[llvm] [NVPTX] Add auto-upgrade rules for fabs.{f, d, ftz.f} (PR #136150)

[Alex MacLean via llvm-commits]

https://github.com/AlexMaclean created https://github.com/llvm/llvm-project/pull/136150

These auto-upgrade rules are required after these intrinsics were removed in https://github.com/llvm/llvm-project/pull/135644/

>From 00ff0e9b65fbdca93d00015d963a9a74a2877f0c Mon Sep 17 00:00:00 2001
From: Alex Maclean <amaclean at nvidia.com>
Date: Thu, 17 Apr 2025 15:45:31 +0000
Subject: [PATCH] [NVPTX] Add auto-upgrade rules for fabs.{f,d,ftz.f}

---
 llvm/lib/IR/AutoUpgrade.cpp                       |  7 +++++++
 .../Assembler/auto_upgrade_nvvm_intrinsics.ll     | 15 +++++++++++++++
 2 files changed, 22 insertions(+)

diff --git a/llvm/lib/IR/AutoUpgrade.cpp b/llvm/lib/IR/AutoUpgrade.cpp
index 7a91da5f9f7eb..2e17b8ce5eb7b 100644
--- a/llvm/lib/IR/AutoUpgrade.cpp
+++ b/llvm/lib/IR/AutoUpgrade.cpp
@@ -1287,6 +1287,9 @@ static bool upgradeIntrinsicFunction1(Function *F, Function *&NewFn,
         // nvvm.abs.{i,ii}
         Expand =
             Name == "i" || Name == "ll" || Name == "bf16" || Name == "bf16x2";
+      else if (Name == "fabs.f" || Name == "fabs.ftz.f" || Name == "fabs.d")
+        // nvvm.fabs.{f,ftz.f,d}
+        Expand = true;
       else if (Name == "clz.ll" || Name == "popc.ll" || Name == "h2f" ||
                Name == "swap.lo.hi.b64")
         Expand = true;
@@ -2318,6 +2321,10 @@ static Value *upgradeNVVMIntrinsicCall(StringRef Name, CallBase *CI,
     Value *Arg = Builder.CreateBitCast(CI->getArgOperand(0), Ty);
     Value *Abs = Builder.CreateUnaryIntrinsic(Intrinsic::nvvm_fabs, Arg);
     Rep = Builder.CreateBitCast(Abs, CI->getType());
+  } else if (Name == "fabs.f" || Name == "fabs.ftz.f" || Name == "fabs.d") {
+    Intrinsic::ID IID = (Name == "fabs.ftz.f") ? Intrinsic::nvvm_fabs_ftz
+                                               : Intrinsic::nvvm_fabs;
+    Rep = Builder.CreateUnaryIntrinsic(IID, CI->getArgOperand(0));
   } else if (Name.starts_with("atomic.load.add.f32.p") ||
              Name.starts_with("atomic.load.add.f64.p")) {
     Value *Ptr = CI->getArgOperand(0);
diff --git a/llvm/test/Assembler/auto_upgrade_nvvm_intrinsics.ll b/llvm/test/Assembler/auto_upgrade_nvvm_intrinsics.ll
index 74b9640df6977..a3ddc62dd34e9 100644
--- a/llvm/test/Assembler/auto_upgrade_nvvm_intrinsics.ll
+++ b/llvm/test/Assembler/auto_upgrade_nvvm_intrinsics.ll
@@ -13,6 +13,10 @@ declare float @llvm.nvvm.h2f(i16)
 declare i32 @llvm.nvvm.abs.i(i32)
 declare i64 @llvm.nvvm.abs.ll(i64)
 
+declare float @llvm.nvvm.fabs.f(float)
+declare float @llvm.nvvm.fabs.ftz.f(float)
+declare double @llvm.nvvm.fabs.d(double)
+
 declare i16 @llvm.nvvm.max.s(i16, i16)
 declare i32 @llvm.nvvm.max.i(i32, i32)
 declare i64 @llvm.nvvm.max.ll(i64, i64)
@@ -97,6 +101,17 @@ define void @abs(i32 %a, i64 %b) {
   ret void
 }
 
+; CHECK-LABEL: @fabs
+define void @fabs(float %a, double %b) {
+; CHECK: call float @llvm.nvvm.fabs.f32(float %a)
+; CHECK: call float @llvm.nvvm.fabs.ftz.f32(float %a)
+; CHECK: call double @llvm.nvvm.fabs.f64(double %b)
+  %r1 = call float @llvm.nvvm.fabs.f(float %a)
+  %r2 = call float @llvm.nvvm.fabs.ftz.f(float %a)
+  %r3 = call double @llvm.nvvm.fabs.d(double %b)
+  ret void
+}
+
 ; CHECK-LABEL: @min_max
 define void @min_max(i16 %a1, i16 %a2, i32 %b1, i32 %b2, i64 %c1, i64 %c2) {
 ; CHECK: [[maxs:%[a-zA-Z0-9.]+]] = icmp sge i16 %a1, %a2