[mlir][math] Use math.h for M_PI_4 on WIN32

[sunshaoce]

Summary

Trying to fix build errors on WIN32 caused by D153656

[sunshaoce]

update

[jpienaar]

https://github.com/llvm/llvm-project/commit/75bd41f61d3247d4394ad0c46d23cc7c62261d63 should resolve the breakage (I saw that previous uses of these constants were also changed to follow that approach).

[yronglin]

Can we add these constants into llvm/Support/MathExtras.h?

[sunshaoce]

In D153706#4446862, @jpienaar wrote:

https://github.com/llvm/llvm-project/commit/75bd41f61d3247d4394ad0c46d23cc7c62261d63 should resolve the breakage (I saw that previous uses of these constants were also changed to follow that approach).

Thanks!

mlir/lib/Dialect/Math/Transforms/PolynomialApproximation.cpp

AtanApproximation::matchAndRewrite(math::AtanOp op,
   d = builder.create<math::FmaOp>(xx, d, q3);
   d = builder.create<math::FmaOp>(xx, d, q4);
 
   // Compute approximation of theta:
   Value ans0 = builder.create<arith::DivFOp>(n, d);
   ans0 = builder.create<math::FmaOp>(ans0, x, x);
 
   // Correct for the input mapping's angles:
+#ifdef _WIN32
+#define _USE_MATH_DEFINES
+#endif
   Value mpi4 = bcast(f32Cst(builder, M_PI_4));
   Value ans2 = builder.create<arith::AddFOp>(mpi4, ans0);
   Value ans = builder.create<arith::SelectOp>(cmp2, ans2, ans0);
 
   Value mpi2 = bcast(f32Cst(builder, M_PI_2));
   Value ans1 = builder.create<arith::SubFOp>(mpi2, ans0);
   ans = builder.create<arith::SelectOp>(cmp1, ans1, ans);