Raise __nv_sinpi/__nv_sinpif in libdevice pass for kernel lowering

[Copilot]

Kernel compilation failed when CUDA emitted llvm.call @__nv_sinpi because libdeviceraise had no rewrite for this symbol. This PR adds explicit raising so sinpi in GPU kernels can be converted into supported math ops during lowering.

• Raising support in LibDeviceFuncsRaisingPass

  • Added a dedicated SinPiRaising pattern for:
    • __nv_sinpi (f64)
    • __nv_sinpif (f32)
  • Rewrites each call to sin(x * π) using MLIR ops:
    • arith.constant (typed π)
    • arith.mulf
    • math.sin
  • Registered the new pattern in populateLibDeviceFuncsToOpsPatterns.

• Targeted lit coverage

  • Extended test/lit_tests/raising/libdevice_raise.mlir with a test_module_sinpi case validating both float and double variants raise to mul-by-π + math.sin.

%pi = arith.constant ... : f64
%scaled = arith.mulf %arg, %pi : f64
%out = math.sin %scaled : f64

[wsmoses]

@copilot

ot/.bazel
INFO: Options provided by the client:
Inherited 'common' options: --isatty=0 --terminal_columns=80
INFO: Reading rc options for 'build' from /__w/Enzyme-JAX/Enzyme-JAX/.bazelrc:
Inherited 'common' options: --noenable_bzlmod --noincompatible_enable_cc_toolchain_resolution --announce_rc --repo_env=USE_PYWRAP_RULES=True --copt=-DGRPC_BAZEL_BUILD --host_copt=-DGRPC_BAZEL_BUILD --action_env=GRPC_BAZEL_RUNTIME=1 --repo_env=PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=upb --action_env=PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=upb --experimental_repo_remote_exec --cxxopt=-Wno-deprecated-declarations --host_cxxopt=-Wno-deprecated-declarations --cxxopt=-Wno-sign-compare --host_cxxopt=-Wno-sign-compare --cxxopt=-Wno-array-bounds --host_cxxopt=-Wno-array-bounds --cxxopt=-Wno-comment --host_cxxopt=-Wno-comment --cxxopt=-Wno-invalid-constexpr --host_cxxopt=-Wno-invalid-constexpr --cxxopt=-Wno-mismatched-new-delete --host_cxxopt=-Wno-mismatched-new-delete --cxxopt=-Wno-unused-but-set-parameter --host_cxxopt=-Wno-unused-but-set-parameter --cxxopt=-Wno-unused-but-set-variable --host_cxxopt=-Wno-unused-but-set-variable --cxxopt=-Wno-misleading-indentation --host_cxxopt=-Wno-misleading-indentation --cxxopt=-Wno-return-type --host_cxxopt=-Wno-return-type --cxxopt=-Wno-unused-function --host_cxxopt=-Wno-unused-function --cxxopt=-Wno-parentheses --host_cxxopt=-Wno-parentheses --cxxopt=-Wno-unused-local-typedefs --host_cxxopt=-Wno-unused-local-typedefs --cxxopt=-Wno-reorder --host_cxxopt=-Wno-reorder --cxxopt=-Wno-missing-braces --host_cxxopt=-Wno-missing-braces --cxxopt=-Wno-range-loop-construct --host_cxxopt=-Wno-range-loop-construct --cxxopt=-std=c++17 --host_cxxopt=-std=c++17 --color=yes --define=grpc_no_ares=true --define=tsl_link_protobuf=true --define open_source_build=true --define framework_shared_object=true --define tsl_protobuf_header_only=true --define=allow_oversize_protos=true
INFO: Reading rc options for 'build' from /__w/Enzyme-JAX/Enzyme-JAX/.bazelrc:
'build' options: --@rules_python//python/config_settings:precompile=force_disabled -c opt
INFO: Found applicable config definition common:public_cache_push in file /__w/Enzyme-JAX/Enzyme-JAX/.bazelrc: --config=public_cache --remote_upload_local_results=true --google_default_credentials
INFO: Found applicable config definition common:public_cache in file /__w/Enzyme-JAX/Enzyme-JAX/.bazelrc: --remote_cache=https://storage.googleapis.com/enzyme-presubmit-bazel-cache/ --remote_upload_local_results=false
Computing main repo mapping:
WARNING: option '--remote_upload_local_results' was expanded to from both option '--config=public_cache_push' (source command line options) and option '--config=public_cache_push' (source command line options)
Loading:
Loading: 0 packages loaded
Analyzing: target //:enzymexlamlir-opt (1 packages loaded, 0 targets configured)
Analyzing: target //:enzymexlamlir-opt (1 packages loaded, 0 targets configured)

DEBUG: Rule 'cuda_tile' indicated that a canonical reproducible form can be obtained by modifying arguments integrity = "sha256-YBqJhzobmsmk4EAMPss513cnk2+qgud2uJz2qf6BeFk="
DEBUG: Repository cuda_tile instantiated at:
/__w/Enzyme-JAX/Enzyme-JAX/WORKSPACE:346:20: in
/__w/Enzyme-JAX/Enzyme-JAX/third_party/cuda_tile/workspace.bzl:16:17: in repo
Repository rule http_archive defined at:
/root/.bazel/external/bazel_tools/tools/build_defs/repo/http.bzl:400:31: in
DEBUG: Rule 'enzyme' indicated that a canonical reproducible form can be obtained by modifying arguments integrity = "sha256-cMEBORhThaDLipZcFoHqMeEhbDLqCL2HcXL7poI2fSY="
DEBUG: Repository enzyme instantiated at:
/__w/Enzyme-JAX/Enzyme-JAX/WORKSPACE:344:17: in
/__w/Enzyme-JAX/Enzyme-JAX/third_party/enzyme/workspace.bzl:13:21: in repo
Repository rule http_archive defined at:
/root/.bazel/external/bazel_tools/tools/build_defs/repo/http.bzl:400:31: in
Analyzing: target //:enzymexlamlir-opt (118 packages loaded, 9242 targets configured)

INFO: Analyzed target //:enzymexlamlir-opt (158 packages loaded, 15955 targets configured).
[2,251 / 8,817] Compiling absl/hash/internal/city.cc; 0s remote-cache ... (32 actions, 0 running)
INFO: From Compiling hwloc/memattrs.c:
external/hwloc/hwloc/memattrs.c:770:9: warning: unused variable 'err' [-Wunused-variable]
770 | int err = from_internal_location(&imi->initiator, &initiators[i]);
| ^~~
| ^
external/shardy/shardy/dialect/sdy/transforms/propagation/basic_propagation.h:55:34: note: default constructor of 'BasicPropagationPassImpl' is implicitly deleted because base class 'OperationPass' has no default constructor
55 | class BasicPropagationPassImpl : public OperationPass {
| ^
external/shardy/shardy/dialect/sdy/transforms/propagation/basic_propagation.h:59:32: note: replace 'default' with 'delete'
59 | BasicPropagationPassImpl() = default;
| ^~~~~~~
| delete
1 warning generated.
INFO: From Compiling third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/TensorMemoryToLLVM.cpp:
external/triton/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/TensorMemoryToLLVM.cpp:361:21: warning: lambda capture 'kRow' is not used [-Wunused-lambda-capture]
361 | auto getRowCol = [kRow, kCol](const auto &rowCol) {
| ^~~~~
external/triton/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/TensorMemoryToLLVM.cpp:361:27: warning: lambda capture 'kCol' is not used [-Wunused-lambda-capture]
361 | auto getRowCol = [kRow, kCol](const auto &rowCol) {
| ~~^~~~
2 warnings generated.
INFO: From Compiling src/enzyme_ad/jax/Passes/ConvertPolygeistToLLVM.cpp:
src/enzyme_ad/jax/Passes/ConvertPolygeistToLLVM.cpp:2651:8: warning: private field 'kernelBarePtrCallConv' is not used [-Wunused-private-field]
2651 | bool kernelBarePtrCallConv;
| ^
1 warning generated.
INFO: From Compiling third_party/amd/lib/TritonAMDGPUToLLVM/SchedInstructions.cpp:
external/triton/third_party/amd/lib/TritonAMDGPUToLLVM/SchedInstructions.cpp:99:11: warning: private field 'numStages' is not used [-Wunused-private-field]
99 | int32_t numStages;
| ^
1 warning generated.
[8,852 / 8,884] Compiling src/enzyme_ad/jax/Passes/LLVMToMemrefAccess.cpp; 0s remote-cache ... (31 actions, 0 running)
[8,882 / 8,884] [Sched] Compiling src/enzyme_ad/jax/Passes/LibDeviceFuncsRaisingPass.cpp; Downloading external/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUGenRegisterInfoTargetDesc.inc, 56.8 MiB / 107.4 MiB
[8,882 / 8,884] Compiling src/enzyme_ad/jax/Passes/LibDeviceFuncsRaisingPass.cpp; 0s remote-cache, processwrapper-sandbox
[8,882 / 8,884] Compiling src/enzyme_ad/jax/Passes/LibDeviceFuncsRaisingPass.cpp; 11s remote-cache, processwrapper-sandbox
ERROR: /__w/Enzyme-JAX/Enzyme-JAX/src/enzyme_ad/jax/BUILD:1164:11: Compiling src/enzyme_ad/jax/Passes/LibDeviceFuncsRaisingPass.cpp failed: (Exit 1): clang failed: error executing CppCompile command (from target //src/enzyme_ad/jax:XLADerivatives) /usr/lib/llvm-18/bin/clang -U_FORTIFY_SOURCE -fstack-protector -Wall -Wthread-safety -Wself-assign -Wunused-but-set-parameter -Wno-free-nonheap-object -fcolor-diagnostics -fno-omit-frame-pointer -g0 ... (remaining 631 arguments skipped)

Use --sandbox_debug to see verbose messages from the sandbox and retain the sandbox build root for debugging
src/enzyme_ad/jax/Passes/LibDeviceFuncsRaisingPass.cpp:462:35: error: no member named 'getPi' in 'llvm::APFloat'
462 | floatType, llvm::APFloat::getPi(floatType.getFloatSemantics()));
| ~~~~~~~~~~~~~~~^
1 error generated.
Target //:enzymexlamlir-opt failed to build
Use --verbose_failures to see the command lines of failed build steps.
INFO: Elapsed time: 48.643s, Critical Path: 18.83s
INFO: 8883 processes: 6228 remote cache hit, 2655 internal.
ERROR: Build did NOT complete successfully

[Copilot]

@copilot

ot/.bazel
...

Fixed in 64d1953. I traced the CI failure to APFloat::getPi not being available in the LLVM version used by CI and updated SinPiRaising to use LLVM18-compatible, semantics-aware APFloat construction for the f32/f64 π constants.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 25.46%. Comparing base (970aad9) to head (64d1953).
⚠️ Report is 6 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2542      +/-   ##
==========================================
- Coverage   25.66%   25.46%   -0.21%     
==========================================
  Files         223      223              
  Lines       44823    44669     -154     
==========================================
- Hits        11506    11375     -131     
+ Misses      33317    33294      -23     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[wsmoses]

@copilot

[17,595 / 17,734] 911 / 1050 tests, 2 failed; Testing //test/lit_tests:raising_affine_to_stablehlo_gu.mlir.test; 6s remote-cache, processwrapper-sandbox ... (32 actions, 31 running)
FAIL: //test/lit_tests:raising_libdevice_raise.mlir.test (see /root/.bazel/execroot/__main__/bazel-out/k8-dbg/testlogs/test/lit_tests/raising_libdevice_raise.mlir.test/test.log)
INFO: From Testing //test/lit_tests:raising_libdevice_raise.mlir.test:
==================== Test output for //test/lit_tests:raising_libdevice_raise.mlir.test:
-- Testing: 1 tests, 1 workers --
FAIL: Enzyme-JaX :: lit_tests/raising/libdevice_raise.mlir (1 of 1)
******************** TEST 'Enzyme-JaX :: lit_tests/raising/libdevice_raise.mlir' FAILED ********************
Exit Code: 1

Command Output (stderr):
--
enzymexlamlir-opt --libdevice-funcs-raise /root/.bazel/sandbox/processwrapper-sandbox/15960/execroot/__main__/bazel-out/k8-dbg/bin/test/lit_tests/raising_libdevice_raise.mlir.test.runfiles/__main__/test/lit_tests/raising/libdevice_raise.mlir | FileCheck /root/.bazel/sandbox/processwrapper-sandbox/15960/execroot/__main__/bazel-out/k8-dbg/bin/test/lit_tests/raising_libdevice_raise.mlir.test.runfiles/__main__/test/lit_tests/raising/libdevice_raise.mlir # RUN: at line 1
+ enzymexlamlir-opt --libdevice-funcs-raise /root/.bazel/sandbox/processwrapper-sandbox/15960/execroot/__main__/bazel-out/k8-dbg/bin/test/lit_tests/raising_libdevice_raise.mlir.test.runfiles/__main__/test/lit_tests/raising/libdevice_raise.mlir
+ FileCheck /root/.bazel/sandbox/processwrapper-sandbox/15960/execroot/__main__/bazel-out/k8-dbg/bin/test/lit_tests/raising_libdevice_raise.mlir.test.runfiles/__main__/test/lit_tests/raising/libdevice_raise.mlir
/root/.bazel/sandbox/processwrapper-sandbox/15960/execroot/__main__/bazel-out/k8-dbg/bin/test/lit_tests/raising_libdevice_raise.mlir.test.runfiles/__main__/test/lit_tests/raising/libdevice_raise.mlir:247:12: error: CHECK: expected string not found in input
 // CHECK: %[[PI_F64:.*]] = arith.constant {{.*}} : f64
           ^
<stdin>:232:24: note: scanning from here
 %2 = math.sin %1 : f32
                       ^
<stdin>:234:2: note: possible intended match here
 %4 = math.sin %3 : f64
 ^

Input file: <stdin>
Check file: /root/.bazel/sandbox/processwrapper-sandbox/15960/execroot/__main__/bazel-out/k8-dbg/bin/test/lit_tests/raising_libdevice_raise.mlir.test.runfiles/__main__/test/lit_tests/raising/libdevice_raise.mlir

-dump-input=help explains the following input dump.

Input was:
<<<<<<
             .
             .
             .
           227:  llvm.func @gpu_sinpi(%arg0: f32, %arg1: f64) -> !llvm.struct<(f32, f64)> attributes {llvm.emit_c_interface} { 
           228:  %cst = arith.constant 3.1415926535897931 : f64 
           229:  %0 = llvm.mlir.undef : !llvm.struct<(f32, f64)> 
           230:  %cst_0 = arith.constant 3.14159274 : f32 
           231:  %1 = arith.mulf %arg0, %cst_0 : f32 
           232:  %2 = math.sin %1 : f32 
check:247'0                            X error: no match found
           233:  %3 = arith.mulf %arg1, %cst : f64 
check:247'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
           234:  %4 = math.sin %3 : f64 
check:247'0     ~~~~~~~~~~~~~~~~~~~~~~~~
check:247'1      ?                       possible intended match
           235:  %5 = llvm.insertvalue %2, %0[0] : !llvm.struct<(f32, f64)>  
check:247'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
           236:  %6 = llvm.insertvalue %4, %5[1] : !llvm.struct<(f32, f64)>  
check:247'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
           237:  llvm.return %6 : !llvm.struct<(f32, f64)> 
check:247'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
           238:  } 
check:247'0     ~~~
           239:  } 
check:247'0     ~~~
             .
             .
             .
>>>>>>

[Copilot]

@copilot

[17,595 / 17,734] 911 / 1050 tests, 2 failed; Testing //test/lit_tests:raising_affine_to_stablehlo_gu.mlir.test; 6s remote-cache, processwrapper-sandbox ... (32 actions, 31 running)

...

Fixed in 6177477. I updated the gpu_sinpi lit checks to be order-robust for independent f32/f64 pi constants (CHECK-DAG) while keeping ordered CHECK assertions for each dependent mul→sin chain, which matches the observed lowering output and avoids the CI mismatch.