +5,18 @@
define void @foo(ptr nocapture readonly %x, ptr nocapture readonly %y, ptr nocapture %q) nounwind {
; CHECK-LLC-LABEL: foo:
; CHECK-LLC: # %bb.0: # %entry
-; CHECK-LLC-NEXT: lu12i.w $a3, -1
+; CHECK-LLC-NEXT: move $a3, $zero
+; CHECK-LLC-NEXT: ori $a4, $zero, 1024
; CHECK-LLC-NEXT: .p2align 4, , 16
; CHECK-LLC-NEXT: .LBB0_1: # %for.body
; CHECK-LLC-NEXT: # =>This Inner Loop Header: Depth=1
-; CHECK-LLC-NEXT: add.d $a4, $a0, $a3
-; CHECK-LLC-NEXT: ldptr.w $a4, $a4, 4096
-; CHECK-LLC-NEXT: add.d $a5, $a1, $a3
-; CHECK-LLC-NEXT: ldptr.w $a5, $a5, 4096
-; CHECK-LLC-NEXT: add.d $a4, $a5, $a4
-; CHECK-LLC-NEXT: add.d $a5, $a2, $a3
+; CHECK-LLC-NEXT: ldx.w $a5, $a0, $a3
+; CHECK-LLC-NEXT: ldx.w $a6, $a1, $a3
+; CHECK-LLC-NEXT: add.d $a5, $a6, $a5
+; CHECK-LLC-NEXT: stx.w $a5, $a2, $a3
+; CHECK-LLC-NEXT: addi.d $a4, $a4, -1
; CHECK-LLC-NEXT: addi.d $a3, $a3, 4
-; CHECK-LLC-NEXT: stptr.w $a4, $a5, 4096
-; CHECK-LLC-NEXT: bnez $a3, .LBB0_1
+; CHECK-LLC-NEXT: bnez $a4, .LBB0_1
; CHECK-LLC-NEXT: # %bb.2: # %for.cond.cleanup
; CHECK-LLC-NEXT: ret
; CHECK-OPT-LABEL: define void @foo(
@@ -27,18 +26,17 @@ define void @foo(ptr nocapture readonly %x, ptr nocapture readonly %y, ptr nocap
; CHECK-OPT: [[FOR_COND_CLEANUP:.*]]:
; CHECK-OPT-NEXT: ret void
; CHECK-OPT: [[FOR_BODY]]:
-; CHECK-OPT-NEXT: [[LSR_IV:%.*]] = phi i64 [ [[LSR_IV_NEXT:%.*]], %[[FOR_BODY]] ], [ -4096, %[[ENTRY]] ]
+; CHECK-OPT-NEXT: [[LSR_IV:%.*]] = phi i64 [ [[LSR_IV_NEXT2:%.*]], %[[FOR_BODY]] ], [ 0, %[[ENTRY]] ]
+; CHECK-OPT-NEXT: [[LSR_IV1:%.*]] = phi i64 [ [[LSR_IV_NEXT:%.*]], %[[FOR_BODY]] ], [ 1024, %[[ENTRY]] ]
; CHECK-OPT-NEXT: [[SCEVGEP4:%.*]] = getelementptr i8, ptr [[X]], i64 [[LSR_IV]]
-; CHECK-OPT-NEXT: [[SCEVGEP5:%.*]] = getelementptr i8, ptr [[SCEVGEP4]], i64 4096
-; CHECK-OPT-NEXT: [[LDTMP:%.*]] = load i32, ptr [[SCEVGEP5]], align 4
+; CHECK-OPT-NEXT: [[LDTMP:%.*]] = load i32, ptr [[SCEVGEP4]], align 4
; CHECK-OPT-NEXT: [[SCEVGEP2:%.*]] = getelementptr i8, ptr [[Y]], i64 [[LSR_IV]]
-; CHECK-OPT-NEXT: [[SCEVGEP3:%.*]] = getelementptr i8, ptr [[SCEVGEP2]], i64 4096
-; CHECK-OPT-NEXT: [[LDTMP1:%.*]] = load i32, ptr [[SCEVGEP3]], align 4
+; CHECK-OPT-NEXT: [[LDTMP1:%.*]] = load i32, ptr [[SCEVGEP2]], align 4
; CHECK-OPT-NEXT: [[ADD:%.*]] = add nsw i32 [[LDTMP1]], [[LDTMP]]
; CHECK-OPT-NEXT: [[SCEVGEP:%.*]] = getelementptr i8, ptr [[Q]], i64 [[LSR_IV]]
-; CHECK-OPT-NEXT: [[SCEVGEP1:%.*]] = getelementptr i8, ptr [[SCEVGEP]], i64 4096
-; CHECK-OPT-NEXT: store i32 [[ADD]], ptr [[SCEVGEP1]], align 4
-; CHECK-OPT-NEXT: [[LSR_IV_NEXT]] = add nsw i64 [[LSR_IV]], 4
+; CHECK-OPT-NEXT: store i32 [[ADD]], ptr [[SCEVGEP]], align 4
+; CHECK-OPT-NEXT: [[LSR_IV_NEXT]] = add nsw i64 [[LSR_IV1]], -1
+; CHECK-OPT-NEXT: [[LSR_IV_NEXT2]] = add nuw nsw i64 [[LSR_IV]], 4
; CHECK-OPT-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[LSR_IV_NEXT]], 0
; CHECK-OPT-NEXT: br i1 [[EXITCOND]], label %[[FOR_COND_CLEANUP]], label %[[FOR_BODY]]
;
@@ -68,19 +66,18 @@ define void @bar(ptr nocapture readonly %x, ptr nocapture readonly %y, ptr nocap
; CHECK-LLC-NEXT: addi.w $a4, $a3, 0
; CHECK-LLC-NEXT: blez $a4, .LBB1_3
; CHECK-LLC-NEXT: # %bb.1: # %for.body.preheader
+; CHECK-LLC-NEXT: move $a4, $zero
; CHECK-LLC-NEXT: bstrpick.d $a3, $a3, 31, 0
+; CHECK-LLC-NEXT: slli.d $a3, $a3, 2
; CHECK-LLC-NEXT: .p2align 4, , 16
; CHECK-LLC-NEXT: .LBB1_2: # %for.body
; CHECK-LLC-NEXT: # =>This Inner Loop Header: Depth=1
-; CHECK-LLC-NEXT: ld.w $a4, $a0, 0
-; CHECK-LLC-NEXT: ld.w $a5, $a1, 0
-; CHECK-LLC-NEXT: add.d $a4, $a5, $a4
-; CHECK-LLC-NEXT: st.w $a4, $a2, 0
-; CHECK-LLC-NEXT: addi.d $a3, $a3, -1
-; CHECK-LLC-NEXT: addi.d $a2, $a2, 4
-; CHECK-LLC-NEXT: addi.d $a1, $a1, 4
-; CHECK-LLC-NEXT: addi.d $a0, $a0, 4
-; CHECK-LLC-NEXT: bnez $a3, .LBB1_2
+; CHECK-LLC-NEXT: ldx.w $a5, $a0, $a4
+; CHECK-LLC-NEXT: ldx.w $a6, $a1, $a4
+; CHECK-LLC-NEXT: add.d $a5, $a6, $a5
+; CHECK-LLC-NEXT: stx.w $a5, $a2, $a4
+; CHECK-LLC-NEXT: addi.d $a4, $a4, 4
+; CHECK-LLC-NEXT: bne $a3, $a4, .LBB1_2
; CHECK-LLC-NEXT: .LBB1_3: # %for.cond.cleanup
; CHECK-LLC-NEXT: ret
; CHECK-OPT-LABEL: define void @bar(
@@ -90,25 +87,23 @@ define void @bar(ptr nocapture readonly %x, ptr nocapture readonly %y, ptr nocap
; CHECK-OPT-NEXT: br i1 [[CMP10]], label %[[FOR_BODY_PREHEADER:.*]], label %[[FOR_COND_CLEANUP:.*]]
; CHECK-OPT: [[FOR_BODY_PREHEADER]]:
; CHECK-OPT-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N]] to i64
+; CHECK-OPT-NEXT: [[TMP0:%.*]] = shl nuw nsw i64 [[WIDE_TRIP_COUNT]], 2
; CHECK-OPT-NEXT: br label %[[FOR_BODY:.*]]
; CHECK-OPT: [[FOR_COND_CLEANUP_LOOPEXIT:.*]]:
; CHECK-OPT-NEXT: br label %[[FOR_COND_CLEANUP]]
; CHECK-OPT: [[FOR_COND_CLEANUP]]:
; CHECK-OPT-NEXT: ret void
; CHECK-OPT: [[FOR_BODY]]:
-; CHECK-OPT-NEXT: [[LSR_IV4:%.*]] = phi ptr [ [[SCEVGEP5:%.*]], %[[FOR_BODY]] ], [ [[X]], %[[FOR_BODY_PREHEADER]] ]
-; CHECK-OPT-NEXT: [[LSR_IV2:%.*]] = phi ptr [ [[SCEVGEP3:%.*]], %[[FOR_BODY]] ], [ [[Y]], %[[FOR_BODY_PREHEADER]] ]
-; CHECK-OPT-NEXT: [[LSR_IV1:%.*]] = phi ptr [ [[SCEVGEP:%.*]], %[[FOR_BODY]] ], [ [[Q]], %[[FOR_BODY_PREHEADER]] ]
-; CHECK-OPT-NEXT: [[LSR_IV:%.*]] = phi i64 [ [[LSR_IV_NEXT:%.*]], %[[FOR_BODY]] ], [ [[WIDE_TRIP_COUNT]], %[[FOR_BODY_PREHEADER]] ]
+; CHECK-OPT-NEXT: [[LSR_IV:%.*]] = phi i64 [ [[LSR_IV_NEXT:%.*]], %[[FOR_BODY]] ], [ 0, %[[FOR_BODY_PREHEADER]] ]
+; CHECK-OPT-NEXT: [[LSR_IV4:%.*]] = getelementptr i8, ptr [[X]], i64 [[LSR_IV]]
; CHECK-OPT-NEXT: [[LDTMP:%.*]] = load i32, ptr [[LSR_IV4]], align 4
+; CHECK-OPT-NEXT: [[LSR_IV2:%.*]] = getelementptr i8, ptr [[Y]], i64 [[LSR_IV]]
; CHECK-OPT-NEXT: [[LDTMP1:%.*]] = load i32, ptr [[LSR_IV2]], align 4
; CHECK-OPT-NEXT: [[ADD:%.*]] = add nsw i32 [[LDTMP1]], [[LDTMP]]
+; CHECK-OPT-NEXT: [[LSR_IV1:%.*]] = getelementptr i8, ptr [[Q]], i64 [[LSR_IV]]
; CHECK-OPT-NEXT: store i32 [[ADD]], ptr [[LSR_IV1]], align 4
-; CHECK-OPT-NEXT: [[LSR_IV_NEXT]] = add nsw i64 [[LSR_IV]], -1
-; CHECK-OPT-NEXT: [[SCEVGEP]] = getelementptr i8, ptr [[LSR_IV1]], i64 4
-; CHECK-OPT-NEXT: [[SCEVGEP3]] = getelementptr i8, ptr [[LSR_IV2]], i64 4
-; CHECK-OPT-NEXT: [[SCEVGEP5]] = getelementptr i8, ptr [[LSR_IV4]], i64 4
-; CHECK-OPT-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[LSR_IV_NEXT]], 0
+; CHECK-OPT-NEXT: [[LSR_IV_NEXT]] = add nuw nsw i64 [[LSR_IV]], 4
+; CHECK-OPT-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[TMP0]], [[LSR_IV_NEXT]]
; CHECK-OPT-NEXT: br i1 [[EXITCOND]], label %[[FOR_COND_CLEANUP_LOOPEXIT]], label %[[FOR_BODY]]
;
entry:
More information about the llvm-branch-commits
mailing list