[mips] Fix an issue with long double when function roundl is defined

[zbuljan]

Summary

Patch adds roundl to a list of library calls in isF128SoftLibCall method (MipsCCState.cpp).
Test file roundl-call.ll is also added to LLVM regression tests.

Problem description:
There was a problem in LLVM backend that is triggered with next C code:

#include <math.h>

int main() {
  long double x, y;
  y = roundl(x);
  return 0;
}

With clang we generate IR code using -emit-llvm option (roundl-call.ll):

clang -static -O0 --target=mips64-img-linux-gnu -EL -mcpu=mips64r6 -mabi=64 -mhard-float -no-integrated-as --gcc-toolchain=/home/rtrk/mips-toolchain/mips-img-linux-gnu/2015.01-7 -S -emit-llvm roundl-test.c

Then we generate assembler code from roundl-call.ll with llc:

llc -march=mips64el -mcpu=mips64r3 -target-abi=n64 -O2 < roundl-call.ll

We can see that generated assembler code is using the wrong convention for roundl call.

Soft-float registers V0_64 and V1_64 are used for storing result instead of hard-float registers:

sd  $3, 8($fp)
sd  $2, 0($fp)

After the patch is applied we generate assembler code again from the same .ll file.
Now we can see that FPU registers D0_64 and D2_64 are used as it is expected for hard-float:

sdc1  $f2, 8($fp)
sdc1  $f0, 0($fp)

*This bug is also reported on bugzilla (https://dmz-portal.mips.com/bugz/show_bug.cgi?id=2303).

[dsanders]

#include <math.h>

int main() {
  long double x, y;
  y = roundl(x);
  return 0;
}

The majority of the function is dead code and x is uninitialized.

clang -static -O0 --target=mips64-img-linux-gnu -EL -mcpu=mips64r6 -mabi=64 -mhard-float -no-integrated-as --gcc-toolchain=/home/rtrk/mips-toolchain/mips-img-linux-gnu/2015.01-7 -S -emit-llvm roundl-test.c

Then we generate assembler code from roundl-call.ll with llc:

llc -march=mips64el -mcpu=mips64r3 -target-abi=n64 -O2 < roundl-call.ll

I notice these two commands use different CPU's and that the ISA's for the two are incompatible.

test/CodeGen/Mips/roundl-call.ll

1 ↗

(On Diff #49476)

Can you put calling convention tests in test/CodeGen/Mips/cconv/?

return-hard-fp128.ll and arguments-hard-fp128.ll are the fp128 hard-float tests. Could you test in a similar way to those two?
In particular:

• We need to cover N32.
• We need to ensure that soft-float still works correctly.
• We don't need to test mips64r2/mips64r3 explicitly since there's no difference compared to mips64.
• We should eliminate as much unnecessary code as possible. There's usually no need for alloca's, loads, stores, attributes, etc. in these tests.

[zbuljan]

Updated .ll with tests for N32 and soft-float.
Removed unnecessary IR code.
.ll file moved to test/CodeGen/Mips/cconv/ folder.

[dsanders]

LGTM

llvm/trunk/lib/Target/Mips/MipsCCState.cpp

const char *const LibCalls[] = {
       "__fixunstfdi",  "__fixunstfsi", "__fixunstfti",  "__floatditf",
       "__floatsitf",   "__floattitf",  "__floatunditf", "__floatunsitf",
       "__floatuntitf", "__getf2",      "__gttf2",       "__letf2",
       "__lttf2",       "__multf3",     "__netf2",       "__powitf2",
       "__subtf3",      "__trunctfdf2", "__trunctfsf2",  "__unordtf2",
       "ceill",         "copysignl",    "cosl",          "exp2l",
       "expl",          "floorl",       "fmal",          "fmodl",
       "log10l",        "log2l",        "logl",          "nearbyintl",
-      "powl",          "rintl",        "sinl",          "sqrtl",
-      "truncl"};
+      "powl",          "rintl",        "roundl",        "sinl",
+      "sqrtl",         "truncl"};
 
   // Check that LibCalls is sorted alphabetically.
   auto Comp = [](const char *S1, const char *S2) { return strcmp(S1, S2) < 0; };
   assert(std::is_sorted(std::begin(LibCalls), std::end(LibCalls), Comp));
   return std::binary_search(std::begin(LibCalls), std::end(LibCalls),
                             CallSym, Comp);
 }
 

llvm/trunk/test/CodeGen/Mips/cconv/roundl-call.ll

+; RUN: llc -march=mips64 -mcpu=mips64 -target-abi=n32 < %s | FileCheck %s \
+; RUN:     -check-prefix=ALL -check-prefix=N32 -check-prefix=HARD-FLOAT
+; RUN: llc -march=mips64el -mcpu=mips64 -target-abi=n32 < %s | FileCheck %s \
+; RUN:     -check-prefix=ALL -check-prefix=N32 -check-prefix=HARD-FLOAT
+
+; RUN: llc -march=mips64 -mcpu=mips64 -target-abi=n64 < %s | FileCheck %s \
+; RUN:     -check-prefix=ALL -check-prefix=N64 -check-prefix=HARD-FLOAT
+; RUN: llc -march=mips64el -mcpu=mips64 -target-abi=n64 < %s | FileCheck %s \
+; RUN:     -check-prefix=ALL -check-prefix=N64 -check-prefix=HARD-FLOAT
+
+; RUN: llc -march=mips64 -mcpu=mips64 -mattr=+soft-float -target-abi=n32 < %s \
+; RUN: | FileCheck %s -check-prefix=ALL -check-prefix=N32 \
+; RUN:                -check-prefix=SOFT-FLOAT
+; RUN: llc -march=mips64el -mcpu=mips64 -mattr=+soft-float -target-abi=n32 < \
+; RUN:     %s | FileCheck %s -check-prefix=ALL -check-prefix=N32 \
+; RUN:                       -check-prefix=SOFT-FLOAT
+
+; RUN: llc -march=mips64 -mcpu=mips64 -mattr=+soft-float -target-abi=n64 < %s \
+; RUN: | FileCheck %s -check-prefix=ALL -check-prefix=N64 \
+; RUN:                -check-prefix=SOFT-FLOAT
+; RUN: llc -march=mips64el -mcpu=mips64 -mattr=+soft-float -target-abi=n64 < \
+; RUN:     %s | FileCheck %s -check-prefix=ALL -check-prefix=N64 \
+; RUN:                       -check-prefix=SOFT-FLOAT
+
+@fp128 = global fp128 zeroinitializer
+
+define void @roundl_call(fp128 %value) {
+entry:
+; ALL-LABEL: roundl_call:
+; N32:          lw      $25, %call16(roundl)($gp)
+; N64:          ld      $25, %call16(roundl)($gp)
+
+; SOFT-FLOAT:   sd      $4, 8(${{[0-9]+}})
+; SOFT-FLOAT:   sd      $2, 0(${{[0-9]+}})
+
+; HARD-FLOAT:   sdc1    $f2, 8(${{[0-9]+}})
+; HARD-FLOAT:   sdc1    $f0, 0(${{[0-9]+}})
+
+  %call = call fp128 @roundl(fp128 %value)
+  store fp128 %call, fp128* @fp128
+  ret void
+}
+
+declare fp128 @roundl(fp128) nounwind readnone