[SYCL][CUDA][HIP] Propagate -foffload-fp32-prec-sqrt (#17044)

-foffload-fp32-prec-sqrt and -fsycl-fp32-prec-sqrt options should be
merged together as they have the same purpose.
In this patch ability of -fsycl-fp32-prec-sqrt to pass appropriate
options to CUDA and HIP compilers was added to
-foffload-fp32-prec-sqrt to allow such merge in the future.

---------

Signed-off-by: Sidorov, Dmitry <[email protected]>

Author: MrSidims

clang/lib/Driver/ToolChains/AMDGPU.cpp

@@ -1056,7 +1056,8 @@ llvm::SmallVector<std::string, 12> ROCMToolChain::getCommonDeviceLibNames(
   bool CorrectSqrt = false;
   if (DeviceOffloadingKind == Action::OFK_SYCL) {
     // When using SYCL, sqrt is only correctly rounded if the flag is specified
-    CorrectSqrt = DriverArgs.hasArg(options::OPT_fsycl_fp32_prec_sqrt);
+    CorrectSqrt = DriverArgs.hasArg(options::OPT_fsycl_fp32_prec_sqrt) ||
+                  DriverArgs.hasArg(options::OPT_foffload_fp32_prec_sqrt);
   } else
     CorrectSqrt = DriverArgs.hasFlag(
         options::OPT_fhip_fp32_correctly_rounded_divide_sqrt,

clang/lib/Driver/ToolChains/Cuda.cpp

@@ -965,7 +965,8 @@ void CudaToolChain::addClangTargetOptions(
   if (DeviceOffloadingKind == Action::OFK_SYCL) {
     SYCLInstallation.addSYCLIncludeArgs(DriverArgs, CC1Args);
 
-    if (DriverArgs.hasArg(options::OPT_fsycl_fp32_prec_sqrt))
+    if (DriverArgs.hasArg(options::OPT_fsycl_fp32_prec_sqrt) ||
+        DriverArgs.hasArg(options::OPT_foffload_fp32_prec_sqrt))
       CC1Args.push_back("-fcuda-prec-sqrt");
 
     bool FastRelaxedMath = DriverArgs.hasFlag(

clang/test/Driver/sycl-amdgcn-sqrt.cpp

@@ -9,6 +9,14 @@
 // RUN:   %s \
 // RUN: 2>&1 | FileCheck  --check-prefix=CHECK-CORRECT %s
 
+// RUN: %clang -### \
+// RUN:   -fsycl -fsycl-targets=amdgcn-amd-amdhsa -fno-sycl-libspirv \
+// RUN:   -Xsycl-target-backend --offload-arch=gfx900 \
+// RUN:   -foffload-fp32-prec-sqrt \
+// RUN:   --rocm-path=%S/Inputs/rocm \
+// RUN:   %s \
+// RUN: 2>&1 | FileCheck  --check-prefix=CHECK-CORRECT %s
+
 // CHECK-CORRECT: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_correctly_rounded_sqrt_on.bc"
 
 // RUN: %clang -### \
@@ -28,6 +36,14 @@
 // RUN:   %s \
 // RUN: 2>&1 | FileCheck  --check-prefix=CHECK-CONFLICT %s
 
+// RUN: %clang -### \
+// RUN:   -fsycl -fsycl-targets=amdgcn-amd-amdhsa  -fno-sycl-libspirv \
+// RUN:   -Xsycl-target-backend --offload-arch=gfx900 \
+// RUN:   -foffload-fp32-prec-sqrt -fno-hip-fp32-correctly-rounded-divide-sqrt \
+// RUN:   --rocm-path=%S/Inputs/rocm \
+// RUN:   %s \
+// RUN: 2>&1 | FileCheck  --check-prefix=CHECK-CONFLICT %s
+
 // CHECK-CONFLICT: warning: argument unused during compilation: '-fno-hip-fp32-correctly-rounded-divide-sqrt'
 // CHECK-CONFLICT: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_correctly_rounded_sqrt_on.bc"
 

clang/test/Driver/sycl-nvptx-sqrt.cpp

@@ -6,6 +6,12 @@
 // RUN:   %s \
 // RUN: 2>&1 | FileCheck --check-prefix=CHECK-CORRECT %s
 
+// RUN: %clang -### -nocudalib \
+// RUN:   -fsycl -fsycl-targets=nvptx64-nvidia-cuda \
+// RUN:   -foffload-fp32-prec-sqrt \
+// RUN:   %s \
+// RUN: 2>&1 | FileCheck --check-prefix=CHECK-CORRECT %s
+
 // CHECK-CORRECT: "-fcuda-prec-sqrt"
 
 // RUN: %clang -### -nocudalib \

llvm/docs/NVPTXUsage.rst

@@ -1049,8 +1049,9 @@ The following sets the ftz flag to 1, and the precise sqrt flag to 1.
 
 .. code-block:: llvm
 
-    !llvm.module.flags = !{!0}
+    !llvm.module.flags = !{!0, !1}
     !0 = !{i32 4, !"nvvm-reflect-ftz", i32 1}
+    !1 = !{i32 4, !"nvvm-reflect-prec-sqrt", i32 1}
 
 (``i32 4`` indicates that the value set here overrides the value in another
 module we link with.  See the `LangRef <LangRef.html#module-flags-metadata>`