[llvm] [NVPTX] Add IR pass for FMA transformation in the llc pipeline (PR #154735)

[Alex MacLean via llvm-commits]

================
@@ -0,0 +1,146 @@
+//===------ NVPTXFoldFMA.cpp - Fold FMA --------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements FMA folding for float/double type for NVPTX. It folds
+// following patterns:
+// 1. fadd(fmul(a, b), c) => fma(a, b, c)
+// 2. fadd(c, fmul(a, b)) => fma(a, b, c)
+// 3. fadd(fmul(a, b), fmul(c, d)) => fma(a, b, fmul(c, d))
+// 4. fsub(fmul(a, b), c) => fma(a, b, fneg(c))
+// 5. fsub(a, fmul(b, c)) => fma(fneg(b), c, a)
+// 6. fsub(fmul(a, b), fmul(c, d)) => fma(a, b, fneg(fmul(c, d)))
+//===----------------------------------------------------------------------===//
+
+#include "NVPTXUtilities.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/InstIterator.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/Intrinsics.h"
+
+#define DEBUG_TYPE "nvptx-fold-fma"
+
+using namespace llvm;
+
+static bool foldFMA(Function &F) {
+  bool Changed = false;
+  SmallVector<BinaryOperator *, 16> FAddFSubInsts;
+
+  // Collect all float/double FAdd/FSub instructions with allow-contract
+  for (auto &I : instructions(F)) {
+    if (auto *BI = dyn_cast<BinaryOperator>(&I)) {
+      // Only FAdd and FSub are supported.
+      if (BI->getOpcode() != Instruction::FAdd &&
+          BI->getOpcode() != Instruction::FSub)
+        continue;
+
+      // At minimum, the instruction should have allow-contract.
+      if (!BI->hasAllowContract())
+        continue;
+
+      // Only float and double are supported.
+      if (!BI->getType()->isFloatTy() && !BI->getType()->isDoubleTy())
+        continue;
+
+      FAddFSubInsts.push_back(BI);
+    }
+  }
+
+  auto tryFoldBinaryFMul = [](BinaryOperator *BI, Value *MulOperand,
----------------
AlexMaclean wrote:

Since this doesn't capture anything can it be pulled out into a static helper?

https://github.com/llvm/llvm-project/pull/154735