[llvm] WIP: [SelectionDAG] Remove `UnsafeFPMath` in LegalizeDAG (PR #146316)

[via llvm-commits]

https://github.com/paperchalice updated https://github.com/llvm/llvm-project/pull/146316

>From 8b22544821f0d232b01db038f9d6fec36eca3e6a Mon Sep 17 00:00:00 2001
From: PaperChalice <liujunchang97 at outlook.com>
Date: Thu, 24 Jul 2025 17:32:48 +0800
Subject: [PATCH 1/3] Consider fastmath flags in `visitFPTrunc`

---
 llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index 163646513918d..6eca7b73a9d76 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -3923,11 +3923,15 @@ void SelectionDAGBuilder::visitFPTrunc(const User &I) {
   // FPTrunc is never a no-op cast, no need to check
   SDValue N = getValue(I.getOperand(0));
   SDLoc dl = getCurSDLoc();
+  SDNodeFlags Flags;
+  if (auto *TruncInst = dyn_cast<FPMathOperator>(&I))
+    Flags.copyFMF(*TruncInst);
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   EVT DestVT = TLI.getValueType(DAG.getDataLayout(), I.getType());
   setValue(&I, DAG.getNode(ISD::FP_ROUND, dl, DestVT, N,
                            DAG.getTargetConstant(
-                               0, dl, TLI.getPointerTy(DAG.getDataLayout()))));
+                               0, dl, TLI.getPointerTy(DAG.getDataLayout())),
+                           Flags));
 }
 
 void SelectionDAGBuilder::visitFPExt(const User &I) {

>From bc2638d2bb90106f1e499885886d484914ec12a7 Mon Sep 17 00:00:00 2001
From: PaperChalice <liujunchang97 at outlook.com>
Date: Mon, 30 Jun 2025 16:19:04 +0800
Subject: [PATCH 2/3] Remove `UnsafeFPMath` in LegalizeDAG

---
 llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp | 2 +-
 llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll       | 6 +++---
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
index 74172b230361d..b7a96cb2dc826 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -3853,7 +3853,7 @@ bool SelectionDAGLegalize::ExpandNode(SDNode *Node) {
     break;
   case ISD::FP_TO_FP16:
     LLVM_DEBUG(dbgs() << "Legalizing FP_TO_FP16\n");
-    if (!TLI.useSoftFloat() && TM.Options.UnsafeFPMath) {
+    if (!TLI.useSoftFloat() && Node->getFlags().hasApproximateFuncs()) {
       SDValue Op = Node->getOperand(0);
       MVT SVT = Op.getSimpleValueType();
       if ((SVT == MVT::f64 || SVT == MVT::f80) &&
diff --git a/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll b/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll
index 0a900f904bec5..89ce0bda41f8e 100644
--- a/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll
@@ -196,7 +196,7 @@ define amdgpu_kernel void @fptrunc_f32_to_f16(
     ptr addrspace(1) %a) {
 entry:
   %a.val = load float, ptr addrspace(1) %a
-  %r.val = fptrunc float %a.val to half
+  %r.val = fptrunc afn float %a.val to half
   store half %r.val, ptr addrspace(1) %r
   ret void
 }
@@ -401,7 +401,7 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(
     ptr addrspace(1) %a) {
 entry:
   %a.val = load double, ptr addrspace(1) %a
-  %r.val = fptrunc double %a.val to half
+  %r.val = fptrunc afn double %a.val to half
   store half %r.val, ptr addrspace(1) %r
   ret void
 }
@@ -863,7 +863,7 @@ define amdgpu_kernel void @fptrunc_v2f64_to_v2f16(
     ptr addrspace(1) %a) {
 entry:
   %a.val = load <2 x double>, ptr addrspace(1) %a
-  %r.val = fptrunc <2 x double> %a.val to <2 x half>
+  %r.val = fptrunc afn <2 x double> %a.val to <2 x half>
   store <2 x half> %r.val, ptr addrspace(1) %r
   ret void
 }

>From c386827795121e54d6fd82493e89ac5873a075f9 Mon Sep 17 00:00:00 2001
From: PaperChalice <liujunchang97 at outlook.com>
Date: Thu, 24 Jul 2025 17:37:44 +0800
Subject: [PATCH 3/3] Fix arm test This change is not the solution, due to the
 type of `llvm.convert.to.fp16.f64`.

---
 llvm/test/CodeGen/ARM/fp16.ll | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/llvm/test/CodeGen/ARM/fp16.ll b/llvm/test/CodeGen/ARM/fp16.ll
index dc35fa34f42c1..9ff701050ac7e 100644
--- a/llvm/test/CodeGen/ARM/fp16.ll
+++ b/llvm/test/CodeGen/ARM/fp16.ll
@@ -86,8 +86,8 @@ define i16 @test_to_fp16(double %in) {
 
 ; CHECK-FP16-SAFE: bl __aeabi_d2h
 
-; CHECK-FP16-UNSAFE:      vcvt.f32.f64 s0, d0
-; CHECK-FP16-UNSAFE-NEXT: vcvtb.f16.f32 s0, s0
+; CHECK-FP16-UNSAFE:      vmov r0, r1, d0
+; CHECK-FP16-UNSAFE-NEXT: bl __aeabi_d2h
 
 ; CHECK-ARMV8: vcvtb.f16.f64 [[TMP:s[0-9]+]], d0
 ; CHECK-ARMV8: vmov r0, [[TMP]]