Skip to content

[LoongArch] Fix broadcast load with extension. #155960

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

[LoongArch] Fix broadcast load with extension. #155960

wants to merge 2 commits into from

Conversation

tangaac
Copy link
Contributor

@tangaac tangaac commented Aug 29, 2025

PR #135896 introduces [x]vldrepl instructions without handling extension.
This patch will fix that.

@llvmbot
Copy link
Member

llvmbot commented Aug 29, 2025

@llvm/pr-subscribers-backend-loongarch

Author: None (tangaac)

Changes

PR #135896 introduces [x]vldrepl instructions without handling extension.
This patch will fix that.

Full diff: https://github.com/llvm/llvm-project/pull/155960.diff

3 Files Affected:

  • (modified) llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp (+2-1)
  • (modified) llvm/test/CodeGen/LoongArch/lasx/broadcast-load.ll (+38-6)
  • (modified) llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll (+39-6)
diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
index ffb6c2980026f..478c335c3f07e 100644
--- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
@@ -2471,8 +2471,9 @@ static SDValue lowerBUILD_VECTORAsBroadCastLoad(BuildVectorSDNode *BVOp,
   if (!IdentitySrc || !BVOp->isOnlyUserOf(IdentitySrc.getNode()))
     return SDValue();
 
-  if (IsIdeneity) {
+  if (IsIdeneity && ISD::isNON_EXTLoad(IdentitySrc.getNode())) {
     auto *LN = cast<LoadSDNode>(IdentitySrc);
+
     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 976924bdca686..89592a0886cc1 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/broadcast-load.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/broadcast-load.ll
@@ -18,6 +18,32 @@ 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) {
 ; CHECK-LABEL: xvldrepl_d_unaligned_offset:
 ; CHECK:       # %bb.0:
@@ -34,7 +60,8 @@ define <4 x i64> @xvldrepl_d_unaligned_offset(ptr %ptr) {
 define <32 x i8> @xvldrepl_b(ptr %ptr) {
 ; CHECK-LABEL: xvldrepl_b:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    xvldrepl.b $xr0, $a0, 0
+; CHECK-NEXT:    ld.b $a0, $a0, 0
+; CHECK-NEXT:    xvreplgr2vr.b $xr0, $a0
 ; CHECK-NEXT:    ret
   %tmp = load i8, ptr %ptr
   %tmp1 = insertelement <32 x i8> zeroinitializer, i8 %tmp, i32 0
@@ -45,7 +72,8 @@ define <32 x i8> @xvldrepl_b(ptr %ptr) {
 define <32 x i8> @xvldrepl_b_offset(ptr %ptr) {
 ; CHECK-LABEL: xvldrepl_b_offset:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    xvldrepl.b $xr0, $a0, 33
+; CHECK-NEXT:    ld.b $a0, $a0, 33
+; CHECK-NEXT:    xvreplgr2vr.b $xr0, $a0
 ; CHECK-NEXT:    ret
   %p = getelementptr i8, ptr %ptr, i64 33
   %tmp = load i8, ptr %p
@@ -58,7 +86,8 @@ define <32 x i8> @xvldrepl_b_offset(ptr %ptr) {
 define <16 x i16> @xvldrepl_h(ptr %ptr) {
 ; CHECK-LABEL: xvldrepl_h:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    xvldrepl.h $xr0, $a0, 0
+; CHECK-NEXT:    ld.h $a0, $a0, 0
+; CHECK-NEXT:    xvreplgr2vr.h $xr0, $a0
 ; CHECK-NEXT:    ret
   %tmp = load i16, ptr %ptr
   %tmp1 = insertelement <16 x i16> zeroinitializer, i16 %tmp, i32 0
@@ -69,7 +98,8 @@ define <16 x i16> @xvldrepl_h(ptr %ptr) {
 define <16 x i16> @xvldrepl_h_offset(ptr %ptr) {
 ; CHECK-LABEL: xvldrepl_h_offset:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    xvldrepl.h $xr0, $a0, 66
+; CHECK-NEXT:    ld.h $a0, $a0, 66
+; CHECK-NEXT:    xvreplgr2vr.h $xr0, $a0
 ; CHECK-NEXT:    ret
   %p = getelementptr i16, ptr %ptr, i64 33
   %tmp = load i16, ptr %p
@@ -81,7 +111,8 @@ define <16 x i16> @xvldrepl_h_offset(ptr %ptr) {
 define <8 x i32> @xvldrepl_w(ptr %ptr) {
 ; CHECK-LABEL: xvldrepl_w:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    xvldrepl.w $xr0, $a0, 0
+; CHECK-NEXT:    ld.w $a0, $a0, 0
+; CHECK-NEXT:    xvreplgr2vr.w $xr0, $a0
 ; CHECK-NEXT:    ret
   %tmp = load i32, ptr %ptr
   %tmp1 = insertelement <8 x i32> zeroinitializer, i32 %tmp, i32 0
@@ -92,7 +123,8 @@ define <8 x i32> @xvldrepl_w(ptr %ptr) {
 define <8 x i32> @xvldrepl_w_offset(ptr %ptr) {
 ; CHECK-LABEL: xvldrepl_w_offset:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    xvldrepl.w $xr0, $a0, 132
+; CHECK-NEXT:    ld.w $a0, $a0, 132
+; CHECK-NEXT:    xvreplgr2vr.w $xr0, $a0
 ; CHECK-NEXT:    ret
   %p = getelementptr i32, ptr %ptr, i64 33
   %tmp = load i32, ptr %p
diff --git a/llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll b/llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll
index c46747ef30509..a8cddbf9e6400 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll
@@ -18,6 +18,32 @@ 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) {
 ; CHECK-LABEL: vldrepl_d_unaligned_offset:
 ; CHECK:       # %bb.0:
@@ -34,7 +60,8 @@ define <2 x i64> @vldrepl_d_unaligned_offset(ptr %ptr) {
 define <16 x i8> @vldrepl_b(ptr %ptr) {
 ; CHECK-LABEL: vldrepl_b:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vldrepl.b $vr0, $a0, 0
+; CHECK-NEXT:    ld.b $a0, $a0, 0
+; CHECK-NEXT:    vreplgr2vr.b $vr0, $a0
 ; CHECK-NEXT:    ret
   %tmp = load i8, ptr %ptr
   %tmp1 = insertelement <16 x i8> zeroinitializer, i8 %tmp, i32 0
@@ -45,7 +72,8 @@ define <16 x i8> @vldrepl_b(ptr %ptr) {
 define <16 x i8> @vldrepl_b_offset(ptr %ptr) {
 ; CHECK-LABEL: vldrepl_b_offset:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vldrepl.b $vr0, $a0, 33
+; CHECK-NEXT:    ld.b $a0, $a0, 33
+; CHECK-NEXT:    vreplgr2vr.b $vr0, $a0
 ; CHECK-NEXT:    ret
   %p = getelementptr i8, ptr %ptr, i64 33
   %tmp = load i8, ptr %p
@@ -58,7 +86,8 @@ define <16 x i8> @vldrepl_b_offset(ptr %ptr) {
 define <8 x i16> @vldrepl_h(ptr %ptr) {
 ; CHECK-LABEL: vldrepl_h:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vldrepl.h $vr0, $a0, 0
+; CHECK-NEXT:    ld.h $a0, $a0, 0
+; CHECK-NEXT:    vreplgr2vr.h $vr0, $a0
 ; CHECK-NEXT:    ret
   %tmp = load i16, ptr %ptr
   %tmp1 = insertelement <8 x i16> zeroinitializer, i16 %tmp, i32 0
@@ -69,7 +98,8 @@ define <8 x i16> @vldrepl_h(ptr %ptr) {
 define <8 x i16> @vldrepl_h_offset(ptr %ptr) {
 ; CHECK-LABEL: vldrepl_h_offset:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vldrepl.h $vr0, $a0, 66
+; CHECK-NEXT:    ld.h $a0, $a0, 66
+; CHECK-NEXT:    vreplgr2vr.h $vr0, $a0
 ; CHECK-NEXT:    ret
   %p = getelementptr i16, ptr %ptr, i64 33
   %tmp = load i16, ptr %p
@@ -81,7 +111,8 @@ define <8 x i16> @vldrepl_h_offset(ptr %ptr) {
 define <4 x i32> @vldrepl_w(ptr %ptr) {
 ; CHECK-LABEL: vldrepl_w:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vldrepl.w $vr0, $a0, 0
+; CHECK-NEXT:    ld.w $a0, $a0, 0
+; CHECK-NEXT:    vreplgr2vr.w $vr0, $a0
 ; CHECK-NEXT:    ret
   %tmp = load i32, ptr %ptr
   %tmp1 = insertelement <4 x i32> zeroinitializer, i32 %tmp, i32 0
@@ -92,7 +123,8 @@ define <4 x i32> @vldrepl_w(ptr %ptr) {
 define <4 x i32> @vldrepl_w_offset(ptr %ptr) {
 ; CHECK-LABEL: vldrepl_w_offset:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vldrepl.w $vr0, $a0, 132
+; CHECK-NEXT:    ld.w $a0, $a0, 132
+; CHECK-NEXT:    vreplgr2vr.w $vr0, $a0
 ; CHECK-NEXT:    ret
   %p = getelementptr i32, ptr %ptr, i64 33
   %tmp = load i32, ptr %p
@@ -169,3 +201,4 @@ define <2 x double> @vldrepl_d_dbl_offset(ptr %ptr) {
   %tmp2 = shufflevector <2 x double> %tmp1, <2 x double> poison, <2 x i32> zeroinitializer
   ret <2 x double> %tmp2
 }
+

zhaoqi5
zhaoqi5 reviewed Aug 29, 2025
View reviewed changes
@SixWeining SixWeining added this to the LLVM 21.x Release milestone Aug 29, 2025
@github-project-automation github-project-automation bot moved this to Needs Triage in LLVM Release Status Aug 29, 2025
@SixWeining SixWeining requested a review from heiher August 29, 2025 03:56
heiher
heiher reviewed Aug 29, 2025
View reviewed changes
llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp Outdated
auto *LN = cast<LoadSDNode>(IdentitySrc);
auto ExtType = LN->getExtensionType();

if (IsIdeneity && (ExtType == ISD::EXTLOAD || ExtType == ISD::NON_EXTLOAD)) {
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Do you think it makes sense to precisely match the sizes of VT.getVectorElementVT() and LN->getMemoryVT()?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It matches the case like this.

    t2: i64,ch = CopyFromReg t0, Register:i64 %0
  t16: i64,ch = load<(load (s8) from %ir.ptr), anyext from i8> t0, t2, undef:i64


    t2: i64,ch = CopyFromReg t0, Register:i64 %0
  t5: i64,ch = load<(load (s64) from %ir.ptr)> t0, t2, undef:i64

@zhaoqi5
Copy link
Contributor

zhaoqi5 commented Aug 29, 2025

The TODO at the top of the test files may can be deleted.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@heiher heiher heiher left review comments

@zhaoqi5 zhaoqi5 zhaoqi5 left review comments

Assignees
No one assigned
Labels
backend:loongarch
Projects
LLVM Release Status
Status: Needs Triage
Milestone
LLVM 21.x Release
Development

Successfully merging this pull request may close these issues.
5 participants
@tangaac @llvmbot @zhaoqi5 @heiher @SixWeining