[llvm] [AMDGPU] Add V_ADD|SUB|MUL_U64 gfx1250 opcodes (PR #150291)
via llvm-commits
llvm-commits at lists.llvm.org
Wed Jul 23 12:21:43 PDT 2025
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-mc
Author: Stanislav Mekhanoshin (rampitec)
<details>
<summary>Changes</summary>
---
Patch is 207.87 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/150291.diff
23 Files Affected:
- (modified) llvm/lib/Target/AMDGPU/AMDGPU.td (+8)
- (modified) llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp (+3)
- (modified) llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp (+7-2)
- (modified) llvm/lib/Target/AMDGPU/GCNSubtarget.h (+7)
- (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+16-1)
- (modified) llvm/lib/Target/AMDGPU/SIInstrInfo.cpp (+4)
- (modified) llvm/lib/Target/AMDGPU/SIInstrInfo.td (+1)
- (modified) llvm/lib/Target/AMDGPU/VOP2Instructions.td (+17)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/mul.ll (+521)
- (added) llvm/test/CodeGen/AMDGPU/add_u64.ll (+129)
- (modified) llvm/test/CodeGen/AMDGPU/branch-relaxation-gfx1250.ll (+1-1)
- (modified) llvm/test/CodeGen/AMDGPU/carryout-selection.ll (+398)
- (modified) llvm/test/CodeGen/AMDGPU/code-size-estimate.ll (+4-6)
- (modified) llvm/test/CodeGen/AMDGPU/flat-saddr-atomics.ll (+126-142)
- (modified) llvm/test/CodeGen/AMDGPU/flat-saddr-load.ll (+5-5)
- (modified) llvm/test/CodeGen/AMDGPU/global-load-xcnt.ll (+59-58)
- (modified) llvm/test/CodeGen/AMDGPU/literal64.ll (+11-33)
- (modified) llvm/test/CodeGen/AMDGPU/mul.ll (+422)
- (modified) llvm/test/CodeGen/AMDGPU/scale-offset-flat.ll (+2-2)
- (added) llvm/test/CodeGen/AMDGPU/sub_u64.ll (+146)
- (modified) llvm/test/MC/AMDGPU/gfx1250_asm_vop2.s (+356)
- (modified) llvm/test/MC/AMDGPU/gfx1250_asm_vop2_err.s (+5)
- (modified) llvm/test/MC/Disassembler/AMDGPU/gfx1250_dasm_vop2.txt (+258)
``````````diff
diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
index 1c7ee724fef09..ff2595ef51869 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPU.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -1355,6 +1355,10 @@ def FeatureLshlAddU64Inst
: SubtargetFeature<"lshl-add-u64-inst", "HasLshlAddU64Inst", "true",
"Has v_lshl_add_u64 instruction">;
+def FeatureAddSubU64Insts
+ : SubtargetFeature<"add-sub-u64-insts", "HasAddSubU64Insts", "true",
+ "Has v_add_u64 and v_sub_u64 instructions">;
+
def FeatureMemToLDSLoad : SubtargetFeature<"vmem-to-lds-load-insts",
"HasVMemToLDSLoad",
"true",
@@ -2010,6 +2014,7 @@ def FeatureISAVersion12_50 : FeatureSet<
FeatureMemoryAtomicFAddF32DenormalSupport,
FeatureKernargPreload,
FeatureLshlAddU64Inst,
+ FeatureAddSubU64Insts,
FeatureLdsBarrierArriveAtomic,
FeatureSetPrioIncWgInst,
]>;
@@ -2787,6 +2792,9 @@ def HasAshrPkInsts : Predicate<"Subtarget->hasAshrPkInsts()">,
def HasLshlAddU64Inst : Predicate<"Subtarget->hasLshlAddU64Inst()">,
AssemblerPredicate<(all_of FeatureLshlAddU64Inst)>;
+def HasAddSubU64Insts : Predicate<"Subtarget->hasAddSubU64Insts()">,
+ AssemblerPredicate<(all_of FeatureAddSubU64Insts)>;
+
def HasLdsBarrierArriveAtomic : Predicate<"Subtarget->hasLdsBarrierArriveAtomic()">,
AssemblerPredicate<(all_of FeatureLdsBarrierArriveAtomic)>;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
index e7bf88d2ee5b6..fedfa3f9dd900 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
@@ -4208,6 +4208,9 @@ bool AMDGPULegalizerInfo::legalizeMul(LegalizerHelper &Helper,
assert(Ty.isScalar());
unsigned Size = Ty.getSizeInBits();
+ if (ST.hasVectorMulU64() && Size == 64)
+ return true;
+
unsigned NumParts = Size / 32;
assert((Size % 32) == 0);
assert(NumParts >= 2);
diff --git a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
index f1caf2478e630..9b05f7c339738 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
@@ -2528,7 +2528,8 @@ void AMDGPURegisterBankInfo::applyMappingImpl(
// Special case for s_mul_u64. There is not a vector equivalent of
// s_mul_u64. Hence, we have to break down s_mul_u64 into 32-bit vector
// multiplications.
- if (Opc == AMDGPU::G_MUL && DstTy.getSizeInBits() == 64) {
+ if (!Subtarget.hasVectorMulU64() && Opc == AMDGPU::G_MUL &&
+ DstTy.getSizeInBits() == 64) {
applyMappingSMULU64(B, OpdMapper);
return;
}
@@ -3973,7 +3974,11 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
OpdsMapping[0] = getValueMappingSGPR64Only(AMDGPU::SGPRRegBankID, Size);
OpdsMapping[1] = OpdsMapping[2] = OpdsMapping[0];
} else {
- OpdsMapping[0] = getValueMappingSGPR64Only(AMDGPU::VGPRRegBankID, Size);
+ if (MI.getOpcode() == AMDGPU::G_MUL && Subtarget.hasVectorMulU64())
+ OpdsMapping[0] = AMDGPU::getValueMapping(AMDGPU::VGPRRegBankID, Size);
+ else
+ OpdsMapping[0] =
+ getValueMappingSGPR64Only(AMDGPU::VGPRRegBankID, Size);
unsigned Bank1 = getRegBankID(MI.getOperand(1).getReg(), MRI /*, DefaultBankID*/);
OpdsMapping[1] = AMDGPU::getValueMapping(Bank1, Size);
diff --git a/llvm/lib/Target/AMDGPU/GCNSubtarget.h b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
index 8b758b011f6ad..5eddde1f72ec7 100644
--- a/llvm/lib/Target/AMDGPU/GCNSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
@@ -267,6 +267,7 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
bool HasMinimum3Maximum3F16 = false;
bool HasMinimum3Maximum3PKF16 = false;
bool HasLshlAddU64Inst = false;
+ bool HasAddSubU64Insts = false;
bool HasPointSampleAccel = false;
bool HasLdsBarrierArriveAtomic = false;
bool HasSetPrioIncWgInst = false;
@@ -1500,6 +1501,12 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
bool hasVOPD3() const { return GFX1250Insts; }
+ // \returns true if the target has V_ADD_U64/V_SUB_U64 instructions.
+ bool hasAddSubU64Insts() const { return HasAddSubU64Insts; }
+
+ // \returns true if the target has V_MUL_U64/V_MUL_I64 instructions.
+ bool hasVectorMulU64() const { return GFX1250Insts; }
+
// \returns true if the target has V_PK_ADD_{MIN|MAX}_{I|U}16 instructions.
bool hasPkAddMinMaxInsts() const { return GFX1250Insts; }
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 441034b508c10..92a56a1d5f492 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -874,7 +874,9 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
setOperationAction({ISD::SMULO, ISD::UMULO}, MVT::i64, Custom);
- if (Subtarget->hasScalarSMulU64())
+ if (Subtarget->hasVectorMulU64())
+ setOperationAction(ISD::MUL, MVT::i64, Legal);
+ else if (Subtarget->hasScalarSMulU64())
setOperationAction(ISD::MUL, MVT::i64, Custom);
if (Subtarget->hasMad64_32())
@@ -5421,6 +5423,19 @@ SITargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
MachineOperand &Src0 = MI.getOperand(1);
MachineOperand &Src1 = MI.getOperand(2);
+ if (ST.hasAddSubU64Insts()) {
+ auto I = BuildMI(*BB, MI, DL,
+ TII->get(IsAdd ? AMDGPU::V_ADD_U64_e64
+ : AMDGPU::V_SUB_U64_e64),
+ Dest.getReg())
+ .add(Src0)
+ .add(Src1)
+ .addImm(0); // clamp
+ TII->legalizeOperands(*I);
+ MI.eraseFromParent();
+ return BB;
+ }
+
if (IsAdd && ST.hasLshlAddU64Inst()) {
auto Add = BuildMI(*BB, MI, DL, TII->get(AMDGPU::V_LSHL_ADD_U64_e64),
Dest.getReg())
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 571f3efd68260..40e687178fb01 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -7361,6 +7361,10 @@ void SIInstrInfo::moveToVALUImpl(SIInstrWorklist &Worklist,
}
case AMDGPU::S_MUL_U64:
+ if (ST.hasVectorMulU64()) {
+ NewOpcode = AMDGPU::V_MUL_U64_e64;
+ break;
+ }
// Split s_mul_u64 in 32-bit vector multiplications.
splitScalarSMulU64(Worklist, Inst, MDT);
Inst.eraseFromParent();
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
index b8537513ce986..485ca78db93a7 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -2914,6 +2914,7 @@ def VOP_I32_F32_I32 : VOPProfile <[i32, f32, i32, untyped]>;
def VOP_I32_I32_I32 : VOPProfile <[i32, i32, i32, untyped]>;
def VOP_I16_F32_F32 : VOPProfile <[i16, f32, f32, untyped]>;
def VOP_I32_I32_I32_ARITH : VOPProfile <[i32, i32, i32, untyped], /*EnableClamp=*/1>;
+def VOP_I64_I64_I64_ARITH : VOPProfile <[i64, i64, i64, untyped], /*EnableClamp=*/1>;
def VOP_V2F16_F32_F32 : VOPProfile <[v2f16, f32, f32, untyped]>;
def VOP_F32_F16_F16_F16 : VOPProfile <[f32, f16, f16, f16]>;
def VOP_V2BF16_F32_F32 : VOPProfile <[v2bf16, f32, f32, untyped]>;
diff --git a/llvm/lib/Target/AMDGPU/VOP2Instructions.td b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
index 030a6e1e978c1..550ec9d3f55ab 100644
--- a/llvm/lib/Target/AMDGPU/VOP2Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
@@ -925,6 +925,17 @@ let isAdd = 1 in {
defm V_ADDC_U32 : VOP2bInst <"v_addc_u32", VOP2b_I32_I1_I32_I32_I1, null_frag, "v_addc_u32">;
}
+let isReMaterializable = 1 in {
+let SubtargetPredicate = HasAddSubU64Insts, SchedRW = [Write64Bit] in {
+defm V_ADD_U64 : VOP2Inst <"v_add_nc_u64", VOP_I64_I64_I64_ARITH>;
+// We don't actually have something like V_SUBREV_U64 so V_SUB_U64 can't be treated as commutable.
+let isCommutable = 0 in
+defm V_SUB_U64 : VOP2Inst <"v_sub_nc_u64", VOP_I64_I64_I64_ARITH>;
+} // End SubtargetPredicate = HasAddSubU64Insts, SchedRW = [Write64Bit]
+let SubtargetPredicate = isGFX1250Plus, SchedRW = [WriteDouble] in
+defm V_MUL_U64 : VOP2Inst <"v_mul_u64", VOP_I64_I64_I64, DivergentBinFrag<mul>>;
+} // End isReMaterializable = 1
+
} // End isCommutable = 1
// These are special and do not read the exec mask.
@@ -1754,6 +1765,9 @@ multiclass VOP2_Real_FULL_with_name<GFXGen Gen, bits<6> op, string opName,
VOP2_Realtriple_e64_with_name<Gen, op, opName, asmName>,
VOP2_Real_NO_VOP3_with_name<Gen, op, opName, asmName>;
+multiclass VOP2_Real_NO_DPP<GFXGen Gen, bits<6> op> :
+ VOP2_Real_e32<Gen, op>, VOP2_Real_e64<Gen, op>;
+
multiclass VOP2_Real_NO_DPP_with_name<GFXGen Gen, bits<6> op, string opName,
string asmName> {
defm NAME : VOP2_Real_e32_with_name<Gen, op, opName, asmName>,
@@ -1843,6 +1857,9 @@ defm V_FMAC_F64 : VOP2_Real_FULL<GFX12Gen, 0x17>;
defm V_FMAMK_F64 : VOP2Only_Real_MADK64<GFX1250Gen, 0x23>;
defm V_FMAAK_F64 : VOP2Only_Real_MADK64<GFX1250Gen, 0x24>;
+defm V_ADD_U64 : VOP2_Real_FULL<GFX1250Gen, 0x28>;
+defm V_SUB_U64 : VOP2_Real_FULL<GFX1250Gen, 0x29>;
+defm V_MUL_U64 : VOP2_Real_NO_DPP<GFX1250Gen, 0x2a>;
//===----------------------------------------------------------------------===//
// GFX11.
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/mul.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/mul.ll
index f7f7e9645fa62..0d571d0e563b5 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/mul.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/mul.ll
@@ -6,6 +6,7 @@
; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16, -amdgpu-enable-delay-alu=0 < %s | FileCheck -check-prefixes=GFX10PLUS,GFX11,GFX11-TRUE16 %s
; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16, -amdgpu-enable-delay-alu=0 < %s | FileCheck -check-prefixes=GFX10PLUS,GFX11,GFX11-FAKE16 %s
; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX12 %s
+; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1250 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX1250 %s
define amdgpu_ps i16 @s_mul_i16(i16 inreg %num, i16 inreg %den) {
; GCN-LABEL: s_mul_i16:
@@ -22,6 +23,11 @@ define amdgpu_ps i16 @s_mul_i16(i16 inreg %num, i16 inreg %den) {
; GFX12: ; %bb.0:
; GFX12-NEXT: s_mul_i32 s0, s0, s1
; GFX12-NEXT: ; return to shader part epilog
+;
+; GFX1250-LABEL: s_mul_i16:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_mul_i32 s0, s0, s1
+; GFX1250-NEXT: ; return to shader part epilog
%result = mul i16 %num, %den
ret i16 %result
}
@@ -74,6 +80,13 @@ define i16 @v_mul_i16(i16 %num, i16 %den) {
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: v_mul_lo_u16 v0, v0, v1
; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1250-LABEL: v_mul_i16:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: v_mul_lo_u16 v0, v0, v1
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
%result = mul i16 %num, %den
ret i16 %result
}
@@ -109,6 +122,13 @@ define amdgpu_ps zeroext i16 @s_mul_i16_zeroext(i16 inreg zeroext %num, i16 inre
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: s_and_b32 s0, 0xffff, s0
; GFX12-NEXT: ; return to shader part epilog
+;
+; GFX1250-LABEL: s_mul_i16_zeroext:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_mul_i32 s0, s0, s1
+; GFX1250-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1250-NEXT: s_and_b32 s0, 0xffff, s0
+; GFX1250-NEXT: ; return to shader part epilog
%result = mul i16 %num, %den
ret i16 %result
}
@@ -165,6 +185,15 @@ define zeroext i16 @v_mul_i16_zeroext(i16 zeroext %num, i16 zeroext %den) {
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1250-LABEL: v_mul_i16_zeroext:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: v_mul_lo_u16 v0, v0, v1
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1250-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
%result = mul i16 %num, %den
ret i16 %result
}
@@ -188,6 +217,13 @@ define amdgpu_ps signext i16 @s_mul_i16_signext(i16 inreg signext %num, i16 inre
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: s_sext_i32_i16 s0, s0
; GFX12-NEXT: ; return to shader part epilog
+;
+; GFX1250-LABEL: s_mul_i16_signext:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_mul_i32 s0, s0, s1
+; GFX1250-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1250-NEXT: s_sext_i32_i16 s0, s0
+; GFX1250-NEXT: ; return to shader part epilog
%result = mul i16 %num, %den
ret i16 %result
}
@@ -248,6 +284,15 @@ define signext i16 @v_mul_i16_signext(i16 signext %num, i16 signext %den) {
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT: v_bfe_i32 v0, v0, 0, 16
; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1250-LABEL: v_mul_i16_signext:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: v_mul_lo_u16 v0, v0, v1
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1250-NEXT: v_bfe_i32 v0, v0, 0, 16
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
%result = mul i16 %num, %den
ret i16 %result
}
@@ -267,6 +312,11 @@ define amdgpu_ps i32 @s_mul_i32(i32 inreg %num, i32 inreg %den) {
; GFX12: ; %bb.0:
; GFX12-NEXT: s_mul_i32 s0, s0, s1
; GFX12-NEXT: ; return to shader part epilog
+;
+; GFX1250-LABEL: s_mul_i32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_mul_i32 s0, s0, s1
+; GFX1250-NEXT: ; return to shader part epilog
%result = mul i32 %num, %den
ret i32 %result
}
@@ -293,6 +343,13 @@ define i32 @v_mul_i32(i32 %num, i32 %den) {
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: v_mul_lo_u32 v0, v0, v1
; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1250-LABEL: v_mul_i32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: v_mul_lo_u32 v0, v0, v1
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
%result = mul i32 %num, %den
ret i32 %result
}
@@ -315,6 +372,12 @@ define amdgpu_ps <2 x i32> @s_mul_v2i32(<2 x i32> inreg %num, <2 x i32> inreg %d
; GFX12-NEXT: s_mul_i32 s0, s0, s2
; GFX12-NEXT: s_mul_i32 s1, s1, s3
; GFX12-NEXT: ; return to shader part epilog
+;
+; GFX1250-LABEL: s_mul_v2i32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_mul_i32 s0, s0, s2
+; GFX1250-NEXT: s_mul_i32 s1, s1, s3
+; GFX1250-NEXT: ; return to shader part epilog
%result = mul <2 x i32> %num, %den
ret <2 x i32> %result
}
@@ -344,6 +407,14 @@ define <2 x i32> @v_mul_v2i32(<2 x i32> %num, <2 x i32> %den) {
; GFX12-NEXT: v_mul_lo_u32 v0, v0, v2
; GFX12-NEXT: v_mul_lo_u32 v1, v1, v3
; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1250-LABEL: v_mul_v2i32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: v_mul_lo_u32 v0, v0, v2
+; GFX1250-NEXT: v_mul_lo_u32 v1, v1, v3
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
%result = mul <2 x i32> %num, %den
ret <2 x i32> %result
}
@@ -400,6 +471,11 @@ define amdgpu_cs i33 @s_mul_i33(i33 inreg %num, i33 inreg %den) {
; GFX12: ; %bb.0:
; GFX12-NEXT: s_mul_u64 s[0:1], s[0:1], s[2:3]
; GFX12-NEXT: ; return to shader part epilog
+;
+; GFX1250-LABEL: s_mul_i33:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_mul_u64 s[0:1], s[0:1], s[2:3]
+; GFX1250-NEXT: ; return to shader part epilog
%result = mul i33 %num, %den
ret i33 %result
}
@@ -456,6 +532,11 @@ define amdgpu_ps i64 @s_mul_i64(i64 inreg %num, i64 inreg %den) {
; GFX12: ; %bb.0:
; GFX12-NEXT: s_mul_u64 s[0:1], s[0:1], s[2:3]
; GFX12-NEXT: ; return to shader part epilog
+;
+; GFX1250-LABEL: s_mul_i64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_mul_u64 s[0:1], s[0:1], s[2:3]
+; GFX1250-NEXT: ; return to shader part epilog
%result = mul i64 %num, %den
ret i64 %result
}
@@ -504,6 +585,13 @@ define i64 @v_mul_i64(i64 %num, i64 %den) {
; GFX12-NEXT: v_mul_lo_u32 v0, v0, v2
; GFX12-NEXT: v_mad_co_u64_u32 v[1:2], null, v1, v2, v[3:4]
; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1250-LABEL: v_mul_i64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: v_mul_u64_e32 v[0:1], v[0:1], v[2:3]
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
%result = mul i64 %num, %den
ret i64 %result
}
@@ -620,6 +708,26 @@ define amdgpu_ps <3 x i32> @s_mul_i96(i96 inreg %num, i96 inreg %den) {
; GFX12-NEXT: s_add_co_ci_u32 s2, s3, s0
; GFX12-NEXT: s_mov_b32 s0, s5
; GFX12-NEXT: ; return to shader part epilog
+;
+; GFX1250-LABEL: s_mul_i96:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_mul_i32 s6, s0, s5
+; GFX1250-NEXT: s_mul_i32 s7, s1, s4
+; GFX1250-NEXT: s_mul_i32 s2, s2, s3
+; GFX1250-NEXT: s_add_co_i32 s6, s6, s7
+; GFX1250-NEXT: s_mul_hi_u32 s7, s0, s3
+; GFX1250-NEXT: s_add_co_i32 s6, s6, s2
+; GFX1250-NEXT: s_mul_i32 s2, s0, s4
+; GFX1250-NEXT: s_mul_i32 s5, s0, s3
+; GFX1250-NEXT: s_mul_hi_u32 s0, s0, s4
+; GFX1250-NEXT: s_add_co_u32 s2, s2, s7
+; GFX1250-NEXT: s_mul_i32 s4, s1, s3
+; GFX1250-NEXT: s_add_co_ci_u32 s0, s0, s6
+; GFX1250-NEXT: s_mul_hi_u32 s3, s1, s3
+; GFX1250-NEXT: s_add_co_u32 s1, s4, s2
+; GFX1250-NEXT: s_add_co_ci_u32 s2, s3, s0
+; GFX1250-NEXT: s_mov_b32 s0, s5
+; GFX1250-NEXT: ; return to shader part epilog
%result = mul i96 %num, %den
%cast = bitcast i96 %result to <3 x i32>
ret <3 x i32> %cast
@@ -686,6 +794,25 @@ define i96 @v_mul_i96(i96 %num, i96 %den) {
; GFX12-NEXT: v_mad_co_u64_u32 v[1:2], null, v6, v4, v[1:2]
; GFX12-NEXT: v_mad_co_u64_u32 v[1:2], null, v7, v3, v[1:2]
; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1250-LABEL: v_mul_i96:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: v_dual_mov_b32 v6, v0 :: v_dual_mov_b32 v7, v1
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1250-NEXT: v_mul_lo_u32 v0, v6, v5
+; GFX1250-NEXT: v_mad_co_u64_u32 v[8:9], null, v7, v4, v[0:1]
+; GFX1250-NEXT: v_mad_co_u64_u32 v[0:1], null, v6, v3, 0
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1250-NEXT: v_mad_co_u64_u32 v[8:9], null, v2, v3, v[8:9]
+; GFX1250-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v11, v8
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1250-NEXT: v_mad_co_u64_u32 v[4:5], null, v6, v4, v[10:11]
+; GFX1250-NEXT: v_mad_co_u64_u32 v[2:3], null, v7, v3, v[4:5]
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1250-NEXT: v_dual_mov_b32 v1, v2 :: v_dual_mov_b32 v2, v3
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
%result = mul i96 %num, %den
ret i96 %result
}
@@ -895,6 +1022,42 @@ define amdgpu_ps <4 x i32> @s_mul_i128(i128 inreg %num, i128 inreg %den) {
; GFX12-NEXT: s_mov_b32 s1, s8
; GFX12-NEXT: s_mov_b32 s2, s7
; GFX12-NEXT: ; return to shader part epilog
+;
+; GFX1250-LABEL: s_mul_i128:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_mul_i32 s9, s0, s6
+; GFX1250-NEXT: s_mul_i32 s11, s1, s5
+; GFX1250-NEXT: s_mul_hi_u32 s10, s0, s6
+; GFX1250-NEXT: s_mul_hi_u32 s12, s1, s5
+; GFX1250-NEXT: s_add_co_u32 s9, s11, s9
+; GFX1250-NEXT: s_mul_i32 s11, s2, s4
+; GFX1250-NEXT: s_add_co_ci_u32 s10, s12, s10
+; GFX1250-NEXT: s_mul_hi_u32 s12, s2, s4
+; GFX1250-NEXT: s_mul_hi_u32 s8, s0, s4
+; GFX1250-NEXT: s_add_co_u32 s9, s11, s9
+; GFX1250-NEXT: s_mul_i32 s11, s0, s5
+; GFX1250-NEXT: s_add_co_ci_u32 s10, s12, s10
+; GFX1250-NEXT: s_mul_hi_u32 s12, s0, s5
+; GFX1250-NEXT: s_add_co_u32 s8, s11, s8
+; GFX1250-NEXT: s_add_co_ci_u32 s9, s12, s9
+; GFX1250-NEXT: s_mul_i32 s12, s1, s4
+; GFX1250-NEXT: s_mul_hi_u32 s13, s1, s4
+; GFX1250-NEXT: s_cselect_b32 s11, 1, 0
+; GFX1250-NEXT: s_add...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/150291
More information about the llvm-commits
mailing list