[llvm] 110c22f - [ExpandLargeFpConvert] Support bfloat. (#87619)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Apr 8 00:07:59 PDT 2024
Author: Bevin Hansson
Date: 2024-04-08T09:07:55+02:00
New Revision: 110c22fe127f0c8c0d8acfddd123b9e9423d087a
URL: https://github.com/llvm/llvm-project/commit/110c22fe127f0c8c0d8acfddd123b9e9423d087a
DIFF: https://github.com/llvm/llvm-project/commit/110c22fe127f0c8c0d8acfddd123b9e9423d087a.diff
LOG: [ExpandLargeFpConvert] Support bfloat. (#87619)
The conversion expansions did not properly handle bfloat types.
I'm not certain that these expansions are completely correct;
I don't have any experience with AMDGPU or the ability to run
anything to test it.
Note that it doesn't seem like AMDGPU with GlobalISel can
handle fptrunc of float to bfloat, which is needed for itofp.
I've omitted the GISEL run for the bfloat case.
This fixes #85379.
Added:
llvm/test/CodeGen/AMDGPU/itofp.i128.bf.ll
Modified:
llvm/lib/CodeGen/ExpandLargeFpConvert.cpp
llvm/test/CodeGen/AMDGPU/fptoi.i128.ll
llvm/test/CodeGen/AMDGPU/itofp.i128.ll
Removed:
################################################################################
diff --git a/llvm/lib/CodeGen/ExpandLargeFpConvert.cpp b/llvm/lib/CodeGen/ExpandLargeFpConvert.cpp
index 62135304e859bd..938dda37b9f631 100644
--- a/llvm/lib/CodeGen/ExpandLargeFpConvert.cpp
+++ b/llvm/lib/CodeGen/ExpandLargeFpConvert.cpp
@@ -116,7 +116,8 @@ static void expandFPToI(Instruction *FPToI) {
// fp80 conversion is implemented by fpext to fp128 first then do the
// conversion.
FPMantissaWidth = FPMantissaWidth == 63 ? 112 : FPMantissaWidth;
- unsigned FloatWidth = PowerOf2Ceil(FPMantissaWidth);
+ unsigned FloatWidth =
+ PowerOf2Ceil(FloatVal->getType()->getScalarSizeInBits());
unsigned ExponentWidth = FloatWidth - FPMantissaWidth - 1;
unsigned ExponentBias = (1 << (ExponentWidth - 1)) - 1;
Value *ImplicitBit = Builder.CreateShl(
@@ -319,6 +320,7 @@ static void expandIToFP(Instruction *IToFP) {
// FIXME: As there is no related builtins added in compliler-rt,
// here currently utilized the fp32 <-> fp16 lib calls to implement.
FPMantissaWidth = FPMantissaWidth == 10 ? 23 : FPMantissaWidth;
+ FPMantissaWidth = FPMantissaWidth == 7 ? 23 : FPMantissaWidth;
unsigned FloatWidth = PowerOf2Ceil(FPMantissaWidth);
bool IsSigned = IToFP->getOpcode() == Instruction::SIToFP;
@@ -547,7 +549,7 @@ static void expandIToFP(Instruction *IToFP) {
Value *A40 =
Builder.CreateBitCast(Or35, Type::getFP128Ty(Builder.getContext()));
A4 = Builder.CreateFPTrunc(A40, IToFP->getType());
- } else if (IToFP->getType()->isHalfTy()) {
+ } else if (IToFP->getType()->isHalfTy() || IToFP->getType()->isBFloatTy()) {
// Deal with "half" situation. This is a workaround since we don't have
// floattihf.c currently as referring.
Value *A40 =
diff --git a/llvm/test/CodeGen/AMDGPU/fptoi.i128.ll b/llvm/test/CodeGen/AMDGPU/fptoi.i128.ll
index 66bf0d5abb732a..99818df6175bdf 100644
--- a/llvm/test/CodeGen/AMDGPU/fptoi.i128.ll
+++ b/llvm/test/CodeGen/AMDGPU/fptoi.i128.ll
@@ -1490,13 +1490,704 @@ define i128 @fptoui_f16_to_i128(half %x) {
ret i128 %cvt
}
-; FIXME: ExpandLargeFpConvert asserts on bfloat
-; define i128 @fptosi_bf16_to_i128(bfloat %x) {
-; %cvt = fptosi bfloat %x to i128
-; ret i128 %cvt
-; }
+define i128 @fptosi_bf16_to_i128(bfloat %x) {
+; SDAG-LABEL: fptosi_bf16_to_i128:
+; SDAG: ; %bb.0: ; %fp-to-i-entry
+; SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SDAG-NEXT: v_mov_b32_e32 v4, v0
+; SDAG-NEXT: v_bfe_u32 v5, v4, 7, 8
+; SDAG-NEXT: s_movk_i32 s4, 0x7e
+; SDAG-NEXT: v_mov_b32_e32 v0, 0
+; SDAG-NEXT: v_mov_b32_e32 v2, 0
+; SDAG-NEXT: v_mov_b32_e32 v6, 0
+; SDAG-NEXT: v_mov_b32_e32 v1, 0
+; SDAG-NEXT: v_mov_b32_e32 v3, 0
+; SDAG-NEXT: v_cmp_lt_u32_e32 vcc, s4, v5
+; SDAG-NEXT: s_and_saveexec_b64 s[8:9], vcc
+; SDAG-NEXT: s_cbranch_execz .LBB6_10
+; SDAG-NEXT: ; %bb.1: ; %fp-to-i-if-end
+; SDAG-NEXT: v_add_co_u32_e32 v0, vcc, 0xffffff01, v5
+; SDAG-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v6, vcc
+; SDAG-NEXT: v_addc_co_u32_e32 v2, vcc, -1, v6, vcc
+; SDAG-NEXT: s_movk_i32 s4, 0xff7f
+; SDAG-NEXT: v_addc_co_u32_e32 v3, vcc, -1, v6, vcc
+; SDAG-NEXT: s_mov_b32 s5, -1
+; SDAG-NEXT: v_cmp_lt_u64_e64 s[4:5], s[4:5], v[0:1]
+; SDAG-NEXT: v_cmp_eq_u64_e64 s[6:7], -1, v[2:3]
+; SDAG-NEXT: v_cmp_lt_i16_e32 vcc, -1, v4
+; SDAG-NEXT: s_and_b64 s[4:5], s[6:7], s[4:5]
+; SDAG-NEXT: ; implicit-def: $vgpr0_vgpr1
+; SDAG-NEXT: ; implicit-def: $vgpr2_vgpr3
+; SDAG-NEXT: s_and_saveexec_b64 s[6:7], s[4:5]
+; SDAG-NEXT: s_xor_b64 s[10:11], exec, s[6:7]
+; SDAG-NEXT: s_cbranch_execz .LBB6_7
+; SDAG-NEXT: ; %bb.2: ; %fp-to-i-if-end9
+; SDAG-NEXT: s_movk_i32 s4, 0x7f
+; SDAG-NEXT: v_and_b32_sdwa v0, v4, s4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; SDAG-NEXT: s_mov_b64 s[4:5], 0x85
+; SDAG-NEXT: v_cmp_lt_u64_e64 s[4:5], s[4:5], v[5:6]
+; SDAG-NEXT: v_mov_b32_e32 v7, 0
+; SDAG-NEXT: v_cndmask_b32_e64 v9, -1, 0, vcc
+; SDAG-NEXT: v_cndmask_b32_e64 v8, -1, 1, vcc
+; SDAG-NEXT: v_or_b32_e32 v6, 0x80, v0
+; SDAG-NEXT: ; implicit-def: $vgpr0_vgpr1
+; SDAG-NEXT: ; implicit-def: $vgpr2_vgpr3
+; SDAG-NEXT: s_and_saveexec_b64 s[6:7], s[4:5]
+; SDAG-NEXT: s_xor_b64 s[12:13], exec, s[6:7]
+; SDAG-NEXT: s_cbranch_execz .LBB6_4
+; SDAG-NEXT: ; %bb.3: ; %fp-to-i-if-else
+; SDAG-NEXT: v_sub_u32_e32 v0, 0xc6, v5
+; SDAG-NEXT: v_add_u32_e32 v2, 0xffffff3a, v5
+; SDAG-NEXT: v_add_u32_e32 v4, 0xffffff7a, v5
+; SDAG-NEXT: v_lshrrev_b64 v[0:1], v0, v[6:7]
+; SDAG-NEXT: v_lshlrev_b64 v[2:3], v2, v[6:7]
+; SDAG-NEXT: v_cmp_gt_u32_e64 s[4:5], 64, v4
+; SDAG-NEXT: v_cndmask_b32_e64 v1, v3, v1, s[4:5]
+; SDAG-NEXT: v_cmp_ne_u32_e64 s[6:7], 0, v4
+; SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v1, s[6:7]
+; SDAG-NEXT: v_cndmask_b32_e64 v2, v2, v0, s[4:5]
+; SDAG-NEXT: v_lshlrev_b64 v[0:1], v4, v[6:7]
+; SDAG-NEXT: v_cndmask_b32_e64 v2, 0, v2, s[6:7]
+; SDAG-NEXT: v_cndmask_b32_e64 v12, 0, v0, s[4:5]
+; SDAG-NEXT: v_cndmask_b32_e64 v11, 0, v1, s[4:5]
+; SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v12, v8, 0
+; SDAG-NEXT: v_cndmask_b32_e64 v10, 0, 1, vcc
+; SDAG-NEXT: v_mul_lo_u32 v13, v9, v2
+; SDAG-NEXT: v_mov_b32_e32 v6, v1
+; SDAG-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v11, v8, v[6:7]
+; SDAG-NEXT: v_mul_lo_u32 v14, v8, v3
+; SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v8, v2, 0
+; SDAG-NEXT: v_add_co_u32_e64 v6, s[4:5], -1, v10
+; SDAG-NEXT: v_mov_b32_e32 v10, v5
+; SDAG-NEXT: v_mov_b32_e32 v5, v7
+; SDAG-NEXT: v_addc_co_u32_e64 v8, s[4:5], 0, -1, s[4:5]
+; SDAG-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v12, v9, v[4:5]
+; SDAG-NEXT: v_add3_u32 v3, v3, v14, v13
+; SDAG-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v6, v12, v[2:3]
+; SDAG-NEXT: v_add_co_u32_e64 v5, s[4:5], v10, v5
+; SDAG-NEXT: v_mul_lo_u32 v3, v6, v11
+; SDAG-NEXT: v_addc_co_u32_e64 v6, s[4:5], 0, 0, s[4:5]
+; SDAG-NEXT: v_mul_lo_u32 v7, v8, v12
+; SDAG-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v11, v9, v[5:6]
+; SDAG-NEXT: ; implicit-def: $vgpr8
+; SDAG-NEXT: v_add3_u32 v3, v7, v2, v3
+; SDAG-NEXT: v_add_co_u32_e64 v2, s[4:5], v5, v1
+; SDAG-NEXT: v_addc_co_u32_e64 v3, s[4:5], v6, v3, s[4:5]
+; SDAG-NEXT: ; implicit-def: $vgpr5_vgpr6
+; SDAG-NEXT: v_mov_b32_e32 v1, v4
+; SDAG-NEXT: ; implicit-def: $vgpr6_vgpr7
+; SDAG-NEXT: .LBB6_4: ; %Flow
+; SDAG-NEXT: s_andn2_saveexec_b64 s[6:7], s[12:13]
+; SDAG-NEXT: ; %bb.5: ; %fp-to-i-if-then12
+; SDAG-NEXT: v_sub_u32_e32 v2, 0x86, v5
+; SDAG-NEXT: v_lshrrev_b64 v[0:1], v2, v[6:7]
+; SDAG-NEXT: v_cmp_gt_u32_e64 s[4:5], 64, v2
+; SDAG-NEXT: v_cndmask_b32_e64 v0, 0, v0, s[4:5]
+; SDAG-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v2
+; SDAG-NEXT: v_cndmask_b32_e64 v0, v0, v6, s[4:5]
+; SDAG-NEXT: v_mul_hi_i32_i24_e32 v1, v0, v8
+; SDAG-NEXT: v_ashrrev_i32_e32 v2, 31, v1
+; SDAG-NEXT: v_mul_i32_i24_e32 v0, v0, v8
+; SDAG-NEXT: v_mov_b32_e32 v3, v2
+; SDAG-NEXT: ; %bb.6: ; %Flow1
+; SDAG-NEXT: s_or_b64 exec, exec, s[6:7]
+; SDAG-NEXT: .LBB6_7: ; %Flow2
+; SDAG-NEXT: s_andn2_saveexec_b64 s[4:5], s[10:11]
+; SDAG-NEXT: ; %bb.8: ; %fp-to-i-if-then5
+; SDAG-NEXT: v_bfrev_b32_e32 v0, 1
+; SDAG-NEXT: v_bfrev_b32_e32 v1, -2
+; SDAG-NEXT: v_cndmask_b32_e64 v2, 0, -1, vcc
+; SDAG-NEXT: v_cndmask_b32_e32 v3, v0, v1, vcc
+; SDAG-NEXT: v_mov_b32_e32 v0, v2
+; SDAG-NEXT: v_mov_b32_e32 v1, v2
+; SDAG-NEXT: ; %bb.9: ; %Flow3
+; SDAG-NEXT: s_or_b64 exec, exec, s[4:5]
+; SDAG-NEXT: .LBB6_10: ; %fp-to-i-cleanup
+; SDAG-NEXT: s_or_b64 exec, exec, s[8:9]
+; SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; GISEL-LABEL: fptosi_bf16_to_i128:
+; GISEL: ; %bb.0: ; %fp-to-i-entry
+; GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GISEL-NEXT: v_mov_b32_e32 v4, v0
+; GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v4
+; GISEL-NEXT: v_mov_b32_e32 v6, 0
+; GISEL-NEXT: v_lshrrev_b64 v[0:1], 7, v[5:6]
+; GISEL-NEXT: v_mov_b32_e32 v1, 0x7f
+; GISEL-NEXT: s_mov_b64 s[4:5], 0
+; GISEL-NEXT: v_mov_b32_e32 v2, 0
+; GISEL-NEXT: v_bfe_u32 v5, v0, 0, 8
+; GISEL-NEXT: v_cmp_ge_u64_e32 vcc, v[5:6], v[1:2]
+; GISEL-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GISEL-NEXT: v_mov_b32_e32 v0, s4
+; GISEL-NEXT: v_mov_b32_e32 v1, s5
+; GISEL-NEXT: v_mov_b32_e32 v2, s6
+; GISEL-NEXT: v_mov_b32_e32 v3, s7
+; GISEL-NEXT: s_and_saveexec_b64 s[12:13], vcc
+; GISEL-NEXT: s_cbranch_execz .LBB6_10
+; GISEL-NEXT: ; %bb.1: ; %fp-to-i-if-end
+; GISEL-NEXT: v_add_co_u32_e32 v0, vcc, 0xffffff01, v5
+; GISEL-NEXT: v_mov_b32_e32 v2, 0xffffff80
+; GISEL-NEXT: v_addc_co_u32_e64 v1, s[6:7], 0, -1, vcc
+; GISEL-NEXT: v_mov_b32_e32 v3, -1
+; GISEL-NEXT: v_addc_co_u32_e64 v7, s[6:7], 0, -1, s[6:7]
+; GISEL-NEXT: v_cmp_ge_u64_e32 vcc, v[0:1], v[2:3]
+; GISEL-NEXT: v_addc_co_u32_e64 v8, s[6:7], 0, -1, s[6:7]
+; GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
+; GISEL-NEXT: v_cmp_le_u64_e32 vcc, -1, v[7:8]
+; GISEL-NEXT: v_cmp_lt_i16_e64 s[4:5], -1, v4
+; GISEL-NEXT: v_cndmask_b32_e64 v1, 0, 1, vcc
+; GISEL-NEXT: v_cmp_eq_u64_e32 vcc, -1, v[7:8]
+; GISEL-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc
+; GISEL-NEXT: v_and_b32_e32 v0, 1, v0
+; GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
+; GISEL-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
+; GISEL-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GISEL-NEXT: s_xor_b64 s[14:15], exec, s[6:7]
+; GISEL-NEXT: s_cbranch_execz .LBB6_7
+; GISEL-NEXT: ; %bb.2: ; %fp-to-i-if-end9
+; GISEL-NEXT: s_xor_b64 s[6:7], s[4:5], -1
+; GISEL-NEXT: v_cndmask_b32_e64 v0, 0, -1, s[6:7]
+; GISEL-NEXT: v_and_b32_e32 v0, 1, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v2, 1, v0
+; GISEL-NEXT: v_cndmask_b32_e64 v1, 0, 1, s[6:7]
+; GISEL-NEXT: v_lshlrev_b16_e32 v3, 2, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v7, 3, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v8, 4, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v9, 5, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v10, 6, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v11, 7, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v12, 8, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v13, 9, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v14, 10, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v15, 11, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v16, 12, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v17, 13, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v18, 14, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v19, 15, v0
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v2
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v2
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v3
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v3
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v7
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v7
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v8
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v8
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v9
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v9
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v10
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v10
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v11
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v11
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v12
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v12
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v13
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v13
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v14
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v14
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v15
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v15
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v16
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v16
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v17
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v17
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v18
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v18
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v19
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v19
+; GISEL-NEXT: v_and_b32_e32 v11, 0xffff, v0
+; GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v0, 16, v11
+; GISEL-NEXT: v_or3_b32 v9, v1, v0, 1
+; GISEL-NEXT: v_or3_b32 v10, v11, v0, 0
+; GISEL-NEXT: v_mov_b32_e32 v0, 0x86
+; GISEL-NEXT: v_mov_b32_e32 v1, 0
+; GISEL-NEXT: v_and_b32_e32 v2, 0x7f, v4
+; GISEL-NEXT: v_cmp_ge_u64_e32 vcc, v[5:6], v[0:1]
+; GISEL-NEXT: v_or_b32_e32 v7, 0x80, v2
+; GISEL-NEXT: v_mov_b32_e32 v8, 0
+; GISEL-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
+; GISEL-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GISEL-NEXT: s_xor_b64 s[16:17], exec, s[6:7]
+; GISEL-NEXT: s_cbranch_execz .LBB6_4
+; GISEL-NEXT: ; %bb.3: ; %fp-to-i-if-else
+; GISEL-NEXT: v_add_u32_e32 v6, 0xffffff7a, v5
+; GISEL-NEXT: v_lshlrev_b64 v[0:1], v6, v[7:8]
+; GISEL-NEXT: v_subrev_u32_e32 v4, 64, v6
+; GISEL-NEXT: v_sub_u32_e32 v2, 64, v6
+; GISEL-NEXT: v_cmp_gt_u32_e32 vcc, 64, v6
+; GISEL-NEXT: v_lshl_or_b32 v11, v11, 16, v11
+; GISEL-NEXT: v_lshrrev_b64 v[2:3], v2, v[7:8]
+; GISEL-NEXT: v_lshlrev_b64 v[4:5], v4, v[7:8]
+; GISEL-NEXT: v_cndmask_b32_e32 v8, 0, v0, vcc
+; GISEL-NEXT: v_cndmask_b32_e32 v12, 0, v1, vcc
+; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v8, v11, 0
+; GISEL-NEXT: v_cmp_eq_u32_e64 s[6:7], 0, v6
+; GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
+; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v10, v[0:1]
+; GISEL-NEXT: v_cndmask_b32_e64 v13, v2, 0, s[6:7]
+; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v8, v9, 0
+; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v13, v9, v[6:7]
+; GISEL-NEXT: v_mul_lo_u32 v4, v12, v11
+; GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
+; GISEL-NEXT: v_mov_b32_e32 v2, v6
+; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[8:9], v8, v10, v[1:2]
+; GISEL-NEXT: v_mul_lo_u32 v6, v8, v11
+; GISEL-NEXT: v_cndmask_b32_e64 v3, v3, 0, s[6:7]
+; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[10:11], v12, v9, v[1:2]
+; GISEL-NEXT: v_addc_co_u32_e64 v6, s[10:11], v7, v6, s[10:11]
+; GISEL-NEXT: v_addc_co_u32_e64 v4, s[8:9], v6, v4, s[8:9]
+; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v13, v10, v[4:5]
+; GISEL-NEXT: ; implicit-def: $vgpr5
+; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v3, v9, v[6:7]
+; GISEL-NEXT: ; implicit-def: $vgpr7_vgpr8
+; GISEL-NEXT: ; implicit-def: $vgpr9
+; GISEL-NEXT: .LBB6_4: ; %Flow
+; GISEL-NEXT: s_andn2_saveexec_b64 s[6:7], s[16:17]
+; GISEL-NEXT: s_cbranch_execz .LBB6_6
+; GISEL-NEXT: ; %bb.5: ; %fp-to-i-if-then12
+; GISEL-NEXT: v_sub_co_u32_e32 v3, vcc, 0x86, v5
+; GISEL-NEXT: v_subrev_u32_e32 v2, 64, v3
+; GISEL-NEXT: v_lshrrev_b64 v[0:1], v3, v[7:8]
+; GISEL-NEXT: v_lshrrev_b64 v[1:2], v2, 0
+; GISEL-NEXT: v_cmp_gt_u32_e32 vcc, 64, v3
+; GISEL-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc
+; GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GISEL-NEXT: v_cndmask_b32_e32 v0, v0, v7, vcc
+; GISEL-NEXT: v_mul_hi_i32_i24_e32 v1, v0, v9
+; GISEL-NEXT: v_ashrrev_i32_e32 v2, 31, v1
+; GISEL-NEXT: v_mul_i32_i24_e32 v0, v0, v9
+; GISEL-NEXT: v_mov_b32_e32 v3, v2
+; GISEL-NEXT: .LBB6_6: ; %Flow1
+; GISEL-NEXT: s_or_b64 exec, exec, s[6:7]
+; GISEL-NEXT: .LBB6_7: ; %Flow2
+; GISEL-NEXT: s_andn2_saveexec_b64 s[6:7], s[14:15]
+; GISEL-NEXT: s_cbranch_execz .LBB6_9
+; GISEL-NEXT: ; %bb.8: ; %fp-to-i-if-then5
+; GISEL-NEXT: v_cndmask_b32_e64 v1, 0, -1, s[4:5]
+; GISEL-NEXT: v_and_b32_e32 v1, 1, v1
+; GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[4:5]
+; GISEL-NEXT: v_lshlrev_b32_e32 v2, 1, v1
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v2
+; GISEL-NEXT: v_lshlrev_b32_e32 v3, 2, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v4, 3, v1
+; GISEL-NEXT: v_or_b32_e32 v2, v1, v2
+; GISEL-NEXT: v_or3_b32 v0, v0, v3, v4
+; GISEL-NEXT: v_lshlrev_b32_e32 v5, 4, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v6, 5, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v3, v4
+; GISEL-NEXT: v_or3_b32 v0, v0, v5, v6
+; GISEL-NEXT: v_lshlrev_b32_e32 v7, 6, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v8, 7, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v5, v6
+; GISEL-NEXT: v_or3_b32 v0, v0, v7, v8
+; GISEL-NEXT: v_lshlrev_b32_e32 v9, 8, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v10, 9, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v7, v8
+; GISEL-NEXT: v_or3_b32 v0, v0, v9, v10
+; GISEL-NEXT: v_lshlrev_b32_e32 v11, 10, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v12, 11, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v9, v10
+; GISEL-NEXT: v_or3_b32 v0, v0, v11, v12
+; GISEL-NEXT: v_lshlrev_b32_e32 v13, 12, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v14, 13, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v11, v12
+; GISEL-NEXT: v_or3_b32 v0, v0, v13, v14
+; GISEL-NEXT: v_lshlrev_b32_e32 v15, 14, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v16, 15, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v13, v14
+; GISEL-NEXT: v_or3_b32 v0, v0, v15, v16
+; GISEL-NEXT: v_lshlrev_b32_e32 v17, 16, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v18, 17, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v15, v16
+; GISEL-NEXT: v_or3_b32 v0, v0, v17, v18
+; GISEL-NEXT: v_lshlrev_b32_e32 v19, 18, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v20, 19, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v17, v18
+; GISEL-NEXT: v_or3_b32 v0, v0, v19, v20
+; GISEL-NEXT: v_lshlrev_b32_e32 v3, 20, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v4, 21, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v19, v20
+; GISEL-NEXT: v_or3_b32 v0, v0, v3, v4
+; GISEL-NEXT: v_lshlrev_b32_e32 v5, 22, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v6, 23, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v3, v4
+; GISEL-NEXT: v_or3_b32 v0, v0, v5, v6
+; GISEL-NEXT: v_lshlrev_b32_e32 v7, 24, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v8, 25, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v5, v6
+; GISEL-NEXT: v_or3_b32 v0, v0, v7, v8
+; GISEL-NEXT: v_lshlrev_b32_e32 v9, 26, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v10, 27, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v7, v8
+; GISEL-NEXT: v_or3_b32 v0, v0, v9, v10
+; GISEL-NEXT: v_lshlrev_b32_e32 v11, 28, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v12, 29, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v9, v10
+; GISEL-NEXT: v_or3_b32 v0, v0, v11, v12
+; GISEL-NEXT: v_lshlrev_b32_e32 v13, 30, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v1, 31, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v11, v12
+; GISEL-NEXT: v_or3_b32 v0, v0, v13, v1
+; GISEL-NEXT: v_or3_b32 v1, v2, v13, v1
+; GISEL-NEXT: v_add_u32_e32 v3, 0x80000000, v1
+; GISEL-NEXT: v_mov_b32_e32 v2, v1
+; GISEL-NEXT: .LBB6_9: ; %Flow3
+; GISEL-NEXT: s_or_b64 exec, exec, s[6:7]
+; GISEL-NEXT: .LBB6_10: ; %fp-to-i-cleanup
+; GISEL-NEXT: s_or_b64 exec, exec, s[12:13]
+; GISEL-NEXT: s_setpc_b64 s[30:31]
+ %cvt = fptosi bfloat %x to i128
+ ret i128 %cvt
+}
-; define i128 @fptoui_bf16_to_i128(bfloat %x) {
-; %cvt = fptoui bfloat %x to i128
-; ret i128 %cvt
-; }
+define i128 @fptoui_bf16_to_i128(bfloat %x) {
+; SDAG-LABEL: fptoui_bf16_to_i128:
+; SDAG: ; %bb.0: ; %fp-to-i-entry
+; SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SDAG-NEXT: v_mov_b32_e32 v4, v0
+; SDAG-NEXT: v_bfe_u32 v5, v4, 7, 8
+; SDAG-NEXT: s_movk_i32 s4, 0x7e
+; SDAG-NEXT: v_mov_b32_e32 v0, 0
+; SDAG-NEXT: v_mov_b32_e32 v2, 0
+; SDAG-NEXT: v_mov_b32_e32 v6, 0
+; SDAG-NEXT: v_mov_b32_e32 v1, 0
+; SDAG-NEXT: v_mov_b32_e32 v3, 0
+; SDAG-NEXT: v_cmp_lt_u32_e32 vcc, s4, v5
+; SDAG-NEXT: s_and_saveexec_b64 s[8:9], vcc
+; SDAG-NEXT: s_cbranch_execz .LBB7_10
+; SDAG-NEXT: ; %bb.1: ; %fp-to-i-if-end
+; SDAG-NEXT: v_add_co_u32_e32 v0, vcc, 0xffffff01, v5
+; SDAG-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v6, vcc
+; SDAG-NEXT: v_addc_co_u32_e32 v2, vcc, -1, v6, vcc
+; SDAG-NEXT: s_movk_i32 s4, 0xff7f
+; SDAG-NEXT: v_addc_co_u32_e32 v3, vcc, -1, v6, vcc
+; SDAG-NEXT: s_mov_b32 s5, -1
+; SDAG-NEXT: v_cmp_lt_u64_e64 s[4:5], s[4:5], v[0:1]
+; SDAG-NEXT: v_cmp_eq_u64_e64 s[6:7], -1, v[2:3]
+; SDAG-NEXT: v_cmp_lt_i16_e32 vcc, -1, v4
+; SDAG-NEXT: s_and_b64 s[4:5], s[6:7], s[4:5]
+; SDAG-NEXT: ; implicit-def: $vgpr0_vgpr1
+; SDAG-NEXT: ; implicit-def: $vgpr2_vgpr3
+; SDAG-NEXT: s_and_saveexec_b64 s[6:7], s[4:5]
+; SDAG-NEXT: s_xor_b64 s[10:11], exec, s[6:7]
+; SDAG-NEXT: s_cbranch_execz .LBB7_7
+; SDAG-NEXT: ; %bb.2: ; %fp-to-i-if-end9
+; SDAG-NEXT: s_movk_i32 s4, 0x7f
+; SDAG-NEXT: v_and_b32_sdwa v0, v4, s4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; SDAG-NEXT: s_mov_b64 s[4:5], 0x85
+; SDAG-NEXT: v_cmp_lt_u64_e64 s[4:5], s[4:5], v[5:6]
+; SDAG-NEXT: v_mov_b32_e32 v7, 0
+; SDAG-NEXT: v_cndmask_b32_e64 v9, -1, 0, vcc
+; SDAG-NEXT: v_cndmask_b32_e64 v8, -1, 1, vcc
+; SDAG-NEXT: v_or_b32_e32 v6, 0x80, v0
+; SDAG-NEXT: ; implicit-def: $vgpr0_vgpr1
+; SDAG-NEXT: ; implicit-def: $vgpr2_vgpr3
+; SDAG-NEXT: s_and_saveexec_b64 s[6:7], s[4:5]
+; SDAG-NEXT: s_xor_b64 s[12:13], exec, s[6:7]
+; SDAG-NEXT: s_cbranch_execz .LBB7_4
+; SDAG-NEXT: ; %bb.3: ; %fp-to-i-if-else
+; SDAG-NEXT: v_sub_u32_e32 v0, 0xc6, v5
+; SDAG-NEXT: v_add_u32_e32 v2, 0xffffff3a, v5
+; SDAG-NEXT: v_add_u32_e32 v4, 0xffffff7a, v5
+; SDAG-NEXT: v_lshrrev_b64 v[0:1], v0, v[6:7]
+; SDAG-NEXT: v_lshlrev_b64 v[2:3], v2, v[6:7]
+; SDAG-NEXT: v_cmp_gt_u32_e64 s[4:5], 64, v4
+; SDAG-NEXT: v_cndmask_b32_e64 v1, v3, v1, s[4:5]
+; SDAG-NEXT: v_cmp_ne_u32_e64 s[6:7], 0, v4
+; SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v1, s[6:7]
+; SDAG-NEXT: v_cndmask_b32_e64 v2, v2, v0, s[4:5]
+; SDAG-NEXT: v_lshlrev_b64 v[0:1], v4, v[6:7]
+; SDAG-NEXT: v_cndmask_b32_e64 v2, 0, v2, s[6:7]
+; SDAG-NEXT: v_cndmask_b32_e64 v12, 0, v0, s[4:5]
+; SDAG-NEXT: v_cndmask_b32_e64 v11, 0, v1, s[4:5]
+; SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v12, v8, 0
+; SDAG-NEXT: v_cndmask_b32_e64 v10, 0, 1, vcc
+; SDAG-NEXT: v_mul_lo_u32 v13, v9, v2
+; SDAG-NEXT: v_mov_b32_e32 v6, v1
+; SDAG-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v11, v8, v[6:7]
+; SDAG-NEXT: v_mul_lo_u32 v14, v8, v3
+; SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v8, v2, 0
+; SDAG-NEXT: v_add_co_u32_e64 v6, s[4:5], -1, v10
+; SDAG-NEXT: v_mov_b32_e32 v10, v5
+; SDAG-NEXT: v_mov_b32_e32 v5, v7
+; SDAG-NEXT: v_addc_co_u32_e64 v8, s[4:5], 0, -1, s[4:5]
+; SDAG-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v12, v9, v[4:5]
+; SDAG-NEXT: v_add3_u32 v3, v3, v14, v13
+; SDAG-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v6, v12, v[2:3]
+; SDAG-NEXT: v_add_co_u32_e64 v5, s[4:5], v10, v5
+; SDAG-NEXT: v_mul_lo_u32 v3, v6, v11
+; SDAG-NEXT: v_addc_co_u32_e64 v6, s[4:5], 0, 0, s[4:5]
+; SDAG-NEXT: v_mul_lo_u32 v7, v8, v12
+; SDAG-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v11, v9, v[5:6]
+; SDAG-NEXT: ; implicit-def: $vgpr8
+; SDAG-NEXT: v_add3_u32 v3, v7, v2, v3
+; SDAG-NEXT: v_add_co_u32_e64 v2, s[4:5], v5, v1
+; SDAG-NEXT: v_addc_co_u32_e64 v3, s[4:5], v6, v3, s[4:5]
+; SDAG-NEXT: ; implicit-def: $vgpr5_vgpr6
+; SDAG-NEXT: v_mov_b32_e32 v1, v4
+; SDAG-NEXT: ; implicit-def: $vgpr6_vgpr7
+; SDAG-NEXT: .LBB7_4: ; %Flow
+; SDAG-NEXT: s_andn2_saveexec_b64 s[6:7], s[12:13]
+; SDAG-NEXT: ; %bb.5: ; %fp-to-i-if-then12
+; SDAG-NEXT: v_sub_u32_e32 v2, 0x86, v5
+; SDAG-NEXT: v_lshrrev_b64 v[0:1], v2, v[6:7]
+; SDAG-NEXT: v_cmp_gt_u32_e64 s[4:5], 64, v2
+; SDAG-NEXT: v_cndmask_b32_e64 v0, 0, v0, s[4:5]
+; SDAG-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v2
+; SDAG-NEXT: v_cndmask_b32_e64 v0, v0, v6, s[4:5]
+; SDAG-NEXT: v_mul_hi_i32_i24_e32 v1, v0, v8
+; SDAG-NEXT: v_ashrrev_i32_e32 v2, 31, v1
+; SDAG-NEXT: v_mul_i32_i24_e32 v0, v0, v8
+; SDAG-NEXT: v_mov_b32_e32 v3, v2
+; SDAG-NEXT: ; %bb.6: ; %Flow1
+; SDAG-NEXT: s_or_b64 exec, exec, s[6:7]
+; SDAG-NEXT: .LBB7_7: ; %Flow2
+; SDAG-NEXT: s_andn2_saveexec_b64 s[4:5], s[10:11]
+; SDAG-NEXT: ; %bb.8: ; %fp-to-i-if-then5
+; SDAG-NEXT: v_bfrev_b32_e32 v0, 1
+; SDAG-NEXT: v_bfrev_b32_e32 v1, -2
+; SDAG-NEXT: v_cndmask_b32_e64 v2, 0, -1, vcc
+; SDAG-NEXT: v_cndmask_b32_e32 v3, v0, v1, vcc
+; SDAG-NEXT: v_mov_b32_e32 v0, v2
+; SDAG-NEXT: v_mov_b32_e32 v1, v2
+; SDAG-NEXT: ; %bb.9: ; %Flow3
+; SDAG-NEXT: s_or_b64 exec, exec, s[4:5]
+; SDAG-NEXT: .LBB7_10: ; %fp-to-i-cleanup
+; SDAG-NEXT: s_or_b64 exec, exec, s[8:9]
+; SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; GISEL-LABEL: fptoui_bf16_to_i128:
+; GISEL: ; %bb.0: ; %fp-to-i-entry
+; GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GISEL-NEXT: v_mov_b32_e32 v4, v0
+; GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v4
+; GISEL-NEXT: v_mov_b32_e32 v6, 0
+; GISEL-NEXT: v_lshrrev_b64 v[0:1], 7, v[5:6]
+; GISEL-NEXT: v_mov_b32_e32 v1, 0x7f
+; GISEL-NEXT: s_mov_b64 s[4:5], 0
+; GISEL-NEXT: v_mov_b32_e32 v2, 0
+; GISEL-NEXT: v_bfe_u32 v5, v0, 0, 8
+; GISEL-NEXT: v_cmp_ge_u64_e32 vcc, v[5:6], v[1:2]
+; GISEL-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GISEL-NEXT: v_mov_b32_e32 v0, s4
+; GISEL-NEXT: v_mov_b32_e32 v1, s5
+; GISEL-NEXT: v_mov_b32_e32 v2, s6
+; GISEL-NEXT: v_mov_b32_e32 v3, s7
+; GISEL-NEXT: s_and_saveexec_b64 s[12:13], vcc
+; GISEL-NEXT: s_cbranch_execz .LBB7_10
+; GISEL-NEXT: ; %bb.1: ; %fp-to-i-if-end
+; GISEL-NEXT: v_add_co_u32_e32 v0, vcc, 0xffffff01, v5
+; GISEL-NEXT: v_mov_b32_e32 v2, 0xffffff80
+; GISEL-NEXT: v_addc_co_u32_e64 v1, s[6:7], 0, -1, vcc
+; GISEL-NEXT: v_mov_b32_e32 v3, -1
+; GISEL-NEXT: v_addc_co_u32_e64 v7, s[6:7], 0, -1, s[6:7]
+; GISEL-NEXT: v_cmp_ge_u64_e32 vcc, v[0:1], v[2:3]
+; GISEL-NEXT: v_addc_co_u32_e64 v8, s[6:7], 0, -1, s[6:7]
+; GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
+; GISEL-NEXT: v_cmp_le_u64_e32 vcc, -1, v[7:8]
+; GISEL-NEXT: v_cmp_lt_i16_e64 s[4:5], -1, v4
+; GISEL-NEXT: v_cndmask_b32_e64 v1, 0, 1, vcc
+; GISEL-NEXT: v_cmp_eq_u64_e32 vcc, -1, v[7:8]
+; GISEL-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc
+; GISEL-NEXT: v_and_b32_e32 v0, 1, v0
+; GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
+; GISEL-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
+; GISEL-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GISEL-NEXT: s_xor_b64 s[14:15], exec, s[6:7]
+; GISEL-NEXT: s_cbranch_execz .LBB7_7
+; GISEL-NEXT: ; %bb.2: ; %fp-to-i-if-end9
+; GISEL-NEXT: s_xor_b64 s[6:7], s[4:5], -1
+; GISEL-NEXT: v_cndmask_b32_e64 v0, 0, -1, s[6:7]
+; GISEL-NEXT: v_and_b32_e32 v0, 1, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v2, 1, v0
+; GISEL-NEXT: v_cndmask_b32_e64 v1, 0, 1, s[6:7]
+; GISEL-NEXT: v_lshlrev_b16_e32 v3, 2, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v7, 3, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v8, 4, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v9, 5, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v10, 6, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v11, 7, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v12, 8, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v13, 9, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v14, 10, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v15, 11, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v16, 12, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v17, 13, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v18, 14, v0
+; GISEL-NEXT: v_lshlrev_b16_e32 v19, 15, v0
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v2
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v2
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v3
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v3
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v7
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v7
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v8
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v8
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v9
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v9
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v10
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v10
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v11
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v11
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v12
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v12
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v13
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v13
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v14
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v14
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v15
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v15
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v16
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v16
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v17
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v17
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v18
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v18
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v19
+; GISEL-NEXT: v_or_b32_e32 v1, v1, v19
+; GISEL-NEXT: v_and_b32_e32 v11, 0xffff, v0
+; GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v0, 16, v11
+; GISEL-NEXT: v_or3_b32 v9, v1, v0, 1
+; GISEL-NEXT: v_or3_b32 v10, v11, v0, 0
+; GISEL-NEXT: v_mov_b32_e32 v0, 0x86
+; GISEL-NEXT: v_mov_b32_e32 v1, 0
+; GISEL-NEXT: v_and_b32_e32 v2, 0x7f, v4
+; GISEL-NEXT: v_cmp_ge_u64_e32 vcc, v[5:6], v[0:1]
+; GISEL-NEXT: v_or_b32_e32 v7, 0x80, v2
+; GISEL-NEXT: v_mov_b32_e32 v8, 0
+; GISEL-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
+; GISEL-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GISEL-NEXT: s_xor_b64 s[16:17], exec, s[6:7]
+; GISEL-NEXT: s_cbranch_execz .LBB7_4
+; GISEL-NEXT: ; %bb.3: ; %fp-to-i-if-else
+; GISEL-NEXT: v_add_u32_e32 v6, 0xffffff7a, v5
+; GISEL-NEXT: v_lshlrev_b64 v[0:1], v6, v[7:8]
+; GISEL-NEXT: v_subrev_u32_e32 v4, 64, v6
+; GISEL-NEXT: v_sub_u32_e32 v2, 64, v6
+; GISEL-NEXT: v_cmp_gt_u32_e32 vcc, 64, v6
+; GISEL-NEXT: v_lshl_or_b32 v11, v11, 16, v11
+; GISEL-NEXT: v_lshrrev_b64 v[2:3], v2, v[7:8]
+; GISEL-NEXT: v_lshlrev_b64 v[4:5], v4, v[7:8]
+; GISEL-NEXT: v_cndmask_b32_e32 v8, 0, v0, vcc
+; GISEL-NEXT: v_cndmask_b32_e32 v12, 0, v1, vcc
+; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v8, v11, 0
+; GISEL-NEXT: v_cmp_eq_u32_e64 s[6:7], 0, v6
+; GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
+; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v10, v[0:1]
+; GISEL-NEXT: v_cndmask_b32_e64 v13, v2, 0, s[6:7]
+; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v8, v9, 0
+; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v13, v9, v[6:7]
+; GISEL-NEXT: v_mul_lo_u32 v4, v12, v11
+; GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
+; GISEL-NEXT: v_mov_b32_e32 v2, v6
+; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[8:9], v8, v10, v[1:2]
+; GISEL-NEXT: v_mul_lo_u32 v6, v8, v11
+; GISEL-NEXT: v_cndmask_b32_e64 v3, v3, 0, s[6:7]
+; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[10:11], v12, v9, v[1:2]
+; GISEL-NEXT: v_addc_co_u32_e64 v6, s[10:11], v7, v6, s[10:11]
+; GISEL-NEXT: v_addc_co_u32_e64 v4, s[8:9], v6, v4, s[8:9]
+; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v13, v10, v[4:5]
+; GISEL-NEXT: ; implicit-def: $vgpr5
+; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v3, v9, v[6:7]
+; GISEL-NEXT: ; implicit-def: $vgpr7_vgpr8
+; GISEL-NEXT: ; implicit-def: $vgpr9
+; GISEL-NEXT: .LBB7_4: ; %Flow
+; GISEL-NEXT: s_andn2_saveexec_b64 s[6:7], s[16:17]
+; GISEL-NEXT: s_cbranch_execz .LBB7_6
+; GISEL-NEXT: ; %bb.5: ; %fp-to-i-if-then12
+; GISEL-NEXT: v_sub_co_u32_e32 v3, vcc, 0x86, v5
+; GISEL-NEXT: v_subrev_u32_e32 v2, 64, v3
+; GISEL-NEXT: v_lshrrev_b64 v[0:1], v3, v[7:8]
+; GISEL-NEXT: v_lshrrev_b64 v[1:2], v2, 0
+; GISEL-NEXT: v_cmp_gt_u32_e32 vcc, 64, v3
+; GISEL-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc
+; GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GISEL-NEXT: v_cndmask_b32_e32 v0, v0, v7, vcc
+; GISEL-NEXT: v_mul_hi_i32_i24_e32 v1, v0, v9
+; GISEL-NEXT: v_ashrrev_i32_e32 v2, 31, v1
+; GISEL-NEXT: v_mul_i32_i24_e32 v0, v0, v9
+; GISEL-NEXT: v_mov_b32_e32 v3, v2
+; GISEL-NEXT: .LBB7_6: ; %Flow1
+; GISEL-NEXT: s_or_b64 exec, exec, s[6:7]
+; GISEL-NEXT: .LBB7_7: ; %Flow2
+; GISEL-NEXT: s_andn2_saveexec_b64 s[6:7], s[14:15]
+; GISEL-NEXT: s_cbranch_execz .LBB7_9
+; GISEL-NEXT: ; %bb.8: ; %fp-to-i-if-then5
+; GISEL-NEXT: v_cndmask_b32_e64 v1, 0, -1, s[4:5]
+; GISEL-NEXT: v_and_b32_e32 v1, 1, v1
+; GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[4:5]
+; GISEL-NEXT: v_lshlrev_b32_e32 v2, 1, v1
+; GISEL-NEXT: v_or_b32_e32 v0, v0, v2
+; GISEL-NEXT: v_lshlrev_b32_e32 v3, 2, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v4, 3, v1
+; GISEL-NEXT: v_or_b32_e32 v2, v1, v2
+; GISEL-NEXT: v_or3_b32 v0, v0, v3, v4
+; GISEL-NEXT: v_lshlrev_b32_e32 v5, 4, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v6, 5, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v3, v4
+; GISEL-NEXT: v_or3_b32 v0, v0, v5, v6
+; GISEL-NEXT: v_lshlrev_b32_e32 v7, 6, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v8, 7, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v5, v6
+; GISEL-NEXT: v_or3_b32 v0, v0, v7, v8
+; GISEL-NEXT: v_lshlrev_b32_e32 v9, 8, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v10, 9, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v7, v8
+; GISEL-NEXT: v_or3_b32 v0, v0, v9, v10
+; GISEL-NEXT: v_lshlrev_b32_e32 v11, 10, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v12, 11, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v9, v10
+; GISEL-NEXT: v_or3_b32 v0, v0, v11, v12
+; GISEL-NEXT: v_lshlrev_b32_e32 v13, 12, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v14, 13, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v11, v12
+; GISEL-NEXT: v_or3_b32 v0, v0, v13, v14
+; GISEL-NEXT: v_lshlrev_b32_e32 v15, 14, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v16, 15, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v13, v14
+; GISEL-NEXT: v_or3_b32 v0, v0, v15, v16
+; GISEL-NEXT: v_lshlrev_b32_e32 v17, 16, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v18, 17, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v15, v16
+; GISEL-NEXT: v_or3_b32 v0, v0, v17, v18
+; GISEL-NEXT: v_lshlrev_b32_e32 v19, 18, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v20, 19, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v17, v18
+; GISEL-NEXT: v_or3_b32 v0, v0, v19, v20
+; GISEL-NEXT: v_lshlrev_b32_e32 v3, 20, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v4, 21, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v19, v20
+; GISEL-NEXT: v_or3_b32 v0, v0, v3, v4
+; GISEL-NEXT: v_lshlrev_b32_e32 v5, 22, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v6, 23, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v3, v4
+; GISEL-NEXT: v_or3_b32 v0, v0, v5, v6
+; GISEL-NEXT: v_lshlrev_b32_e32 v7, 24, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v8, 25, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v5, v6
+; GISEL-NEXT: v_or3_b32 v0, v0, v7, v8
+; GISEL-NEXT: v_lshlrev_b32_e32 v9, 26, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v10, 27, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v7, v8
+; GISEL-NEXT: v_or3_b32 v0, v0, v9, v10
+; GISEL-NEXT: v_lshlrev_b32_e32 v11, 28, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v12, 29, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v9, v10
+; GISEL-NEXT: v_or3_b32 v0, v0, v11, v12
+; GISEL-NEXT: v_lshlrev_b32_e32 v13, 30, v1
+; GISEL-NEXT: v_lshlrev_b32_e32 v1, 31, v1
+; GISEL-NEXT: v_or3_b32 v2, v2, v11, v12
+; GISEL-NEXT: v_or3_b32 v0, v0, v13, v1
+; GISEL-NEXT: v_or3_b32 v1, v2, v13, v1
+; GISEL-NEXT: v_add_u32_e32 v3, 0x80000000, v1
+; GISEL-NEXT: v_mov_b32_e32 v2, v1
+; GISEL-NEXT: .LBB7_9: ; %Flow3
+; GISEL-NEXT: s_or_b64 exec, exec, s[6:7]
+; GISEL-NEXT: .LBB7_10: ; %fp-to-i-cleanup
+; GISEL-NEXT: s_or_b64 exec, exec, s[12:13]
+; GISEL-NEXT: s_setpc_b64 s[30:31]
+ %cvt = fptoui bfloat %x to i128
+ ret i128 %cvt
+}
diff --git a/llvm/test/CodeGen/AMDGPU/itofp.i128.bf.ll b/llvm/test/CodeGen/AMDGPU/itofp.i128.bf.ll
new file mode 100644
index 00000000000000..f950717c591a96
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/itofp.i128.bf.ll
@@ -0,0 +1,275 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
+; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,SDAG %s
+; RUN: not --crash llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s 2>&1 | FileCheck -check-prefix=GISEL %s
+
+; FIXME: GISEL can't handle the "fptrunc float to bfloat" that expand-large-fp-convert emits.
+
+; GISEL: unable to translate instruction: fptrunc
+
+define bfloat @sitofp_i128_to_bf16(i128 %x) {
+; GCN-LABEL: sitofp_i128_to_bf16:
+; GCN: ; %bb.0: ; %itofp-entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_or_b32_e32 v5, v1, v3
+; GCN-NEXT: v_or_b32_e32 v4, v0, v2
+; GCN-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5]
+; GCN-NEXT: v_mov_b32_e32 v4, 0
+; GCN-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GCN-NEXT: s_cbranch_execz .LBB0_14
+; GCN-NEXT: ; %bb.1: ; %itofp-if-end
+; GCN-NEXT: v_ashrrev_i32_e32 v5, 31, v3
+; GCN-NEXT: v_xor_b32_e32 v0, v5, v0
+; GCN-NEXT: v_xor_b32_e32 v1, v5, v1
+; GCN-NEXT: v_sub_co_u32_e32 v0, vcc, v0, v5
+; GCN-NEXT: v_xor_b32_e32 v2, v5, v2
+; GCN-NEXT: v_subb_co_u32_e32 v1, vcc, v1, v5, vcc
+; GCN-NEXT: v_xor_b32_e32 v6, v5, v3
+; GCN-NEXT: v_subb_co_u32_e32 v4, vcc, v2, v5, vcc
+; GCN-NEXT: v_subb_co_u32_e32 v5, vcc, v6, v5, vcc
+; GCN-NEXT: v_ffbh_u32_e32 v2, v4
+; GCN-NEXT: v_add_u32_e32 v2, 32, v2
+; GCN-NEXT: v_ffbh_u32_e32 v6, v5
+; GCN-NEXT: v_min_u32_e32 v2, v2, v6
+; GCN-NEXT: v_ffbh_u32_e32 v6, v0
+; GCN-NEXT: v_add_u32_e32 v6, 32, v6
+; GCN-NEXT: v_ffbh_u32_e32 v7, v1
+; GCN-NEXT: v_min_u32_e32 v6, v6, v7
+; GCN-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5]
+; GCN-NEXT: v_add_u32_e32 v6, 64, v6
+; GCN-NEXT: v_cndmask_b32_e32 v7, v6, v2, vcc
+; GCN-NEXT: v_sub_u32_e32 v6, 0x80, v7
+; GCN-NEXT: v_sub_u32_e32 v2, 0x7f, v7
+; GCN-NEXT: v_cmp_gt_i32_e32 vcc, 25, v6
+; GCN-NEXT: ; implicit-def: $vgpr8
+; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GCN-NEXT: ; %bb.2: ; %itofp-if-else
+; GCN-NEXT: v_add_u32_e32 v4, 0xffffff98, v7
+; GCN-NEXT: v_lshlrev_b64 v[0:1], v4, v[0:1]
+; GCN-NEXT: v_cmp_gt_u32_e32 vcc, 64, v4
+; GCN-NEXT: v_cndmask_b32_e32 v8, 0, v0, vcc
+; GCN-NEXT: ; implicit-def: $vgpr6
+; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GCN-NEXT: ; implicit-def: $vgpr7
+; GCN-NEXT: ; implicit-def: $vgpr4_vgpr5
+; GCN-NEXT: ; %bb.3: ; %Flow3
+; GCN-NEXT: s_andn2_saveexec_b64 s[8:9], s[4:5]
+; GCN-NEXT: s_cbranch_execz .LBB0_13
+; GCN-NEXT: ; %bb.4: ; %NodeBlock
+; GCN-NEXT: v_cmp_lt_i32_e32 vcc, 25, v6
+; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GCN-NEXT: s_xor_b64 s[10:11], exec, s[4:5]
+; GCN-NEXT: s_cbranch_execz .LBB0_8
+; GCN-NEXT: ; %bb.5: ; %LeafBlock
+; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 26, v6
+; GCN-NEXT: s_and_saveexec_b64 s[12:13], vcc
+; GCN-NEXT: s_cbranch_execz .LBB0_7
+; GCN-NEXT: ; %bb.6: ; %itofp-sw-default
+; GCN-NEXT: v_sub_u32_e32 v12, 0x66, v7
+; GCN-NEXT: v_sub_u32_e32 v10, 64, v12
+; GCN-NEXT: v_lshrrev_b64 v[8:9], v12, v[0:1]
+; GCN-NEXT: v_lshlrev_b64 v[10:11], v10, v[4:5]
+; GCN-NEXT: v_sub_u32_e32 v13, 38, v7
+; GCN-NEXT: v_or_b32_e32 v11, v9, v11
+; GCN-NEXT: v_or_b32_e32 v10, v8, v10
+; GCN-NEXT: v_lshrrev_b64 v[8:9], v13, v[4:5]
+; GCN-NEXT: v_cmp_gt_u32_e32 vcc, 64, v12
+; GCN-NEXT: v_add_u32_e32 v14, 26, v7
+; GCN-NEXT: v_cndmask_b32_e32 v9, v9, v11, vcc
+; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v12
+; GCN-NEXT: v_cndmask_b32_e32 v8, v8, v10, vcc
+; GCN-NEXT: v_lshrrev_b64 v[10:11], v13, v[0:1]
+; GCN-NEXT: v_lshlrev_b64 v[12:13], v14, v[4:5]
+; GCN-NEXT: v_subrev_u32_e32 v7, 38, v7
+; GCN-NEXT: v_cndmask_b32_e64 v15, v8, v0, s[4:5]
+; GCN-NEXT: v_lshlrev_b64 v[7:8], v7, v[0:1]
+; GCN-NEXT: v_cndmask_b32_e64 v9, v9, v1, s[4:5]
+; GCN-NEXT: v_or_b32_e32 v11, v13, v11
+; GCN-NEXT: v_or_b32_e32 v10, v12, v10
+; GCN-NEXT: v_cmp_gt_u32_e32 vcc, 64, v14
+; GCN-NEXT: v_lshlrev_b64 v[0:1], v14, v[0:1]
+; GCN-NEXT: v_cndmask_b32_e32 v8, v8, v11, vcc
+; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v14
+; GCN-NEXT: v_cndmask_b32_e32 v7, v7, v10, vcc
+; GCN-NEXT: v_cndmask_b32_e64 v5, v8, v5, s[4:5]
+; GCN-NEXT: v_cndmask_b32_e64 v4, v7, v4, s[4:5]
+; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
+; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
+; GCN-NEXT: v_or_b32_e32 v1, v1, v5
+; GCN-NEXT: v_or_b32_e32 v0, v0, v4
+; GCN-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
+; GCN-NEXT: v_or_b32_e32 v8, v15, v0
+; GCN-NEXT: v_mov_b32_e32 v0, v8
+; GCN-NEXT: v_mov_b32_e32 v1, v9
+; GCN-NEXT: .LBB0_7: ; %Flow1
+; GCN-NEXT: s_or_b64 exec, exec, s[12:13]
+; GCN-NEXT: .LBB0_8: ; %Flow2
+; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[10:11]
+; GCN-NEXT: ; %bb.9: ; %itofp-sw-bb
+; GCN-NEXT: v_lshlrev_b64 v[0:1], 1, v[0:1]
+; GCN-NEXT: ; %bb.10: ; %itofp-sw-epilog
+; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
+; GCN-NEXT: v_lshrrev_b32_e32 v4, 2, v0
+; GCN-NEXT: v_and_or_b32 v0, v4, 1, v0
+; GCN-NEXT: v_add_co_u32_e32 v0, vcc, 1, v0
+; GCN-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GCN-NEXT: v_and_b32_e32 v4, 0x4000000, v0
+; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
+; GCN-NEXT: v_alignbit_b32 v8, v1, v0, 2
+; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GCN-NEXT: ; %bb.11: ; %itofp-if-then20
+; GCN-NEXT: v_alignbit_b32 v8, v1, v0, 3
+; GCN-NEXT: v_mov_b32_e32 v2, v6
+; GCN-NEXT: ; %bb.12: ; %Flow
+; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
+; GCN-NEXT: .LBB0_13: ; %Flow4
+; GCN-NEXT: s_or_b64 exec, exec, s[8:9]
+; GCN-NEXT: v_and_b32_e32 v0, 0x80000000, v3
+; GCN-NEXT: v_lshl_add_u32 v1, v2, 23, 1.0
+; GCN-NEXT: v_and_b32_e32 v2, 0x7fffff, v8
+; GCN-NEXT: v_or3_b32 v0, v2, v0, v1
+; GCN-NEXT: v_bfe_u32 v1, v8, 16, 1
+; GCN-NEXT: s_movk_i32 s4, 0x7fff
+; GCN-NEXT: v_add3_u32 v1, v1, v0, s4
+; GCN-NEXT: v_or_b32_e32 v2, 0x400000, v0
+; GCN-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
+; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v0
+; GCN-NEXT: .LBB0_14: ; %Flow5
+; GCN-NEXT: s_or_b64 exec, exec, s[6:7]
+; GCN-NEXT: v_mov_b32_e32 v0, v4
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %cvt = sitofp i128 %x to bfloat
+ ret bfloat %cvt
+}
+
+define bfloat @uitofp_i128_to_bf16(i128 %x) {
+; GCN-LABEL: uitofp_i128_to_bf16:
+; GCN: ; %bb.0: ; %itofp-entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_or_b32_e32 v5, v1, v3
+; GCN-NEXT: v_or_b32_e32 v4, v0, v2
+; GCN-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5]
+; GCN-NEXT: v_mov_b32_e32 v4, 0
+; GCN-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GCN-NEXT: s_cbranch_execz .LBB1_14
+; GCN-NEXT: ; %bb.1: ; %itofp-if-end
+; GCN-NEXT: v_ffbh_u32_e32 v4, v2
+; GCN-NEXT: v_add_u32_e32 v4, 32, v4
+; GCN-NEXT: v_ffbh_u32_e32 v5, v3
+; GCN-NEXT: v_min_u32_e32 v4, v4, v5
+; GCN-NEXT: v_ffbh_u32_e32 v5, v0
+; GCN-NEXT: v_add_u32_e32 v5, 32, v5
+; GCN-NEXT: v_ffbh_u32_e32 v6, v1
+; GCN-NEXT: v_min_u32_e32 v5, v5, v6
+; GCN-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[2:3]
+; GCN-NEXT: v_add_u32_e32 v5, 64, v5
+; GCN-NEXT: v_cndmask_b32_e32 v6, v5, v4, vcc
+; GCN-NEXT: v_sub_u32_e32 v5, 0x80, v6
+; GCN-NEXT: v_sub_u32_e32 v4, 0x7f, v6
+; GCN-NEXT: v_cmp_gt_i32_e32 vcc, 25, v5
+; GCN-NEXT: ; implicit-def: $vgpr7
+; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GCN-NEXT: ; %bb.2: ; %itofp-if-else
+; GCN-NEXT: v_add_u32_e32 v2, 0xffffff98, v6
+; GCN-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
+; GCN-NEXT: v_cmp_gt_u32_e32 vcc, 64, v2
+; GCN-NEXT: v_cndmask_b32_e32 v7, 0, v0, vcc
+; GCN-NEXT: ; implicit-def: $vgpr5
+; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GCN-NEXT: ; implicit-def: $vgpr6
+; GCN-NEXT: ; implicit-def: $vgpr2_vgpr3
+; GCN-NEXT: ; %bb.3: ; %Flow3
+; GCN-NEXT: s_andn2_saveexec_b64 s[8:9], s[4:5]
+; GCN-NEXT: s_cbranch_execz .LBB1_13
+; GCN-NEXT: ; %bb.4: ; %NodeBlock
+; GCN-NEXT: v_cmp_lt_i32_e32 vcc, 25, v5
+; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GCN-NEXT: s_xor_b64 s[10:11], exec, s[4:5]
+; GCN-NEXT: s_cbranch_execz .LBB1_8
+; GCN-NEXT: ; %bb.5: ; %LeafBlock
+; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 26, v5
+; GCN-NEXT: s_and_saveexec_b64 s[12:13], vcc
+; GCN-NEXT: s_cbranch_execz .LBB1_7
+; GCN-NEXT: ; %bb.6: ; %itofp-sw-default
+; GCN-NEXT: v_sub_u32_e32 v11, 0x66, v6
+; GCN-NEXT: v_sub_u32_e32 v9, 64, v11
+; GCN-NEXT: v_lshrrev_b64 v[7:8], v11, v[0:1]
+; GCN-NEXT: v_lshlrev_b64 v[9:10], v9, v[2:3]
+; GCN-NEXT: v_sub_u32_e32 v12, 38, v6
+; GCN-NEXT: v_or_b32_e32 v10, v8, v10
+; GCN-NEXT: v_or_b32_e32 v9, v7, v9
+; GCN-NEXT: v_lshrrev_b64 v[7:8], v12, v[2:3]
+; GCN-NEXT: v_cmp_gt_u32_e32 vcc, 64, v11
+; GCN-NEXT: v_add_u32_e32 v13, 26, v6
+; GCN-NEXT: v_cndmask_b32_e32 v8, v8, v10, vcc
+; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v11
+; GCN-NEXT: v_cndmask_b32_e32 v7, v7, v9, vcc
+; GCN-NEXT: v_lshrrev_b64 v[9:10], v12, v[0:1]
+; GCN-NEXT: v_lshlrev_b64 v[11:12], v13, v[2:3]
+; GCN-NEXT: v_subrev_u32_e32 v6, 38, v6
+; GCN-NEXT: v_cndmask_b32_e64 v14, v7, v0, s[4:5]
+; GCN-NEXT: v_lshlrev_b64 v[6:7], v6, v[0:1]
+; GCN-NEXT: v_cndmask_b32_e64 v8, v8, v1, s[4:5]
+; GCN-NEXT: v_or_b32_e32 v10, v12, v10
+; GCN-NEXT: v_or_b32_e32 v9, v11, v9
+; GCN-NEXT: v_cmp_gt_u32_e32 vcc, 64, v13
+; GCN-NEXT: v_lshlrev_b64 v[0:1], v13, v[0:1]
+; GCN-NEXT: v_cndmask_b32_e32 v7, v7, v10, vcc
+; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v13
+; GCN-NEXT: v_cndmask_b32_e32 v6, v6, v9, vcc
+; GCN-NEXT: v_cndmask_b32_e64 v3, v7, v3, s[4:5]
+; GCN-NEXT: v_cndmask_b32_e64 v2, v6, v2, s[4:5]
+; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
+; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
+; GCN-NEXT: v_or_b32_e32 v1, v1, v3
+; GCN-NEXT: v_or_b32_e32 v0, v0, v2
+; GCN-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
+; GCN-NEXT: v_or_b32_e32 v7, v14, v0
+; GCN-NEXT: v_mov_b32_e32 v0, v7
+; GCN-NEXT: v_mov_b32_e32 v1, v8
+; GCN-NEXT: .LBB1_7: ; %Flow1
+; GCN-NEXT: s_or_b64 exec, exec, s[12:13]
+; GCN-NEXT: .LBB1_8: ; %Flow2
+; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[10:11]
+; GCN-NEXT: ; %bb.9: ; %itofp-sw-bb
+; GCN-NEXT: v_lshlrev_b64 v[0:1], 1, v[0:1]
+; GCN-NEXT: ; %bb.10: ; %itofp-sw-epilog
+; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
+; GCN-NEXT: v_lshrrev_b32_e32 v2, 2, v0
+; GCN-NEXT: v_and_or_b32 v0, v2, 1, v0
+; GCN-NEXT: v_add_co_u32_e32 v0, vcc, 1, v0
+; GCN-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GCN-NEXT: v_and_b32_e32 v2, 0x4000000, v0
+; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
+; GCN-NEXT: v_alignbit_b32 v7, v1, v0, 2
+; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GCN-NEXT: ; %bb.11: ; %itofp-if-then20
+; GCN-NEXT: v_alignbit_b32 v7, v1, v0, 3
+; GCN-NEXT: v_mov_b32_e32 v4, v5
+; GCN-NEXT: ; %bb.12: ; %Flow
+; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
+; GCN-NEXT: .LBB1_13: ; %Flow4
+; GCN-NEXT: s_or_b64 exec, exec, s[8:9]
+; GCN-NEXT: v_and_b32_e32 v0, 0x7fffff, v7
+; GCN-NEXT: v_lshl_or_b32 v0, v4, 23, v0
+; GCN-NEXT: v_add_u32_e32 v0, 1.0, v0
+; GCN-NEXT: v_bfe_u32 v1, v0, 16, 1
+; GCN-NEXT: s_movk_i32 s4, 0x7fff
+; GCN-NEXT: v_add3_u32 v1, v1, v0, s4
+; GCN-NEXT: v_or_b32_e32 v2, 0x400000, v0
+; GCN-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
+; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v0
+; GCN-NEXT: .LBB1_14: ; %Flow5
+; GCN-NEXT: s_or_b64 exec, exec, s[6:7]
+; GCN-NEXT: v_mov_b32_e32 v0, v4
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %cvt = uitofp i128 %x to bfloat
+ ret bfloat %cvt
+}
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; SDAG: {{.*}}
diff --git a/llvm/test/CodeGen/AMDGPU/itofp.i128.ll b/llvm/test/CodeGen/AMDGPU/itofp.i128.ll
index bfeb214c5af8f4..c6aa2182aec80c 100644
--- a/llvm/test/CodeGen/AMDGPU/itofp.i128.ll
+++ b/llvm/test/CodeGen/AMDGPU/itofp.i128.ll
@@ -1604,15 +1604,5 @@ define half @uitofp_i128_to_f16(i128 %x) {
ret half %cvt
}
-; FIXME: ExpandLargeFpConvert asserts on bfloat
-; define bfloat @sitofp_i128_to_bf16(i128 %x) {
-; %cvt = sitofp i128 %x to bfloat
-; ret bfloat %cvt
-; }
-
-; define bfloat @uitofp_i128_to_bf16(i128 %x) {
-; %cvt = uitofp i128 %x to bfloat
-; ret bfloat %cvt
-; }
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
; GCN: {{.*}}
More information about the llvm-commits
mailing list