[mips] bnec/beqc register constraint fix

[sdardis]

Summary

beqc and bnec cannot have $rs == $rt. Inhibit compact branch creation
if that would occur.

[dsanders]

This needs a test case. Going by the original 7MB reproducer I think we need a test case with a conditional branch where both operands are evaluated in different basic blocks but can be CSE'd. I think the compiler should have eliminated the branch as well though since 'bnec $5, $5, ...' is never going to be taken.

[sdardis]

Added a test reduced down from bug report.

[dsanders]

test/CodeGen/Mips/compactbranches/beqc-bnec-register-constraint.ll

35–36

; CHECK-NOT: bnec $[[R0:[0-9]+]], R0

The R0 and R1 should be [[R0]] and [[R1]]

Also, I'm not sure variables can be used in the same directive that defines them. I'll double check that.

[dsanders]

LGTM with the missing [[ and ]] added.

test/CodeGen/Mips/compactbranches/beqc-bnec-register-constraint.ll

35–36

Also, I'm not sure variables can be used in the same directive that defines them. I'll double check that.

This works.

[sdardis]

Thanks for the review.

lib/Target/Mips/MipsInstrInfo.cpp

if (Subtarget.hasMips32r6() || canUseShortMicroMipsCTI) {
     switch (Opcode) {
     case Mips::B:
       return Mips::BC;
     case Mips::BAL:
       return Mips::BALC;
     case Mips::BEQ:
       if (canUseShortMicroMipsCTI)
         return Mips::BEQZC_MM;
-      else
-        return Mips::BEQC;
+      else if (I->getOperand(0).getReg() == I->getOperand(1).getReg())
+        return 0;
     case Mips::BNE:
       if (canUseShortMicroMipsCTI)
         return Mips::BNEZC_MM;
-      else
-        return Mips::BNEC;
+      else if (I->getOperand(0).getReg() == I->getOperand(1).getReg())
+        return 0;
     case Mips::BGE:
       if (I->getOperand(0).getReg() == I->getOperand(1).getReg())
         return 0;
       return Mips::BGEC;
     case Mips::BGEU:
       if (I->getOperand(0).getReg() == I->getOperand(1).getReg())
         return 0;
       return Mips::BGEUC;

test/CodeGen/Mips/compactbranches/beqc-bnec-register-constraint.ll

+; RUN: llc -march=mips -mcpu=mips32r6 -O1 -start-after=dwarfehprepare < %s | FileCheck %s
+
+; beqc/bnec have the constraint that $rs < $rt && $rs != 0 && $rt != 0
+; Cases where $rs == 0 and $rt != 0 should be transformed into beqzc/bnezc.
+; Cases where $rs > $rt can have the operands swapped as ==,!= are commutative.
+
+; Cases where beq & bne where $rs == $rt have to inhibited from being turned
+; into compact branches but arguably should not occur. This test covers the
+; $rs == $rt case.
+
+; Starting from dwarf exception handling preparation skips optimizations that
+; may simplify out crucical bnec $4, $4 instruction.
+
+define internal void @_ZL14TestRemoveLastv(i32* %alist.sroa.0.4) {
+entry:
+  %ascevgep = getelementptr i32, i32* %alist.sroa.0.4, i64 99
+  br label %do.body121
+
+for.cond117:
+  %alsr.iv.next = add nsw i32 %alsr.iv, -1
+  %ascevgep340 = getelementptr i32, i32* %alsr.iv339, i64 -1
+  %acmp118 = icmp sgt i32 %alsr.iv.next, 0
+  br i1 %acmp118, label %do.body121, label %if.then143
+
+do.body121:
+  %alsr.iv339 = phi i32* [ %ascevgep, %entry ], [ %ascevgep340, %for.cond117 ]
+  %alsr.iv = phi i32 [ 100, %entry ], [ %alsr.iv.next, %for.cond117 ]
+  %a9 = add i32 %alsr.iv, -1
+  %alnot124 = icmp eq i32 %alsr.iv, %alsr.iv
+  br i1 %alnot124, label %do.body134, label %if.then143, !prof !11
+
+do.body134:
+  %a10 = add i32 %alsr.iv, -1
+  %a11 = load i32, i32* %alsr.iv339, align 4, !tbaa !5
+; CHECK-NOT: bnec $[[R0:[0-9]+]], R0
+; CHECK-NOT: beqc $[[R1:[0-9]+]], R1
+  %alnot137 = icmp eq i32 %a9, %a11
+  br i1 %alnot137, label %do.end146, label %if.then143, !prof !11
+
+if.then143:
+ ret void
+ unreachable
+
+do.end146:
+  %alnot151 = icmp eq i32 %a9, %a10
+  br i1 %alnot151, label %for.cond117, label %if.then143, !prof !11
+
+}
+!3 = !{!"omnipotent char", !4, i64 0}
+!4 = !{!"Simple C++ TBAA"}
+!5 = !{!6, !6, i64 0}
+!6 = !{!"int", !3, i64 0}
+!11 = !{!"branch_weights", i32 2000, i32 1}
+!12 = !{!"branch_weights", i32 -388717296, i32 7818360}
+