[Driver][AMDGPU][HIP][SPIRV] Disable optimizations for AMDGCN SPIR-V (#154765)

SPIR-V specific optimizations can inadvertently remove information that
is important for the AMDGPU BE / break certain code patterns we rely on.
Therefore, for AMDGCN flavoured SPIR-V we disable optimizations over IR,
to ensure that we operate directly on the output of Clang CodeGen when
we finalise.

Author: AlexVlx

clang/lib/Driver/ToolChains/AMDGPU.cpp

@@ -858,6 +858,20 @@ void AMDGPUToolChain::addClangTargetOptions(
     CC1Args.push_back("-fapply-global-visibility-to-externs");
   }
 
+  // For SPIR-V we want to retain the pristine output of Clang CodeGen, since
+  // optimizations might lose structure / information that is necessary for
+  // generating optimal concrete AMDGPU code.
+  // TODO: using the below option is a temporary placeholder until Clang
+  //       provides the required functionality, which essentially boils down to
+  //       -O0 being refactored / reworked to not imply optnone / remove TBAA.
+  //       Once that is added, we should pivot to that functionality, being
+  //       mindful to not corrupt the user provided and subsequently embedded
+  //       command-line (i.e. if the user asks for -O3 this is what the
+  //       finalisation should use).
+  if (getTriple().isSPIRV() &&
+      !DriverArgs.hasArg(options::OPT_disable_llvm_optzns))
+    CC1Args.push_back("-disable-llvm-optzns");
+
   if (DeviceOffloadingKind == Action::OFK_None)
     addOpenCLBuiltinsLib(getDriver(), DriverArgs, CC1Args);
 }

clang/lib/Driver/ToolChains/HIPAMD.cpp

@@ -261,6 +261,12 @@ void HIPAMDToolChain::addClangTargetOptions(
     // with options that match the user-supplied ones.
     if (!DriverArgs.hasArg(options::OPT_fembed_bitcode_marker))
       CC1Args.push_back("-fembed-bitcode=marker");
+    // For SPIR-V we want to retain the pristine output of Clang CodeGen, since
+    // optimizations might lose structure / information that is necessary for
+    // generating optimal concrete AMDGPU code. We duplicate this because the
+    // HIP TC doesn't invoke the base AMDGPU TC addClangTargetOptions.
+    if (!DriverArgs.hasArg(options::OPT_disable_llvm_passes))
+      CC1Args.push_back("-disable-llvm-passes");
     return; // No DeviceLibs for SPIR-V.
   }
 

clang/test/Driver/hip-toolchain-no-rdc.hip

@@ -206,7 +206,7 @@
 // Check mixed AMDGCNSPIRV and concrete GPU arch.
 //
 
-// AMDGCNSPIRV: "-cc1" "-triple" "spirv64-amd-amdhsa" {{.*}}"-emit-llvm-bc" {{.*}} "-o" "[[AMDGCNSPV_BC:.*bc]]"
+// AMDGCNSPIRV: "-cc1" "-triple" "spirv64-amd-amdhsa" {{.*}}"-emit-llvm-bc" {{.*}}"-fembed-bitcode=marker" "-disable-llvm-passes" {{.*}} "-o" "[[AMDGCNSPV_BC:.*bc]]"
 // AMDGCNSPIRV: {{".*llvm-link.*"}} "-o" "[[AMDGCNSPV_TMP:.*out]]" "[[AMDGCNSPV_BC]]"
 // AMDGCNSPIRV: {{".*llvm-spirv.*"}} "--spirv-max-version=1.6" "--spirv-ext=+all" {{.*}} "[[AMDGCNSPV_TMP]]" {{.*}}"-o" "[[AMDGCNSPV_CO:.*out]]"
 // AMDGCNSPIRV: "-cc1" "-triple" "amdgcn-amd-amdhsa" {{.*}}"-emit-obj" {{.*}}"-target-cpu" "gfx900"{{.*}} "-o" "[[GFX900_OBJ:.*o]]"

clang/test/Driver/spirv-amd-toolchain.c

@@ -14,6 +14,6 @@
 
 // RUN: %clang -### --target=spirv64-amd-amdhsa %s -nogpulib -nogpuinc 2>&1 \
 // RUN:   | FileCheck %s --check-prefix=INVOCATION
-// INVOCATION: "-cc1" "-triple" "spirv64-amd-amdhsa" {{.*}} "-o" "[[OUTPUT:.+]]" "-x" "c"
+// INVOCATION: "-cc1" "-triple" "spirv64-amd-amdhsa" {{.*}}"-disable-llvm-optzns" {{.*}} "-o" "[[OUTPUT:.+]]" "-x" "c"
 // INVOCATION: "{{.*}}llvm-link" "-o" "a.out" "[[OUTPUT]]"
 // INVOCATION: "{{.*}}llvm-spirv" "--spirv-max-version=1.6" "--spirv-ext=+all" "--spirv-allow-unknown-intrinsics" "--spirv-lower-const-expr" "--spirv-preserve-auxdata" "--spirv-debug-info-version=nonsemantic-shader-200" "a.out" "-o" "a.out"