Add a regression test for mlir-stage-05-v1

This adds one LIT test covering 3 lines and 3 branch outcomes that the existing suite does not reach, across 2 source files, at 48615bc5925ef4b9db8b4550b5d4322933cf4b7b.

Coverage this test adds

Source file	Newly covered lines	Newly covered branch outcomes
lib/Frontend/Raising/ExactStandardSpelling.h	128–129	127:12–127:25 true; 128:9–128:33 true
lib/Frontend/Raising/MaterializeFMulAddPass.cpp	178	177:9–177:25 true

lib/Frontend/Raising/ExactStandardSpelling.h

 125      if (!statesDefaultDenormalEnvironment(env)) 126        return true;+127    if (auto noSignedZeros = funcOp.getNoSignedZerosFpMathAttr())    [true branch at 127:12]+128      if (noSignedZeros.getValue())    [true branch at 128:9]+129        return true; 130    if (auto contraction = funcOp.getFpContractAttr()) 131      if (contraction.getValue() != "off")

lib/Frontend/Raising/MaterializeFMulAddPass.cpp

 175          }); 176  +177      if (selected.empty())    [true branch at 177:9]+178        return markAllAnalysesPreserved(); 179   180      materializeSelected(getContext(), selected, shape.getValue());

The test

The input is 7 lines; reduction found nothing further to remove that kept the coverage above and a compiler that still accepted it. The CHECK lines are the compiler's exact output on this input at 48615bc5925ef4b9db8b4550b5d4322933cf4b7b.

// RUN: loom-raise-opt --loom-materialize-fmuladd=shape=fused --mlir-print-op-generic %s | FileCheck %s

module {
llvm.func @blocked_0(%x: f32, %y: f32, %z: f32) -> f32 attributes {no_signed_zeros_fp_math = true} {
  %b0 = llvm.intr.fmuladd(%x, %y, %z) : (f32, f32, f32) -> f32 loc("fmuladd_blocked_0")
  llvm.return %x : f32
}

}

// CHECK: "builtin.module"() ({
// CHECK-NEXT:   "llvm.func"() <{CConv = #llvm.cconv<ccc>, function_type = !llvm.func<f32 (f32, f32, f32)>, linkage = #llvm.linkage<external>, no_signed_zeros_fp_math = true, sym_name = "blocked_0", unnamed_addr = 0 : i64, visibility_ = 0 : i64}> ({
// CHECK-NEXT:   ^bb0(%arg0: f32, %arg1: f32, %arg2: f32):
// CHECK-NEXT:     %0 = "llvm.intr.fmuladd"(%arg0, %arg1, %arg2) <{fastmathFlags = #llvm.fastmath<none>}> : (f32, f32, f32) -> f32
// CHECK-NEXT:     "llvm.return"(%arg0) : (f32) -> ()
// CHECK-NEXT:   }) : () -> ()
// CHECK-NEXT: }) : () -> ()
// CHECK-EMPTY:

Validation

• Native compiler and FileCheck replay: passed at 48615bc5925ef4b9db8b4550b5d4322933cf4b7b.
• Property checker verdict on the reduced pair: PASS (postcondition.spct, SHA-256 cd479d53d3e454031c9da3cb497446c59c45d7b6625a2f2ecdd40196bdabaee4).
• Reduction criterion: accepted checker PASS and every original line/branch increment over the suite baseline.

Where this test came from

PBT mlir-stage-05-v1, run 20260911-043029, seed 62.

The property under test is anchored on documentation:

• selected output: docs/spec-compiler-part-2-scf.md lines 1036–1039 (output side)
• linked input 61: docs/spec-compiler-part-2-scf.md lines 1018–1021 (input side)
• linked input 101: docs/spec-compiler-part-2-scf.md lines 178–181 (input side)
• linked input 127: docs/spec-compiler-part-2-scf.md lines 1021–1022 (input side)
• linked input 165: docs/spec-compiler-part-2-scf.md lines 1029–1034 (input side)
• linked input 209: docs/spec-compiler-part-2-scf.md lines 181–185 (input side)

---

Coverage is measured against the recorded baseline suite at this revision, and counts unique mapped file locations including generated code. Reduction retains every added line or branch outcome and a passing property check; it is not a claim of global minimality, nor of a defect.