[Mips] Do not emit teq if NoZeroDivCheck in FastISel

[xry111]

Fix the issue: -mno-check-zero-division is not respected at -O0 for some test cases.

[llvmorg-github-actions]

@llvm/pr-subscribers-backend-mips

Author: Xi Ruoyao (xry111)

Changes

Fix the issue: -mno-check-zero-division is not respected at -O0 for some test cases.

---

Full diff: https://github.com/llvm/llvm-project/pull/204386.diff

2 Files Affected:

• (modified) llvm/lib/Target/Mips/MipsFastISel.cpp (+4-2)
• (modified) llvm/test/CodeGen/Mips/divrem.ll (+2)

diff --git a/llvm/lib/Target/Mips/MipsFastISel.cpp b/llvm/lib/Target/Mips/MipsFastISel.cpp
index 9645fb5293609..ab76ee069fb15 100644
--- a/llvm/lib/Target/Mips/MipsFastISel.cpp
+++ b/llvm/lib/Target/Mips/MipsFastISel.cpp
@@ -74,6 +74,7 @@
 using namespace llvm;
 
 extern cl::opt<bool> EmitJalrReloc;
+extern cl::opt<bool> NoZeroDivCheck;
 
 namespace {
 
@@ -1952,8 +1953,9 @@ bool MipsFastISel::selectDivRem(const Instruction *I, unsigned ISDOpcode) {
     return false;
 
   emitInst(DivOpc).addReg(Src0Reg).addReg(Src1Reg);
-  if (!isa<ConstantInt>(I->getOperand(1)) ||
-      dyn_cast<ConstantInt>(I->getOperand(1))->isZero()) {
+  if (!NoZeroDivCheck &&
+      (!isa<ConstantInt>(I->getOperand(1)) ||
+       dyn_cast<ConstantInt>(I->getOperand(1))->isZero())) {
     emitInst(Mips::TEQ).addReg(Src1Reg).addReg(Mips::ZERO).addImm(7);
   }
 
diff --git a/llvm/test/CodeGen/Mips/divrem.ll b/llvm/test/CodeGen/Mips/divrem.ll
index a29f2793c081f..e30764fef9042 100644
--- a/llvm/test/CodeGen/Mips/divrem.ll
+++ b/llvm/test/CodeGen/Mips/divrem.ll
@@ -12,6 +12,8 @@
 ; RUN: llc -mtriple=mips64 -mcpu=mips64r2 -mno-check-zero-division -relocation-model=pic < %s | FileCheck %s -check-prefixes=ALL,ACC64,NOCHECK
 ; RUN: llc -mtriple=mips64 -mcpu=mips64r6 -mno-check-zero-division -relocation-model=pic < %s | FileCheck %s -check-prefixes=ALL,GPR64,NOCHECK
 
+; RUN: llc -mtriple=mipsel -mcpu=mips32 -O0 -mno-check-zero-division -relocation-model=pic < %s | FileCheck %s -check-prefixes=NOCHECK
+
 ; FileCheck Prefixes:
 ;   ALL - All targets
 ;   ACC32 - Accumulator based multiply/divide on 32-bit targets

[github-actions]

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

[github-actions]

🐧 Linux x64 Test Results

• 200200 tests passed
• 5438 tests skipped

✅ The build succeeded and all tests passed.

[github-actions]

🪟 Windows x64 Test Results

• 138468 tests passed
• 3541 tests skipped

✅ The build succeeded and all tests passed.

[folkertdev]

Looks like LLVM main is currently broken. That'll probably get resolved within a couple of hours (or maybe it has been already). You can merge current main to re-trigger a CI build (also a rebase is probably fine for a PR this small).

[yingopq]

LGTM.

[xry111]

Ping.

[folkertdev]

Looks like you're really unlucky with the merge commits, there is some issue with SystemZ which can't possibly be related to your changes.

The change has already been approved, unfortunately I think you just have to time the sync right for CI to pass.