[TypePromotion] Don't treat bitcast as a Source

[dmgreen]

Summary

This removes BitCasts from isSource in Type Promotion, as I don't believe they need to be treated as Sources. They will usually be from floats or hoisted constants, where constants will be handled already.

This fixes #62513, but didn't otherwise cause any differences in the tests I ran.

[samparker]

Okay!

[fhahn]

LGTM, thanks for looking into this!

[srj]

This appears to have broken arm64 codegen for macOS in Halide:

Assertion failed: (VT.getSizeInBits() == Operand.getValueSizeInBits() && "Cannot BITCAST between types of different sizes!"), function getNode, file SelectionDAG.cpp, line 5715.

I'm investigating to see if this is a latent pre-existing issue in Halide (ie, it just happened to work before) or an actual injection.

[paulkirth]

I think this is also causing failures in our CI for the Fuchsia Toolchain. I'm bisecting now to confirm, but it seems most likely given the blamelist + the earlier report. Can you revert this until a fix is ready?

See the bot here: https://ci.chromium.org/ui/p/fuchsia/builders/toolchain.ci/clang-linux-arm64/b8779118297575483793/overview

Unfortunately our CI isn't set up to capture the reproducer in the runtimes tests correctly, so we can't provide you w/ something easy to reduce.

FAILED: libcxx/src/CMakeFiles/cxx_static.dir/locale.cpp.o 
/b/s/w/ir/x/w/staging/llvm_build/./bin/clang++ --target=aarch64-unknown-linux-gnu --sysroot=/b/s/w/ir/x/w/cipd/linux -DLIBCXX_BUILDING_LIBCXXABI -D_LIBCPP_BUILDING_LIBRARY -D_LIBCPP_DISABLE_NEW_DELETE_DEFINITIONS -D_LIBCPP_DISABLE_VISIBILITY_ANNOTATIONS="" -D_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER -D_LIBCPP_LINK_PTHREAD_LIB -D_LIBCPP_LINK_RT_LIB -D_LIBCPP_REMOVE_TRANSITIVE_INCLUDES -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS -I/b/s/w/ir/x/w/llvm-llvm-project/libcxx/src -I/b/s/w/ir/x/w/staging/llvm_build/runtimes/runtimes-aarch64-unknown-linux-gnu-bins/compiler-rt/lib/msan/libcxx_msan_aarch64/include/c++/v1 -I/b/s/w/ir/x/w/llvm-llvm-project/libcxxabi/include -fsanitize=memory -fsanitize-memory-track-origins -fno-sanitize-memory-param-retval  -Xclang -fdepfile-entry=/b/s/w/ir/x/w/staging/llvm_build/./lib/../lib/clang/17/share/msan_ignorelist.txt --target=aarch64-unknown-linux-gnu -fvisibility-inlines-hidden -Werror=date-time -Werror=unguarded-availability-new -Wall -Wextra -Wno-unused-parameter -Wwrite-strings -Wcast-qual -Wmissing-field-initializers -Wimplicit-fallthrough -Wcovered-switch-default -Wno-noexcept-type -Wnon-virtual-dtor -Wdelete-non-virtual-dtor -Wsuggest-override -Wstring-conversion -Wmisleading-indentation -Wctad-maybe-unsupported -fdiagnostics-color -ffunction-sections -fdata-sections -O3 -DNDEBUG -std=c++20 -faligned-allocation -nostdinc++ -fvisibility-inlines-hidden -fvisibility=hidden -Wall -Wextra -Wnewline-eof -Wshadow -Wwrite-strings -Wno-unused-parameter -Wno-long-long -Werror=return-type -Wextra-semi -Wundef -Wunused-template -Wformat-nonliteral -Wno-user-defined-literals -Wno-covered-switch-default -Wno-suggest-override -Wno-error -MD -MT libcxx/src/CMakeFiles/cxx_static.dir/locale.cpp.o -MF libcxx/src/CMakeFiles/cxx_static.dir/locale.cpp.o.d -o libcxx/src/CMakeFiles/cxx_static.dir/locale.cpp.o -c /b/s/w/ir/x/w/llvm-llvm-project/libcxx/src/locale.cpp
Invalid bitcast
  %173 = bitcast <8 x i1> %172 to i32
Invalid bitcast
  %174 = bitcast <8 x i1> %_msprop_icmp_s66 to i32
in function _ZNKSt3__15ctypeIcE9do_narrowEPKcS3_cPc
fatal error: error in backend: Broken function found, compilation aborted!
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace, preprocessed source, and associated run script.
Stack dump:
0.	Program arguments: /b/s/w/ir/x/w/staging/llvm_build/./bin/clang++ --target=aarch64-unknown-linux-gnu --sysroot=/b/s/w/ir/x/w/cipd/linux -DLIBCXX_BUILDING_LIBCXXABI -D_LIBCPP_BUILDING_LIBRARY -D_LIBCPP_DISABLE_NEW_DELETE_DEFINITIONS -D_LIBCPP_DISABLE_VISIBILITY_ANNOTATIONS= -D_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER -D_LIBCPP_LINK_PTHREAD_LIB -D_LIBCPP_LINK_RT_LIB -D_LIBCPP_REMOVE_TRANSITIVE_INCLUDES -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS -I/b/s/w/ir/x/w/llvm-llvm-project/libcxx/src -I/b/s/w/ir/x/w/staging/llvm_build/runtimes/runtimes-aarch64-unknown-linux-gnu-bins/compiler-rt/lib/msan/libcxx_msan_aarch64/include/c++/v1 -I/b/s/w/ir/x/w/llvm-llvm-project/libcxxabi/include -fsanitize=memory -fsanitize-memory-track-origins -fno-sanitize-memory-param-retval -Xclang -fdepfile-entry=/b/s/w/ir/x/w/staging/llvm_build/./lib/../lib/clang/17/share/msan_ignorelist.txt --target=aarch64-unknown-linux-gnu -fvisibility-inlines-hidden -Werror=date-time -Werror=unguarded-availability-new -Wall -Wextra -Wno-unused-parameter -Wwrite-strings -Wcast-qual -Wmissing-field-initializers -Wimplicit-fallthrough -Wcovered-switch-default -Wno-noexcept-type -Wnon-virtual-dtor -Wdelete-non-virtual-dtor -Wsuggest-override -Wstring-conversion -Wmisleading-indentation -Wctad-maybe-unsupported -fdiagnostics-color -ffunction-sections -fdata-sections -O3 -DNDEBUG -std=c++20 -faligned-allocation -nostdinc++ -fvisibility-inlines-hidden -fvisibility=hidden -Wall -Wextra -Wnewline-eof -Wshadow -Wwrite-strings -Wno-unused-parameter -Wno-long-long -Werror=return-type -Wextra-semi -Wundef -Wunused-template -Wformat-nonliteral -Wno-user-defined-literals -Wno-covered-switch-default -Wno-suggest-override -Wno-error -MD -MT libcxx/src/CMakeFiles/cxx_static.dir/locale.cpp.o -MF libcxx/src/CMakeFiles/cxx_static.dir/locale.cpp.o.d -o libcxx/src/CMakeFiles/cxx_static.dir/locale.cpp.o -c /b/s/w/ir/x/w/llvm-llvm-project/libcxx/src/locale.cpp
1.	<eof> parser at end of file
2.	Code generation
3.	Running pass 'Function Pass Manager' on module '/b/s/w/ir/x/w/llvm-llvm-project/libcxx/src/locale.cpp'.
4.	Running pass 'Module Verifier' on function '@_ZNKSt3__15ctypeIcE9do_narrowEPKcS3_cPc'
#0 0x0000aaaac7ed7194 llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) (/b/s/w/ir/x/w/staging/llvm_build/./bin/clang+++0x7927194)
clang++: error: clang frontend command failed with exit code 70 (use -v to see invocation)
Fuchsia clang version 17.0.0 (https://llvm.googlesource.com/llvm-project c9998ec145972714d581e19fd1c9cb276b836387)
Target: aarch64-unknown-linux-gnu
Thread model: posix
InstalledDir: /b/s/w/ir/x/w/staging/llvm_build/./bin
clang++: note: diagnostic msg: 
********************

PLEASE ATTACH THE FOLLOWING FILES TO THE BUG REPORT:
Preprocessed source(s) and associated run script(s) are located at:
clang++: note: diagnostic msg: /b/s/w/ir/x/w/staging/llvm_build/clang-crashreports/locale-e001fb.cpp
clang++: note: diagnostic msg: /b/s/w/ir/x/w/staging/llvm_build/clang-crashreports/locale-e001fb.sh
clang++: note: diagnostic msg: 

********************
ninja: build stopped: subcommand failed.

[srj]

In D152112#4401561, @paulkirth wrote:

I think this is also causing failures in our CI for the Fuchsia Toolchain. I'm bisecting now to confirm, but it seems most likely given the blamelist + the earlier report. Can you revert this until a fix is ready?

Agreed, this should probably be reverted pending a fix.

[paulkirth]

I tested reverting this patch in our CI and can confirm it fixes the issue:
https://luci-milo.appspot.com/raw/build/logs.chromium.org/fuchsia/led/haowei_google.com/2d6f29abab973fbc281161503543f2e20c73276338e0cd5499a9db6d1e73f2ff/+/build.proto

[paulkirth]

I've reverted this for now.

[dmgreen]

Thanks, I was able to reproduce the problem form the error message. I'll take a look and put up a new version.

[dmgreen]

I've limited bitcasts to only when the input and output types are the same, in order to capture constants but leave other bitcasts alone. This prevents v8i1->i8 bitcasts being promoted. (Also half->i16 but I didn't see those causing a problem, and ptrs will be too large to be promoted).

[samparker]

Cheers Dave.

llvm/lib/CodeGen/TypePromotion.cpp

bool TypePromotionImpl::isSource(Value *V) {
   if (!isa<IntegerType>(V->getType()))
     return false;
 
   // TODO Allow zext to be sources.
   if (isa<Argument>(V))
     return true;
   else if (isa<LoadInst>(V))
     return true;
-  else if (isa<BitCastInst>(V))
-    return true;
   else if (auto *Call = dyn_cast<CallInst>(V))
     return Call->hasRetAttr(Attribute::AttrKind::ZExt);
   else if (auto *Trunc = dyn_cast<TruncInst>(V))
     return EqualTypeSize(Trunc);
   return false;
 }
 
 /// Return true if V will require any promoted values to be truncated for the
if (auto *I = dyn_cast<Instruction>(V)) {
     case Instruction::Br:
     case Instruction::Switch:
       return true;
     case Instruction::PHI:
     case Instruction::Select:
     case Instruction::Ret:
     case Instruction::Load:
     case Instruction::Trunc:
-    case Instruction::BitCast:
       return isSupportedType(I);
+    case Instruction::BitCast:
+      return I->getOperand(0)->getType() == I->getType();
     case Instruction::ZExt:
       return isSupportedType(I->getOperand(0));
     case Instruction::ICmp:
       // Now that we allow small types than TypeSize, only allow icmp of
       // TypeSize because they will require a trunc to be legalised.
       // TODO: Allow icmp of smaller types, and calculate at the end
       // whether the transform would be beneficial.
       if (isa<PointerType>(I->getOperand(0)->getType()))

llvm/test/Transforms/TypePromotion/AArch64/bitcast.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
+; RUN: opt -mtriple=aarch64 -passes=typepromotion,verify,dce -S %s -o - | FileCheck %s
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+
+define i1 @v8i8bitcast(<8 x i8> %a) {
+; CHECK-LABEL: define i1 @v8i8bitcast
+; CHECK-SAME: (<8 x i8> [[A:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[C:%.*]] = icmp slt <8 x i8> [[A]], zeroinitializer
+; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <8 x i1> [[C]] to i8
+; CHECK-NEXT:    [[N:%.*]] = icmp eq i8 [[TMP0]], 0
+; CHECK-NEXT:    ret i1 [[N]]
+;
+entry:
+  %c = icmp slt <8 x i8> %a, zeroinitializer
+  %0 = bitcast <8 x i1> %c to i8
+  %n = icmp eq i8 %0, 0
+  ret i1 %n
+}
+
+define i1 @halfbitcast() {
+; CHECK-LABEL: define i1 @halfbitcast() {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = bitcast half 0xH8000 to i16
+; CHECK-NEXT:    [[DOTNOT114:%.*]] = icmp eq i16 [[TMP0]], 0
+; CHECK-NEXT:    ret i1 [[DOTNOT114]]
+;
+entry:
+  %0 = bitcast half 0xH8000 to i16
+  %.not114 = icmp eq i16 %0, 0
+  ret i1 %.not114
+}
+
+define i1 @v8i8constant(<8 x i1> %a) {
+; CHECK-LABEL: define i1 @v8i8constant
+; CHECK-SAME: (<8 x i1> [[A:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <8 x i1> <i1 true, i1 true, i1 true, i1 false, i1 true, i1 true, i1 true, i1 true> to i8
+; CHECK-NEXT:    [[DOTNOT114:%.*]] = icmp eq i8 [[TMP0]], 0
+; CHECK-NEXT:    ret i1 [[DOTNOT114]]
+;
+entry:
+  %0 = bitcast <8 x i1> <i1 1, i1 1, i1 1, i1 0, i1 1, i1 1, i1 1, i1 1> to i8
+  %.not114 = icmp eq i8 %0, 0
+  ret i1 %.not114
+}

llvm/test/Transforms/TypePromotion/ARM/casts.ll

exit:
   %exitval = phi i32 [ %call, %if.then ], [ 0, %entry  ]
   ret i32 %exitval
 }
 
 define i16 @bitcast_i16(i16 zeroext %arg0, i16 zeroext %arg1) {
 ; CHECK-LABEL: @bitcast_i16(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[TMP0:%.*]] = zext i16 [[ARG0:%.*]] to i32
-; CHECK-NEXT:    [[CAST:%.*]] = bitcast i16 12345 to i16
-; CHECK-NEXT:    [[TMP1:%.*]] = zext i16 [[CAST]] to i32
+; CHECK-NEXT:    [[CAST:%.*]] = bitcast i32 12345 to i32
 ; CHECK-NEXT:    [[ADD:%.*]] = add nuw i32 [[TMP0]], 1
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ule i32 [[ADD]], [[TMP1]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ule i32 [[ADD]], [[CAST]]
 ; CHECK-NEXT:    [[RES:%.*]] = select i1 [[CMP]], i16 [[ARG1:%.*]], i16 32657
 ; CHECK-NEXT:    ret i16 [[RES]]
 ;
 entry:
   %cast = bitcast i16 12345 to i16
   %add = add nuw i16 %arg0, 1
   %cmp = icmp ule i16 %add, %cast
   %res = select i1 %cmp, i16 %arg1, i16 32657
if.end:
   %conv = sitofp i16 %arg to float
   %div = fdiv float %conv, 2.000000e+01
   br label %exit
 
 exit:
   %res = phi float [ 0.0, %entry ], [ %div, %if.end ]
   ret float %res
 }
+
+define i32 @bitcasted() {
+; CHECK-LABEL: @bitcasted(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[LOOP:%.*]]
+; CHECK:       loop:
+; CHECK-NEXT:    [[T157_PH:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[T145_OLD:%.*]], [[LATCH:%.*]] ]
+; CHECK-NEXT:    [[EXT:%.*]] = zext i32 [[T157_PH]] to i64
+; CHECK-NEXT:    br label [[LATCH]]
+; CHECK:       latch:
+; CHECK-NEXT:    [[T145_OLD]] = bitcast i32 8 to i32
+; CHECK-NEXT:    [[T146_OLD:%.*]] = bitcast i32 [[T145_OLD]] to i32
+; CHECK-NEXT:    br label [[LOOP]]
+;
+entry:
+  br label %loop
+
+loop:
+  %t157.ph = phi i32 [ 0, %entry ], [ %t145.old, %latch ]
+  %ext = zext i32 %t157.ph to i64
+  br label %latch
+
+latch:
+  %t145.old = bitcast i32 8 to i32
+  %t146.old = bitcast i32 %t145.old to i32
+  br label %loop
+}