[GlobalISel] Support saturated truncate #150219

jyli0116 · 2025-07-23T13:27:06Z

Implements combining and legalization of G_TRUNC_SSAT_S, G_TRUNC_SSAT_U, and G_TRUNC_USAT_U, which where previously added to SDAG with the below patterns:

truncate(smin(smax(x, C1), C2)) -> trunc_ssat_s(x)
truncate(smax(smin(x, C2), C1)) -> trunc_ssat_s(x)

truncate(smax(smin(x, C), 0)) -> trunc_ssat_u(x)
truncate(smin(smax(x, 0), C)) -> trunc_ssat_u(x)
truncate(umin(smax(x, 0), C)) -> trunc_ssat_u(x)

truncate(umin(x, C)) -> trunc_usat_u(x)

llvmbot · 2025-07-23T13:27:39Z

@llvm/pr-subscribers-backend-aarch64

Author: None (jyli0116)

Changes

Implements combining and legalization of G_TRUNC_SSAT_S, G_TRUNC_SSAT_U, and G_TRUNC_USAT_U, which where previously added to SDAG with the below patterns:

truncate(smin(smax(x, C1), C2)) -&gt; trunc_ssat_s(x)
truncate(smax(smin(x, C2), C1)) -&gt; trunc_ssat_s(x)

truncate(smax(smin(x, C), 0)) -&gt; trunc_ssat_u(x)
truncate(smin(smax(x, 0), C)) -&gt; trunc_ssat_u(x)
truncate(umin(smax(x, 0), C)) -&gt; trunc_ssat_u(x)

truncate(umin(x, C)) -&gt; trunc_usat_u(x)

Full diff: https://github.com/llvm/llvm-project/pull/150219.diff

9 Files Affected:

(modified) llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h (+19)
(modified) llvm/include/llvm/CodeGen/GlobalISel/GenericMachineInstrs.h (+27)
(modified) llvm/include/llvm/CodeGen/GlobalISel/MIPatternMatch.h (+12)
(modified) llvm/include/llvm/Target/GlobalISel/Combine.td (+27-1)
(modified) llvm/include/llvm/Target/GlobalISel/SelectionDAGCompat.td (+3)
(modified) llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp (+128)
(modified) llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp (+3)
(modified) llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir (+8-6)
(added) llvm/test/CodeGen/AArch64/truncsat.ll (+79)

diff --git a/llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h b/llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h
index 31f1197b9723b..12a0fc4a2e1c4 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h
@@ -725,6 +725,25 @@ class CombinerHelper {
   bool matchUMulHToLShr(MachineInstr &MI) const;
   void applyUMulHToLShr(MachineInstr &MI) const;
 
+  // Combine trunc(smin(smax(x, C1), C2)) -> truncssat_s(x)
+  // or      trunc(smax(smin(x, C2), C1)) -> truncssat_s(x).
+  bool matchTruncSSatS(MachineInstr &MI, Register &MatchInfo) const;
+  void applyTruncSSatS(MachineInstr &MI, Register &MatchInfo) const;
+
+  // Combine trunc(smin(smax(x, 0), C)) -> truncssat_u(x)
+  // or      trunc(smax(smin(x, C), 0)) -> truncssat_u(x)
+  // or      trunc(umin(smax(x, 0), C)) -> truncssat_u(x)
+  bool matchTruncSSatU(MachineInstr &MI, Register &MatchInfo) const;
+  void applyTruncSSatU(MachineInstr &MI, Register &MatchInfo) const;
+
+  // Combine trunc(umin(x, C)) -> truncusat_u(x).
+  bool matchTruncUSatU(MachineInstr &MI, Register &MatchInfo) const;
+  void applyTruncUSatU(MachineInstr &MI, Register &MatchInfo) const;
+
+  // Combine truncusat_u(fptoui(x)) -> fptoui_sat(x)
+  bool matchTruncUSatUToFPTOUISat(MachineInstr &MI, Register &MatchInfo) const;
+  void applyTruncUSatUToFPTOUISat(MachineInstr &MI, Register &MatchInfo) const;
+
   /// Try to transform \p MI by using all of the above
   /// combine functions. Returns true if changed.
   bool tryCombine(MachineInstr &MI) const;
diff --git a/llvm/include/llvm/CodeGen/GlobalISel/GenericMachineInstrs.h b/llvm/include/llvm/CodeGen/GlobalISel/GenericMachineInstrs.h
index 571ec6dd7e9ba..f5cd5beb43767 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/GenericMachineInstrs.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/GenericMachineInstrs.h
@@ -874,6 +874,9 @@ class GCastOp : public GenericMachineInstr {
     case TargetOpcode::G_SEXT:
     case TargetOpcode::G_SITOFP:
     case TargetOpcode::G_TRUNC:
+    case TargetOpcode::G_TRUNC_SSAT_S:
+    case TargetOpcode::G_TRUNC_SSAT_U:
+    case TargetOpcode::G_TRUNC_USAT_U:
     case TargetOpcode::G_UITOFP:
     case TargetOpcode::G_ZEXT:
     case TargetOpcode::G_ANYEXT:
@@ -916,6 +919,30 @@ class GTrunc : public GCastOp {
   };
 };
 
+/// Represents a saturated trunc from a signed input to a signed result.
+class GTruncSSatS : public GCastOp {
+public:
+  static bool classof(const MachineInstr *MI) {
+    return MI->getOpcode() == TargetOpcode::G_TRUNC_SSAT_S;
+  };
+};
+
+/// Represents a saturated trunc from a signed input to an unsigned result.
+class GTruncSSatU : public GCastOp {
+public:
+  static bool classof(const MachineInstr *MI) {
+    return MI->getOpcode() == TargetOpcode::G_TRUNC_SSAT_U;
+  };
+};
+
+/// Represents a saturated trunc from an unsigned input to an unsigned result.
+class GTruncUSatU : public GCastOp {
+public:
+  static bool classof(const MachineInstr *MI) {
+    return MI->getOpcode() == TargetOpcode::G_TRUNC_USAT_U;
+  };
+};
+
 /// Represents a vscale.
 class GVScale : public GenericMachineInstr {
 public:
diff --git a/llvm/include/llvm/CodeGen/GlobalISel/MIPatternMatch.h b/llvm/include/llvm/CodeGen/GlobalISel/MIPatternMatch.h
index c0d3a12cbcb41..a27c79af87c02 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/MIPatternMatch.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/MIPatternMatch.h
@@ -709,6 +709,18 @@ m_GFPTrunc(const SrcTy &Src) {
   return UnaryOp_match<SrcTy, TargetOpcode::G_FPTRUNC>(Src);
 }
 
+template <typename SrcTy>
+inline UnaryOp_match<SrcTy, TargetOpcode::G_FPTOSI>
+m_GFPToSI(const SrcTy &Src) {
+  return UnaryOp_match<SrcTy, TargetOpcode::G_FPTOSI>(Src);
+}
+
+template <typename SrcTy>
+inline UnaryOp_match<SrcTy, TargetOpcode::G_FPTOUI>
+m_GFPToUI(const SrcTy &Src) {
+  return UnaryOp_match<SrcTy, TargetOpcode::G_FPTOUI>(Src);
+}
+
 template <typename SrcTy>
 inline UnaryOp_match<SrcTy, TargetOpcode::G_FABS> m_GFabs(const SrcTy &Src) {
   return UnaryOp_match<SrcTy, TargetOpcode::G_FABS>(Src);
diff --git a/llvm/include/llvm/Target/GlobalISel/Combine.td b/llvm/include/llvm/Target/GlobalISel/Combine.td
index 66051d756c808..61df21630a3b4 100644
--- a/llvm/include/llvm/Target/GlobalISel/Combine.td
+++ b/llvm/include/llvm/Target/GlobalISel/Combine.td
@@ -1236,6 +1236,32 @@ def mulh_to_lshr : GICombineRule<
 
 def mulh_combines : GICombineGroup<[mulh_to_lshr]>;
 
+def trunc_ssats : GICombineRule<
+  (defs root:$root, register_matchinfo:$matchinfo),
+  (match (G_TRUNC $dst, $src):$root,
+         [{ return Helper.matchTruncSSatS(*${root}, ${matchinfo}); }]),
+  (apply [{ Helper.applyTruncSSatS(*${root}, ${matchinfo}); }])>;
+
+def trunc_ssatu : GICombineRule<
+  (defs root:$root, register_matchinfo:$matchinfo),
+  (match (G_TRUNC $dst, $src):$root,
+    [{ return Helper.matchTruncSSatU(*${root}, ${matchinfo}); }]),
+  (apply [{ Helper.applyTruncSSatU(*${root}, ${matchinfo}); }])>;
+
+def trunc_usatu : GICombineRule<
+  (defs root:$root, register_matchinfo:$matchinfo),
+  (match (G_TRUNC $dst, $src):$root,
+    [{ return Helper.matchTruncUSatU(*${root}, ${matchinfo}); }]),
+  (apply [{ Helper.applyTruncUSatU(*${root}, ${matchinfo}); }])>;
+
+def truncusatu_to_fptouisat : GICombineRule<
+  (defs root:$root, register_matchinfo:$matchinfo),
+  (match (G_TRUNC_USAT_U $dst, $src):$root,
+    [{ return Helper.matchTruncUSatUToFPTOUISat(*${root}, ${matchinfo}); }]),
+  (apply [{ Helper.applyTruncUSatUToFPTOUISat(*${root}, ${matchinfo}); }])>;
+
+def truncsat_combines : GICombineGroup<[trunc_ssats, trunc_ssatu, trunc_usatu, truncusatu_to_fptouisat]>;
+
 def redundant_neg_operands: GICombineRule<
   (defs root:$root, build_fn_matchinfo:$matchinfo),
   (match (wip_match_opcode G_FADD, G_FSUB, G_FMUL, G_FDIV, G_FMAD, G_FMA):$root,
@@ -2060,7 +2086,7 @@ def all_combines : GICombineGroup<[integer_reassoc_combines, trivial_combines,
     fsub_to_fneg, commute_constant_to_rhs, match_ands, match_ors,
     simplify_neg_minmax, combine_concat_vector,
     sext_trunc, zext_trunc, prefer_sign_combines, shuffle_combines,
-    combine_use_vector_truncate, merge_combines, overflow_combines]>;
+    combine_use_vector_truncate, merge_combines, overflow_combines, truncsat_combines]>;
 
 // A combine group used to for prelegalizer combiners at -O0. The combines in
 // this group have been selected based on experiments to balance code size and
diff --git a/llvm/include/llvm/Target/GlobalISel/SelectionDAGCompat.td b/llvm/include/llvm/Target/GlobalISel/SelectionDAGCompat.td
index 7577792003d2e..e38a814b6e7b9 100644
--- a/llvm/include/llvm/Target/GlobalISel/SelectionDAGCompat.td
+++ b/llvm/include/llvm/Target/GlobalISel/SelectionDAGCompat.td
@@ -51,6 +51,9 @@ def : GINodeEquiv<G_SEXT, sext>;
 def : GINodeEquiv<G_SEXT_INREG, sext_inreg>;
 def : GINodeEquiv<G_ZEXT, zext>;
 def : GINodeEquiv<G_TRUNC, trunc>;
+def : GINodeEquiv<G_TRUNC_SSAT_S, truncssat_s>;
+def : GINodeEquiv<G_TRUNC_SSAT_U, truncssat_u>;
+def : GINodeEquiv<G_TRUNC_USAT_U, truncusat_u>;
 def : GINodeEquiv<G_BITCAST, bitconvert>;
 // G_INTTOPTR - SelectionDAG has no equivalent.
 // G_PTRTOINT - SelectionDAG has no equivalent.
diff --git a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
index 3922eba55e195..45914e717c3e6 100644
--- a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
@@ -5831,6 +5831,134 @@ void CombinerHelper::applyUMulHToLShr(MachineInstr &MI) const {
   MI.eraseFromParent();
 }
 
+bool CombinerHelper::matchTruncSSatS(MachineInstr &MI,
+                                     Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Register Src = MI.getOperand(1).getReg();
+  LLT DstTy = MRI.getType(Dst);
+  LLT SrcTy = MRI.getType(Src);
+  unsigned NumDstBits = DstTy.getScalarSizeInBits();
+  unsigned NumSrcBits = SrcTy.getScalarSizeInBits();
+  assert(NumSrcBits > NumDstBits && "Unexpected types for truncate operation");
+
+  APInt MinConst, MaxConst;
+  APInt SignedMax = APInt::getSignedMaxValue(NumDstBits).sext(NumSrcBits);
+  APInt SignedMin = APInt::getSignedMinValue(NumDstBits).sext(NumSrcBits);
+
+  if (isLegal({TargetOpcode::G_TRUNC_SSAT_S, {DstTy, SrcTy}})) {
+    if (mi_match(Src, MRI,
+                 m_GSMin(m_GSMax(m_Reg(MatchInfo), m_ICstOrSplat(MinConst)),
+                         m_ICstOrSplat(MaxConst))) &&
+        APInt::isSameValue(MinConst, SignedMin) &&
+        APInt::isSameValue(MaxConst, SignedMax))
+      return true;
+    if (mi_match(Src, MRI,
+                 m_GSMax(m_GSMin(m_Reg(MatchInfo), m_ICstOrSplat(MaxConst)),
+                         m_ICstOrSplat(MinConst))) &&
+        APInt::isSameValue(MinConst, SignedMin) &&
+        APInt::isSameValue(MaxConst, SignedMax))
+      return true;
+  }
+  return false;
+}
+
+void CombinerHelper::applyTruncSSatS(MachineInstr &MI,
+                                     Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Builder.buildTruncSSatS(Dst, MatchInfo);
+  MI.eraseFromParent();
+}
+
+bool CombinerHelper::matchTruncSSatU(MachineInstr &MI,
+                                     Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Register Src = MI.getOperand(1).getReg();
+  LLT DstTy = MRI.getType(Dst);
+  LLT SrcTy = MRI.getType(Src);
+  unsigned NumDstBits = DstTy.getScalarSizeInBits();
+  unsigned NumSrcBits = SrcTy.getScalarSizeInBits();
+  assert(NumSrcBits > NumDstBits && "Unexpected types for truncate operation");
+
+  APInt MaxConst;
+  APInt UnsignedMax = APInt::getMaxValue(NumDstBits).zext(NumSrcBits);
+
+  if (isLegal({TargetOpcode::G_TRUNC_SSAT_U, {DstTy, SrcTy}})) {
+    if (mi_match(Src, MRI,
+                 m_GSMin(m_GSMax(m_Reg(MatchInfo), m_SpecificICstOrSplat(0)),
+                         m_ICstOrSplat(MaxConst))) &&
+        APInt::isSameValue(MaxConst, UnsignedMax))
+      return true;
+    if (mi_match(Src, MRI,
+                 m_GSMax(m_GSMin(m_Reg(MatchInfo), m_ICstOrSplat(MaxConst)),
+                         m_SpecificICstOrSplat(0))) &&
+        APInt::isSameValue(MaxConst, UnsignedMax))
+      return true;
+    if (mi_match(Src, MRI,
+                 m_GUMin(m_GSMax(m_Reg(MatchInfo), m_SpecificICstOrSplat(0)),
+                         m_ICstOrSplat(MaxConst))) &&
+        APInt::isSameValue(MaxConst, UnsignedMax))
+      return true;
+  }
+  return false;
+}
+
+void CombinerHelper::applyTruncSSatU(MachineInstr &MI,
+                                     Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Builder.buildTruncSSatU(Dst, MatchInfo);
+  MI.eraseFromParent();
+}
+
+bool CombinerHelper::matchTruncUSatU(MachineInstr &MI,
+                                     Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Register Src = MI.getOperand(1).getReg();
+  LLT DstTy = MRI.getType(Dst);
+  LLT SrcTy = MRI.getType(Src);
+  unsigned NumDstBits = DstTy.getScalarSizeInBits();
+  unsigned NumSrcBits = SrcTy.getScalarSizeInBits();
+  assert(NumSrcBits > NumDstBits && "Unexpected types for truncate operation");
+
+  APInt MaxConst;
+  APInt UnsignedMax = APInt::getMaxValue(NumDstBits).zext(NumSrcBits);
+
+  if (isLegal({TargetOpcode::G_TRUNC_SSAT_U, {DstTy, SrcTy}})) {
+    if (mi_match(Src, MRI,
+                 m_GUMin(m_Reg(MatchInfo), m_ICstOrSplat(MaxConst))) &&
+        APInt::isSameValue(MaxConst, UnsignedMax))
+      return true;
+  }
+  return false;
+}
+
+void CombinerHelper::applyTruncUSatU(MachineInstr &MI,
+                                     Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Builder.buildTruncUSatU(Dst, MatchInfo);
+  MI.eraseFromParent();
+}
+
+bool CombinerHelper::matchTruncUSatUToFPTOUISat(MachineInstr &MI,
+                                                Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Register Src = MI.getOperand(1).getReg();
+  LLT DstTy = MRI.getType(Dst);
+  LLT SrcTy = MRI.getType(Src);
+
+  if (isLegalOrBeforeLegalizer({TargetOpcode::G_FPTOUI_SAT, {DstTy, SrcTy}})) {
+    if (mi_match(Src, MRI, m_GFPToUI((m_Reg(MatchInfo)))))
+      return true;
+  }
+  return false;
+}
+
+void CombinerHelper::applyTruncUSatUToFPTOUISat(MachineInstr &MI,
+                                                Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Builder.buildFPTOUI_SAT(Dst, MatchInfo);
+  MI.eraseFromParent();
+}
+
 bool CombinerHelper::matchRedundantNegOperands(MachineInstr &MI,
                                                BuildFnTy &MatchInfo) const {
   unsigned Opc = MI.getOpcode();
diff --git a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
index 473ba5e2fe11e..75805246c117d 100644
--- a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
@@ -793,6 +793,9 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
       .clampMinNumElements(0, s16, 4)
       .alwaysLegal();
 
+  getActionDefinitionsBuilder({G_TRUNC_SSAT_S, G_TRUNC_SSAT_U, G_TRUNC_USAT_U})
+      .legalFor({{v8s8, v8s16}, {v4s16, v4s32}, {v2s32, v2s64}});
+
   getActionDefinitionsBuilder(G_SEXT_INREG)
       .legalFor({s32, s64})
       .legalFor(PackedVectorAllTypeList)
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir b/llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir
index bd2d8c095831b..d4eec682a6276 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir
@@ -320,14 +320,16 @@
 # DEBUG-NEXT: .. type index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: .. imm index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: G_TRUNC_SSAT_S (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
-# DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
-# DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
+# DEBUG-NEXT: .. the first uncovered type index: 2, OK
+# DEBUG-NEXT: .. the first uncovered imm index: 0, OK
 # DEBUG-NEXT: G_TRUNC_SSAT_U (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
-# DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
-# DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
+# DEBUG-NEXT: .. opcode {{[0-9]+}} is aliased to {{[0-9]+}}
+# DEBUG-NEXT: .. the first uncovered type index: 2, OK
+# DEBUG-NEXT: .. the first uncovered imm index: 0, OK
 # DEBUG-NEXT: G_TRUNC_USAT_U (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
-# DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
-# DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
+# DEBUG-NEXT: .. opcode {{[0-9]+}} is aliased to {{[0-9]+}}
+# DEBUG-NEXT: .. the first uncovered type index: 2, OK
+# DEBUG-NEXT: .. the first uncovered imm index: 0, OK
 # DEBUG-NEXT: G_CONSTANT (opcode {{[0-9]+}}): 1 type index, 0 imm indices
 # DEBUG-NEXT: .. the first uncovered type index: 1, OK
 # DEBUG-NEXT: .. the first uncovered imm index: 0, OK
diff --git a/llvm/test/CodeGen/AArch64/truncsat.ll b/llvm/test/CodeGen/AArch64/truncsat.ll
new file mode 100644
index 0000000000000..19779aba2697e
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/truncsat.ll
@@ -0,0 +1,79 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -mtriple=aarch64-unknown-unknown -global-isel=0 | FileCheck %s --check-prefixes=CHECK,CHECK-SD
+; RUN: llc < %s -mtriple=aarch64-unknown-unknown -global-isel=1 | FileCheck %s --check-prefixes=CHECK,CHECK-GI
+
+
+define <4 x i16> @ssats_1(<4 x i32> %x) {
+; CHECK-LABEL: ssats_1:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sqxtn v0.4h, v0.4s
+; CHECK-NEXT:    ret
+entry:
+  %spec.store.select = call <4 x i32> @llvm.smin.v4i32(<4 x i32> %x, <4 x i32> <i32 32767, i32 32767, i32 32767, i32 32767>)
+  %spec.store.select7 = call <4 x i32> @llvm.smax.v4i32(<4 x i32> %spec.store.select, <4 x i32> <i32 -32768, i32 -32768, i32 -32768, i32 -32768>)
+  %conv6 = trunc <4 x i32> %spec.store.select7 to <4 x i16>
+  ret <4 x i16> %conv6
+}
+
+define <4 x i16> @ssats_2(<4 x i32> %x) {
+; CHECK-LABEL: ssats_2:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sqxtn v0.4h, v0.4s
+; CHECK-NEXT:    ret
+entry:
+  %spec.store.select = call <4 x i32> @llvm.smax.v4i32(<4 x i32> %x, <4 x i32> <i32 -32768, i32 -32768, i32 -32768, i32 -32768>)
+  %spec.store.select7 = call <4 x i32> @llvm.smin.v4i32(<4 x i32> %spec.store.select, <4 x i32> <i32 32767, i32 32767, i32 32767, i32 32767>)
+  %conv6 = trunc <4 x i32> %spec.store.select7 to <4 x i16>
+  ret <4 x i16> %conv6
+}
+
+define <4 x i16> @ssatu_1(<4 x i32> %x) {
+; CHECK-LABEL: ssatu_1:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sqxtun v0.4h, v0.4s
+; CHECK-NEXT:    ret
+entry:
+  %spec.store.select = call <4 x i32> @llvm.smin.v4i32(<4 x i32> %x, <4 x i32> <i32 65535, i32 65535, i32 65535, i32 65535>)
+  %spec.store.select7 = call <4 x i32> @llvm.smax.v4i32(<4 x i32> %spec.store.select, <4 x i32> zeroinitializer)
+  %conv6 = trunc <4 x i32> %spec.store.select7 to <4 x i16>
+  ret <4 x i16> %conv6
+}
+
+define <4 x i16> @ssatu_2(<4 x i32> %x) {
+; CHECK-LABEL: ssatu_2:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sqxtun v0.4h, v0.4s
+; CHECK-NEXT:    ret
+entry:
+  %spec.store.select = call <4 x i32> @llvm.smax.v4i32(<4 x i32> %x, <4 x i32> zeroinitializer)
+  %spec.store.select7 = call <4 x i32> @llvm.smin.v4i32(<4 x i32> %spec.store.select, <4 x i32> <i32 65535, i32 65535, i32 65535, i32 65535>)
+  %conv6 = trunc <4 x i32> %spec.store.select7 to <4 x i16>
+  ret <4 x i16> %conv6
+}
+
+define <4 x i16> @ssatu_3(<4 x i32> %x) {
+; CHECK-LABEL: ssatu_3:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sqxtun v0.4h, v0.4s
+; CHECK-NEXT:    ret
+entry:
+  %spec.store.select = call <4 x i32> @llvm.smax.v4i32(<4 x i32> %x, <4 x i32> zeroinitializer)
+  %spec.store.select7 = call <4 x i32> @llvm.umin.v4i32(<4 x i32> %spec.store.select, <4 x i32> <i32 65535, i32 65535, i32 65535, i32 65535>)
+  %conv6 = trunc <4 x i32> %spec.store.select7 to <4 x i16>
+  ret <4 x i16> %conv6
+}
+
+define <4 x i16> @usatu(<4 x i32> %x) {
+; CHECK-LABEL: usatu:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    uqxtn v0.4h, v0.4s
+; CHECK-NEXT:    ret
+entry:
+  %spec.store.select = call <4 x i32> @llvm.umin.v4i32(<4 x i32> %x, <4 x i32> <i32 65535, i32 65535, i32 65535, i32 65535>)
+  %conv6 = trunc <4 x i32> %spec.store.select to <4 x i16>
+  ret <4 x i16> %conv6
+}
+
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; CHECK-GI: {{.*}}
+; CHECK-SD: {{.*}}

llvmbot · 2025-07-23T13:27:40Z

@llvm/pr-subscribers-llvm-globalisel

Author: None (jyli0116)

Changes

Implements combining and legalization of G_TRUNC_SSAT_S, G_TRUNC_SSAT_U, and G_TRUNC_USAT_U, which where previously added to SDAG with the below patterns:

truncate(smin(smax(x, C1), C2)) -&gt; trunc_ssat_s(x)
truncate(smax(smin(x, C2), C1)) -&gt; trunc_ssat_s(x)

truncate(smax(smin(x, C), 0)) -&gt; trunc_ssat_u(x)
truncate(smin(smax(x, 0), C)) -&gt; trunc_ssat_u(x)
truncate(umin(smax(x, 0), C)) -&gt; trunc_ssat_u(x)

truncate(umin(x, C)) -&gt; trunc_usat_u(x)

Full diff: https://github.com/llvm/llvm-project/pull/150219.diff

9 Files Affected:

(modified) llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h (+19)
(modified) llvm/include/llvm/CodeGen/GlobalISel/GenericMachineInstrs.h (+27)
(modified) llvm/include/llvm/CodeGen/GlobalISel/MIPatternMatch.h (+12)
(modified) llvm/include/llvm/Target/GlobalISel/Combine.td (+27-1)
(modified) llvm/include/llvm/Target/GlobalISel/SelectionDAGCompat.td (+3)
(modified) llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp (+128)
(modified) llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp (+3)
(modified) llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir (+8-6)
(added) llvm/test/CodeGen/AArch64/truncsat.ll (+79)

diff --git a/llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h b/llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h
index 31f1197b9723b..12a0fc4a2e1c4 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h
@@ -725,6 +725,25 @@ class CombinerHelper {
   bool matchUMulHToLShr(MachineInstr &MI) const;
   void applyUMulHToLShr(MachineInstr &MI) const;
 
+  // Combine trunc(smin(smax(x, C1), C2)) -> truncssat_s(x)
+  // or      trunc(smax(smin(x, C2), C1)) -> truncssat_s(x).
+  bool matchTruncSSatS(MachineInstr &MI, Register &MatchInfo) const;
+  void applyTruncSSatS(MachineInstr &MI, Register &MatchInfo) const;
+
+  // Combine trunc(smin(smax(x, 0), C)) -> truncssat_u(x)
+  // or      trunc(smax(smin(x, C), 0)) -> truncssat_u(x)
+  // or      trunc(umin(smax(x, 0), C)) -> truncssat_u(x)
+  bool matchTruncSSatU(MachineInstr &MI, Register &MatchInfo) const;
+  void applyTruncSSatU(MachineInstr &MI, Register &MatchInfo) const;
+
+  // Combine trunc(umin(x, C)) -> truncusat_u(x).
+  bool matchTruncUSatU(MachineInstr &MI, Register &MatchInfo) const;
+  void applyTruncUSatU(MachineInstr &MI, Register &MatchInfo) const;
+
+  // Combine truncusat_u(fptoui(x)) -> fptoui_sat(x)
+  bool matchTruncUSatUToFPTOUISat(MachineInstr &MI, Register &MatchInfo) const;
+  void applyTruncUSatUToFPTOUISat(MachineInstr &MI, Register &MatchInfo) const;
+
   /// Try to transform \p MI by using all of the above
   /// combine functions. Returns true if changed.
   bool tryCombine(MachineInstr &MI) const;
diff --git a/llvm/include/llvm/CodeGen/GlobalISel/GenericMachineInstrs.h b/llvm/include/llvm/CodeGen/GlobalISel/GenericMachineInstrs.h
index 571ec6dd7e9ba..f5cd5beb43767 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/GenericMachineInstrs.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/GenericMachineInstrs.h
@@ -874,6 +874,9 @@ class GCastOp : public GenericMachineInstr {
     case TargetOpcode::G_SEXT:
     case TargetOpcode::G_SITOFP:
     case TargetOpcode::G_TRUNC:
+    case TargetOpcode::G_TRUNC_SSAT_S:
+    case TargetOpcode::G_TRUNC_SSAT_U:
+    case TargetOpcode::G_TRUNC_USAT_U:
     case TargetOpcode::G_UITOFP:
     case TargetOpcode::G_ZEXT:
     case TargetOpcode::G_ANYEXT:
@@ -916,6 +919,30 @@ class GTrunc : public GCastOp {
   };
 };
 
+/// Represents a saturated trunc from a signed input to a signed result.
+class GTruncSSatS : public GCastOp {
+public:
+  static bool classof(const MachineInstr *MI) {
+    return MI->getOpcode() == TargetOpcode::G_TRUNC_SSAT_S;
+  };
+};
+
+/// Represents a saturated trunc from a signed input to an unsigned result.
+class GTruncSSatU : public GCastOp {
+public:
+  static bool classof(const MachineInstr *MI) {
+    return MI->getOpcode() == TargetOpcode::G_TRUNC_SSAT_U;
+  };
+};
+
+/// Represents a saturated trunc from an unsigned input to an unsigned result.
+class GTruncUSatU : public GCastOp {
+public:
+  static bool classof(const MachineInstr *MI) {
+    return MI->getOpcode() == TargetOpcode::G_TRUNC_USAT_U;
+  };
+};
+
 /// Represents a vscale.
 class GVScale : public GenericMachineInstr {
 public:
diff --git a/llvm/include/llvm/CodeGen/GlobalISel/MIPatternMatch.h b/llvm/include/llvm/CodeGen/GlobalISel/MIPatternMatch.h
index c0d3a12cbcb41..a27c79af87c02 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/MIPatternMatch.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/MIPatternMatch.h
@@ -709,6 +709,18 @@ m_GFPTrunc(const SrcTy &Src) {
   return UnaryOp_match<SrcTy, TargetOpcode::G_FPTRUNC>(Src);
 }
 
+template <typename SrcTy>
+inline UnaryOp_match<SrcTy, TargetOpcode::G_FPTOSI>
+m_GFPToSI(const SrcTy &Src) {
+  return UnaryOp_match<SrcTy, TargetOpcode::G_FPTOSI>(Src);
+}
+
+template <typename SrcTy>
+inline UnaryOp_match<SrcTy, TargetOpcode::G_FPTOUI>
+m_GFPToUI(const SrcTy &Src) {
+  return UnaryOp_match<SrcTy, TargetOpcode::G_FPTOUI>(Src);
+}
+
 template <typename SrcTy>
 inline UnaryOp_match<SrcTy, TargetOpcode::G_FABS> m_GFabs(const SrcTy &Src) {
   return UnaryOp_match<SrcTy, TargetOpcode::G_FABS>(Src);
diff --git a/llvm/include/llvm/Target/GlobalISel/Combine.td b/llvm/include/llvm/Target/GlobalISel/Combine.td
index 66051d756c808..61df21630a3b4 100644
--- a/llvm/include/llvm/Target/GlobalISel/Combine.td
+++ b/llvm/include/llvm/Target/GlobalISel/Combine.td
@@ -1236,6 +1236,32 @@ def mulh_to_lshr : GICombineRule<
 
 def mulh_combines : GICombineGroup<[mulh_to_lshr]>;
 
+def trunc_ssats : GICombineRule<
+  (defs root:$root, register_matchinfo:$matchinfo),
+  (match (G_TRUNC $dst, $src):$root,
+         [{ return Helper.matchTruncSSatS(*${root}, ${matchinfo}); }]),
+  (apply [{ Helper.applyTruncSSatS(*${root}, ${matchinfo}); }])>;
+
+def trunc_ssatu : GICombineRule<
+  (defs root:$root, register_matchinfo:$matchinfo),
+  (match (G_TRUNC $dst, $src):$root,
+    [{ return Helper.matchTruncSSatU(*${root}, ${matchinfo}); }]),
+  (apply [{ Helper.applyTruncSSatU(*${root}, ${matchinfo}); }])>;
+
+def trunc_usatu : GICombineRule<
+  (defs root:$root, register_matchinfo:$matchinfo),
+  (match (G_TRUNC $dst, $src):$root,
+    [{ return Helper.matchTruncUSatU(*${root}, ${matchinfo}); }]),
+  (apply [{ Helper.applyTruncUSatU(*${root}, ${matchinfo}); }])>;
+
+def truncusatu_to_fptouisat : GICombineRule<
+  (defs root:$root, register_matchinfo:$matchinfo),
+  (match (G_TRUNC_USAT_U $dst, $src):$root,
+    [{ return Helper.matchTruncUSatUToFPTOUISat(*${root}, ${matchinfo}); }]),
+  (apply [{ Helper.applyTruncUSatUToFPTOUISat(*${root}, ${matchinfo}); }])>;
+
+def truncsat_combines : GICombineGroup<[trunc_ssats, trunc_ssatu, trunc_usatu, truncusatu_to_fptouisat]>;
+
 def redundant_neg_operands: GICombineRule<
   (defs root:$root, build_fn_matchinfo:$matchinfo),
   (match (wip_match_opcode G_FADD, G_FSUB, G_FMUL, G_FDIV, G_FMAD, G_FMA):$root,
@@ -2060,7 +2086,7 @@ def all_combines : GICombineGroup<[integer_reassoc_combines, trivial_combines,
     fsub_to_fneg, commute_constant_to_rhs, match_ands, match_ors,
     simplify_neg_minmax, combine_concat_vector,
     sext_trunc, zext_trunc, prefer_sign_combines, shuffle_combines,
-    combine_use_vector_truncate, merge_combines, overflow_combines]>;
+    combine_use_vector_truncate, merge_combines, overflow_combines, truncsat_combines]>;
 
 // A combine group used to for prelegalizer combiners at -O0. The combines in
 // this group have been selected based on experiments to balance code size and
diff --git a/llvm/include/llvm/Target/GlobalISel/SelectionDAGCompat.td b/llvm/include/llvm/Target/GlobalISel/SelectionDAGCompat.td
index 7577792003d2e..e38a814b6e7b9 100644
--- a/llvm/include/llvm/Target/GlobalISel/SelectionDAGCompat.td
+++ b/llvm/include/llvm/Target/GlobalISel/SelectionDAGCompat.td
@@ -51,6 +51,9 @@ def : GINodeEquiv<G_SEXT, sext>;
 def : GINodeEquiv<G_SEXT_INREG, sext_inreg>;
 def : GINodeEquiv<G_ZEXT, zext>;
 def : GINodeEquiv<G_TRUNC, trunc>;
+def : GINodeEquiv<G_TRUNC_SSAT_S, truncssat_s>;
+def : GINodeEquiv<G_TRUNC_SSAT_U, truncssat_u>;
+def : GINodeEquiv<G_TRUNC_USAT_U, truncusat_u>;
 def : GINodeEquiv<G_BITCAST, bitconvert>;
 // G_INTTOPTR - SelectionDAG has no equivalent.
 // G_PTRTOINT - SelectionDAG has no equivalent.
diff --git a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
index 3922eba55e195..45914e717c3e6 100644
--- a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
@@ -5831,6 +5831,134 @@ void CombinerHelper::applyUMulHToLShr(MachineInstr &MI) const {
   MI.eraseFromParent();
 }
 
+bool CombinerHelper::matchTruncSSatS(MachineInstr &MI,
+                                     Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Register Src = MI.getOperand(1).getReg();
+  LLT DstTy = MRI.getType(Dst);
+  LLT SrcTy = MRI.getType(Src);
+  unsigned NumDstBits = DstTy.getScalarSizeInBits();
+  unsigned NumSrcBits = SrcTy.getScalarSizeInBits();
+  assert(NumSrcBits > NumDstBits && "Unexpected types for truncate operation");
+
+  APInt MinConst, MaxConst;
+  APInt SignedMax = APInt::getSignedMaxValue(NumDstBits).sext(NumSrcBits);
+  APInt SignedMin = APInt::getSignedMinValue(NumDstBits).sext(NumSrcBits);
+
+  if (isLegal({TargetOpcode::G_TRUNC_SSAT_S, {DstTy, SrcTy}})) {
+    if (mi_match(Src, MRI,
+                 m_GSMin(m_GSMax(m_Reg(MatchInfo), m_ICstOrSplat(MinConst)),
+                         m_ICstOrSplat(MaxConst))) &&
+        APInt::isSameValue(MinConst, SignedMin) &&
+        APInt::isSameValue(MaxConst, SignedMax))
+      return true;
+    if (mi_match(Src, MRI,
+                 m_GSMax(m_GSMin(m_Reg(MatchInfo), m_ICstOrSplat(MaxConst)),
+                         m_ICstOrSplat(MinConst))) &&
+        APInt::isSameValue(MinConst, SignedMin) &&
+        APInt::isSameValue(MaxConst, SignedMax))
+      return true;
+  }
+  return false;
+}
+
+void CombinerHelper::applyTruncSSatS(MachineInstr &MI,
+                                     Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Builder.buildTruncSSatS(Dst, MatchInfo);
+  MI.eraseFromParent();
+}
+
+bool CombinerHelper::matchTruncSSatU(MachineInstr &MI,
+                                     Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Register Src = MI.getOperand(1).getReg();
+  LLT DstTy = MRI.getType(Dst);
+  LLT SrcTy = MRI.getType(Src);
+  unsigned NumDstBits = DstTy.getScalarSizeInBits();
+  unsigned NumSrcBits = SrcTy.getScalarSizeInBits();
+  assert(NumSrcBits > NumDstBits && "Unexpected types for truncate operation");
+
+  APInt MaxConst;
+  APInt UnsignedMax = APInt::getMaxValue(NumDstBits).zext(NumSrcBits);
+
+  if (isLegal({TargetOpcode::G_TRUNC_SSAT_U, {DstTy, SrcTy}})) {
+    if (mi_match(Src, MRI,
+                 m_GSMin(m_GSMax(m_Reg(MatchInfo), m_SpecificICstOrSplat(0)),
+                         m_ICstOrSplat(MaxConst))) &&
+        APInt::isSameValue(MaxConst, UnsignedMax))
+      return true;
+    if (mi_match(Src, MRI,
+                 m_GSMax(m_GSMin(m_Reg(MatchInfo), m_ICstOrSplat(MaxConst)),
+                         m_SpecificICstOrSplat(0))) &&
+        APInt::isSameValue(MaxConst, UnsignedMax))
+      return true;
+    if (mi_match(Src, MRI,
+                 m_GUMin(m_GSMax(m_Reg(MatchInfo), m_SpecificICstOrSplat(0)),
+                         m_ICstOrSplat(MaxConst))) &&
+        APInt::isSameValue(MaxConst, UnsignedMax))
+      return true;
+  }
+  return false;
+}
+
+void CombinerHelper::applyTruncSSatU(MachineInstr &MI,
+                                     Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Builder.buildTruncSSatU(Dst, MatchInfo);
+  MI.eraseFromParent();
+}
+
+bool CombinerHelper::matchTruncUSatU(MachineInstr &MI,
+                                     Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Register Src = MI.getOperand(1).getReg();
+  LLT DstTy = MRI.getType(Dst);
+  LLT SrcTy = MRI.getType(Src);
+  unsigned NumDstBits = DstTy.getScalarSizeInBits();
+  unsigned NumSrcBits = SrcTy.getScalarSizeInBits();
+  assert(NumSrcBits > NumDstBits && "Unexpected types for truncate operation");
+
+  APInt MaxConst;
+  APInt UnsignedMax = APInt::getMaxValue(NumDstBits).zext(NumSrcBits);
+
+  if (isLegal({TargetOpcode::G_TRUNC_SSAT_U, {DstTy, SrcTy}})) {
+    if (mi_match(Src, MRI,
+                 m_GUMin(m_Reg(MatchInfo), m_ICstOrSplat(MaxConst))) &&
+        APInt::isSameValue(MaxConst, UnsignedMax))
+      return true;
+  }
+  return false;
+}
+
+void CombinerHelper::applyTruncUSatU(MachineInstr &MI,
+                                     Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Builder.buildTruncUSatU(Dst, MatchInfo);
+  MI.eraseFromParent();
+}
+
+bool CombinerHelper::matchTruncUSatUToFPTOUISat(MachineInstr &MI,
+                                                Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Register Src = MI.getOperand(1).getReg();
+  LLT DstTy = MRI.getType(Dst);
+  LLT SrcTy = MRI.getType(Src);
+
+  if (isLegalOrBeforeLegalizer({TargetOpcode::G_FPTOUI_SAT, {DstTy, SrcTy}})) {
+    if (mi_match(Src, MRI, m_GFPToUI((m_Reg(MatchInfo)))))
+      return true;
+  }
+  return false;
+}
+
+void CombinerHelper::applyTruncUSatUToFPTOUISat(MachineInstr &MI,
+                                                Register &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  Builder.buildFPTOUI_SAT(Dst, MatchInfo);
+  MI.eraseFromParent();
+}
+
 bool CombinerHelper::matchRedundantNegOperands(MachineInstr &MI,
                                                BuildFnTy &MatchInfo) const {
   unsigned Opc = MI.getOpcode();
diff --git a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
index 473ba5e2fe11e..75805246c117d 100644
--- a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
@@ -793,6 +793,9 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
       .clampMinNumElements(0, s16, 4)
       .alwaysLegal();
 
+  getActionDefinitionsBuilder({G_TRUNC_SSAT_S, G_TRUNC_SSAT_U, G_TRUNC_USAT_U})
+      .legalFor({{v8s8, v8s16}, {v4s16, v4s32}, {v2s32, v2s64}});
+
   getActionDefinitionsBuilder(G_SEXT_INREG)
       .legalFor({s32, s64})
       .legalFor(PackedVectorAllTypeList)
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir b/llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir
index bd2d8c095831b..d4eec682a6276 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir
@@ -320,14 +320,16 @@
 # DEBUG-NEXT: .. type index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: .. imm index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: G_TRUNC_SSAT_S (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
-# DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
-# DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
+# DEBUG-NEXT: .. the first uncovered type index: 2, OK
+# DEBUG-NEXT: .. the first uncovered imm index: 0, OK
 # DEBUG-NEXT: G_TRUNC_SSAT_U (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
-# DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
-# DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
+# DEBUG-NEXT: .. opcode {{[0-9]+}} is aliased to {{[0-9]+}}
+# DEBUG-NEXT: .. the first uncovered type index: 2, OK
+# DEBUG-NEXT: .. the first uncovered imm index: 0, OK
 # DEBUG-NEXT: G_TRUNC_USAT_U (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
-# DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
-# DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
+# DEBUG-NEXT: .. opcode {{[0-9]+}} is aliased to {{[0-9]+}}
+# DEBUG-NEXT: .. the first uncovered type index: 2, OK
+# DEBUG-NEXT: .. the first uncovered imm index: 0, OK
 # DEBUG-NEXT: G_CONSTANT (opcode {{[0-9]+}}): 1 type index, 0 imm indices
 # DEBUG-NEXT: .. the first uncovered type index: 1, OK
 # DEBUG-NEXT: .. the first uncovered imm index: 0, OK
diff --git a/llvm/test/CodeGen/AArch64/truncsat.ll b/llvm/test/CodeGen/AArch64/truncsat.ll
new file mode 100644
index 0000000000000..19779aba2697e
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/truncsat.ll
@@ -0,0 +1,79 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -mtriple=aarch64-unknown-unknown -global-isel=0 | FileCheck %s --check-prefixes=CHECK,CHECK-SD
+; RUN: llc < %s -mtriple=aarch64-unknown-unknown -global-isel=1 | FileCheck %s --check-prefixes=CHECK,CHECK-GI
+
+
+define <4 x i16> @ssats_1(<4 x i32> %x) {
+; CHECK-LABEL: ssats_1:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sqxtn v0.4h, v0.4s
+; CHECK-NEXT:    ret
+entry:
+  %spec.store.select = call <4 x i32> @llvm.smin.v4i32(<4 x i32> %x, <4 x i32> <i32 32767, i32 32767, i32 32767, i32 32767>)
+  %spec.store.select7 = call <4 x i32> @llvm.smax.v4i32(<4 x i32> %spec.store.select, <4 x i32> <i32 -32768, i32 -32768, i32 -32768, i32 -32768>)
+  %conv6 = trunc <4 x i32> %spec.store.select7 to <4 x i16>
+  ret <4 x i16> %conv6
+}
+
+define <4 x i16> @ssats_2(<4 x i32> %x) {
+; CHECK-LABEL: ssats_2:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sqxtn v0.4h, v0.4s
+; CHECK-NEXT:    ret
+entry:
+  %spec.store.select = call <4 x i32> @llvm.smax.v4i32(<4 x i32> %x, <4 x i32> <i32 -32768, i32 -32768, i32 -32768, i32 -32768>)
+  %spec.store.select7 = call <4 x i32> @llvm.smin.v4i32(<4 x i32> %spec.store.select, <4 x i32> <i32 32767, i32 32767, i32 32767, i32 32767>)
+  %conv6 = trunc <4 x i32> %spec.store.select7 to <4 x i16>
+  ret <4 x i16> %conv6
+}
+
+define <4 x i16> @ssatu_1(<4 x i32> %x) {
+; CHECK-LABEL: ssatu_1:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sqxtun v0.4h, v0.4s
+; CHECK-NEXT:    ret
+entry:
+  %spec.store.select = call <4 x i32> @llvm.smin.v4i32(<4 x i32> %x, <4 x i32> <i32 65535, i32 65535, i32 65535, i32 65535>)
+  %spec.store.select7 = call <4 x i32> @llvm.smax.v4i32(<4 x i32> %spec.store.select, <4 x i32> zeroinitializer)
+  %conv6 = trunc <4 x i32> %spec.store.select7 to <4 x i16>
+  ret <4 x i16> %conv6
+}
+
+define <4 x i16> @ssatu_2(<4 x i32> %x) {
+; CHECK-LABEL: ssatu_2:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sqxtun v0.4h, v0.4s
+; CHECK-NEXT:    ret
+entry:
+  %spec.store.select = call <4 x i32> @llvm.smax.v4i32(<4 x i32> %x, <4 x i32> zeroinitializer)
+  %spec.store.select7 = call <4 x i32> @llvm.smin.v4i32(<4 x i32> %spec.store.select, <4 x i32> <i32 65535, i32 65535, i32 65535, i32 65535>)
+  %conv6 = trunc <4 x i32> %spec.store.select7 to <4 x i16>
+  ret <4 x i16> %conv6
+}
+
+define <4 x i16> @ssatu_3(<4 x i32> %x) {
+; CHECK-LABEL: ssatu_3:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sqxtun v0.4h, v0.4s
+; CHECK-NEXT:    ret
+entry:
+  %spec.store.select = call <4 x i32> @llvm.smax.v4i32(<4 x i32> %x, <4 x i32> zeroinitializer)
+  %spec.store.select7 = call <4 x i32> @llvm.umin.v4i32(<4 x i32> %spec.store.select, <4 x i32> <i32 65535, i32 65535, i32 65535, i32 65535>)
+  %conv6 = trunc <4 x i32> %spec.store.select7 to <4 x i16>
+  ret <4 x i16> %conv6
+}
+
+define <4 x i16> @usatu(<4 x i32> %x) {
+; CHECK-LABEL: usatu:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    uqxtn v0.4h, v0.4s
+; CHECK-NEXT:    ret
+entry:
+  %spec.store.select = call <4 x i32> @llvm.umin.v4i32(<4 x i32> %x, <4 x i32> <i32 65535, i32 65535, i32 65535, i32 65535>)
+  %conv6 = trunc <4 x i32> %spec.store.select to <4 x i16>
+  ret <4 x i16> %conv6
+}
+
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; CHECK-GI: {{.*}}
+; CHECK-SD: {{.*}}

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

llvm/test/CodeGen/AArch64/truncsat.ll

github-actions · 2025-08-06T09:44:50Z

✅ With the latest revision this PR passed the C/C++ code formatter.

llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

llvm/test/CodeGen/AArch64/fpclamptosat_vec.ll

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

llvm/test/CodeGen/AArch64/neon-qmovn.ll

llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp

llvm/include/llvm/CodeGen/GlobalISel/GenericMachineInstrs.h

llvm/include/llvm/CodeGen/GlobalISel/MIPatternMatch.h

jyli0116 · 2025-08-19T13:01:43Z

The size of the patch is getting rather large - I'm going to split it between three patches: two NFC patches for adding the tests and the legalizer respectively, with the combiner code being left within this patch

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

davemgreen

Other than the comments below, LGTM.

llvm/lib/Target/AArch64/AArch64Combine.td

llvm/test/CodeGen/AArch64/neon-qmovn.ll

llvmbot added backend:AArch64 llvm:globalisel labels Jul 23, 2025

jyli0116 requested a review from davemgreen July 23, 2025 14:59

jyli0116 force-pushed the GISel_truncsat branch from da22a7b to 5dfd1e6 Compare July 23, 2025 15:30

davemgreen requested review from aemerson, arsenm and jayfoad July 23, 2025 15:31

davemgreen reviewed Jul 23, 2025

View reviewed changes

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp Outdated Show resolved Hide resolved

llvm/test/CodeGen/AArch64/truncsat.ll Outdated Show resolved Hide resolved

jyli0116 force-pushed the GISel_truncsat branch from 5dfd1e6 to 4cc82c8 Compare August 6, 2025 09:42

jyli0116 requested a review from davemgreen August 6, 2025 12:02

davemgreen reviewed Aug 7, 2025

View reviewed changes

llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp Outdated Show resolved Hide resolved

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp Outdated Show resolved Hide resolved

llvm/test/CodeGen/AArch64/fpclamptosat_vec.ll Outdated Show resolved Hide resolved

arsenm reviewed Aug 7, 2025

View reviewed changes

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp Outdated Show resolved Hide resolved

arsenm requested a review from Pierre-vh August 7, 2025 08:10

arsenm reviewed Aug 8, 2025

View reviewed changes

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp Outdated Show resolved Hide resolved

jyli0116 force-pushed the GISel_truncsat branch from d4345fd to e2eede7 Compare August 12, 2025 12:41

jyli0116 requested review from arsenm and davemgreen August 12, 2025 13:47

davemgreen reviewed Aug 13, 2025

View reviewed changes

arsenm reviewed Aug 20, 2025

View reviewed changes

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp Outdated Show resolved Hide resolved

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp Outdated Show resolved Hide resolved

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp Outdated Show resolved Hide resolved

jyli0116 added 5 commits August 20, 2025 10:03

[GlobalISel] Support saturated truncate

e1e1b98

resolved comments

684ffb1

formatting

6d1f9c7

resolved comments

69af41f

rebased and resolved comments

bba8956

jyli0116 force-pushed the GISel_truncsat branch from 9872a19 to bba8956 Compare August 20, 2025 14:57

arsenm reviewed Aug 20, 2025

View reviewed changes

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp Outdated Show resolved Hide resolved

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp Outdated Show resolved Hide resolved

early return and reduce indent

4b7fb8b

davemgreen approved these changes Aug 21, 2025

View reviewed changes

llvm/lib/Target/AArch64/AArch64Combine.td Outdated Show resolved Hide resolved

llvm/test/CodeGen/AArch64/neon-qmovn.ll Outdated Show resolved Hide resolved

formatting and test coverage

abd32cd

jyli0116 merged commit dbadab9 into llvm:main Aug 21, 2025
9 checks passed

[GlobalISel] Support saturated truncate #150219

[GlobalISel] Support saturated truncate #150219

Uh oh!

Conversation

jyli0116 commented Jul 23, 2025

Uh oh!

llvmbot commented Jul 23, 2025

Uh oh!

llvmbot commented Jul 23, 2025

Uh oh!

Uh oh!

Uh oh!

github-actions bot commented Aug 6, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

jyli0116 commented Aug 19, 2025

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

davemgreen left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

github-actions bot commented Aug 6, 2025 •

edited

Loading