[llvm] [LegalizeIntegerTypes] Simplify ExpandIntRes_FP_TO_XINT when operand needs to be SoftPromoted. (PR #107634)

[Craig Topper via llvm-commits]

https://github.com/topperc created https://github.com/llvm/llvm-project/pull/107634

Create an FP_EXTEND instead of handling the soft promote directly. This FP_EXTEND will be visited and soft promoted itself.

This removes a zero extend from the generated code when the f32 type is itself softened. Previously we softened it as an fp16_to_fp which sees the operand as an integer type so we extend it. When we soften the result as an fp_extend we see the source as f16 and don't extend. It only becomes an integer inside call lowering not by type legalization.

If this extend is really necessary, then we have an issue when an f16->f32 fp_extend exists in the source and f32 needs to be softened.

This simplifies part of #102503.

>From 9c508b40f12e2fb0aff47181d635765197cfc027 Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper at sifive.com>
Date: Fri, 6 Sep 2024 12:27:29 -0700
Subject: [PATCH] [LegalizeIntegerTypes] Simplify ExpandIntRes_FP_TO_XINT when
 operand needs to be SoftPromoted.

Create an FP_EXTEND instead of handling the soft promote directly. This
FP_EXTEND will be visited and soft promoted itself.

This removes zero extend from the generated code when the f32 type
is itself softened. Previously we softened it as an fp16_to_fp which
sees the operand as an integer type and we extend it. When we soften
the result as an fp_extend we see the source as f16 and don't extend.
It only becomes an integer inside call lowering not by type legalization.

If this extend is really necessary, then we have an issue when an
f16->f32 fp_extend exists in the source and f32 needs to be softened.

This simplifies part of #102503.
---
 .../SelectionDAG/LegalizeIntegerTypes.cpp        | 16 +++-------------
 llvm/test/CodeGen/RISCV/half-convert.ll          |  4 ----
 2 files changed, 3 insertions(+), 17 deletions(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
index a1cb74f43e6050..530570ee815b39 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
@@ -3951,19 +3951,9 @@ void DAGTypeLegalizer::ExpandIntRes_FP_TO_XINT(SDNode *N, SDValue &Lo,
   if (getTypeAction(Op.getValueType()) == TargetLowering::TypePromoteFloat)
     Op = GetPromotedFloat(Op);
 
-  if (getTypeAction(Op.getValueType()) == TargetLowering::TypeSoftPromoteHalf) {
-    EVT OFPVT = Op.getValueType();
-    EVT NFPVT = TLI.getTypeToTransformTo(*DAG.getContext(), OFPVT);
-    Op = GetSoftPromotedHalf(Op);
-    Op = DAG.getNode(OFPVT == MVT::f16 ? ISD::FP16_TO_FP : ISD::BF16_TO_FP, dl,
-                     NFPVT, Op);
-    Op = DAG.getNode(IsSigned ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, dl, VT, Op);
-    SplitInteger(Op, Lo, Hi);
-    return;
-  }
-
-  if (Op.getValueType() == MVT::bf16) {
-    // Extend to f32 as there is no bf16 libcall.
+  // If the input is bf16 or needs to be soft promoted, extend to f32.
+  if (getTypeAction(Op.getValueType()) == TargetLowering::TypeSoftPromoteHalf ||
+      Op.getValueType() == MVT::bf16) {
     Op = fpExtendHelper(Op, Chain, IsStrict, MVT::f32, dl, DAG);
   }
 
diff --git a/llvm/test/CodeGen/RISCV/half-convert.ll b/llvm/test/CodeGen/RISCV/half-convert.ll
index 32f7dfaee8837c..d5373f847643c8 100644
--- a/llvm/test/CodeGen/RISCV/half-convert.ll
+++ b/llvm/test/CodeGen/RISCV/half-convert.ll
@@ -2033,8 +2033,6 @@ define i64 @fcvt_l_h(half %a) nounwind {
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    addi sp, sp, -16
 ; RV32I-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
-; RV32I-NEXT:    slli a0, a0, 16
-; RV32I-NEXT:    srli a0, a0, 16
 ; RV32I-NEXT:    call __extendhfsf2
 ; RV32I-NEXT:    call __fixsfdi
 ; RV32I-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
@@ -2816,8 +2814,6 @@ define i64 @fcvt_lu_h(half %a) nounwind {
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    addi sp, sp, -16
 ; RV32I-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
-; RV32I-NEXT:    slli a0, a0, 16
-; RV32I-NEXT:    srli a0, a0, 16
 ; RV32I-NEXT:    call __extendhfsf2
 ; RV32I-NEXT:    call __fixunssfdi
 ; RV32I-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload