[X86][SelectionDAG] Register type affect DAG ISel by woruyu · Pull Request #179446 · llvm/llvm-project

woruyu · 2026-02-03T12:12:07Z

Summary

This PR resolves #179100, the related PR: #164565

llvmbot · 2026-02-03T12:12:40Z

@llvm/pr-subscribers-llvm-selectiondag

Author: woruyu (woruyu)

Changes

Summary

This PR resolves #179100, the related PR: #164565

Full diff: https://github.com/llvm/llvm-project/pull/179446.diff

1 Files Affected:

(modified) llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp (+1-6)

diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index 63c9d193421ea..ea3192208a81e 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -2017,13 +2017,8 @@ SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
   if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
     Register InReg = FuncInfo.InitializeRegForValue(Inst);
 
-    std::optional<CallingConv::ID> CallConv;
-    auto *CB = dyn_cast<CallBase>(Inst);
-    if (CB && !CB->isInlineAsm())
-      CallConv = CB->getCallingConv();
-
     RegsForValue RFV(*DAG.getContext(), TLI, DAG.getDataLayout(), InReg,
-                     Inst->getType(), CallConv);
+                     Inst->getType(), std::nullopt);
     SDValue Chain = DAG.getEntryNode();
     return RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
   }

arsenm

Needs tests

phoebewang · 2026-02-03T12:15:44Z

Then we will make problems like #134222 and #151692 again. Is there a way we can fix both?

woruyu · 2026-02-03T12:26:23Z

// Old DAG
SelectionDAG has 25 nodes:
  t0: ch,glue = EntryToken
  t3: v4i1 = BUILD_VECTOR Constant:i1<0>, Constant:i1<0>, Constant:i1<0>, Constant:i1<0>
    t10: ch,glue = callseq_start t0, TargetConstant:i64<0>, TargetConstant:i64<0>
  t12: ch,glue = CopyToReg t10, Register:i64 $rdi, Constant:i64<0>
    t7: v4i32 = BUILD_VECTOR Constant:i32<0>, Constant:i32<0>, Constant:i32<0>, Constant:i32<0>
  t14: ch,glue = CopyToReg t12, Register:v4i32 $xmm0, t7, t12:1
      t5: v4i1,ch = CopyFromReg t0, Register:v4i1 %3
    t8: v4i32 = any_extend t5
  t16: ch,glue = CopyToReg t14, Register:v4i32 $xmm1, t8, t14:1
  t18: ch,glue = X86ISD::CALL t16, Constant:i64<0>, Register:i64 $rdi, Register:v4i32 $xmm0, Register:v4i32 $xmm1, RegisterMask:Untyped, t16:1
  t19: ch,glue = callseq_end t18, TargetConstant:i64<0>, TargetConstant:i64<0>, t18:1
  t21: i16,ch,glue = CopyFromReg t19, Register:i16 $ax, t19:1
    t23: ch = CopyToReg t0, Register:i16 %0, t21
  t24: ch = TokenFactor t23, t21:1

// Cur DAG
SelectionDAG has 25 nodes:
  t0: ch,glue = EntryToken
  t3: v4i1 = BUILD_VECTOR Constant:i1<0>, Constant:i1<0>, Constant:i1<0>, Constant:i1<0>
  t5: v4i32,ch = CopyFromReg t0, Register:v4i32 %3
  t6: v4i1 = truncate t5
    t10: ch,glue = callseq_start t0, TargetConstant:i64<0>, TargetConstant:i64<0>
  t12: ch,glue = CopyToReg t10, Register:i64 $rdi, Constant:i64<0>
    t8: v4i32 = BUILD_VECTOR Constant:i32<0>, Constant:i32<0>, Constant:i32<0>, Constant:i32<0>
  t14: ch,glue = CopyToReg t12, Register:v4i32 $xmm0, t8, t12:1
  t16: ch,glue = CopyToReg t14, Register:v4i32 $xmm1, t5, t14:1
  t18: ch,glue = X86ISD::CALL t16, Constant:i64<0>, Register:i64 $rdi, Register:v4i32 $xmm0, Register:v4i32 $xmm1, RegisterMask:Untyped, t16:1
  t19: ch,glue = callseq_end t18, TargetConstant:i64<0>, TargetConstant:i64<0>, t18:1
  t21: i16,ch,glue = CopyFromReg t19, Register:i16 $ax, t19:1
    t23: ch = CopyToReg t0, Register:i16 %0, t21
  t24: ch = TokenFactor t23, t21:1

The difference comes from the virtual register type. In the old DAG, the value is carried as v4i1, which matches the source-level semantics. In the new DAG it is carried as v4i32. As a result, the new lowering no longer produces the explicit step:

t8: v4i32 = any_extend t5

However, that any_extend is not “free”: it still has to be legalized/selected later, typically into something like:

t25: v4i32 = VPMOVM2DZ128rk t5

I understand the overall pipeline, but I’m not sure what the best approach is to avoid or handle this missing any_extend step. Any suggestions?

woruyu · 2026-02-03T12:30:00Z

Needs tests

understand, but I don't come up with a suitable method to fix both

woruyu · 2026-02-03T12:30:42Z

Then we will make problems like #134222 and #151692 again. Is there a way we can fix both?

maybe I need spent some time to understand fast Isel parts.

github-actions · 2026-02-03T12:50:39Z

🐧 Linux x64 Test Results

169116 tests passed
3028 tests skipped
1 test failed

Failed Tests

(click on a test name to see its output)

LLVM

LLVM.CodeGen/X86/bf16-fast-isel.ll

Exit Code: 1

Command Output (stdout):
--
# RUN: at line 2
/home/gha/actions-runner/_work/llvm-project/llvm-project/build/bin/llc --fast-isel < /home/gha/actions-runner/_work/llvm-project/llvm-project/llvm/test/CodeGen/X86/bf16-fast-isel.ll -mtriple=x86_64-unknown-unknown | /home/gha/actions-runner/_work/llvm-project/llvm-project/build/bin/FileCheck /home/gha/actions-runner/_work/llvm-project/llvm-project/llvm/test/CodeGen/X86/bf16-fast-isel.ll
# executed command: /home/gha/actions-runner/_work/llvm-project/llvm-project/build/bin/llc --fast-isel -mtriple=x86_64-unknown-unknown
# note: command had no output on stdout or stderr
# executed command: /home/gha/actions-runner/_work/llvm-project/llvm-project/build/bin/FileCheck /home/gha/actions-runner/_work/llvm-project/llvm-project/llvm/test/CodeGen/X86/bf16-fast-isel.ll
# .---command stderr------------
# | /home/gha/actions-runner/_work/llvm-project/llvm-project/llvm/test/CodeGen/X86/bf16-fast-isel.ll:9:15: error: CHECK-NEXT: is not on the line after the previous match
# | ; CHECK-NEXT: callq bar@PLT
# |               ^
# | <stdin>:12:2: note: 'next' match was here
# |  callq bar@PLT
# |  ^
# | <stdin>:10:15: note: previous match ended here
# |  callq foo@PLT
# |               ^
# | <stdin>:11:1: note: non-matching line after previous match is here
# |  callq __truncsfbf2@PLT
# | ^
# | /home/gha/actions-runner/_work/llvm-project/llvm-project/llvm/test/CodeGen/X86/bf16-fast-isel.ll:23:15: error: CHECK-NEXT: is not on the line after the previous match
# | ; CHECK-NEXT: callq bar@PLT
# |               ^
# | <stdin>:26:2: note: 'next' match was here
# |  callq bar@PLT
# |  ^
# | <stdin>:24:20: note: previous match ended here
# |  callq foo_fast@PLT
# |                    ^
# | <stdin>:25:1: note: non-matching line after previous match is here
# |  callq __truncsfbf2@PLT
# | ^
# | /home/gha/actions-runner/_work/llvm-project/llvm-project/llvm/test/CodeGen/X86/bf16-fast-isel.ll:39:15: error: CHECK-NEXT: is not on the line after the previous match
# | ; CHECK-NEXT: callq *%rbx
# |               ^
# | <stdin>:42:2: note: 'next' match was here
# |  callq *%rbx
# |  ^
# | <stdin>:40:15: note: previous match ended here
# |  callq foo@PLT
# |               ^
# | <stdin>:41:1: note: non-matching line after previous match is here
# |  callq __truncsfbf2@PLT
# | ^
# | /home/gha/actions-runner/_work/llvm-project/llvm-project/llvm/test/CodeGen/X86/bf16-fast-isel.ll:55:15: error: CHECK-NEXT: is not on the line after the previous match
# | ; CHECK-NEXT: callq *%rbx
# |               ^
# | <stdin>:58:2: note: 'next' match was here
# |  callq *%rbx
# |  ^
# | <stdin>:56:15: note: previous match ended here
# |  callq foo@PLT
# |               ^
# | <stdin>:57:1: note: non-matching line after previous match is here
# |  callq __truncsfbf2@PLT
# | ^
# | 
# | Input file: <stdin>
# | Check file: /home/gha/actions-runner/_work/llvm-project/llvm-project/llvm/test/CodeGen/X86/bf16-fast-isel.ll
# | 
# | -dump-input=help explains the following input dump.
# | 
# | Input was:
# | <<<<<<
# |          .
# |          .
# |          .
# |          7: test_direct_call: # @test_direct_call 
# |          8: # %bb.0: # %entry 
# |          9:  pushq %rax 
# |         10:  callq foo@PLT 
# |         11:  callq __truncsfbf2@PLT 
# |         12:  callq bar@PLT 
# | next:9       !~~~~~~~~~~~~  error: match on wrong line
# |         13:  popq %rcx 
# |         14:  retq 
# |         15: .Lfunc_end0: 
# |         16:  .size test_direct_call, .Lfunc_end0-test_direct_call 
# |         17:  # -- End function 
# |         18:  .globl test_fast_direct_call # -- Begin function test_fast_direct_call 
# |         19:  .p2align 4 
# |         20:  .type test_fast_direct_call,@function 
# |         21: test_fast_direct_call: # @test_fast_direct_call 
# |         22: # %bb.0: # %entry 
# |         23:  pushq %rax 
# |         24:  callq foo_fast@PLT 
# |         25:  callq __truncsfbf2@PLT 
# |         26:  callq bar@PLT 
# | next:23      !~~~~~~~~~~~~  error: match on wrong line
# |         27:  popq %rcx 
# |         28:  retq 
# |         29: .Lfunc_end1: 
# |         30:  .size test_fast_direct_call, .Lfunc_end1-test_fast_direct_call 
# |         31:  # -- End function 
# |          .
# |          .
# |          .
# |         37:  pushq %rbx 
# |         38:  movq %rdi, %rbx 
# |         39:  movq %rsi, %rdi 
# |         40:  callq foo@PLT 
# |         41:  callq __truncsfbf2@PLT 
# |         42:  callq *%rbx 
# | next:39      !~~~~~~~~~~  error: match on wrong line
# |         43:  popq %rbx 
# |         44:  retq 
# |         45: .Lfunc_end2: 
# |         46:  .size test_indirect_all, .Lfunc_end2-test_indirect_all 
# |         47:  # -- End function 
# |          .
# |          .
# |          .
# |         53:  pushq %rbx 
# |         54:  movq %rdi, %rbx 
# |         55:  movq %rsi, %rdi 
# |         56:  callq foo@PLT 
# |         57:  callq __truncsfbf2@PLT 
# |         58:  callq *%rbx 
# | next:55      !~~~~~~~~~~  error: match on wrong line
# |         59:  popq %rbx 
# |         60:  retq 
# |         61: .Lfunc_end3: 
# |         62:  .size test_fast_indirect_all, .Lfunc_end3-test_fast_indirect_all 
# |         63:  # -- End function 
# |         64:  .section ".note.GNU-stack","",@progbits 
# | >>>>>>
# `-----------------------------
# error: command failed with exit status: 1

--

If these failures are unrelated to your changes (for example tests are broken or flaky at HEAD), please open an issue at https://github.com/llvm/llvm-project/issues and add the infrastructure label.

github-actions · 2026-02-03T12:50:39Z

🪟 Windows x64 Test Results

130401 tests passed
2904 tests skipped
1 test failed

Failed Tests

(click on a test name to see its output)

LLVM

LLVM.CodeGen/X86/bf16-fast-isel.ll

Exit Code: 1

Command Output (stdout):
--
# RUN: at line 2
c:\_work\llvm-project\llvm-project\build\bin\llc.exe --fast-isel < C:\_work\llvm-project\llvm-project\llvm\test\CodeGen\X86\bf16-fast-isel.ll -mtriple=x86_64-unknown-unknown | c:\_work\llvm-project\llvm-project\build\bin\filecheck.exe C:\_work\llvm-project\llvm-project\llvm\test\CodeGen\X86\bf16-fast-isel.ll
# executed command: 'c:\_work\llvm-project\llvm-project\build\bin\llc.exe' --fast-isel -mtriple=x86_64-unknown-unknown
# note: command had no output on stdout or stderr
# executed command: 'c:\_work\llvm-project\llvm-project\build\bin\filecheck.exe' 'C:\_work\llvm-project\llvm-project\llvm\test\CodeGen\X86\bf16-fast-isel.ll'
# .---command stderr------------
# | C:\_work\llvm-project\llvm-project\llvm\test\CodeGen\X86\bf16-fast-isel.ll:9:15: error: CHECK-NEXT: is not on the line after the previous match
# | ; CHECK-NEXT: callq bar@PLT
# |               ^
# | <stdin>:12:2: note: 'next' match was here
# |  callq bar@PLT
# |  ^
# | <stdin>:10:15: note: previous match ended here
# |  callq foo@PLT
# |               ^
# | <stdin>:11:1: note: non-matching line after previous match is here
# |  callq __truncsfbf2@PLT
# | ^
# | C:\_work\llvm-project\llvm-project\llvm\test\CodeGen\X86\bf16-fast-isel.ll:23:15: error: CHECK-NEXT: is not on the line after the previous match
# | ; CHECK-NEXT: callq bar@PLT
# |               ^
# | <stdin>:26:2: note: 'next' match was here
# |  callq bar@PLT
# |  ^
# | <stdin>:24:20: note: previous match ended here
# |  callq foo_fast@PLT
# |                    ^
# | <stdin>:25:1: note: non-matching line after previous match is here
# |  callq __truncsfbf2@PLT
# | ^
# | C:\_work\llvm-project\llvm-project\llvm\test\CodeGen\X86\bf16-fast-isel.ll:39:15: error: CHECK-NEXT: is not on the line after the previous match
# | ; CHECK-NEXT: callq *%rbx
# |               ^
# | <stdin>:42:2: note: 'next' match was here
# |  callq *%rbx
# |  ^
# | <stdin>:40:15: note: previous match ended here
# |  callq foo@PLT
# |               ^
# | <stdin>:41:1: note: non-matching line after previous match is here
# |  callq __truncsfbf2@PLT
# | ^
# | C:\_work\llvm-project\llvm-project\llvm\test\CodeGen\X86\bf16-fast-isel.ll:55:15: error: CHECK-NEXT: is not on the line after the previous match
# | ; CHECK-NEXT: callq *%rbx
# |               ^
# | <stdin>:58:2: note: 'next' match was here
# |  callq *%rbx
# |  ^
# | <stdin>:56:15: note: previous match ended here
# |  callq foo@PLT
# |               ^
# | <stdin>:57:1: note: non-matching line after previous match is here
# |  callq __truncsfbf2@PLT
# | ^
# | 
# | Input file: <stdin>
# | Check file: C:\_work\llvm-project\llvm-project\llvm\test\CodeGen\X86\bf16-fast-isel.ll
# | 
# | -dump-input=help explains the following input dump.
# | 
# | Input was:
# | <<<<<<
# |          .
# |          .
# |          .
# |          7: test_direct_call: # @test_direct_call 
# |          8: # %bb.0: # %entry 
# |          9:  pushq %rax 
# |         10:  callq foo@PLT 
# |         11:  callq __truncsfbf2@PLT 
# |         12:  callq bar@PLT 
# | next:9       !~~~~~~~~~~~~  error: match on wrong line
# |         13:  popq %rcx 
# |         14:  retq 
# |         15: .Lfunc_end0: 
# |         16:  .size test_direct_call, .Lfunc_end0-test_direct_call 
# |         17:  # -- End function 
# |         18:  .globl test_fast_direct_call # -- Begin function test_fast_direct_call 
# |         19:  .p2align 4 
# |         20:  .type test_fast_direct_call,@function 
# |         21: test_fast_direct_call: # @test_fast_direct_call 
# |         22: # %bb.0: # %entry 
# |         23:  pushq %rax 
# |         24:  callq foo_fast@PLT 
# |         25:  callq __truncsfbf2@PLT 
# |         26:  callq bar@PLT 
# | next:23      !~~~~~~~~~~~~  error: match on wrong line
# |         27:  popq %rcx 
# |         28:  retq 
# |         29: .Lfunc_end1: 
# |         30:  .size test_fast_direct_call, .Lfunc_end1-test_fast_direct_call 
# |         31:  # -- End function 
# |          .
# |          .
# |          .
# |         37:  pushq %rbx 
# |         38:  movq %rdi, %rbx 
# |         39:  movq %rsi, %rdi 
# |         40:  callq foo@PLT 
# |         41:  callq __truncsfbf2@PLT 
# |         42:  callq *%rbx 
# | next:39      !~~~~~~~~~~  error: match on wrong line
# |         43:  popq %rbx 
# |         44:  retq 
# |         45: .Lfunc_end2: 
# |         46:  .size test_indirect_all, .Lfunc_end2-test_indirect_all 
# |         47:  # -- End function 
# |          .
# |          .
# |          .
# |         53:  pushq %rbx 
# |         54:  movq %rdi, %rbx 
# |         55:  movq %rsi, %rdi 
# |         56:  callq foo@PLT 
# |         57:  callq __truncsfbf2@PLT 
# |         58:  callq *%rbx 
# | next:55      !~~~~~~~~~~  error: match on wrong line
# |         59:  popq %rbx 
# |         60:  retq 
# |         61: .Lfunc_end3: 
# |         62:  .size test_fast_indirect_all, .Lfunc_end3-test_fast_indirect_all 
# |         63:  # -- End function 
# |         64:  .section ".note.GNU-stack","",@progbits 
# | >>>>>>
# `-----------------------------
# error: command failed with exit status: 1

--

If these failures are unrelated to your changes (for example tests are broken or flaky at HEAD), please open an issue at https://github.com/llvm/llvm-project/issues and add the infrastructure label.

phoebewang · 2026-02-03T12:55:28Z

Then we will make problems like #134222 and #151692 again. Is there a way we can fix both?

maybe I need spent some time to understand fast Isel parts.

That would be great. I'd say #134222 is more important, because bfloat is a legal ABI type that used by front end, while <4 x i1> is not.

alexrp · 2026-02-05T12:28:05Z

Then we will make problems like #134222 and #151692 again. Is there a way we can fix both?

maybe I need spent some time to understand fast Isel parts.

That would be great. I'd say #134222 is more important, because bfloat is a legal ABI type that used by front end, while <4 x i1> is not.

I'm not too worried about this situation -- I'm sure #179100 will be resolved one way or another in time for the 21.1.0 tag -- but just as a point of order, I think it's worth noting that per the developer policy, it is actually the responsibility of a committer to resolve any regressions caused by a commit - ideally with a follow-up fix, but failing that, a revert. Barring exceptional circumstances (probably requiring an RFC), we can't get into a situation where we justify breaking existing valid code to fix some other code; that approach will quickly break down in a project the scale of LLVM.

LuoYuanke · 2026-02-07T06:06:29Z

I got a rough patch #180322 to fix this issue. I need more time to know the fastISel behavior.

register type affect DAG ISel

1d01da7

llvmbot added the llvm:SelectionDAG SelectionDAGISel as well label Feb 3, 2026

woruyu requested review from LuoYuanke, alexrp, arsenm and phoebewang February 3, 2026 12:12

arsenm requested changes Feb 3, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[X86][SelectionDAG] Register type affect DAG ISel#179446

[X86][SelectionDAG] Register type affect DAG ISel#179446
woruyu wants to merge 1 commit intollvm:mainfrom
woruyu:fix/register-type-for-call

woruyu commented Feb 3, 2026

Uh oh!

llvmbot commented Feb 3, 2026

Summary

Uh oh!

arsenm left a comment

Uh oh!

phoebewang commented Feb 3, 2026

Uh oh!

woruyu commented Feb 3, 2026

Uh oh!

woruyu commented Feb 3, 2026

Uh oh!

woruyu commented Feb 3, 2026

Uh oh!

github-actions bot commented Feb 3, 2026

Uh oh!

github-actions bot commented Feb 3, 2026

Uh oh!

phoebewang commented Feb 3, 2026

Uh oh!

alexrp commented Feb 5, 2026

Uh oh!

LuoYuanke commented Feb 7, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

Conversation

woruyu commented Feb 3, 2026

Summary

Uh oh!

llvmbot commented Feb 3, 2026

Summary

Uh oh!

arsenm left a comment

Choose a reason for hiding this comment

Uh oh!

phoebewang commented Feb 3, 2026

Uh oh!

woruyu commented Feb 3, 2026

Uh oh!

woruyu commented Feb 3, 2026

Uh oh!

woruyu commented Feb 3, 2026

Uh oh!

github-actions bot commented Feb 3, 2026

🐧 Linux x64 Test Results

Failed Tests

LLVM

Uh oh!

github-actions bot commented Feb 3, 2026

🪟 Windows x64 Test Results

Failed Tests

LLVM

Uh oh!

phoebewang commented Feb 3, 2026

Uh oh!

alexrp commented Feb 5, 2026

Uh oh!

LuoYuanke commented Feb 7, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants