From 798189e004badbcdcab907663cf7831afe06d575 Mon Sep 17 00:00:00 2001
From: "Wang, Phoebe" <phoebe.wang@intel.com>
Date: Thu, 5 Jun 2025 20:18:12 +0800
Subject: [PATCH] [X86][FP16] Winden UI2FP for FP16 when VLX not enabled

Fixes: https://godbolt.org/z/5vc8oMhxz
---
 llvm/lib/Target/X86/X86ISelLowering.cpp      | 12 +++++--
 llvm/test/CodeGen/X86/avx512fp16-cvt-novl.ll | 35 ++++++++++++--------
 2 files changed, 30 insertions(+), 17 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 9cff3d76913ab..760119bc62604 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -20361,10 +20361,16 @@ static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, const SDLoc &DL,
     if (VT == MVT::v8f64)
       return Op;
 
-    assert((VT == MVT::v4f32 || VT == MVT::v8f32 || VT == MVT::v4f64) &&
+    assert((VT == MVT::v4f32 || VT == MVT::v8f32 || VT == MVT::v4f64 ||
+            VT == MVT::v8f16) &&
            "Unexpected VT!");
-    MVT WideVT = VT == MVT::v4f64 ? MVT::v8f64 : MVT::v16f32;
-    MVT WideIntVT = VT == MVT::v4f64 ? MVT::v8i32 : MVT::v16i32;
+    MVT WideVT = VT == MVT::v8f16 ? MVT::v16f16 : MVT::v16f32;
+    MVT WideIntVT = MVT::v16i32;
+    if (VT == MVT::v4f64) {
+      WideVT = MVT::v8f64;
+      WideIntVT = MVT::v8i32;
+    }
+
     // Need to concat with zero vector for strict fp to avoid spurious
     // exceptions.
     SDValue Tmp =
diff --git a/llvm/test/CodeGen/X86/avx512fp16-cvt-novl.ll b/llvm/test/CodeGen/X86/avx512fp16-cvt-novl.ll
index 38c833ee89bc0..26947b5eb3022 100644
--- a/llvm/test/CodeGen/X86/avx512fp16-cvt-novl.ll
+++ b/llvm/test/CodeGen/X86/avx512fp16-cvt-novl.ll
@@ -8,8 +8,8 @@ define <2 x half> @vector_sint64ToHalf(<2 x i64> %int64) {
 ; CHECK-NEXT:    vcvtqq2ph %zmm0, %xmm0
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
-    %fp16 = sitofp <2 x i64> %int64 to <2 x half>
-    ret <2 x half> %fp16
+  %fp16 = sitofp <2 x i64> %int64 to <2 x half>
+  ret <2 x half> %fp16
 }
 
 define <4 x half> @vector_sint32ToHalf(<4 x i32> %int32) {
@@ -27,8 +27,8 @@ define <4 x half> @vector_sint32ToHalf(<4 x i32> %int32) {
 ; CHECK-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
 ; CHECK-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0],xmm1[0],zero,zero
 ; CHECK-NEXT:    retq
-    %fp16 = sitofp <4 x i32> %int32 to <4 x half>
-    ret <4 x half> %fp16
+  %fp16 = sitofp <4 x i32> %int32 to <4 x half>
+  ret <4 x half> %fp16
 }
 
 define <8 x half> @vector_sint16ToHalf(<8 x i16> %int16) {
@@ -66,8 +66,8 @@ define <8 x half> @vector_sint16ToHalf(<8 x i16> %int16) {
 ; CHECK-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
 ; CHECK-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 ; CHECK-NEXT:    retq
-    %fp16 = sitofp <8 x i16> %int16 to <8 x half>
-    ret <8 x half> %fp16
+  %fp16 = sitofp <8 x i16> %int16 to <8 x half>
+  ret <8 x half> %fp16
 }
 
 define <2 x half> @vector_uint64ToHalf(<2 x i64> %int64) {
@@ -77,14 +77,21 @@ define <2 x half> @vector_uint64ToHalf(<2 x i64> %int64) {
 ; CHECK-NEXT:    vcvtuqq2ph %zmm0, %xmm0
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
-    %fp16 = uitofp <2 x i64> %int64 to <2 x half>
-    ret <2 x half> %fp16
+  %fp16 = uitofp <2 x i64> %int64 to <2 x half>
+  ret <2 x half> %fp16
 }
 
-; define <4 x half> @vector_uint32ToHalf(<4 x i32> %int32) {
-;     %fp16 = uitofp <4 x i32> %int32 to <4 x half>
-;     ret <4 x half> %fp16
-; }
+define <4 x half> @vector_uint32ToHalf(<4 x i32> %int32) {
+; CHECK-LABEL: vector_uint32ToHalf:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
+; CHECK-NEXT:    vcvtudq2ph %zmm0, %ymm0
+; CHECK-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; CHECK-NEXT:    vzeroupper
+; CHECK-NEXT:    retq
+  %fp16 = uitofp <4 x i32> %int32 to <4 x half>
+  ret <4 x half> %fp16
+}
 
 define <8 x half> @vector_uint16ToHalf(<8 x i16> %int16) {
 ; CHECK-LABEL: vector_uint16ToHalf:
@@ -113,6 +120,6 @@ define <8 x half> @vector_uint16ToHalf(<8 x i16> %int16) {
 ; CHECK-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
 ; CHECK-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 ; CHECK-NEXT:    retq
-    %fp16 = uitofp <8 x i16> %int16 to <8 x half>
-    ret <8 x half> %fp16
+  %fp16 = uitofp <8 x i16> %int16 to <8 x half>
+  ret <8 x half> %fp16
 }