[clang] [clang][CIR]: X86 movnti/movntsd/movntss CIR implementation. (PR #206388)

[Lucas Ribeiro Lima via cfe-commits]

https://github.com/lucaslive974 updated https://github.com/llvm/llvm-project/pull/206388

>From f3aacf7a9a1ed28a7eb7fe4487bf789272bc9696 Mon Sep 17 00:00:00 2001
From: Lucas-Ribeiro-Lima <lucasribeirolima974 at gmail.com>
Date: Sun, 28 Jun 2026 22:46:31 -0300
Subject: [PATCH] [clang][CIR]: X86 movnti/movntsd/movntss CIR implementation.

---
 clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp | 16 ++++++++++
 clang/test/CIR/CodeGen/builtins-x86.c      | 34 ++++++++++++++++++++--
 2 files changed, 47 insertions(+), 3 deletions(-)

diff --git a/clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp b/clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp
index 94fd16c6837a6..673ba95bccadb 100644
--- a/clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp
+++ b/clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp
@@ -1859,6 +1859,22 @@ CIRGenFunction::emitX86BuiltinExpr(unsigned builtinID, const CallExpr *expr) {
     mlir::Value sv = builder.createVecShuffle(loc, in, zero, indices);
     return builder.createBitcast(sv, ops[0].getType());
   }
+  case X86::BI__builtin_ia32_movnti:
+  case X86::BI__builtin_ia32_movnti64:
+  case X86::BI__builtin_ia32_movntsd:
+  case X86::BI__builtin_ia32_movntss: {
+    mlir::Location loc = getLoc(expr->getExprLoc());
+
+    Address dest = Address{ops[0], CharUnits::One()};
+    mlir::Value src = ops[1];
+
+    if (builtinID == X86::BI__builtin_ia32_movntsd ||
+        builtinID == X86::BI__builtin_ia32_movntss)
+      src = builder.createExtractElement(loc, ops[1], 0);
+
+    cir::StoreOp so = builder.createStore(loc, src, dest, false, true);
+    return so.getValue();
+  }
   case X86::BI__builtin_ia32_vprotbi:
   case X86::BI__builtin_ia32_vprotwi:
   case X86::BI__builtin_ia32_vprotdi:
diff --git a/clang/test/CIR/CodeGen/builtins-x86.c b/clang/test/CIR/CodeGen/builtins-x86.c
index f6991b4493f7f..2c8ebfd27eefb 100644
--- a/clang/test/CIR/CodeGen/builtins-x86.c
+++ b/clang/test/CIR/CodeGen/builtins-x86.c
@@ -1,8 +1,8 @@
-// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -target-feature +avx512fp16 -target-feature +avx512bf16 -target-feature +avx512vl -fclangir -emit-cir %s -o %t.cir
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -target-feature +avx512fp16 -target-feature +avx512bf16 -target-feature +avx512vl -target-feature +sse4a -fclangir -emit-cir %s -o %t.cir
 // RUN: FileCheck --input-file=%t.cir %s -check-prefix=CIR
-// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -target-feature +avx512fp16 -target-feature +avx512bf16 -target-feature +avx512vl -fclangir -emit-llvm %s -o %t.ll
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -target-feature +avx512fp16 -target-feature +avx512bf16 -target-feature +avx512vl -target-feature +sse4a -fclangir -emit-llvm %s -o %t.ll
 // RUN: FileCheck --input-file=%t.ll %s -check-prefix=LLVM
-// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -target-feature +avx512fp16 -target-feature +avx512bf16 -target-feature +avx512vl -emit-llvm %s -o %t-ogcg.ll
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -target-feature +avx512fp16 -target-feature +avx512bf16 -target-feature +avx512vl -target-feature +sse4a -emit-llvm %s -o %t-ogcg.ll
 // RUN: FileCheck --input-file=%t-ogcg.ll %s -check-prefix=OGCG
 
 void test_sfence(void) {
@@ -261,3 +261,31 @@ void test_conditional_bzero(void) {
               : foo())
           : __builtin_bzero(dst, len));
 }
+
+void test_movnti(int dest, int src) {
+  // CIR-LABEL: test_movnti
+  // CIR: %[[VAL:.*]] = cir.load
+  // CIR: cir.store nontemporal {{.*}} %[[VAL]]
+  
+  // LLVM-LABEL: @test_movnti
+  // LLVM: store i32 %{{.*}}, ptr %{{.*}}, align 1, !nontemporal
+  
+  // OGCG-LABEL: @test_movnti
+  // OGCG: store i32 %{{.*}}, ptr %{{.*}}, align 1, !nontemporal
+  return __builtin_ia32_movnti(&dest, src);
+}
+
+void test_movntss(float *dest, v4f src) {
+  // CIR-LABEL: test_movntss
+  // CIR: %[[ARR:.*]] = cir.load align(16) {{.*}} : !cir.ptr<!cir.vector<4 x !cir.float>>
+  // CIR: %[[IDX:.*]] = cir.const #cir.int<0>
+  // CIR: %[[SCALAR:.*]] = cir.vec.extract %[[ARR]][%[[IDX]] : !u64i]
+  // CIR: cir.store nontemporal align(1) %[[SCALAR]], {{.*}} : !cir.float, !cir.ptr<!cir.float>
+
+  // LLVM-LABEL: @test_movntss
+  // LLVM: store float %{{.*}}, ptr %{{.*}}, align 1, !nontemporal
+
+  // OGCG-LABEL: @test_movntss
+  // OGCG: store float %{{.*}}, ptr %{{.*}}, align 1, !nontemporal 
+  return __builtin_ia32_movntss(dest, src);
+}
\ No newline at end of file