From 7e48ec701ea857f630c8bed58868b5646be4220c Mon Sep 17 00:00:00 2001
From: William Moses <gh@wsmoses.com>
Date: Tue, 18 Mar 2025 01:18:43 +0100
Subject: [PATCH 1/4] [MLIR][GPUToNVVM] Support 32-bit isfinite

---
 mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
index 159a6afd4a917..9e00bb34f4cb4 100644
--- a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
+++ b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
@@ -552,8 +552,8 @@ void mlir::populateLibDeviceConversionPatterns(
                                     "__nv_floor");
   populateOpPatterns<math::FmaOp>(converter, patterns, benefit, "__nv_fmaf",
                                   "__nv_fma");
-  // Note: libdevice does not provide `__nv_isfinitef` as of moment of writing.
-  populateOpPatterns<math::IsFiniteOp>(converter, patterns, benefit, "",
+  // Note: libdevice uses a different name for 32-bit finite checking
+  populateOpPatterns<math::IsFiniteOp>(converter, patterns, benefit, "__nv_finitef",
                                        "__nv_isfinited");
   populateOpPatterns<math::IsInfOp>(converter, patterns, benefit, "__nv_isinff",
                                     "__nv_isinfd");

From 287e687d4b6e603c8485bcbb44b5c2626dd141ba Mon Sep 17 00:00:00 2001
From: Ivan Radanov Ivanov <ivanov.i.aa@m.titech.ac.jp>
Date: Tue, 18 Mar 2025 09:43:57 +0900
Subject: [PATCH 2/4] Add test

---
 mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir b/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
index 14594cd6badb1..ce8da38f24179 100644
--- a/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
+++ b/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
@@ -1083,9 +1083,9 @@ gpu.module @test_module_54 {
     // CHECK: llvm.mlir.constant(0
     // CHECK: llvm.icmp "ne"
     %3 = math.isnan %f64 : f64
-    // Note: for some reason, libdevice does not provide isfinite for f32, so
-    // this should fail to convert.
-    // CHECK: math.isfinite {{.*}} : f32
+    // CHECK: llvm.call @__nv_finitef(%{{.*}}) : (f32) -> i32
+    // CHECK: llvm.mlir.constant(0
+    // CHECK: llvm.icmp "ne"
     %4 = math.isfinite %f32 : f32
     // CHECK: llvm.call @__nv_isfinited(%{{.*}}) : (f64) -> i32
     // CHECK: llvm.mlir.constant(0

From b824afd9e8f52c83f8c0ab2e78e38f7629a4d291 Mon Sep 17 00:00:00 2001
From: Ivan Radanov Ivanov <ivanov.i.aa@m.titech.ac.jp>
Date: Tue, 18 Mar 2025 09:44:08 +0900
Subject: [PATCH 3/4] clang-format

---
 mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
index 9e00bb34f4cb4..958d0d085fce1 100644
--- a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
+++ b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
@@ -553,8 +553,8 @@ void mlir::populateLibDeviceConversionPatterns(
   populateOpPatterns<math::FmaOp>(converter, patterns, benefit, "__nv_fmaf",
                                   "__nv_fma");
   // Note: libdevice uses a different name for 32-bit finite checking
-  populateOpPatterns<math::IsFiniteOp>(converter, patterns, benefit, "__nv_finitef",
-                                       "__nv_isfinited");
+  populateOpPatterns<math::IsFiniteOp>(converter, patterns, benefit,
+                                       "__nv_finitef", "__nv_isfinited");
   populateOpPatterns<math::IsInfOp>(converter, patterns, benefit, "__nv_isinff",
                                     "__nv_isinfd");
   populateOpPatterns<math::IsNaNOp>(converter, patterns, benefit, "__nv_isnanf",

From ee95fa6b2336e947bcd10075c52333251d7d62ce Mon Sep 17 00:00:00 2001
From: Ivan Radanov Ivanov <ivanov.i.aa@m.titech.ac.jp>
Date: Tue, 18 Mar 2025 09:49:14 +0900
Subject: [PATCH 4/4] Also check function signature

---
 mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir | 1 +
 1 file changed, 1 insertion(+)

diff --git a/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir b/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
index ce8da38f24179..6d4555e815b66 100644
--- a/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
+++ b/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
@@ -1064,6 +1064,7 @@ gpu.module @test_module_54 {
   // CHECK: llvm.func @__nv_isinfd(f64) -> i32
   // CHECK: llvm.func @__nv_isnanf(f32) -> i32
   // CHECK: llvm.func @__nv_isnand(f64) -> i32
+  // CHECK: llvm.func @__nv_finitef(f32) -> i32
   // CHECK: llvm.func @__nv_isfinited(f64) -> i32
   // CHECK-LABEL: @fpclassify
   func.func @fpclassify(%f32: f32, %f64: f64) -> (i1, i1, i1, i1, i1, i1) {