From 83a25a101051b404bec1a5ba9cb867705f31262d Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Fri, 26 Mar 2021 17:29:36 -0400
Subject: [PATCH] GlobalISel: Restrict narrow scalar for fptoui/fptosi results

This practically only works for the f16 case AMDGPU uses, not wider
types.

Fixes bug 49710 by failing legalization.
---
 .../llvm/CodeGen/GlobalISel/LegalizerHelper.h      |  1 +
 llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp    | 44 ++++++++++++++--------
 .../CodeGen/AArch64/GlobalISel/legalize-fptoi.mir  | 28 ++++++++++++++
 3 files changed, 57 insertions(+), 16 deletions(-)

diff --git a/llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h b/llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h
index d276fab..e40bdbc 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h
@@ -330,6 +330,7 @@ public:
   LegalizeResult narrowScalarAddSub(MachineInstr &MI, unsigned TypeIdx,
                                     LLT NarrowTy);
   LegalizeResult narrowScalarMul(MachineInstr &MI, LLT Ty);
+  LegalizeResult narrowScalarFPTOI(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
   LegalizeResult narrowScalarExtract(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
   LegalizeResult narrowScalarInsert(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
 
diff --git a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
index 9eb4c80..99b6ea8 100644
--- a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
@@ -1216,22 +1216,9 @@ LegalizerHelper::LegalizeResult LegalizerHelper::narrowScalar(MachineInstr &MI,
     Observer.changedInstr(MI);
     return Legalized;
   }
-  case TargetOpcode::G_FPTOUI: {
-    if (TypeIdx != 0)
-      return UnableToLegalize;
-    Observer.changingInstr(MI);
-    narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_ZEXT);
-    Observer.changedInstr(MI);
-    return Legalized;
-  }
-  case TargetOpcode::G_FPTOSI: {
-    if (TypeIdx != 0)
-      return UnableToLegalize;
-    Observer.changingInstr(MI);
-    narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_SEXT);
-    Observer.changedInstr(MI);
-    return Legalized;
-  }
+  case TargetOpcode::G_FPTOUI:
+  case TargetOpcode::G_FPTOSI:
+    return narrowScalarFPTOI(MI, TypeIdx, NarrowTy);
   case TargetOpcode::G_FPEXT:
     if (TypeIdx != 0)
       return UnableToLegalize;
@@ -4847,6 +4834,31 @@ LegalizerHelper::narrowScalarMul(MachineInstr &MI, LLT NarrowTy) {
 }
 
 LegalizerHelper::LegalizeResult
+LegalizerHelper::narrowScalarFPTOI(MachineInstr &MI, unsigned TypeIdx,
+                                   LLT NarrowTy) {
+  if (TypeIdx != 0)
+    return UnableToLegalize;
+
+  bool IsSigned = MI.getOpcode() == TargetOpcode::G_FPTOSI;
+
+  Register Src = MI.getOperand(1).getReg();
+  LLT SrcTy = MRI.getType(Src);
+
+  // If all finite floats fit into the narrowed integer type, we can just swap
+  // out the result type. This is practically only useful for conversions from
+  // half to at least 16-bits, so just handle the one case.
+  if (SrcTy.getScalarType() != LLT::scalar(16) ||
+      NarrowTy.getScalarSizeInBits() < (IsSigned ? 17 : 16))
+    return UnableToLegalize;
+
+  Observer.changingInstr(MI);
+  narrowScalarDst(MI, NarrowTy, 0,
+                  IsSigned ? TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT);
+  Observer.changedInstr(MI);
+  return Legalized;
+}
+
+LegalizerHelper::LegalizeResult
 LegalizerHelper::narrowScalarExtract(MachineInstr &MI, unsigned TypeIdx,
                                      LLT NarrowTy) {
   if (TypeIdx != 1)
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/legalize-fptoi.mir b/llvm/test/CodeGen/AArch64/GlobalISel/legalize-fptoi.mir
index 9439056..b2ee3a6 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/legalize-fptoi.mir
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/legalize-fptoi.mir
@@ -238,3 +238,31 @@ body: |
     %1:_(<4 x s32>) = G_FPTOSI %0
     $q0 = COPY %1
 ...
+
+---
+name:            test_fptoui_s128_s32
+body: |
+  bb.0:
+    liveins: $w0
+    ; CHECK-LABEL: name: test_fptoui_s128_s32
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[FPTOUI:%[0-9]+]]:_(s128) = G_FPTOUI [[COPY]](s32)
+    ; CHECK: $q0 = COPY [[FPTOUI]](s128)
+    %0:_(s32) = COPY $w0
+    %1:_(s128) = G_FPTOUI %0
+    $q0 = COPY %1
+...
+
+---
+name:            test_fptosi_s128_s32
+body: |
+  bb.0:
+    liveins: $w0
+    ; CHECK-LABEL: name: test_fptosi_s128_s32
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[FPTOSI:%[0-9]+]]:_(s128) = G_FPTOSI [[COPY]](s32)
+    ; CHECK: $q0 = COPY [[FPTOSI]](s128)
+    %0:_(s32) = COPY $w0
+    %1:_(s128) = G_FPTOSI %0
+    $q0 = COPY %1
+...
-- 
2.7.4