ARM: Avoid adding default libcalls overridden by AEABI functions

[arsenm]

Avoids adding alternative libcall impls for the same libcall.

I'm not sure if the default names exist or not, or are just not
preferred. compiler-rt appears to define aliases for all of these,
so I'm not sure why we bother distinguishing these in the first place.

[arsenm]

• ExpandFp: Require RuntimeLibcallsInfo analysis #165197
• Analysis: Add RuntimeLibcall analysis pass #165196
• RuntimeLibcalls: Split lowering decisions into LibcallLoweringInfo #164987
• ARM: Avoid adding default libcalls overridden by AEABI functions #164983 👈 (View in Graphite)
• X86: Make sure compiler-rt div calls are not added for msvc #164591
• RuntimeLibcalls: Make sure _Unwind_Resume entries are mutually exclusive #164195
• RuntimeLibcalls: Avoid reporting __stack_chk_guard as available for msvc #164133
• CodeGen: Remove overrides of getSSPStackGuardCheck (NFC) #164044
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[llvmbot]

@llvm/pr-subscribers-backend-x86
@llvm/pr-subscribers-backend-aarch64

@llvm/pr-subscribers-llvm-ir

Author: Matt Arsenault (arsenm)

Changes

Avoids adding alternative libcall impls for the same libcall.

I'm not sure if the default names exist or not, or are just not
preferred. compiler-rt appears to define aliases for all of these,
so I'm not sure why we bother distinguishing these in the first place.

---

Full diff: https://github.com/llvm/llvm-project/pull/164983.diff

1 Files Affected:

• (modified) llvm/include/llvm/IR/RuntimeLibcalls.td (+42-1)

diff --git a/llvm/include/llvm/IR/RuntimeLibcalls.td b/llvm/include/llvm/IR/RuntimeLibcalls.td
index 04e0ea3ee75a9..7be1b654ca727 100644
--- a/llvm/include/llvm/IR/RuntimeLibcalls.td
+++ b/llvm/include/llvm/IR/RuntimeLibcalls.td
@@ -1508,6 +1508,41 @@ def __aeabi_ddiv : RuntimeLibcallImpl<DIV_F64>; // CallingConv::ARM_AAPCS
 def __aeabi_dmul : RuntimeLibcallImpl<MUL_F64>; // CallingConv::ARM_AAPCS
 def __aeabi_dsub : RuntimeLibcallImpl<SUB_F64>; // CallingConv::ARM_AAPCS
 
+defvar AEABIOverrides = [
+  __eqsf2, __eqdf2,
+  __nesf2, __nedf2,
+  __ltsf2, __ltdf2,
+  __lesf2, __ledf2,
+  __gesf2, __gedf2,
+  __gtsf2, __gtdf2,
+  __unordsf2, __unorddf2,
+
+  __addsf3, __adddf3,
+  __divsf3, __divdf3,
+  __mulsf3, __muldf3,
+  __subsf3, __subdf3,
+
+  __fixdfsi, __fixunsdfsi,
+  __fixdfdi, __fixunsdfdi,
+  __fixsfsi, __fixunssfsi,
+  __fixsfdi, __fixunssfdi,
+
+  __floatsidf, __floatunsidf,
+  __floatdidf, __floatundidf,
+  __floatsisf, __floatunsisf,
+  __floatdisf, __floatundisf,
+
+  __muldi3, __ashldi3,
+  __lshrdi3, __ashrdi3,
+
+  __divsi3, __udivsi3
+
+  // Half conversion cases are a mess and handled separately.
+  //  __truncdfsf2, __truncdfhf2,
+  //  __extendsfdf2,
+  //  __truncsfhf2, __extendhfsf2
+];
+
 // Double-precision floating-point comparison helper functions
 // RTABI chapter 4.1.2, Table 3
 def __aeabi_dcmpeq__oeq : RuntimeLibcallImpl<OEQ_F64, "__aeabi_dcmpeq">; // CallingConv::ARM_AAPCS, CmpInst::ICMP_NE
@@ -1793,7 +1828,8 @@ def ARMSystemLibrary
     : SystemRuntimeLibrary<isARMOrThumb,
       (add (sub WinDefaultLibcallImpls, ARMLibgccHalfConvertCalls,
                                         GNUEABIHalfConvertCalls,
-                                        ARMDoubleToHalfCalls),
+                                        ARMDoubleToHalfCalls,
+                                        AEABIOverrides),
            LibcallImpls<(add __powisf2, __powidf2), isNotOSMSVCRT>,
            LibmHasFrexpF32, LibmHasLdexpF32,
            LibmHasFrexpF128, LibmHasLdexpF128,
@@ -1812,6 +1848,11 @@ def ARMSystemLibrary
            GNUEABIHalfConvertCalls,
            ARMDoubleToHalfCalls,
 
+           LibcallImpls<(add AEABIOverrides),
+             RuntimeLibcallPredicate<[{
+               (!hasAEABILibcalls(TT) || !isAAPCS_ABI(TT, ABIName)) &&
+               !TT.isOSWindows()
+            }]>>,
            // Use divmod compiler-rt calls for iOS 5.0 and later.
            LibcallImpls<(add __divmodsi4, __udivmodsi4),
                         RuntimeLibcallPredicate<[{TT.isOSBinFormatMachO() &&

[compnerd]

The AEABI names should be preferred. The aliases are for easing migration I believe. Switching to the GNU names for non-AEABI or non-AAPCS calling conventions on non-Windows targets is the correct behaviour.