[llvm] [AMDGPU] Add wave reduce intrinsics for double types - 1 (PR #170811)
via llvm-commits
llvm-commits at lists.llvm.org
Tue Dec 9 10:36:36 PST 2025
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: Aaditya (easyonaadit)
<details>
<summary>Changes</summary>
Supported Ops: `min`, `max`
---
Patch is 133.26 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/170811.diff
4 Files Affected:
- (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+65-1)
- (modified) llvm/lib/Target/AMDGPU/SIInstructions.td (+2)
- (modified) llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.fmax.ll (+1280)
- (modified) llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.fmin.ll (+1280)
``````````diff
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 301f2fc8dab45..80978c6a00a9c 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -5522,6 +5522,9 @@ static uint64_t getIdentityValueFor64BitWaveReduction(unsigned Opc) {
return std::numeric_limits<uint64_t>::min();
case AMDGPU::V_CMP_GT_I64_e64: // max.i64
return std::numeric_limits<int64_t>::min();
+ case AMDGPU::V_MIN_F64_e64:
+ case AMDGPU::V_MAX_F64_e64:
+ return 0x7FF8000000000000; // qNAN
case AMDGPU::S_ADD_U64_PSEUDO:
case AMDGPU::S_SUB_U64_PSEUDO:
case AMDGPU::S_OR_B64:
@@ -5547,7 +5550,8 @@ static bool is32bitWaveReduceOperation(unsigned Opc) {
static bool isFloatingPointWaveReduceOperation(unsigned Opc) {
return Opc == AMDGPU::V_MIN_F32_e64 || Opc == AMDGPU::V_MAX_F32_e64 ||
- Opc == AMDGPU::V_ADD_F32_e64 || Opc == AMDGPU::V_SUB_F32_e64;
+ Opc == AMDGPU::V_ADD_F32_e64 || Opc == AMDGPU::V_SUB_F32_e64 ||
+ Opc == AMDGPU::V_MIN_F64_e64 || Opc == AMDGPU::V_MAX_F64_e64;
}
static MachineBasicBlock *lowerWaveReduce(MachineInstr &MI,
@@ -5583,6 +5587,8 @@ static MachineBasicBlock *lowerWaveReduce(MachineInstr &MI,
case AMDGPU::V_CMP_LT_I64_e64: // min
case AMDGPU::V_CMP_GT_U64_e64: // umax
case AMDGPU::V_CMP_GT_I64_e64: // max
+ case AMDGPU::V_MIN_F64_e64:
+ case AMDGPU::V_MAX_F64_e64:
case AMDGPU::S_AND_B64:
case AMDGPU::S_OR_B64: {
// Idempotent operations.
@@ -5952,6 +5958,60 @@ static MachineBasicBlock *lowerWaveReduce(MachineInstr &MI,
.addReg(Accumulator->getOperand(0).getReg());
break;
}
+ case AMDGPU::V_MIN_F64_e64:
+ case AMDGPU::V_MAX_F64_e64: {
+ const TargetRegisterClass *VregRC = TRI->getVGPR64Class();
+ const TargetRegisterClass *VregSubRC =
+ TRI->getSubRegisterClass(VregRC, AMDGPU::sub0);
+ Register AccumulatorVReg = MRI.createVirtualRegister(VregRC);
+ Register DstVreg = MRI.createVirtualRegister(VregRC);
+ Register LaneValLo =
+ MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
+ Register LaneValHi =
+ MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
+ BuildMI(*ComputeLoop, I, DL, TII->get(AMDGPU::V_MOV_B64_PSEUDO),
+ AccumulatorVReg)
+ .addReg(Accumulator->getOperand(0).getReg());
+ if (ST.getGeneration() == AMDGPUSubtarget::Generation::GFX12) {
+ switch (Opc) {
+ case AMDGPU::V_MIN_F64_e64:
+ Opc = AMDGPU::V_MIN_NUM_F64_e64;
+ break;
+ case AMDGPU::V_MAX_F64_e64:
+ Opc = AMDGPU::V_MAX_NUM_F64_e64;
+ break;
+ }
+ }
+ auto DstVregInst = BuildMI(*ComputeLoop, I, DL, TII->get(Opc), DstVreg)
+ .addImm(0) // src0 modifiers
+ .addReg(LaneValue->getOperand(0).getReg())
+ .addImm(0) // src1 modifiers
+ .addReg(AccumulatorVReg)
+ .addImm(0) // clamp
+ .addImm(0); // omod
+ auto ReadLaneLo =
+ BuildMI(*ComputeLoop, I, DL, TII->get(AMDGPU::V_READFIRSTLANE_B32),
+ LaneValLo);
+ auto ReadLaneHi =
+ BuildMI(*ComputeLoop, I, DL, TII->get(AMDGPU::V_READFIRSTLANE_B32),
+ LaneValHi);
+ MachineBasicBlock::iterator Iters = *ReadLaneLo;
+ MachineOperand Op1L =
+ TII->buildExtractSubRegOrImm(Iters, MRI, DstVregInst->getOperand(0),
+ VregRC, AMDGPU::sub0, VregSubRC);
+ MachineOperand Op1H =
+ TII->buildExtractSubRegOrImm(Iters, MRI, DstVregInst->getOperand(0),
+ VregRC, AMDGPU::sub1, VregSubRC);
+ ReadLaneLo.add(Op1L);
+ ReadLaneHi.add(Op1H);
+ NewAccumulator = BuildMI(*ComputeLoop, I, DL,
+ TII->get(TargetOpcode::REG_SEQUENCE), DstReg)
+ .addReg(LaneValLo)
+ .addImm(AMDGPU::sub0)
+ .addReg(LaneValHi)
+ .addImm(AMDGPU::sub1);
+ break;
+ }
case AMDGPU::S_ADD_U64_PSEUDO:
case AMDGPU::S_SUB_U64_PSEUDO: {
NewAccumulator = BuildMI(*ComputeLoop, I, DL, TII->get(Opc), DstReg)
@@ -6009,6 +6069,8 @@ SITargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
return lowerWaveReduce(MI, *BB, *getSubtarget(), AMDGPU::V_CMP_LT_I64_e64);
case AMDGPU::WAVE_REDUCE_FMIN_PSEUDO_F32:
return lowerWaveReduce(MI, *BB, *getSubtarget(), AMDGPU::V_MIN_F32_e64);
+ case AMDGPU::WAVE_REDUCE_FMIN_PSEUDO_F64:
+ return lowerWaveReduce(MI, *BB, *getSubtarget(), AMDGPU::V_MIN_F64_e64);
case AMDGPU::WAVE_REDUCE_UMAX_PSEUDO_U32:
return lowerWaveReduce(MI, *BB, *getSubtarget(), AMDGPU::S_MAX_U32);
case AMDGPU::WAVE_REDUCE_UMAX_PSEUDO_U64:
@@ -6019,6 +6081,8 @@ SITargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
return lowerWaveReduce(MI, *BB, *getSubtarget(), AMDGPU::V_CMP_GT_I64_e64);
case AMDGPU::WAVE_REDUCE_FMAX_PSEUDO_F32:
return lowerWaveReduce(MI, *BB, *getSubtarget(), AMDGPU::V_MAX_F32_e64);
+ case AMDGPU::WAVE_REDUCE_FMAX_PSEUDO_F64:
+ return lowerWaveReduce(MI, *BB, *getSubtarget(), AMDGPU::V_MAX_F64_e64);
case AMDGPU::WAVE_REDUCE_ADD_PSEUDO_I32:
return lowerWaveReduce(MI, *BB, *getSubtarget(), AMDGPU::S_ADD_I32);
case AMDGPU::WAVE_REDUCE_ADD_PSEUDO_U64:
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index ca5a4d7301bda..82a83e12649fb 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -373,7 +373,9 @@ defvar Operations = [
WaveReduceOp<"xor", "B64", i64, SGPR_64, VSrc_b64>,
WaveReduceOp<"fmin", "F32", f32, SGPR_32, VSrc_b32>,
+ WaveReduceOp<"fmin", "F64", f64, SGPR_64, VSrc_b64>,
WaveReduceOp<"fmax", "F32", f32, SGPR_32, VSrc_b32>,
+ WaveReduceOp<"fmax", "F64", f64, SGPR_64, VSrc_b64>,
WaveReduceOp<"fadd", "F32", f32, SGPR_32, VSrc_b32>,
WaveReduceOp<"fsub", "F32", f32, SGPR_32, VSrc_b32>,
];
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.fmax.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.fmax.ll
index f02fd876f1aac..ca4091bcd3366 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.fmax.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.fmax.ll
@@ -11,6 +11,7 @@
; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=1 -mattr=+wavefrontsize64 < %s | FileCheck -check-prefixes=GFX11GISEL,GFX1164GISEL %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=0 < %s | FileCheck -check-prefixes=GFX11DAGISEL,GFX1132DAGISEL %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=1 < %s | FileCheck -check-prefixes=GFX11GISEL,GFX1132GISEL %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -global-isel=0 < %s | FileCheck -check-prefixes=GFX12DAGISEL %s
define amdgpu_kernel void @uniform_value_float(ptr addrspace(1) %out, float %in) {
@@ -119,6 +120,14 @@ define amdgpu_kernel void @uniform_value_float(ptr addrspace(1) %out, float %in)
; GFX1132GISEL-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s2
; GFX1132GISEL-NEXT: global_store_b32 v1, v0, s[0:1]
; GFX1132GISEL-NEXT: s_endpgm
+;
+; GFX12DAGISEL-LABEL: uniform_value_float:
+; GFX12DAGISEL: ; %bb.0: ; %entry
+; GFX12DAGISEL-NEXT: s_load_b96 s[0:2], s[4:5], 0x24
+; GFX12DAGISEL-NEXT: s_wait_kmcnt 0x0
+; GFX12DAGISEL-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
+; GFX12DAGISEL-NEXT: global_store_b32 v0, v1, s[0:1]
+; GFX12DAGISEL-NEXT: s_endpgm
entry:
%result = call float @llvm.amdgcn.wave.reduce.fmax(float %in, i32 1)
store float %result, ptr addrspace(1) %out
@@ -357,6 +366,33 @@ define void @divergent_value_float(ptr addrspace(1) %out, float %in) {
; GFX1132GISEL-NEXT: v_mov_b32_e32 v2, s1
; GFX1132GISEL-NEXT: global_store_b32 v[0:1], v2, off
; GFX1132GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX12DAGISEL-LABEL: divergent_value_float:
+; GFX12DAGISEL: ; %bb.0: ; %entry
+; GFX12DAGISEL-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12DAGISEL-NEXT: s_wait_expcnt 0x0
+; GFX12DAGISEL-NEXT: s_wait_samplecnt 0x0
+; GFX12DAGISEL-NEXT: s_wait_bvhcnt 0x0
+; GFX12DAGISEL-NEXT: s_wait_kmcnt 0x0
+; GFX12DAGISEL-NEXT: s_mov_b32 s0, exec_lo
+; GFX12DAGISEL-NEXT: s_mov_b32 s1, 0x7fc00000
+; GFX12DAGISEL-NEXT: .LBB1_1: ; =>This Inner Loop Header: Depth=1
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_sa_sdst(0)
+; GFX12DAGISEL-NEXT: s_ctz_i32_b32 s2, s0
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_sa_sdst(0)
+; GFX12DAGISEL-NEXT: v_readlane_b32 s3, v2, s2
+; GFX12DAGISEL-NEXT: s_bitset0_b32 s0, s2
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_sa_sdst(0)
+; GFX12DAGISEL-NEXT: s_cmp_lg_u32 s0, 0
+; GFX12DAGISEL-NEXT: v_max_num_f32_e64 v3, s1, s3
+; GFX12DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12DAGISEL-NEXT: v_readfirstlane_b32 s1, v3
+; GFX12DAGISEL-NEXT: s_cbranch_scc1 .LBB1_1
+; GFX12DAGISEL-NEXT: ; %bb.2:
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_va_sdst(0)
+; GFX12DAGISEL-NEXT: v_mov_b32_e32 v2, s1
+; GFX12DAGISEL-NEXT: global_store_b32 v[0:1], v2, off
+; GFX12DAGISEL-NEXT: s_setpc_b64 s[30:31]
entry:
%result = call float @llvm.amdgcn.wave.reduce.fmax(float %in, i32 1)
store float %result, ptr addrspace(1) %out
@@ -905,6 +941,68 @@ define void @divergent_cfg_float(ptr addrspace(1) %out, float %in, float %in2) {
; GFX1132GISEL-NEXT: v_mov_b32_e32 v2, s0
; GFX1132GISEL-NEXT: global_store_b32 v[0:1], v2, off
; GFX1132GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX12DAGISEL-LABEL: divergent_cfg_float:
+; GFX12DAGISEL: ; %bb.0: ; %entry
+; GFX12DAGISEL-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12DAGISEL-NEXT: s_wait_expcnt 0x0
+; GFX12DAGISEL-NEXT: s_wait_samplecnt 0x0
+; GFX12DAGISEL-NEXT: s_wait_bvhcnt 0x0
+; GFX12DAGISEL-NEXT: s_wait_kmcnt 0x0
+; GFX12DAGISEL-NEXT: v_and_b32_e32 v4, 0x3ff, v31
+; GFX12DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12DAGISEL-NEXT: v_cmp_lt_u32_e32 vcc_lo, 15, v4
+; GFX12DAGISEL-NEXT: ; implicit-def: $vgpr4
+; GFX12DAGISEL-NEXT: s_and_saveexec_b32 s0, vcc_lo
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_sa_sdst(0)
+; GFX12DAGISEL-NEXT: s_xor_b32 s0, exec_lo, s0
+; GFX12DAGISEL-NEXT: s_cbranch_execz .LBB2_4
+; GFX12DAGISEL-NEXT: ; %bb.1: ; %else
+; GFX12DAGISEL-NEXT: s_mov_b32 s1, exec_lo
+; GFX12DAGISEL-NEXT: s_mov_b32 s2, 0x7fc00000
+; GFX12DAGISEL-NEXT: .LBB2_2: ; =>This Inner Loop Header: Depth=1
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_sa_sdst(0)
+; GFX12DAGISEL-NEXT: s_ctz_i32_b32 s3, s1
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_sa_sdst(0)
+; GFX12DAGISEL-NEXT: v_readlane_b32 s4, v2, s3
+; GFX12DAGISEL-NEXT: s_bitset0_b32 s1, s3
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_sa_sdst(0)
+; GFX12DAGISEL-NEXT: s_cmp_lg_u32 s1, 0
+; GFX12DAGISEL-NEXT: v_max_num_f32_e64 v3, s2, s4
+; GFX12DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12DAGISEL-NEXT: v_readfirstlane_b32 s2, v3
+; GFX12DAGISEL-NEXT: s_cbranch_scc1 .LBB2_2
+; GFX12DAGISEL-NEXT: ; %bb.3:
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_va_sdst(0)
+; GFX12DAGISEL-NEXT: v_mov_b32_e32 v4, s2
+; GFX12DAGISEL-NEXT: ; implicit-def: $vgpr3
+; GFX12DAGISEL-NEXT: .LBB2_4: ; %Flow
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_sa_sdst(0)
+; GFX12DAGISEL-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX12DAGISEL-NEXT: s_cbranch_execz .LBB2_8
+; GFX12DAGISEL-NEXT: ; %bb.5: ; %if
+; GFX12DAGISEL-NEXT: s_mov_b32 s1, exec_lo
+; GFX12DAGISEL-NEXT: s_mov_b32 s2, 0x7fc00000
+; GFX12DAGISEL-NEXT: .LBB2_6: ; =>This Inner Loop Header: Depth=1
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_sa_sdst(0)
+; GFX12DAGISEL-NEXT: s_ctz_i32_b32 s3, s1
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_sa_sdst(0)
+; GFX12DAGISEL-NEXT: v_readlane_b32 s4, v3, s3
+; GFX12DAGISEL-NEXT: s_bitset0_b32 s1, s3
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_sa_sdst(0)
+; GFX12DAGISEL-NEXT: s_cmp_lg_u32 s1, 0
+; GFX12DAGISEL-NEXT: v_max_num_f32_e64 v2, s2, s4
+; GFX12DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12DAGISEL-NEXT: v_readfirstlane_b32 s2, v2
+; GFX12DAGISEL-NEXT: s_cbranch_scc1 .LBB2_6
+; GFX12DAGISEL-NEXT: ; %bb.7:
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_va_sdst(0)
+; GFX12DAGISEL-NEXT: v_mov_b32_e32 v4, s2
+; GFX12DAGISEL-NEXT: .LBB2_8: ; %endif
+; GFX12DAGISEL-NEXT: s_wait_alu depctr_sa_sdst(0)
+; GFX12DAGISEL-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12DAGISEL-NEXT: global_store_b32 v[0:1], v4, off
+; GFX12DAGISEL-NEXT: s_setpc_b64 s[30:31]
entry:
%tid = call i32 @llvm.amdgcn.workitem.id.x()
%d_cmp = icmp ult i32 %tid, 16
@@ -923,6 +1021,1188 @@ endif:
store float %combine, ptr addrspace(1) %out
ret void
}
+
+define amdgpu_kernel void @uniform_value_double(ptr addrspace(1) %out, double %in) {
+; GFX8DAGISEL-LABEL: uniform_value_double:
+; GFX8DAGISEL: ; %bb.0: ; %entry
+; GFX8DAGISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX8DAGISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8DAGISEL-NEXT: v_mov_b32_e32 v2, s2
+; GFX8DAGISEL-NEXT: v_mov_b32_e32 v0, s0
+; GFX8DAGISEL-NEXT: v_mov_b32_e32 v1, s1
+; GFX8DAGISEL-NEXT: v_mov_b32_e32 v3, s3
+; GFX8DAGISEL-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
+; GFX8DAGISEL-NEXT: s_endpgm
+;
+; GFX8GISEL-LABEL: uniform_value_double:
+; GFX8GISEL: ; %bb.0: ; %entry
+; GFX8GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX8GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8GISEL-NEXT: v_mov_b32_e32 v0, s2
+; GFX8GISEL-NEXT: v_mov_b32_e32 v3, s1
+; GFX8GISEL-NEXT: v_mov_b32_e32 v1, s3
+; GFX8GISEL-NEXT: v_mov_b32_e32 v2, s0
+; GFX8GISEL-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
+; GFX8GISEL-NEXT: s_endpgm
+;
+; GFX9DAGISEL-LABEL: uniform_value_double:
+; GFX9DAGISEL: ; %bb.0: ; %entry
+; GFX9DAGISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX9DAGISEL-NEXT: v_mov_b32_e32 v2, 0
+; GFX9DAGISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9DAGISEL-NEXT: v_mov_b32_e32 v0, s2
+; GFX9DAGISEL-NEXT: v_mov_b32_e32 v1, s3
+; GFX9DAGISEL-NEXT: global_store_dwordx2 v2, v[0:1], s[0:1]
+; GFX9DAGISEL-NEXT: s_endpgm
+;
+; GFX9GISEL-LABEL: uniform_value_double:
+; GFX9GISEL: ; %bb.0: ; %entry
+; GFX9GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX9GISEL-NEXT: v_mov_b32_e32 v2, 0
+; GFX9GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9GISEL-NEXT: v_mov_b32_e32 v0, s2
+; GFX9GISEL-NEXT: v_mov_b32_e32 v1, s3
+; GFX9GISEL-NEXT: global_store_dwordx2 v2, v[0:1], s[0:1]
+; GFX9GISEL-NEXT: s_endpgm
+;
+; GFX10DAGISEL-LABEL: uniform_value_double:
+; GFX10DAGISEL: ; %bb.0: ; %entry
+; GFX10DAGISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX10DAGISEL-NEXT: v_mov_b32_e32 v2, 0
+; GFX10DAGISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10DAGISEL-NEXT: v_mov_b32_e32 v0, s2
+; GFX10DAGISEL-NEXT: v_mov_b32_e32 v1, s3
+; GFX10DAGISEL-NEXT: global_store_dwordx2 v2, v[0:1], s[0:1]
+; GFX10DAGISEL-NEXT: s_endpgm
+;
+; GFX10GISEL-LABEL: uniform_value_double:
+; GFX10GISEL: ; %bb.0: ; %entry
+; GFX10GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX10GISEL-NEXT: v_mov_b32_e32 v2, 0
+; GFX10GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10GISEL-NEXT: v_mov_b32_e32 v0, s2
+; GFX10GISEL-NEXT: v_mov_b32_e32 v1, s3
+; GFX10GISEL-NEXT: global_store_dwordx2 v2, v[0:1], s[0:1]
+; GFX10GISEL-NEXT: s_endpgm
+;
+; GFX1164DAGISEL-LABEL: uniform_value_double:
+; GFX1164DAGISEL: ; %bb.0: ; %entry
+; GFX1164DAGISEL-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
+; GFX1164DAGISEL-NEXT: v_mov_b32_e32 v2, 0
+; GFX1164DAGISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1164DAGISEL-NEXT: v_mov_b32_e32 v0, s2
+; GFX1164DAGISEL-NEXT: v_mov_b32_e32 v1, s3
+; GFX1164DAGISEL-NEXT: global_store_b64 v2, v[0:1], s[0:1]
+; GFX1164DAGISEL-NEXT: s_endpgm
+;
+; GFX1164GISEL-LABEL: uniform_value_double:
+; GFX1164GISEL: ; %bb.0: ; %entry
+; GFX1164GISEL-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
+; GFX1164GISEL-NEXT: v_mov_b32_e32 v2, 0
+; GFX1164GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1164GISEL-NEXT: v_mov_b32_e32 v0, s2
+; GFX1164GISEL-NEXT: v_mov_b32_e32 v1, s3
+; GFX1164GISEL-NEXT: global_store_b64 v2, v[0:1], s[0:1]
+; GFX1164GISEL-NEXT: s_endpgm
+;
+; GFX1132DAGISEL-LABEL: uniform_value_double:
+; GFX1132DAGISEL: ; %bb.0: ; %entry
+; GFX1132DAGISEL-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
+; GFX1132DAGISEL-NEXT: v_mov_b32_e32 v2, 0
+; GFX1132DAGISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1132DAGISEL-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX1132DAGISEL-NEXT: global_store_b64 v2, v[0:1], s[0:1]
+; GFX1132DAGISEL-NEXT: s_endpgm
+;
+; GFX1132GISEL-LABEL: uniform_value_double:
+; GFX1132GISEL: ; %bb.0: ; %entry
+; GFX1132GISEL-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
+; GFX1132GISEL-NEXT: v_mov_b32_e32 v2, 0
+; GFX1132GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1132GISEL-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX1132GISEL-NEXT: global_store_b64 v2, v[0:1], s[0:1]
+; GFX1132GISEL-NEXT: s_endpgm
+;
+; GFX12DAGISEL-LABEL: uniform_value_double:
+; GFX12DAGISEL: ; %bb.0: ; %entry
+; GFX12DAGISEL-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
+; GFX12DAGISEL-NEXT: v_mov_b32_e32 v2, 0
+; GFX12DAGISEL-NEXT: s_wait_kmcnt 0x0
+; GFX12DAGISEL-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX12DAGISEL-NEXT: global_store_b64 v2, v[0:1], s[0:1]
+; GFX12DAGISEL-NEXT: s_endpgm
+entry:
+ %result = call double @llvm.amdgcn.wave.reduce.fmax(double %in, i32 1)
+ store double %result, ptr addrspace(1) %out
+ ret void
+}
+
+define void @divergent_value_double(ptr addrspace(1) %out, double %in) {
+; GFX8DAGISEL-LABEL: divergent_value_double:
+; GFX8DAGISEL: ; %bb.0: ; %entry
+; GFX8DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8DAGISEL-NEXT: s_mov_b32 s6, 0
+; GFX8DAGISEL-NEXT: s_mov_b32 s7, 0x7ff80000
+; GFX8DAGISEL-NEXT: s_mov_b64 s[4:5], exec
+; GFX8DAGISEL-NEXT: .LBB4_1: ; =>This Inner Loop Header: Depth=1
+; GFX8DAGISEL-NEXT: s_ff1_i32_b64 s10, s[4:5]
+; GFX8DAGISEL-NEXT: v_mov_b32_e32 v4, s6
+; GFX8DAGISEL-NEXT: v_readlane_b32 s8, v2, s10
+; GFX8DAGISEL-NEXT: v_mov_b32_e32 v5, s7
+; GFX8DAGISEL-NEXT: v_readlane_b32 s9, v3, s10
+; GFX8DAGISEL-NEXT: v_max_f64 v[4:5], s[8:9], v[4:5]
+; GFX8DAGISEL-NEXT: s_bitset0_b64 s[4:5], s10
+; GFX8DAGISEL-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX8DAGISEL-NEXT: v_readfirstlane_b32 s6, v4
+; GFX8DAGISEL-NEXT: v_readfirstlane_b32 s7, v5
+; GFX8DAGISEL-NEXT: s_cbranch_scc1 .LBB4_1
+; GFX8DAGISEL-NEXT: ; %bb.2:
+; GFX8DAGISEL-NEXT: v_mov_b32_e32 v2, s6
+; GFX8DAGISEL-NEXT: v_mov_b32_e32 v3, s7
+; GFX8DAGISEL-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
+; GFX8DAGISEL-NEXT: s_waitcnt vmcnt(0)
+; GFX8DAGISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8GISEL-LABEL: divergent_value_double:
+; GFX8GISEL: ; %bb.0: ; %entry
+; GFX8GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8GISEL-NEXT: s_mov_b32 s6, 0
+; GFX8GISEL-NEXT: s_mov_b32 s7, 0x7ff80000
+; GFX8GISEL-NEXT: s_mov_b64 s[4:5], exec
+; GFX8GISEL-NEXT: .LBB4_1: ; =>This Inner Loop Header: Depth=1
+; GFX8GISEL-NEXT: s_ff1_i32_b64 s10, s[4:5]
+; GFX8GISEL-NEXT: v_mov_b32_e32 v4, s6
+; GFX8GISEL-NEXT: v_readlane_b32 s8, v2, s10
+; GFX8GISEL-NEXT: v_mov_b32_e32 v5, s7
+; GFX8GISEL-NEXT: v_readlane_b32 s9, v3, s10
+; GFX8GISEL-NEXT: v_max_f64 v[4:5], s[8:9], v[4:5]
+; GFX8GISEL-NEXT: s_bitset0_b64 s[4:5], s10
+; GFX8GISEL-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX8GISEL-NEXT: v_readfirstlane_b32 s6, v4
+; GFX8GISEL-NEXT: v_readfirstlane_b32 s7, v5
+; GFX8GISEL-NEXT: s_cbranch_scc1 .LBB4_1
+; GFX8GISEL-NEXT: ; %bb.2:
+; GFX8GISEL-NEXT: v_mov_b32_e32 v2, s6
+; GFX8GISEL-NEXT: v_mov_b32_e32 v3, s7
+; GFX8GISEL-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
+; GFX8GISEL-NEXT: s_waitcnt vmcnt(0)
+; GFX8GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9DAGISEL-LABEL: divergent_value_double:
+; GFX9DAGISEL: ; %bb.0: ; %entry
+; GFX9DAGISEL-NEXT: s_waitcnt vmcnt(0...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/170811
More information about the llvm-commits
mailing list