[clang] [Matrix] Copy Row data from padded cbuffer offsets before swizzle (PR #185346)

[Farzon Lotfi via cfe-commits]

https://github.com/farzonl created https://github.com/llvm/llvm-project/pull/185346

fixes https://github.com/llvm/llvm-project/issues/184849

The fix is just to copy the data before a swizzle can happen

>From a4af2ae560b273b925b839d74db4ef8ed196514e Mon Sep 17 00:00:00 2001
From: Farzon Lotfi <farzonlotfi at microsoft.com>
Date: Mon, 9 Mar 2026 00:44:58 -0400
Subject: [PATCH 1/2] [Matrix] Copy Row data from padded cbuffer offsets before
 swizzle

fixes https://github.com/llvm/llvm-project/issues/184849

The fix is just to copy the data before a swizzle can happen
---
 clang/lib/CodeGen/CGExpr.cpp                  | 10 +++++++++-
 .../CBufferMatrixSingleSubscriptSwizzle.hlsl  | 19 +++++++++++++++++++
 2 files changed, 28 insertions(+), 1 deletion(-)
 create mode 100644 clang/test/CodeGenHLSL/BasicFeatures/CBufferMatrixSingleSubscriptSwizzle.hlsl

diff --git a/clang/lib/CodeGen/CGExpr.cpp b/clang/lib/CodeGen/CGExpr.cpp
index eebb36276e0eb..6afb39f194de1 100644
--- a/clang/lib/CodeGen/CGExpr.cpp
+++ b/clang/lib/CodeGen/CGExpr.cpp
@@ -5135,8 +5135,16 @@ LValue CodeGenFunction::EmitMatrixSingleSubscriptExpr(
     const MatrixSingleSubscriptExpr *E) {
   LValue Base = EmitLValue(E->getBase());
   llvm::Value *RowIdx = EmitMatrixIndexExpr(E->getRowIdx());
+
+  RawAddress MatAddr = Base.getAddress();
+  if (getLangOpts().HLSL &&
+      E->getBase()->getType().getAddressSpace() == LangAS::hlsl_constant)
+    MatAddr =
+        CGM.getHLSLRuntime().createBufferMatrixTempAddress(Base, E->getExprLoc(),
+                                                           *this);
+
   return LValue::MakeMatrixRow(
-      MaybeConvertMatrixAddress(Base.getAddress(), *this), RowIdx,
+      MaybeConvertMatrixAddress(MatAddr, *this), RowIdx,
       E->getBase()->getType(), Base.getBaseInfo(), TBAAAccessInfo());
 }
 
diff --git a/clang/test/CodeGenHLSL/BasicFeatures/CBufferMatrixSingleSubscriptSwizzle.hlsl b/clang/test/CodeGenHLSL/BasicFeatures/CBufferMatrixSingleSubscriptSwizzle.hlsl
new file mode 100644
index 0000000000000..46d51725827a5
--- /dev/null
+++ b/clang/test/CodeGenHLSL/BasicFeatures/CBufferMatrixSingleSubscriptSwizzle.hlsl
@@ -0,0 +1,19 @@
+// RUN: %clang_cc1 -triple dxil-pc-shadermodel6.7-library -disable-llvm-passes -emit-llvm -finclude-default-header -fmatrix-memory-layout=row-major -o - %s | FileCheck %s
+
+cbuffer CB {
+  float3x2 Mat;
+}
+
+// CHECK: @Mat = external hidden addrspace(2) global <{ [2 x <{ <2 x float>, target("dx.Padding", 8) }>], <2 x float> }>, align 4
+
+// CHECK-LABEL: define hidden noundef nofpclass(nan inf) <2 x float> @_Z15get_row_swizzlev()
+// CHECK: %matrix.buf.copy = alloca [3 x <2 x float>], align 8
+// CHECK: %cbuf.load = load <2 x float>, ptr addrspace(2) @Mat, align 8
+// CHECK: %cbuf.load2 = load <2 x float>, ptr addrspace(2) getelementptr inbounds nuw (i8, ptr addrspace(2) @Mat, i32 16), align 8
+// CHECK: %cbuf.load4 = load <2 x float>, ptr addrspace(2) getelementptr inbounds nuw (i8, ptr addrspace(2) @Mat, i32 32), align 8
+// CHECK: %0 = load <6 x float>, ptr %matrix.buf.copy, align 8
+// CHECK: %1 = shufflevector <6 x float> %0, <6 x float> poison, <2 x i32> <i32 4, i32 1>
+
+float2 get_row_swizzle() {
+  return Mat[1].gr;
+}

>From e7facbfaf924efb7183da3327c9d33b67bb8a175 Mon Sep 17 00:00:00 2001
From: Farzon Lotfi <farzonlotfi at microsoft.com>
Date: Mon, 9 Mar 2026 00:47:54 -0400
Subject: [PATCH 2/2] run clang-format

---
 clang/lib/CodeGen/CGExpr.cpp | 11 +++++------
 1 file changed, 5 insertions(+), 6 deletions(-)

diff --git a/clang/lib/CodeGen/CGExpr.cpp b/clang/lib/CodeGen/CGExpr.cpp
index 6afb39f194de1..bb6708569cb38 100644
--- a/clang/lib/CodeGen/CGExpr.cpp
+++ b/clang/lib/CodeGen/CGExpr.cpp
@@ -5139,13 +5139,12 @@ LValue CodeGenFunction::EmitMatrixSingleSubscriptExpr(
   RawAddress MatAddr = Base.getAddress();
   if (getLangOpts().HLSL &&
       E->getBase()->getType().getAddressSpace() == LangAS::hlsl_constant)
-    MatAddr =
-        CGM.getHLSLRuntime().createBufferMatrixTempAddress(Base, E->getExprLoc(),
-                                                           *this);
+    MatAddr = CGM.getHLSLRuntime().createBufferMatrixTempAddress(
+        Base, E->getExprLoc(), *this);
 
-  return LValue::MakeMatrixRow(
-      MaybeConvertMatrixAddress(MatAddr, *this), RowIdx,
-      E->getBase()->getType(), Base.getBaseInfo(), TBAAAccessInfo());
+  return LValue::MakeMatrixRow(MaybeConvertMatrixAddress(MatAddr, *this),
+                               RowIdx, E->getBase()->getType(),
+                               Base.getBaseInfo(), TBAAAccessInfo());
 }
 
 LValue CodeGenFunction::EmitMatrixSubscriptExpr(const MatrixSubscriptExpr *E) {