[AArch64] Make machine combiner patterns preserve MIFlags

This is mainly done so that we don't lose the nofpexcept flag once we start emitting it. Differential Revision: https://reviews.llvm.org/D118621
llvm · Feb 3, 2022 · 94843ea · 94843ea
1 parent c39969e
commit 94843ea
Show file tree

Hide file tree

Showing 4 changed files with 150 additions and 8 deletions.
diff --git a/llvm/lib/CodeGen/TargetInstrInfo.cpp b/llvm/lib/CodeGen/TargetInstrInfo.cpp
@@ -873,11 +873,13 @@ void TargetInstrInfo::reassociateOps(
   MachineInstrBuilder MIB1 =
       BuildMI(*MF, Prev.getDebugLoc(), TII->get(Opcode), NewVR)
           .addReg(RegX, getKillRegState(KillX))
-          .addReg(RegY, getKillRegState(KillY));
+          .addReg(RegY, getKillRegState(KillY))
+          .setMIFlags(Prev.getFlags());
   MachineInstrBuilder MIB2 =
       BuildMI(*MF, Root.getDebugLoc(), TII->get(Opcode), RegC)
           .addReg(RegA, getKillRegState(KillA))
-          .addReg(NewVR, getKillRegState(true));
+          .addReg(NewVR, getKillRegState(true))
+          .setMIFlags(Root.getFlags());
 
   setSpecialOperandAttr(Root, Prev, *MIB1, *MIB2);
 

diff --git a/llvm/lib/Target/AArch64/AArch64InstrInfo.cpp b/llvm/lib/Target/AArch64/AArch64InstrInfo.cpp
@@ -6214,6 +6214,14 @@ void AArch64InstrInfo::genAlternativeCodeSequence(
   if (MUL)
     DelInstrs.push_back(MUL);
   DelInstrs.push_back(&Root);
+
+  // Set the flags on the inserted instructions to be the merged flags of the
+  // instructions that we have combined.
+  uint16_t Flags = Root.getFlags();
+  if (MUL)
+    Flags = Root.mergeFlagsWith(*MUL);
+  for (auto *MI : InsInstrs)
+    MI->setFlags(Flags);
 }
 
 /// Replace csincr-branch sequence by simple conditional branch

diff --git a/llvm/test/CodeGen/AArch64/machine-combiner-instr-fmf.mir b/llvm/test/CodeGen/AArch64/machine-combiner-instr-fmf.mir
@@ -6,7 +6,7 @@
 # CHECK:        [[C:%.*]]:fpr32 = COPY $s2
 # CHECK-NEXT:   [[B:%.*]]:fpr32 = COPY $s1
 # CHECK-NEXT:   [[A:%.*]]:fpr32 = COPY $s0
-# CHECK-NEXT:   :fpr32 = FMADDSrrr [[B]], [[A]], [[C]]
+# CHECK-NEXT:   :fpr32 = nnan ninf nsz arcp contract afn reassoc FMADDSrrr [[B]], [[A]], [[C]]
 ---
 name:            scalar_fmadd_fast
 alignment:       4
@@ -45,7 +45,7 @@ body:             |
 # CHECK:        [[C:%.*]]:fpr32 = COPY $s2
 # CHECK-NEXT:   [[B:%.*]]:fpr32 = COPY $s1
 # CHECK-NEXT:   [[A:%.*]]:fpr32 = COPY $s0
-# CHECK-NEXT:   :fpr32 = FMADDSrrr [[B]], [[A]], [[C]]
+# CHECK-NEXT:   :fpr32 = contract FMADDSrrr [[B]], [[A]], [[C]]
 
 ---
 name:            scalar_fmadd_contract
@@ -125,7 +125,7 @@ body:             |
 # CHECK:        [[C:%.*]]:fpr32 = COPY $s2
 # CHECK-NEXT:   [[B:%.*]]:fpr32 = COPY $s1
 # CHECK-NEXT:   [[A:%.*]]:fpr32 = COPY $s0
-# CHECK-NEXT:   :fpr32 = FMADDSrrr [[B]], [[A]], [[C]]
+# CHECK-NEXT:   :fpr32 = contract FMADDSrrr [[B]], [[A]], [[C]]
 
 ---
 name:            scalar_fmadd_contract_op1
@@ -206,7 +206,7 @@ body:             |
 # CHECK:       [[C:%.*]]:fpr128 = COPY $q2
 # CHECK-NEXT:  [[B:%.*]]:fpr128 = COPY $q1
 # CHECK-NEXT:  [[A:%.*]]:fpr128 = COPY $q0
-# CHECK-NEXT:  fpr128 = FMLAv2f64 [[C]], [[B]], [[A]]
+# CHECK-NEXT:  fpr128 = nnan ninf nsz arcp contract afn reassoc FMLAv2f64 [[C]], [[B]], [[A]]
 ---
 name:            vector_fmadd_fast
 alignment:       4
@@ -245,7 +245,7 @@ body:             |
 # CHECK:       [[C:%.*]]:fpr128 = COPY $q2
 # CHECK-NEXT:  [[B:%.*]]:fpr128 = COPY $q1
 # CHECK-NEXT:  [[A:%.*]]:fpr128 = COPY $q0
-# CHECK-NEXT:  fpr128 = FMLAv2f64 [[C]], [[B]], [[A]]
+# CHECK-NEXT:  fpr128 = contract FMLAv2f64 [[C]], [[B]], [[A]]
 ---
 name:            vector_fmadd_contract
 alignment:       4
@@ -324,7 +324,7 @@ body:             |
 # CHECK:       [[C:%.*]]:fpr128 = COPY $q2
 # CHECK-NEXT:  [[B:%.*]]:fpr128 = COPY $q1
 # CHECK-NEXT:  [[A:%.*]]:fpr128 = COPY $q0
-# CHECK-NEXT:  fpr128 = FMLAv2f64 [[C]], [[B]], [[A]]
+# CHECK-NEXT:  fpr128 = contract FMLAv2f64 [[C]], [[B]], [[A]]
 
 ---
 name:            vector_fmadd_contract_op1

diff --git a/llvm/test/CodeGen/AArch64/machine-combiner-reassociate.mir b/llvm/test/CodeGen/AArch64/machine-combiner-reassociate.mir
@@ -0,0 +1,132 @@
+# RUN: llc -run-pass=machine-combiner -mtriple=aarch64-unknown-linux-gnu %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-SAFE
+# RUN: llc -run-pass=machine-combiner -mtriple=aarch64-unknown-linux-gnu -enable-unsafe-fp-math %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-UNSAFE
+
+# fadd without the reassoc flags can be reassociate only when unsafe fp math is
+# enabled.
+# CHECK-LABEL: name: fadd_no_reassoc
+# CHECK:             %4:fpr32 = FADDSrr %0, %1
+# CHECK-SAFE-NEXT:   %5:fpr32 = FADDSrr killed %4, %2
+# CHECK-SAFE-NEXT:   %6:fpr32 = FADDSrr killed %5, %3
+# CHECK-UNSAFE-NEXT: %9:fpr32 = FADDSrr %2, %3
+# CHECK-UNSAFE-NEXT: %6:fpr32 = FADDSrr killed %4, killed %9
+---
+name:            fadd_no_reassoc
+alignment:       4
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: fpr32 }
+  - { id: 1, class: fpr32 }
+  - { id: 2, class: fpr32 }
+  - { id: 3, class: fpr32 }
+  - { id: 4, class: fpr32 }
+  - { id: 5, class: fpr32 }
+  - { id: 6, class: fpr32 }
+liveins:
+  - { reg: '$s0', virtual-reg: '%0' }
+  - { reg: '$s1', virtual-reg: '%1' }
+  - { reg: '$s2', virtual-reg: '%2' }
+  - { reg: '$s3', virtual-reg: '%3' }
+frameInfo:
+  maxAlignment:    1
+  maxCallFrameSize: 0
+machineFunctionInfo: {}
+body:             |
+  bb.0.entry:
+    liveins: $s0, $s1, $s2, $s3
+
+    %3:fpr32 = COPY $s3
+    %2:fpr32 = COPY $s2
+    %1:fpr32 = COPY $s1
+    %0:fpr32 = COPY $s0
+    %4:fpr32 = FADDSrr %0, %1
+    %5:fpr32 = FADDSrr killed %4, %2
+    %6:fpr32 = FADDSrr killed %5, %3
+    $s0 = COPY %6
+    RET_ReallyLR implicit $s0
+
+
+# FIXME: We should be able to reassociate without unsafe fp math, but currently
+# the reassoc flag is ignored.
+# CHECK-LABEL: name: fadd_reassoc
+# CHECK:             %4:fpr32 = reassoc FADDSrr %0, %1
+# CHECK-SAFE-NEXT:   %5:fpr32 = reassoc FADDSrr killed %4, %2
+# CHECK-SAFE-NEXT:   %6:fpr32 = reassoc FADDSrr killed %5, %3
+# CHECK-UNSAFE-NEXT: %9:fpr32 = reassoc FADDSrr %2, %3
+# CHECK-UNSAFE-NEXT: %6:fpr32 = reassoc FADDSrr killed %4, killed %9
+---
+name:            fadd_reassoc
+alignment:       4
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: fpr32 }
+  - { id: 1, class: fpr32 }
+  - { id: 2, class: fpr32 }
+  - { id: 3, class: fpr32 }
+  - { id: 4, class: fpr32 }
+  - { id: 5, class: fpr32 }
+  - { id: 6, class: fpr32 }
+liveins:
+  - { reg: '$s0', virtual-reg: '%0' }
+  - { reg: '$s1', virtual-reg: '%1' }
+  - { reg: '$s2', virtual-reg: '%2' }
+  - { reg: '$s3', virtual-reg: '%3' }
+frameInfo:
+  maxAlignment:    1
+  maxCallFrameSize: 0
+machineFunctionInfo: {}
+body:             |
+  bb.0.entry:
+    liveins: $s0, $s1, $s2, $s3
+
+    %3:fpr32 = COPY $s3
+    %2:fpr32 = COPY $s2
+    %1:fpr32 = COPY $s1
+    %0:fpr32 = COPY $s0
+    %4:fpr32 = reassoc FADDSrr %0, %1
+    %5:fpr32 = reassoc FADDSrr killed %4, %2
+    %6:fpr32 = reassoc FADDSrr killed %5, %3
+    $s0 = COPY %6
+    RET_ReallyLR implicit $s0
+
+
+# Check that flags on the instructions are preserved after reassociation.
+# CHECK-LABEL: name: fadd_flags
+# CHECK:             %4:fpr32 = nsz FADDSrr %0, %1
+# CHECK-SAFE-NEXT:   %5:fpr32 = nnan FADDSrr killed %4, %2
+# CHECK-SAFE-NEXT:   %6:fpr32 = ninf FADDSrr killed %5, %3
+# CHECK-UNSAFE-NEXT: %9:fpr32 = nnan FADDSrr %2, %3
+# CHECK-UNSAFE-NEXT: %6:fpr32 = ninf FADDSrr killed %4, killed %9
+---
+name:            fadd_flags
+alignment:       4
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: fpr32 }
+  - { id: 1, class: fpr32 }
+  - { id: 2, class: fpr32 }
+  - { id: 3, class: fpr32 }
+  - { id: 4, class: fpr32 }
+  - { id: 5, class: fpr32 }
+  - { id: 6, class: fpr32 }
+liveins:
+  - { reg: '$s0', virtual-reg: '%0' }
+  - { reg: '$s1', virtual-reg: '%1' }
+  - { reg: '$s2', virtual-reg: '%2' }
+  - { reg: '$s3', virtual-reg: '%3' }
+frameInfo:
+  maxAlignment:    1
+  maxCallFrameSize: 0
+machineFunctionInfo: {}
+body:             |
+  bb.0.entry:
+    liveins: $s0, $s1, $s2, $s3
+
+    %3:fpr32 = COPY $s3
+    %2:fpr32 = COPY $s2
+    %1:fpr32 = COPY $s1
+    %0:fpr32 = COPY $s0
+    %4:fpr32 = nsz FADDSrr %0, %1
+    %5:fpr32 = nnan FADDSrr killed %4, %2
+    %6:fpr32 = ninf FADDSrr killed %5, %3
+    $s0 = COPY %6
+    RET_ReallyLR implicit $s0