[Flang] Force lowering to Complex for AMDGPU

[TIFitis]

Avoid libm when targeting AMDGPU as those symbols are not available on the device and we rely on MLIR complex operations to later map to OCML calls.

PRs #144924 & #144926 add additional related support.

[Copilot]

Pull Request Overview

This PR implements a change to avoid using libm when targeting AMDGPU by forcing lowering to complex operations which later map to OCML calls.

• Add a new check to detect AMDGPU targets using fir::getTargetTriple().isAMDGCN().
• Bypass the invocation of the libm call when AMDGPU is detected.

Comments suppressed due to low confidence (1)

flang/lib/Optimizer/Builder/IntrinsicCall.cpp:1234

• [nitpick] Consider adding unit tests to validate the AMDGPU branch, ensuring that libm is correctly bypassed when targeting AMDGPU.

    bool isAMDGPU = fir::getTargetTriple(builder.getModule()).isAMDGCN();

[llvmbot]

@llvm/pr-subscribers-flang-fir-hlfir

Author: Akash Banerjee (TIFitis)

Changes

Avoid libm when targeting AMDGPU as those symbols are not available on the device and we rely on MLIR complex operations to later map to OCML calls.

PRs #144924 & #144926 add additional related support.

---

Full diff: https://github.com/llvm/llvm-project/pull/144927.diff

1 Files Affected:

• (modified) flang/lib/Optimizer/Builder/IntrinsicCall.cpp (+5-2)

diff --git a/flang/lib/Optimizer/Builder/IntrinsicCall.cpp b/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
index 178b6770d6b53..2ca7f2784aab3 100644
--- a/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
+++ b/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
@@ -1228,8 +1228,11 @@ mlir::Value genComplexMathOp(fir::FirOpBuilder &builder, mlir::Location loc,
   llvm::StringRef mathLibFuncName = mathOp.runtimeFunc;
   if (!mathLibFuncName.empty()) {
     // If we enabled MLIR complex or can use approximate operations, we should
-    // NOT use libm.
-    if (!forceMlirComplex && !canUseApprox) {
+    // NOT use libm. Avoid libm when targeting AMDGPU as those symbols are not
+    // available on the device and we rely on MLIR complex operations to
+    // later map to OCML calls.
+    bool isAMDGPU = fir::getTargetTriple(builder.getModule()).isAMDGCN();
+    if (!forceMlirComplex && !canUseApprox && !isAMDGPU) {
       result = genLibCall(builder, loc, mathOp, mathLibFuncType, args);
       LLVM_DEBUG(result.dump(); llvm::dbgs() << "\n");
       return result;

[jsjodin]

LGTM