From be1adeaeafafc059c26166bd25a25b067d0ea567 Mon Sep 17 00:00:00 2001 From: tangaac Date: Fri, 29 Aug 2025 09:45:33 +0800 Subject: [PATCH 1/3] Fix broadcast load with extension. --- .../LoongArch/LoongArchISelLowering.cpp | 6 +++-- .../CodeGen/LoongArch/lasx/broadcast-load.ll | 26 +++++++++++++++++++ .../CodeGen/LoongArch/lsx/broadcast-load.ll | 26 +++++++++++++++++++ 3 files changed, 56 insertions(+), 2 deletions(-) diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp index fe650a0b90263..3a965f1499cb1 100644 --- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp +++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp @@ -2519,8 +2519,10 @@ static SDValue lowerBUILD_VECTORAsBroadCastLoad(BuildVectorSDNode *BVOp, if (!IdentitySrc || !BVOp->isOnlyUserOf(IdentitySrc.getNode())) return SDValue(); - if (IsIdeneity) { - auto *LN = cast(IdentitySrc); + auto *LN = cast(IdentitySrc); + auto ExtType = LN->getExtensionType(); + + if (IsIdeneity && (ExtType == ISD::EXTLOAD || ExtType == ISD::NON_EXTLOAD)) { SDVTList Tys = LN->isIndexed() ? DAG.getVTList(VT, LN->getBasePtr().getValueType(), MVT::Other) diff --git a/llvm/test/CodeGen/LoongArch/lasx/broadcast-load.ll b/llvm/test/CodeGen/LoongArch/lasx/broadcast-load.ll index 205e59a18bf9d..4aa2bd76ab80c 100644 --- a/llvm/test/CodeGen/LoongArch/lasx/broadcast-load.ll +++ b/llvm/test/CodeGen/LoongArch/lasx/broadcast-load.ll @@ -33,6 +33,32 @@ define <4 x i64> @should_not_be_optimized(ptr %ptr, ptr %dst) { ret <4 x i64> %tmp2 } +define <16 x i16> @should_not_be_optimized_sext_load(ptr %ptr) { +; CHECK-LABEL: should_not_be_optimized_sext_load: +; CHECK: # %bb.0: +; CHECK-NEXT: ld.b $a0, $a0, 0 +; CHECK-NEXT: xvreplgr2vr.h $xr0, $a0 +; CHECK-NEXT: ret + %tmp = load i8, ptr %ptr + %tmp1 = sext i8 %tmp to i16 + %tmp2 = insertelement <16 x i16> zeroinitializer, i16 %tmp1, i32 0 + %tmp3 = shufflevector <16 x i16> %tmp2, <16 x i16> poison, <16 x i32> zeroinitializer + ret <16 x i16> %tmp3 +} + +define <16 x i16> @should_not_be_optimized_zext_load(ptr %ptr) { +; CHECK-LABEL: should_not_be_optimized_zext_load: +; CHECK: # %bb.0: +; CHECK-NEXT: ld.bu $a0, $a0, 0 +; CHECK-NEXT: xvreplgr2vr.h $xr0, $a0 +; CHECK-NEXT: ret + %tmp = load i8, ptr %ptr + %tmp1 = zext i8 %tmp to i16 + %tmp2 = insertelement <16 x i16> zeroinitializer, i16 %tmp1, i32 0 + %tmp3 = shufflevector <16 x i16> %tmp2, <16 x i16> poison, <16 x i32> zeroinitializer + ret <16 x i16> %tmp3 +} + define <4 x i64> @xvldrepl_d_unaligned_offset(ptr %ptr) { ; LA32-LABEL: xvldrepl_d_unaligned_offset: ; LA32: # %bb.0: diff --git a/llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll b/llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll index 12224f8d59b9f..eed149957f6ab 100644 --- a/llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll +++ b/llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll @@ -31,6 +31,32 @@ define <2 x i64> @should_not_be_optimized(ptr %ptr, ptr %dst){ ret <2 x i64> %tmp2 } +define <8 x i16> @should_not_be_optimized_sext_load(ptr %ptr) { +; CHECK-LABEL: should_not_be_optimized_sext_load: +; CHECK: # %bb.0: +; CHECK-NEXT: ld.b $a0, $a0, 0 +; CHECK-NEXT: vreplgr2vr.h $vr0, $a0 +; CHECK-NEXT: ret + %tmp = load i8, ptr %ptr + %tmp1 = sext i8 %tmp to i16 + %tmp2 = insertelement <8 x i16> zeroinitializer, i16 %tmp1, i32 0 + %tmp3 = shufflevector <8 x i16> %tmp2, <8 x i16> poison, <8 x i32> zeroinitializer + ret <8 x i16> %tmp3 +} + +define <8 x i16> @should_not_be_optimized_zext_load(ptr %ptr) { +; CHECK-LABEL: should_not_be_optimized_zext_load: +; CHECK: # %bb.0: +; CHECK-NEXT: ld.bu $a0, $a0, 0 +; CHECK-NEXT: vreplgr2vr.h $vr0, $a0 +; CHECK-NEXT: ret + %tmp = load i8, ptr %ptr + %tmp1 = zext i8 %tmp to i16 + %tmp2 = insertelement <8 x i16> zeroinitializer, i16 %tmp1, i32 0 + %tmp3 = shufflevector <8 x i16> %tmp2, <8 x i16> poison, <8 x i32> zeroinitializer + ret <8 x i16> %tmp3 +} + define <2 x i64> @vldrepl_d_unaligned_offset(ptr %ptr) { ; LA32-LABEL: vldrepl_d_unaligned_offset: ; LA32: # %bb.0: From f78f7d34c34528ce6686eeb69772c0bdb5a36bf1 Mon Sep 17 00:00:00 2001 From: tangaac Date: Fri, 29 Aug 2025 16:51:02 +0800 Subject: [PATCH 2/3] address comments --- llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp | 4 ++-- llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll | 2 -- 2 files changed, 2 insertions(+), 4 deletions(-) diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp index 3a965f1499cb1..099fa878c5f8d 100644 --- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp +++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp @@ -2516,13 +2516,13 @@ static SDValue lowerBUILD_VECTORAsBroadCastLoad(BuildVectorSDNode *BVOp, } // make sure that this load is valid and only has one user. - if (!IdentitySrc || !BVOp->isOnlyUserOf(IdentitySrc.getNode())) + if (!IsIdeneity || !IdentitySrc || !BVOp->isOnlyUserOf(IdentitySrc.getNode())) return SDValue(); auto *LN = cast(IdentitySrc); auto ExtType = LN->getExtensionType(); - if (IsIdeneity && (ExtType == ISD::EXTLOAD || ExtType == ISD::NON_EXTLOAD)) { + if (ExtType == ISD::EXTLOAD || ExtType == ISD::NON_EXTLOAD) { SDVTList Tys = LN->isIndexed() ? DAG.getVTList(VT, LN->getBasePtr().getValueType(), MVT::Other) diff --git a/llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll b/llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll index eed149957f6ab..349684ff22be2 100644 --- a/llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll +++ b/llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll @@ -2,8 +2,6 @@ ; RUN: llc --mtriple=loongarch32 -mattr=+32s,+lsx < %s | FileCheck %s --check-prefixes=CHECK,LA32 ; RUN: llc --mtriple=loongarch64 -mattr=+lsx < %s | FileCheck %s --check-prefixes=CHECK,LA64 -; TODO: Load a element and splat it to a vector could be lowerd to vldrepl - ; A load has more than one user shouldn't be lowered to vldrepl define <2 x i64> @should_not_be_optimized(ptr %ptr, ptr %dst){ ; LA32-LABEL: should_not_be_optimized: From dcaf590aad9355ff1d514fd8ac0752d8d95d3eb6 Mon Sep 17 00:00:00 2001 From: tangaac Date: Mon, 1 Sep 2025 15:59:14 +0800 Subject: [PATCH 3/3] address hev's comment --- llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp index 099fa878c5f8d..8313869c1016d 100644 --- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp +++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp @@ -2522,7 +2522,8 @@ static SDValue lowerBUILD_VECTORAsBroadCastLoad(BuildVectorSDNode *BVOp, auto *LN = cast(IdentitySrc); auto ExtType = LN->getExtensionType(); - if (ExtType == ISD::EXTLOAD || ExtType == ISD::NON_EXTLOAD) { + if ((ExtType == ISD::EXTLOAD || ExtType == ISD::NON_EXTLOAD) && + VT.getScalarSizeInBits() == LN->getMemoryVT().getScalarSizeInBits()) { SDVTList Tys = LN->isIndexed() ? DAG.getVTList(VT, LN->getBasePtr().getValueType(), MVT::Other)