[MIPS] Fix -msingle-float doesn't work with double on O32

[yingopq]

Skip the following function 'CustomLowerNode' when the operand had done SoftenFloatResult.

Fix #93052

[llvmbot]

@llvm/pr-subscribers-llvm-selectiondag

Author: None (yingopq)

Changes

Skip the following function 'CustomLowerNode' when the operand had done SoftenFloatResult.

Fix #93052

---

Full diff: https://github.com/llvm/llvm-project/pull/107543.diff

1 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp (+11-1)

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
index fefb0844f1ab53..b879f40dfd85a9 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
@@ -2760,9 +2760,19 @@ void DAGTypeLegalizer::ExpandIntegerResult(SDNode *N, unsigned ResNo) {
   LLVM_DEBUG(dbgs() << "Expand integer result: "; N->dump(&DAG));
   SDValue Lo, Hi;
   Lo = Hi = SDValue();
+  bool NeedCustomLower = true;
+
+  // Skip the following function 'CustomLowerNode' when the operand had done
+  // `SoftenFloatResult`.
+  if (N->getOpcode() == ISD::BITCAST &&
+      getTypeAction(N->getOperand(0).getValueType()) ==
+          TargetLowering::TypeSoftenFloat &&
+      N->getValueType(ResNo) ==
+          GetSoftenedFloat(N->getOperand(0)).getValueType())
+    NeedCustomLower = false;
 
   // See if the target wants to custom expand this node.
-  if (CustomLowerNode(N, N->getValueType(ResNo), true))
+  if (NeedCustomLower && CustomLowerNode(N, N->getValueType(ResNo), true))
     return;
 
   switch (N->getOpcode()) {

[arsenm]

Needs tests

[yingopq]

Needs tests

OK, added.

[wzssyqa]

No. This problem is about clang not llc.
The llvm/test/CodeGen/Mips/2008-07-06-fadd64.ll is already OK.
Please see the bug report.

[yingopq]

No. This problem is about clang not llc. The llvm/test/CodeGen/Mips/2008-07-06-fadd64.ll is already OK. Please see the bug report.

The lvm/test/CodeGen/Mips/2008-07-06-fadd64.ll file was discovered when I was looking for test cases that covered this problem reported.

This fix can resolve the bug reported, and when configured -mcpu=mips32r2 -mattr=single-float, llc and clang have same error info, debug information and issue reason. So I chose to improve this test case.

$ sudo ./build/bin/llc -march=mips -mcpu=mips32r2 -mattr=single-float -O2 < llvm/test/CodeGen/Mips/2008-07-06-fadd64.ll
	.text
	.abicalls
	.option	pic0
	.section	.mdebug.abi32,"",@progbits
	.nan	legacy
	.text
	.file	"<stdin>"
SoftenFloatOperand Op #0: t46: i32 = MipsISD::ExtractElementF64 t13, Constant:i32<1>

LLVM ERROR: Do not know how to soften this operator's operand!
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
Stack dump:
0.	Program arguments: ./build/bin/llc -march=mips -mcpu=mips32r2 -mattr=single-float -O2
1.	Running pass 'Function Pass Manager' on module '<stdin>'.
2.	Running pass 'MIPS DAG->DAG Pattern Instruction Selection' on function '@dofloat'

[arsenm]

Shouldn't be special casing a specific opcode here, and shouldn't need to check the type of the actual softened result. It is also OK to request custom lowering on illegal types

[wzssyqa]

Ohh. Thanks. Your patch fixes this problem.

And I will approve this PR once you fix the problem Matt found.
I guess he meant that you can just remove N->getOpcode() == ISD::BITCAST

[arsenm]

Ohh. Thanks. Your patch fixes this problem.

And I will approve this PR once you fix the problem Matt found. I guess he meant that you can just remove N->getOpcode() == ISD::BITCAST

I mean the entire thing is suspect, you can request custom lowering on illegal types so CustomLowerNode should never be skipped

[yingopq]

Ohh. Thanks. Your patch fixes this problem.
And I will approve this PR once you fix the problem Matt found. I guess he meant that you can just remove N->getOpcode() == ISD::BITCAST

I mean the entire thing is suspect, you can request custom lowering on illegal types so CustomLowerNode should never be skipped

OK, I would find another solution.

[yingopq]

@wzssyqa I updated the code, please help review. Thanks!

[RKSimon]

LGTM - cheers

[arsenm]

Don't need -O2. The mattr is also missing +, and -march should be mtriple

[yingopq]

OK, applied, thanks!