[AArch64][GlobalISel] Combine MUL(AND(LSHR(X, 15), 0x10001), 0xffff) to CMLTz

[chuongg3]

No description provided.

[llvmbot]

@llvm/pr-subscribers-backend-aarch64

Author: None (chuongg3)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/92915.diff

3 Files Affected:

• (modified) llvm/lib/Target/AArch64/AArch64Combine.td (+10-1)
• (modified) llvm/lib/Target/AArch64/GISel/AArch64PostLegalizerCombiner.cpp (+54)
• (modified) llvm/test/CodeGen/AArch64/mulcmle.ll (+16-5)

diff --git a/llvm/lib/Target/AArch64/AArch64Combine.td b/llvm/lib/Target/AArch64/AArch64Combine.td
index 10cad6d192440..bee9b07b9d230 100644
--- a/llvm/lib/Target/AArch64/AArch64Combine.td
+++ b/llvm/lib/Target/AArch64/AArch64Combine.td
@@ -265,6 +265,14 @@ def or_to_bsp: GICombineRule <
   (apply [{ applyOrToBSP(*${root}, MRI, B, ${matchinfo}); }])
 >;
 
+// Combines Mul(And(Srl(X, 15), 0x10001), 0xffff) into CMLTz
+def combine_mul_cmlt : GICombineRule<
+  (defs root:$root, register_matchinfo:$matchinfo),
+  (match (wip_match_opcode G_MUL):$root,
+        [{ return matchCombineMulCMLT(*${root}, MRI, ${matchinfo}); }]),
+  (apply [{ applyCombineMulCMLT(*${root}, MRI, B, ${matchinfo}); }])
+>;
+
 // Post-legalization combines which should happen at all optimization levels.
 // (E.g. ones that facilitate matching for the selector) For example, matching
 // pseudos.
@@ -295,5 +303,6 @@ def AArch64PostLegalizerCombiner
                         ptr_add_immed_chain, overlapping_and,
                         split_store_zero_128, undef_combines,
                         select_to_minmax, or_to_bsp, combine_concat_vector,
-                        commute_constant_to_rhs]> {
+                        commute_constant_to_rhs,
+                        combine_mul_cmlt]> {
 }
diff --git a/llvm/lib/Target/AArch64/GISel/AArch64PostLegalizerCombiner.cpp b/llvm/lib/Target/AArch64/GISel/AArch64PostLegalizerCombiner.cpp
index d8ca5494ba50a..82f2904ad8d43 100644
--- a/llvm/lib/Target/AArch64/GISel/AArch64PostLegalizerCombiner.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64PostLegalizerCombiner.cpp
@@ -381,6 +381,60 @@ void applyOrToBSP(MachineInstr &MI, MachineRegisterInfo &MRI,
   MI.eraseFromParent();
 }
 
+bool matchCombineMulCMLT(MachineInstr &MI, MachineRegisterInfo &MRI,
+                         Register &SrcReg) {
+  LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
+
+  if (DstTy != LLT::fixed_vector(2, 64) && DstTy != LLT::fixed_vector(2, 32) &&
+      DstTy != LLT::fixed_vector(4, 32) && DstTy != LLT::fixed_vector(4, 16) &&
+      DstTy != LLT::fixed_vector(8, 16))
+    return false;
+
+  auto AndMI = getDefIgnoringCopies(MI.getOperand(1).getReg(), MRI);
+  if (AndMI->getOpcode() != TargetOpcode::G_AND)
+    return false;
+  auto LShrMI = getDefIgnoringCopies(AndMI->getOperand(1).getReg(), MRI);
+  if (LShrMI->getOpcode() != TargetOpcode::G_LSHR)
+    return false;
+
+  // Check the constant splat values
+  auto V1 = isConstantOrConstantSplatVector(
+      *MRI.getVRegDef(MI.getOperand(2).getReg()), MRI);
+  auto V2 = isConstantOrConstantSplatVector(
+      *MRI.getVRegDef(AndMI->getOperand(2).getReg()), MRI);
+  auto V3 = isConstantOrConstantSplatVector(
+      *MRI.getVRegDef(LShrMI->getOperand(2).getReg()), MRI);
+  if (!V1.has_value() || !V2.has_value() || !V3.has_value())
+    return false;
+  unsigned HalfSize = DstTy.getScalarSizeInBits() / 2;
+  if (!V1.value().isMask(HalfSize) || V2.value() != (1ULL | 1ULL << HalfSize) ||
+      V3 != (HalfSize - 1))
+    return false;
+
+  SrcReg = LShrMI->getOperand(1).getReg();
+
+  return true;
+}
+
+void applyCombineMulCMLT(MachineInstr &MI, MachineRegisterInfo &MRI,
+                         MachineIRBuilder &B, Register &SrcReg) {
+  Register DstReg = MI.getOperand(0).getReg();
+  LLT DstTy = MRI.getType(DstReg);
+  LLT HalfTy =
+      DstTy.changeElementCount(DstTy.getElementCount().multiplyCoefficientBy(2))
+          .changeElementSize(DstTy.getScalarSizeInBits() / 2);
+
+  Register ZeroVec = B.buildConstant(HalfTy, 0).getReg(0);
+  Register CastReg =
+      B.buildInstr(TargetOpcode::G_BITCAST, {HalfTy}, {SrcReg}).getReg(0);
+  Register CMLTReg =
+      B.buildICmp(CmpInst::Predicate::ICMP_SLT, HalfTy, CastReg, ZeroVec)
+          .getReg(0);
+
+  B.buildInstr(TargetOpcode::G_BITCAST, {DstReg}, {CMLTReg}).getReg(0);
+  MI.eraseFromParent();
+}
+
 class AArch64PostLegalizerCombinerImpl : public Combiner {
 protected:
   // TODO: Make CombinerHelper methods const.
diff --git a/llvm/test/CodeGen/AArch64/mulcmle.ll b/llvm/test/CodeGen/AArch64/mulcmle.ll
index 5c216b8550080..32bc5c5e63b3e 100644
--- a/llvm/test/CodeGen/AArch64/mulcmle.ll
+++ b/llvm/test/CodeGen/AArch64/mulcmle.ll
@@ -1,11 +1,22 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=aarch64 %s -o - | FileCheck %s
+; RUN: llc -mtriple=aarch64 %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-SD
+; RUN: llc -mtriple=aarch64 %s -o - -global-isel -global-isel-abort=2 2>&1 | FileCheck %s --check-prefixes=CHECK,CHECK-GI
 
 define <1 x i64> @v1i64(<1 x i64> %a) {
-; CHECK-LABEL: v1i64:
-; CHECK:       // %bb.0:
-; CHECK-NEXT:    cmlt v0.2s, v0.2s, #0
-; CHECK-NEXT:    ret
+; CHECK-SD-LABEL: v1i64:
+; CHECK-SD:       // %bb.0:
+; CHECK-SD-NEXT:    cmlt v0.2s, v0.2s, #0
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: v1i64:
+; CHECK-GI:       // %bb.0:
+; CHECK-GI-NEXT:    fmov x8, d0
+; CHECK-GI-NEXT:    lsr x8, x8, #31
+; CHECK-GI-NEXT:    and x8, x8, #0x100000001
+; CHECK-GI-NEXT:    lsl x9, x8, #32
+; CHECK-GI-NEXT:    sub x8, x9, x8
+; CHECK-GI-NEXT:    fmov d0, x8
+; CHECK-GI-NEXT:    ret
   %b = lshr <1 x i64> %a, <i64 31>
   %c = and <1 x i64> %b, <i64 4294967297>
   %d = mul nuw <1 x i64> %c, <i64 4294967295>

[davemgreen]

It might be good to say this is the equivalent of https://reviews.llvm.org/D130874 in the commit message.

[madhur13490]

Are you planning to optimize this sequence in some further patch? I see this is sub-optimal compared to SDAG.

[chuongg3]

In GlobalISel, all v1 types are treated as scalar values and we do not necessarily want the same optimization for scalar types.

[davemgreen]

LGTM, thanks