[llvm] Fix generating permute bytes from register pair when the initial values are undefined (PR #74437)

[via llvm-commits]

https://github.com/mmoadeli created https://github.com/llvm/llvm-project/pull/74437

When generating the permute bytes for the prmt instruction, the existence of an undefined initial value initialises the int32 that holds the mask with all 1's (0xFFFFFFFF). That initialization subsequently leads to complications during the subsequent OR operation, leading to inaccuracies in populating mask values for the following bytes. Consequently, the final value persists as a constant -1, irrespective of the actual mask values that succeed the initial set value.

>From 27afcf4c84f8cbd37ecdf532bc15c7a23edb19c2 Mon Sep 17 00:00:00 2001
From: m moadeli <mahmoud.moadeli at codeplay.com>
Date: Tue, 5 Dec 2023 09:20:07 +0000
Subject: [PATCH] Fix generating permute bytes from register pair when the
 initial values are undefined.

---
 llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp    |  6 ++++--
 .../CodeGen/NVPTX/shuffle-vec-undef-init.ll    | 18 ++++++++++++++++++
 2 files changed, 22 insertions(+), 2 deletions(-)
 create mode 100644 llvm/test/CodeGen/NVPTX/shuffle-vec-undef-init.ll

diff --git a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
index b975825dae4b6..a13e2db853846 100644
--- a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
+++ b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
@@ -2360,8 +2360,10 @@ SDValue NVPTXTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
   const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
   SDValue V2 = Op.getOperand(1);
   uint32_t Selector = 0;
-  for (auto I : llvm::enumerate(SVN->getMask()))
-    Selector |= (I.value() << (I.index() * 4));
+  for (auto I : llvm::enumerate(SVN->getMask())) {
+    if (I.value() != -1)
+      Selector |= (I.value() << (I.index() * 4));
+  }
 
   SDLoc DL(Op);
   return DAG.getNode(NVPTXISD::PRMT, DL, MVT::v4i8, V1, V2,
diff --git a/llvm/test/CodeGen/NVPTX/shuffle-vec-undef-init.ll b/llvm/test/CodeGen/NVPTX/shuffle-vec-undef-init.ll
new file mode 100644
index 0000000000000..4f147f28e1a57
--- /dev/null
+++ b/llvm/test/CodeGen/NVPTX/shuffle-vec-undef-init.ll
@@ -0,0 +1,18 @@
+; RUN: llc < %s -march=nvptx -mcpu=sm_20 -verify-machineinstrs | FileCheck %s  
+; RUN: llc < %s -march=nvptx -mcpu=sm_20 -verify-machineinstrs | FileCheck %s   -check-prefix=CHECK-FOUND
+
+define void @kernel_func(ptr %in.vec, ptr %out.vec0) nounwind {
+  entry:
+  %wide.vec = load <32 x i8>, ptr %in.vec, align 64
+  %vec0 = shufflevector <32 x i8> %wide.vec, <32 x i8> undef, <4 x i32> <i32 0, i32 8, i32 16, i32 24>
+  store <4 x i8> %vec0, ptr %out.vec0, align 64
+  ret void
+
+; CHECK-FOUND: prmt.b32 	{{.*}} 16384;
+; CHECK-FOUND: prmt.b32 	{{.*}} 64;
+; CHECK-FOUND: prmt.b32 	{{.*}} 30224;
+
+; CHECK:  @kernel_func
+; CHECK-NOT: 	prmt.b32 	{{.*}} -1;
+; CHECK:  -- End function
+}