[AIX][TLS] Add target attribute for -maix-small-local-exec-tls option.

[amyk]

Summary

This patch adds a target attribute for an AIX-specific option that
informs the compiler that it can use a faster access sequence for the
local-exec TLS model (formally named aix-small-local-exec-tls).

The Clang portion of this option is in D155544.
The initial implementation to generate the faster access sequence is in
D155600.

[nemanjai]

llvm/lib/Target/PowerPC/PPC.td

323–327

I don't think it's useful to add the entire help text here. I think something like:

Produce TOC-free TLS sequence for this function for 64-bit AIX

And describe in the comment that this is not really a Subtarget feature, but it it is used for convenience.

// Specifies that TLS access in any function with this target attribute should
// use the optimized TOC-free sequence (where the offset is an immediate
// off of R13 for which the linker might add fix-up code if the immediate
// is too large). Clearly, this isn't really a feature of the subtarget, but is
// used as a convenient way to affect code generation for individual functions.

Also, I don't think we need to use such a verbose function here. This isn't really a user-facing option, so I think we can safely abbreviate it a bit. Maybe aix-smalltls.

llvm/lib/Target/PowerPC/PPCSubtarget.cpp

128–135

I think it suffices to just fold this into a single condition:

if (HasAIXSmallLocalExecTLS && (!TargetTriple.isOSAIX() || !IsPPC64))
  report_fatal_error(
    "The aix-small-local-tls attribute is only supported on AIX in 64-bit mode");

llvm/test/CodeGen/PowerPC/check-aix-small-local-exec-tls-opt.ll

9

Please add another function to this test where the attribute is specified in the IR and it is compiled without -mattr=+aix-small-local-exec-tls but it is being compiled for something other than 64-bit AIX.

[hubert.reinterpretcast]

llvm/lib/Target/PowerPC/PPC.td

323–327

aix-smalltls is ambiguous with potential future development of the analogous local-dynamic access sequence.

[hubert.reinterpretcast]

llvm/lib/Target/PowerPC/PPC.td

323–327

Produce TOC-free TLS sequence for this function for 64-bit AIX

Similarly, this is only true for local-exec cases.

[amyk]

Address review comments:

• Updated comments
• Simplified conditions
• add a new test where the attribute is in the IR

llvm/lib/Target/PowerPC/PPC.td

323–327

I've updated the comment a bit to address both Nemanja and Hubert's concerns.
As Hubert mentioned, I will probably just keep the option as aix-small-local-exec-tls.

llvm/lib/Target/PowerPC/PPCSubtarget.cpp

128–135

Yeah, I agree.
Done.

[hubert.reinterpretcast]

llvm/test/CodeGen/PowerPC/check-aix-small-local-exec-tls-opt-IRattribute.ll

13–14

Same comment re: check prefixes.

llvm/test/CodeGen/PowerPC/check-aix-small-local-exec-tls-opt.ll

16–17

No need for different check prefixes if the expected output is the same.

[amyk]

Address Hubert's review comment of removing the duplicate CHECK.

[hubert.reinterpretcast]

LGTM; thanks!

llvm/lib/Target/PowerPC/PPC.td

 def FeatureROPProtect :
   SubtargetFeature<"rop-protect", "HasROPProtect", "true",
                    "Add ROP protect">;
 
 def FeaturePrivileged :
   SubtargetFeature<"privileged", "HasPrivileged", "true",
                    "Add privileged instructions">;
 
+// Specifies that local-exec TLS accesses in any function with this target
+// attribute should use the optimized TOC-free sequence (where the offset is an
+// immediate off of R13 for which the linker might add fix-up code if the
+// immediate is too large).
+// Clearly, this isn't really a feature of the subtarget, but is used as a
+// convenient way to affect code generation for individual functions.
+def FeatureAIXLocalExecTLS :
+  SubtargetFeature<"aix-small-local-exec-tls", "HasAIXSmallLocalExecTLS", "true",
+                   "Produce a TOC-free local-exec TLS sequence for this function "
+                   "for 64-bit AIX">;
+
 def FeaturePredictableSelectIsExpensive :
   SubtargetFeature<"predictable-select-expensive",
                    "PredictableSelectIsExpensive",
                    "true",
                    "Prefer likely predicted branches over selects">;
 
 def FeatureFastMFLR : SubtargetFeature<"fast-MFLR", "HasFastMFLR", "true",
                                        "MFLR is a fast instruction">;

llvm/lib/Target/PowerPC/PPCSubtarget.cpp

void PPCSubtarget::initSubtargetFeatures(StringRef CPU, StringRef TuneCPU,
   // If not SPE, set standard FPU
   if (!HasSPE)
     HasFPU = true;
 
   StackAlignment = getPlatformStackAlignment();
 
   // Determine endianness.
   IsLittleEndian = TM.isLittleEndian();
+
+  if (HasAIXSmallLocalExecTLS && (!TargetTriple.isOSAIX() || !IsPPC64))
+    report_fatal_error(
+      "The aix-small-local-exec-tls attribute is only supported on AIX in "
+      "64-bit mode.\n", false);
 }
 
 bool PPCSubtarget::enableMachineScheduler() const { return true; }
 
 bool PPCSubtarget::enableMachinePipeliner() const {
   return getSchedModel().hasInstrSchedModel() && EnableMachinePipeliner;
 }
 
 bool PPCSubtarget::useDFAforSMS() const { return false; }
 
 // This overrides the PostRAScheduler bit in the SchedModel for each CPU.
 bool PPCSubtarget::enablePostRAScheduler() const { return true; }
 

llvm/test/CodeGen/PowerPC/check-aix-small-local-exec-tls-opt-IRattribute.ll

+; RUN: llc -mtriple powerpc64-ibm-aix-xcoff -ppc-asm-full-reg-names \
+; RUN:   < %s | FileCheck %s
+; RUN: not llc -mtriple powerpc-ibm-aix-xcoff -ppc-asm-full-reg-names \
+; RUN:   < %s 2>&1 | FileCheck %s --check-prefix=CHECK-NOT-SUPPORTED-AIX32
+; RUN: not llc -mtriple powerpc64le-unknown-linux-gnu -ppc-asm-full-reg-names \
+; RUN:   < %s 2>&1 | FileCheck %s --check-prefix=CHECK-NOT-SUPPORTED
+
+define dso_local signext i32 @testWithIRAttr() #0 {
+entry:
+  ret i32 0
+}
+; Check that the aix-small-local-exec-tls attribute is not supported on Linux and AIX (32-bit).
+; CHECK-NOT-SUPPORTED: The aix-small-local-exec-tls attribute is only supported on AIX in 64-bit mode.
+; CHECK-NOT-SUPPORTED-AIX32: The aix-small-local-exec-tls attribute is only supported on AIX in 64-bit mode.
+
+; Make sure that the test was actually compiled successfully after using the
+; aix-small-local-exec-tls attribute.
+; CHECK-LABEL: testWithIRAttr:
+; CHECK:        li r3, 0
+; CHECK-NEXT:   blr
+
+
+attributes #0 = { "target-features"="+aix-small-local-exec-tls" }
+

llvm/test/CodeGen/PowerPC/check-aix-small-local-exec-tls-opt.ll

+; RUN: llc -mtriple powerpc64-ibm-aix-xcoff -mattr=+aix-small-local-exec-tls \
+; RUN:   -ppc-asm-full-reg-names < %s | FileCheck %s
+; RUN: not llc -mtriple powerpc-ibm-aix-xcoff -mattr=+aix-small-local-exec-tls \
+; RUN:   -ppc-asm-full-reg-names < %s 2>&1 | \
+; RUN:   FileCheck %s --check-prefix=CHECK-NOT-SUPPORTED-AIX32
+; RUN: not llc -mtriple powerpc64le-unknown-linux-gnu -mattr=+aix-small-local-exec-tls \
+; RUN:   -ppc-asm-full-reg-names < %s 2>&1 | \
+; RUN:   FileCheck %s --check-prefix=CHECK-NOT-SUPPORTED
+
+define dso_local signext i32 @testNoIRAttr() {
+entry:
+  ret i32 0
+}
+
+; Check that the aix-small-local-exec-tls attribute is not supported on Linux and AIX (32-bit).
+; CHECK-NOT-SUPPORTED: The aix-small-local-exec-tls attribute is only supported on AIX in 64-bit mode.
+; CHECK-NOT-SUPPORTED-AIX32: The aix-small-local-exec-tls attribute is only supported on AIX in 64-bit mode.
+
+; Make sure that the test was actually compiled successfully after using the
+; aix-small-local-exec-tls attribute.
+; CHECK-LABEL: testNoIRAttr:
+; CHECK:        li r3, 0
+; CHECK-NEXT:   blr