X86 backend crashed with assertion Do not know how to soft promote this operator's operand! for bf16 constants

[joker-eph]

SoftPromoteHalfOperand Op #1: t7: v32bf16 = insert_vector_elt t5, t3, Constant:i64<1>

Do not know how to soft promote this operator's operand!
UNREACHABLE executed at llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp:2940!

Repro with

bin/llc -mtriple=x86_64-unknown-linux-gnu   -mcpu="znver4"

target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"
declare <32 x bfloat> @llvm.masked.load.v32bf16.p0(ptr nocapture, i32 immarg, <32 x i1>, <32 x bfloat>)

define void @dump_vec() {
  %1 = call <32 x bfloat> @llvm.masked.load.v32bf16.p0(ptr poison, i32 2, <32 x i1> poison, <32 x bfloat> <bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80, bfloat 0xRBF80>)
  ret void
}

[llvmbot]

@llvm/issue-subscribers-backend-x86

[phoebewang]

Thanks for reporting it. I'll take a look.

[phoebewang]

Candidate patch: https://reviews.llvm.org/D151778

[joker-eph]

@phoebewang I have the same crash with a slight different case:

SoftPromoteHalfOperand Op #1: t11: v32bf16 = insert_vector_elt t9, t7, Constant:i64<1>

Do not know how to soft promote this operator's operand!
UNREACHABLE executed at llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp:3015!

Let me know if you need more info?

[RKSimon]

@phoebewang A similar crash occurs when the D152928 wip patch is applied and we run ninja check-llvm-codegen-x86

[phoebewang]

@phoebewang I have the same crash with a slight different case:

SoftPromoteHalfOperand Op #1: t11: v32bf16 = insert_vector_elt t9, t7, Constant:i64<1>

Do not know how to soft promote this operator's operand!
UNREACHABLE executed at llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp:3015!

Let me know if you need more info?

No idea if I forgot to verify it with this case or it's new regression caused by latter patches. Anyway, it can be fixed by https://reviews.llvm.org/D155952
If you still see the same crash, please provide the new reproducer. Thanks!

[phoebewang]

@phoebewang A similar crash occurs when the D152928 wip patch is applied and we run ninja check-llvm-codegen-x86

Thanks @RKSimon , but I failed to apply D152928 in my local. Would you mind to have a try with D155952 ?

[phoebewang]

@phoebewang A similar crash occurs when the D152928 wip patch is applied and we run ninja check-llvm-codegen-x86

Thanks @RKSimon , but I failed to apply D152928 in my local. Would you mind to have a try with D155952 ?

I can build it locally now, will try it myself.

[phoebewang]

@phoebewang A similar crash occurs when the D152928 wip patch is applied and we run ninja check-llvm-codegen-x86

Thanks @RKSimon , but I failed to apply D152928 in my local. Would you mind to have a try with D155952 ?

I can build it locally now, will try it myself.

D155961 will solve it ✌️