[llvm] [SDISel][Combine] Constant fold FP16_TO_FP (PR #94790)

[Quentin Colombet via llvm-commits]

https://github.com/qcolombet updated https://github.com/llvm/llvm-project/pull/94790

>From 595741ddf918406b76fd35745d9ee0841579636f Mon Sep 17 00:00:00 2001
From: Quentin Colombet <quentin.colombet at gmail.com>
Date: Fri, 7 Jun 2024 13:01:33 +0200
Subject: [PATCH 1/2] [SDISel][Combine] Constant fold FP16_TO_FP

In some case, constant can survive early constant folding optimization
because they are hidden behind several layers of type changes.

E.g., consider the following sequence (extracted from the arm test
that this commit changes):
```
    t2: v1f16 = BUILD_VECTOR ConstantFP:f16<APFloat(0)>
    t4: v1f16 = insert_vector_elt t2, ConstantFP:f16<APFloat(0)>, Constant:i32<0>
  t5: f16 = bitcast t4
t6: f32 = fp_extend t5
```

Because the constant (APFloat(0)) is hidden behind a <1 x ty> type, all the
constant folding that normally happen for scalar nodes when using
`SelectionDAG::getNode` are blocked.

As a result the constant manages to survive as an actual conversion
instruction down to the select phase:
```
t11: f32 = fp16_to_fp Constant:i32<0>
```

With the change in this patch, we try to do constant folding one more
time during dag combine, which in the motivating example result in
the much better sequence:
```
t7: ch = CopyToReg t0, Register:f32 %0, ConstantFP:f32<0.000000e+00>
```

Note: I'm sure we have this problem in a lot of other places. Generally
speaking I believe SDISel is not that good with <1 x ty> compared to pure
scalar. However, I only changed what I could easily test.
---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 7 ++++++-
 llvm/test/CodeGen/AMDGPU/clamp-modifier.ll    | 3 +--
 llvm/test/CodeGen/AMDGPU/select-phi-s16-fp.ll | 3 +--
 llvm/test/CodeGen/ARM/arm-half-promote.ll     | 4 +---
 4 files changed, 9 insertions(+), 8 deletions(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 2d5968bf5c2ea..ed6eefbaf89c5 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -26586,7 +26586,12 @@ SDValue DAGCombiner::visitFP16_TO_FP(SDNode *N) {
     }
   }
 
-  return SDValue();
+  // Sometimes constants manage to survive very late in the pipeline, e.g.,
+  // because they are wrapped inside the <1 x f16> type. Try one last time to
+  // get rid of them.
+  SDValue Folded = DAG.FoldConstantArithmetic(
+      N->getOpcode(), SDLoc(N), N->getValueType(0), ArrayRef<SDValue>(&N0, 1));
+  return Folded;
 }
 
 SDValue DAGCombiner::visitFP_TO_BF16(SDNode *N) {
diff --git a/llvm/test/CodeGen/AMDGPU/clamp-modifier.ll b/llvm/test/CodeGen/AMDGPU/clamp-modifier.ll
index 0a0179e866cd3..84bd9b6f6c5d4 100644
--- a/llvm/test/CodeGen/AMDGPU/clamp-modifier.ll
+++ b/llvm/test/CodeGen/AMDGPU/clamp-modifier.ll
@@ -1489,9 +1489,8 @@ define amdgpu_kernel void @v_no_clamp_add_src_v2f16_f16_src(ptr addrspace(1) %ou
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    s_mov_b64 s[4:5], s[2:3]
 ; SI-NEXT:    buffer_load_ushort v1, v[1:2], s[4:7], 0 addr64
-; SI-NEXT:    v_cvt_f32_f16_e64 v3, s6 clamp
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, 0
 ; SI-NEXT:    s_mov_b64 s[2:3], s[6:7]
-; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
 ; SI-NEXT:    v_add_f32_e32 v1, 1.0, v1
diff --git a/llvm/test/CodeGen/AMDGPU/select-phi-s16-fp.ll b/llvm/test/CodeGen/AMDGPU/select-phi-s16-fp.ll
index 0d6e987165d87..ba04cdb795ce3 100644
--- a/llvm/test/CodeGen/AMDGPU/select-phi-s16-fp.ll
+++ b/llvm/test/CodeGen/AMDGPU/select-phi-s16-fp.ll
@@ -14,9 +14,8 @@ define void @phi_vec1half_to_f32_with_const_folding(ptr addrspace(1) %dst) #0 {
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_mov_b32 s4, 0
-; CHECK-NEXT:    v_cvt_f32_f16_e64 v2, s4
 ; CHECK-NEXT:  ; %bb.1: ; %bb
-; CHECK-NEXT:    v_cvt_f16_f32_e64 v2, v2
+; CHECK-NEXT:    v_cvt_f16_f32_e64 v2, s4
 ; CHECK-NEXT:    s_mov_b32 s7, 0xf000
 ; CHECK-NEXT:    s_mov_b32 s6, 0
 ; CHECK-NEXT:    s_mov_b32 s4, s6
diff --git a/llvm/test/CodeGen/ARM/arm-half-promote.ll b/llvm/test/CodeGen/ARM/arm-half-promote.ll
index a5fafd4238616..e1ab75b2ac7f1 100644
--- a/llvm/test/CodeGen/ARM/arm-half-promote.ll
+++ b/llvm/test/CodeGen/ARM/arm-half-promote.ll
@@ -116,9 +116,7 @@ define fastcc { <8 x half>, <8 x half> } @f3() {
 
 define void @extract_insert(ptr %dst) optnone noinline {
 ; CHECK-LABEL: extract_insert:
-; CHECK: movs r1, #0
-; CHECK: vmov s0, r1
-; CHECK: vcvtb.f32.f16 s0, s0
+; CHECK: vmov.i32 d0, #0x0
 ; CHECK: vcvtb.f16.f32 s0, s0
 ; CHECK: vmov r1, s0
 ; CHECK: strh r1, [r0]

>From c712e12351adea9f035ec91247a6cb822acc4c1f Mon Sep 17 00:00:00 2001
From: Quentin Colombet <quentin.colombet at gmail.com>
Date: Fri, 7 Jun 2024 22:24:47 +0200
Subject: [PATCH 2/2] use a simpler syntax

---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index ed6eefbaf89c5..25061f1a48474 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -26590,7 +26590,7 @@ SDValue DAGCombiner::visitFP16_TO_FP(SDNode *N) {
   // because they are wrapped inside the <1 x f16> type. Try one last time to
   // get rid of them.
   SDValue Folded = DAG.FoldConstantArithmetic(
-      N->getOpcode(), SDLoc(N), N->getValueType(0), ArrayRef<SDValue>(&N0, 1));
+      N->getOpcode(), SDLoc(N), N->getValueType(0), {N0});
   return Folded;
 }