RegisterCoalescer: Relax assert for super register def rematerialization (#69088)

Author: arsenm

llvm/lib/CodeGen/RegisterCoalescer.cpp

@@ -1415,6 +1415,9 @@ bool RegisterCoalescer::reMaterializeTrivialDef(const CoalescerPair &CP,
   // from SUBREG_TO_REG, such as:
   // $edi = MOV32r0 implicit-def dead $eflags, implicit-def $rdi
   // undef %0.sub_32bit = MOV32r0 implicit-def dead $eflags, implicit-def %0
+  //
+  // The implicit-def of the super register may have been reduced to
+  // subregisters depending on the uses.
 
   bool NewMIDefinesFullReg = false;
 
@@ -1432,12 +1435,14 @@ bool RegisterCoalescer::reMaterializeTrivialDef(const CoalescerPair &CP,
         assert(MO.isImplicit() && MO.getReg().isPhysical() &&
                (MO.isDead() ||
                 (DefSubIdx &&
-                 (TRI->getSubReg(MO.getReg(), DefSubIdx) ==
-                  MCRegister((unsigned)NewMI.getOperand(0).getReg())))));
+                 ((TRI->getSubReg(MO.getReg(), DefSubIdx) ==
+                   MCRegister((unsigned)NewMI.getOperand(0).getReg())) ||
+                  TRI->isSubRegisterEq(NewMI.getOperand(0).getReg(),
+                                       MO.getReg())))));
         NewMIImplDefs.push_back(MO.getReg().asMCReg());
       } else {
-        assert(MO.getReg() == NewMI.getOperand(0).getReg() &&
-               MO.getSubReg() == 0);
+        assert(MO.getReg() == NewMI.getOperand(0).getReg());
+
         // We're only expecting another def of the main output, so the range
         // should get updated with the regular output range.
         //

llvm/test/CodeGen/X86/coalescer-implicit-def-regression-imp-operand-assert.mir

@@ -0,0 +1,39 @@
+# RUN: llc -mtriple=x86_64-unknown-linux-gnu -run-pass=register-coalescer -o - %s
+---
+name:  rematerialize_subreg_to_reg_added_impdef_1
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    successors: %bb.1(0x2aaaaaab), %bb.2(0x55555555)
+    liveins: $edi
+
+    %0:gr32 = MOV32r0 implicit-def dead $eflags
+    %1:gr8 = COPY %0.sub_8bit
+    %2:gr64 = SUBREG_TO_REG 0, killed %0, %subreg.sub_32bit
+    JCC_1 %bb.2, 5, implicit killed undef $eflags
+
+  bb.1:
+    successors: %bb.3(0x80000000)
+
+    JMP_1 %bb.3
+
+  bb.2:
+    successors: %bb.3(0x80000000)
+
+    %5:gr64 = IMPLICIT_DEF
+    %2:gr64 = COPY killed %5
+
+  bb.3:
+    successors: %bb.4(0x30000000), %bb.5(0x50000000)
+
+    JCC_1 %bb.5, 5, implicit killed undef $eflags
+
+  bb.4:
+    $al = COPY killed %1
+    RET 0, killed undef $al
+
+  bb.5:
+    MOV64mr undef $noreg, 1, undef $noreg, 0, undef $noreg, killed %2 :: (store (s64))
+    RET 0, killed undef $al
+
+...

llvm/test/CodeGen/X86/coalescer-implicit-def-regression.mir

@@ -0,0 +1,91 @@
+# RUN: llc -mtriple=x86_64-unknown-linux-gnu -run-pass=register-coalescer -o - %s
+
+# An implicit-def will be added to SUBREG_TO_REG during coalescing
+---
+name:  rematerialize_subreg_to_reg_added_impdef_0
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    successors: %bb.1(0x2aaaaaab), %bb.2(0x55555555)
+    liveins: $edi
+
+    %0:gr32 = MOV32r0 implicit-def dead $eflags
+    %1:gr8 = COPY %0.sub_8bit
+    %2:gr64 = SUBREG_TO_REG 0, killed %0, %subreg.sub_32bit
+    JCC_1 %bb.2, 5, implicit killed undef $eflags
+
+  bb.1:
+    %4:gr8 = COPY %1
+    %5:gr8 = COPY killed undef %1
+    JMP_1 %bb.5
+
+  bb.2:
+    %6:gr64 = IMPLICIT_DEF
+    %2:gr64 = COPY killed %6
+    %5:gr8 = MOV8ri 1
+
+  bb.5:
+    successors: %bb.6(0x30000000), %bb.7(0x50000000)
+
+    TEST8rr killed undef %5, %5, implicit-def $eflags
+    JCC_1 %bb.7, 5, implicit killed undef $eflags
+
+  bb.6:
+    $al = COPY killed %1
+    RET 0, killed undef $al
+
+  bb.7:
+    MOV64mr undef $noreg, 1, undef $noreg, 0, undef $noreg, killed %2 :: (store (s64))
+    RET 0, killed undef $al
+
+...
+
+
+# Reduced version of previous with the SUBREG_TO_REG already folded
+# away.
+#
+# The mov32r0 defines a subregister and has an implicit-def of the
+# super register. After coalescing, the full register implicit def of
+# %2 becomes a different subregister def.
+
+---
+name:  rematerialize_subreg_to_reg_coalesces_to_subreg_impdef
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    successors: %bb.1(0x2aaaaaab), %bb.2(0x55555555)
+    liveins: $edi
+
+    undef %2.sub_32bit:gr64_with_sub_8bit = MOV32r0 implicit-def dead $eflags, implicit-def %2
+    %1:gr8 = COPY %2.sub_8bit
+    JCC_1 %bb.2, 5, implicit killed undef $eflags
+
+  bb.1:
+    successors: %bb.3(0x80000000)
+
+    dead %3:gr8 = COPY %1
+    %4:gr8 = COPY undef %1
+    JMP_1 %bb.3
+
+  bb.2:
+    successors: %bb.3(0x80000000)
+
+    %5:gr64 = IMPLICIT_DEF
+    %2:gr64_with_sub_8bit = COPY %5
+    %4:gr8 = MOV8ri 1
+
+  bb.3:
+    successors: %bb.4(0x30000000), %bb.5(0x50000000)
+
+    TEST8rr undef %4, %4, implicit-def $eflags
+    JCC_1 %bb.5, 5, implicit killed undef $eflags
+
+  bb.4:
+    $al = COPY %1
+    RET 0, killed undef $al
+
+  bb.5:
+    MOV64mr undef $noreg, 1, undef $noreg, 0, undef $noreg, %2 :: (store (s64))
+    RET 0, killed undef $al
+
+...