[llvm-branch-commits] [llvm] [AMDGPU] Support Wave Reduction for i16 types - 3 (PR #194812)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Wed Apr 29 01:47:12 PDT 2026
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: Aaditya (easyonaadit)
<details>
<summary>Changes</summary>
Supported Ops: `and`, `or`, `xor`.
Supports only the iterative stratergy, DPP is yet
to be supported.
Supports only Fake-16 versions of the lowering.
True-16 support is yet to be added.
---
Patch is 219.95 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/194812.diff
5 Files Affected:
- (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+24-2)
- (modified) llvm/lib/Target/AMDGPU/SIInstructions.td (+3)
- (modified) llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.and.ll (+487-136)
- (modified) llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.or.ll (+487-136)
- (modified) llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.xor.ll (+587-160)
``````````diff
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 72c563ba7c0e8..67975b7a7a45a 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -5631,6 +5631,7 @@ static uint64_t getIdentityValueForWaveReduction(unsigned Opc) {
case AMDGPU::V_MIN_U16_opsel_e64:
case AMDGPU::V_MIN_U16_fake16_e64:
case AMDGPU::V_MIN_U16_t16_e64:
+ case AMDGPU::V_AND_B16_fake16_e64:
return 0xffff;
case AMDGPU::V_MIN_I16_e64:
case AMDGPU::V_MIN_I16_opsel_e64:
@@ -5647,6 +5648,8 @@ static uint64_t getIdentityValueForWaveReduction(unsigned Opc) {
case AMDGPU::V_SUB_I16_fake16_e64:
case AMDGPU::V_ADD_I16_t16_e64:
case AMDGPU::V_SUB_I16_t16_e64:
+ case AMDGPU::V_OR_B16_fake16_e64:
+ case AMDGPU::V_XOR_B16_fake16_e64:
return 0x0;
case AMDGPU::V_MAX_I16_e64:
case AMDGPU::V_MAX_I16_opsel_e64:
@@ -5716,7 +5719,10 @@ static bool is16bitWaveReduceOperation(unsigned Opc) {
Opc == AMDGPU::V_MAX_U16_t16_e64 || Opc == AMDGPU::V_MAX_I16_t16_e64 ||
Opc == AMDGPU::V_ADD_I16_e64 || Opc == AMDGPU::V_ADD_I16_fake16_e64 ||
Opc == AMDGPU::V_SUB_I16_e64 || Opc == AMDGPU::V_SUB_I16_fake16_e64 ||
- Opc == AMDGPU::V_ADD_I16_t16_e64 || Opc == AMDGPU::V_SUB_I16_t16_e64;
+ Opc == AMDGPU::V_ADD_I16_t16_e64 || Opc == AMDGPU::V_SUB_I16_t16_e64 ||
+ Opc == AMDGPU::V_AND_B16_fake16_e64 ||
+ Opc == AMDGPU::V_OR_B16_fake16_e64 ||
+ Opc == AMDGPU::V_XOR_B16_fake16_e64;
}
static bool is32bitWaveReduceOperation(unsigned Opc) {
@@ -5882,6 +5888,8 @@ static MachineBasicBlock *lowerWaveReduce(MachineInstr &MI,
case AMDGPU::S_MAX_U32:
case AMDGPU::S_MAX_I32:
case AMDGPU::V_MAX_F32_e64:
+ case AMDGPU::V_AND_B16_fake16_e64:
+ case AMDGPU::V_OR_B16_fake16_e64:
case AMDGPU::S_AND_B32:
case AMDGPU::S_OR_B32: {
// Idempotent operations.
@@ -5904,6 +5912,7 @@ static MachineBasicBlock *lowerWaveReduce(MachineInstr &MI,
RetBB = &BB;
break;
}
+ case AMDGPU::V_XOR_B16_fake16_e64:
case AMDGPU::S_XOR_B32:
case AMDGPU::S_XOR_B64:
case AMDGPU::V_ADD_I16_e64:
@@ -5939,6 +5948,7 @@ static MachineBasicBlock *lowerWaveReduce(MachineInstr &MI,
.addReg(ExecMask);
switch (Opc) {
+ case AMDGPU::V_XOR_B16_fake16_e64:
case AMDGPU::S_XOR_B32:
case AMDGPU::S_XOR_B64: {
// Performing an XOR operation on a uniform value
@@ -5952,7 +5962,7 @@ static MachineBasicBlock *lowerWaveReduce(MachineInstr &MI,
.addReg(NewAccumulator->getOperand(0).getReg())
.addImm(1)
.setOperandDead(3); // Dead scc
- if (Opc == AMDGPU::S_XOR_B32) {
+ if (Opc == AMDGPU::S_XOR_B32 || Opc == AMDGPU::V_XOR_B16_fake16_e64) {
BuildMI(BB, MI, DL, TII->get(AMDGPU::S_MUL_I32), DstReg)
.addReg(SrcReg)
.addReg(ParityRegister);
@@ -6894,14 +6904,26 @@ SITargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
ST.getGeneration() >= AMDGPUSubtarget::GFX12
? AMDGPU::V_ADD_F64_pseudo_e64
: AMDGPU::V_ADD_F64_e64);
+ case AMDGPU::WAVE_REDUCE_AND_PSEUDO_B16:
+ return lowerWaveReduce(MI, *BB, *getSubtarget(),
+ ST.hasTrue16BitInsts() ? AMDGPU::V_AND_B16_fake16_e64
+ : AMDGPU::S_AND_B32);
case AMDGPU::WAVE_REDUCE_AND_PSEUDO_B32:
return lowerWaveReduce(MI, *BB, *getSubtarget(), AMDGPU::S_AND_B32);
case AMDGPU::WAVE_REDUCE_AND_PSEUDO_B64:
return lowerWaveReduce(MI, *BB, *getSubtarget(), AMDGPU::S_AND_B64);
+ case AMDGPU::WAVE_REDUCE_OR_PSEUDO_B16:
+ return lowerWaveReduce(MI, *BB, *getSubtarget(),
+ ST.hasTrue16BitInsts() ? AMDGPU::V_OR_B16_fake16_e64
+ : AMDGPU::S_OR_B32);
case AMDGPU::WAVE_REDUCE_OR_PSEUDO_B32:
return lowerWaveReduce(MI, *BB, *getSubtarget(), AMDGPU::S_OR_B32);
case AMDGPU::WAVE_REDUCE_OR_PSEUDO_B64:
return lowerWaveReduce(MI, *BB, *getSubtarget(), AMDGPU::S_OR_B64);
+ case AMDGPU::WAVE_REDUCE_XOR_PSEUDO_B16:
+ return lowerWaveReduce(MI, *BB, *getSubtarget(),
+ ST.hasTrue16BitInsts() ? AMDGPU::V_XOR_B16_fake16_e64
+ : AMDGPU::S_XOR_B32);
case AMDGPU::WAVE_REDUCE_XOR_PSEUDO_B32:
return lowerWaveReduce(MI, *BB, *getSubtarget(), AMDGPU::S_XOR_B32);
case AMDGPU::WAVE_REDUCE_XOR_PSEUDO_B64:
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index 1afcb3ecc38ba..8d2432e680a5b 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -414,6 +414,9 @@ defvar Operations = [
WaveReduceOp<"max", "I16", i16, SGPR_32, VSrc_b16, NotUseRealTrue16Insts>,
WaveReduceOp<"add", "I16", i16, SGPR_32, VSrc_b16, NotUseRealTrue16Insts>,
WaveReduceOp<"sub", "I16", i16, SGPR_32, VSrc_b16, NotUseRealTrue16Insts>,
+ WaveReduceOp<"and", "B16", i16, SGPR_32, VSrc_b16, NotUseRealTrue16Insts>,
+ WaveReduceOp<"or", "B16", i16, SGPR_32, VSrc_b16, NotUseRealTrue16Insts>,
+ WaveReduceOp<"xor", "B16", i16, SGPR_32, VSrc_b16, NotUseRealTrue16Insts>,
WaveReduceOp<"umin", "U16_t16", i16, SGPR_32, VSrcT_b16, UseRealTrue16Insts>,
WaveReduceOp<"min", "I16_t16", i16, SGPR_32, VSrcT_b16, UseRealTrue16Insts>,
WaveReduceOp<"umax", "U16_t16", i16, SGPR_32, VSrcT_b16, UseRealTrue16Insts>,
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.and.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.and.ll
index 0774b8b5617cc..0cc033be63b32 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.and.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.and.ll
@@ -7,10 +7,356 @@
; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -global-isel=1 -new-reg-bank-select -mattr=+wavefrontsize64 < %s | FileCheck -check-prefixes=GFX10GISEL,GFX1064GISEL %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -global-isel=0 < %s | FileCheck -check-prefixes=GFX10DAGISEL,GFX1032DAGISEL %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -global-isel=1 -new-reg-bank-select < %s | FileCheck -check-prefixes=GFX10GISEL,GFX1032GISEL %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=0 -mattr=+wavefrontsize64 < %s | FileCheck -check-prefixes=GFX1164DAGISEL %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=1 -new-reg-bank-select -mattr=+wavefrontsize64 < %s | FileCheck -check-prefixes=GFX1164GISEL %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=0 < %s | FileCheck -check-prefixes=GFX1132DAGISEL %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=1 -new-reg-bank-select < %s | FileCheck -check-prefixes=GFX1132GISEL %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -global-isel=0 -mattr=+wavefrontsize64 < %s | FileCheck -check-prefixes=GFX1164DAGISEL,GFX1164DAGISEL-FAKE16 %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -global-isel=1 -new-reg-bank-select -mattr=+wavefrontsize64 < %s | FileCheck -check-prefixes=GFX1164GISEL,GFX1164GISEL-FAKE16 %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -global-isel=0 < %s | FileCheck -check-prefixes=GFX1132DAGISEL,GFX1132DAGISEL-FAKE16 %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -global-isel=1 -new-reg-bank-select < %s | FileCheck -check-prefixes=GFX1132GISEL,GFX1132GISEL-FAKE16 %s
+
+define amdgpu_kernel void @uniform_value_i16(ptr addrspace(1) %out, i16 %in) {
+; GFX8DAGISEL-LABEL: uniform_value_i16:
+; GFX8DAGISEL: ; %bb.0: ; %entry
+; GFX8DAGISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX8DAGISEL-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GFX8DAGISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8DAGISEL-NEXT: v_mov_b32_e32 v0, s0
+; GFX8DAGISEL-NEXT: v_mov_b32_e32 v1, s1
+; GFX8DAGISEL-NEXT: v_mov_b32_e32 v2, s2
+; GFX8DAGISEL-NEXT: flat_store_short v[0:1], v2
+; GFX8DAGISEL-NEXT: s_endpgm
+;
+; GFX8GISEL-LABEL: uniform_value_i16:
+; GFX8GISEL: ; %bb.0: ; %entry
+; GFX8GISEL-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GFX8GISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX8GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8GISEL-NEXT: s_and_b32 s2, 0xffff, s2
+; GFX8GISEL-NEXT: v_mov_b32_e32 v0, s0
+; GFX8GISEL-NEXT: v_mov_b32_e32 v2, s2
+; GFX8GISEL-NEXT: v_mov_b32_e32 v1, s1
+; GFX8GISEL-NEXT: flat_store_short v[0:1], v2
+; GFX8GISEL-NEXT: s_endpgm
+;
+; GFX9DAGISEL-LABEL: uniform_value_i16:
+; GFX9DAGISEL: ; %bb.0: ; %entry
+; GFX9DAGISEL-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GFX9DAGISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9DAGISEL-NEXT: v_mov_b32_e32 v0, 0
+; GFX9DAGISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9DAGISEL-NEXT: v_mov_b32_e32 v1, s2
+; GFX9DAGISEL-NEXT: global_store_short v0, v1, s[0:1]
+; GFX9DAGISEL-NEXT: s_endpgm
+;
+; GFX9GISEL-LABEL: uniform_value_i16:
+; GFX9GISEL: ; %bb.0: ; %entry
+; GFX9GISEL-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GFX9GISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9GISEL-NEXT: v_mov_b32_e32 v1, 0
+; GFX9GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9GISEL-NEXT: s_and_b32 s2, 0xffff, s2
+; GFX9GISEL-NEXT: v_mov_b32_e32 v0, s2
+; GFX9GISEL-NEXT: global_store_short v1, v0, s[0:1]
+; GFX9GISEL-NEXT: s_endpgm
+;
+; GFX10DAGISEL-LABEL: uniform_value_i16:
+; GFX10DAGISEL: ; %bb.0: ; %entry
+; GFX10DAGISEL-NEXT: s_clause 0x1
+; GFX10DAGISEL-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GFX10DAGISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX10DAGISEL-NEXT: v_mov_b32_e32 v0, 0
+; GFX10DAGISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10DAGISEL-NEXT: v_mov_b32_e32 v1, s2
+; GFX10DAGISEL-NEXT: global_store_short v0, v1, s[0:1]
+; GFX10DAGISEL-NEXT: s_endpgm
+;
+; GFX10GISEL-LABEL: uniform_value_i16:
+; GFX10GISEL: ; %bb.0: ; %entry
+; GFX10GISEL-NEXT: s_clause 0x1
+; GFX10GISEL-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GFX10GISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX10GISEL-NEXT: v_mov_b32_e32 v1, 0
+; GFX10GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10GISEL-NEXT: s_and_b32 s2, 0xffff, s2
+; GFX10GISEL-NEXT: v_mov_b32_e32 v0, s2
+; GFX10GISEL-NEXT: global_store_short v1, v0, s[0:1]
+; GFX10GISEL-NEXT: s_endpgm
+;
+; GFX1164DAGISEL-LABEL: uniform_value_i16:
+; GFX1164DAGISEL: ; %bb.0: ; %entry
+; GFX1164DAGISEL-NEXT: s_clause 0x1
+; GFX1164DAGISEL-NEXT: s_load_b32 s2, s[4:5], 0x2c
+; GFX1164DAGISEL-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1164DAGISEL-NEXT: v_mov_b32_e32 v0, 0
+; GFX1164DAGISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1164DAGISEL-NEXT: v_mov_b32_e32 v1, s2
+; GFX1164DAGISEL-NEXT: global_store_b16 v0, v1, s[0:1]
+; GFX1164DAGISEL-NEXT: s_endpgm
+;
+; GFX1164GISEL-LABEL: uniform_value_i16:
+; GFX1164GISEL: ; %bb.0: ; %entry
+; GFX1164GISEL-NEXT: s_clause 0x1
+; GFX1164GISEL-NEXT: s_load_b32 s2, s[4:5], 0x2c
+; GFX1164GISEL-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1164GISEL-NEXT: v_mov_b32_e32 v1, 0
+; GFX1164GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1164GISEL-NEXT: s_and_b32 s2, 0xffff, s2
+; GFX1164GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164GISEL-NEXT: v_mov_b32_e32 v0, s2
+; GFX1164GISEL-NEXT: global_store_b16 v1, v0, s[0:1]
+; GFX1164GISEL-NEXT: s_endpgm
+;
+; GFX1132DAGISEL-LABEL: uniform_value_i16:
+; GFX1132DAGISEL: ; %bb.0: ; %entry
+; GFX1132DAGISEL-NEXT: s_clause 0x1
+; GFX1132DAGISEL-NEXT: s_load_b32 s2, s[4:5], 0x2c
+; GFX1132DAGISEL-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1132DAGISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1132DAGISEL-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
+; GFX1132DAGISEL-NEXT: global_store_b16 v0, v1, s[0:1]
+; GFX1132DAGISEL-NEXT: s_endpgm
+;
+; GFX1132GISEL-LABEL: uniform_value_i16:
+; GFX1132GISEL: ; %bb.0: ; %entry
+; GFX1132GISEL-NEXT: s_clause 0x1
+; GFX1132GISEL-NEXT: s_load_b32 s2, s[4:5], 0x2c
+; GFX1132GISEL-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1132GISEL-NEXT: v_mov_b32_e32 v1, 0
+; GFX1132GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1132GISEL-NEXT: s_and_b32 s2, 0xffff, s2
+; GFX1132GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132GISEL-NEXT: v_mov_b32_e32 v0, s2
+; GFX1132GISEL-NEXT: global_store_b16 v1, v0, s[0:1]
+; GFX1132GISEL-NEXT: s_endpgm
+ entry:
+ %result = call i16 @llvm.amdgcn.wave.reduce.and.i16(i16 %in, i32 1)
+ store i16 %result, ptr addrspace(1) %out
+ ret void
+}
+
+define void @divergnet_value_i16(ptr addrspace(1) %out, i16 %in) {
+; GFX8DAGISEL-LABEL: divergnet_value_i16:
+; GFX8DAGISEL: ; %bb.0: ; %entry
+; GFX8DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8DAGISEL-NEXT: s_mov_b64 s[4:5], exec
+; GFX8DAGISEL-NEXT: s_mov_b32 s6, -1
+; GFX8DAGISEL-NEXT: .LBB1_1: ; =>This Inner Loop Header: Depth=1
+; GFX8DAGISEL-NEXT: s_ff1_i32_b64 s7, s[4:5]
+; GFX8DAGISEL-NEXT: v_readlane_b32 s8, v2, s7
+; GFX8DAGISEL-NEXT: s_bitset0_b64 s[4:5], s7
+; GFX8DAGISEL-NEXT: s_and_b32 s6, s6, s8
+; GFX8DAGISEL-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX8DAGISEL-NEXT: s_cbranch_scc1 .LBB1_1
+; GFX8DAGISEL-NEXT: ; %bb.2:
+; GFX8DAGISEL-NEXT: v_mov_b32_e32 v2, s6
+; GFX8DAGISEL-NEXT: flat_store_short v[0:1], v2
+; GFX8DAGISEL-NEXT: s_waitcnt vmcnt(0)
+; GFX8DAGISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8GISEL-LABEL: divergnet_value_i16:
+; GFX8GISEL: ; %bb.0: ; %entry
+; GFX8GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8GISEL-NEXT: s_mov_b64 s[4:5], exec
+; GFX8GISEL-NEXT: s_mov_b32 s6, -1
+; GFX8GISEL-NEXT: .LBB1_1: ; =>This Inner Loop Header: Depth=1
+; GFX8GISEL-NEXT: s_ff1_i32_b64 s7, s[4:5]
+; GFX8GISEL-NEXT: v_readlane_b32 s8, v2, s7
+; GFX8GISEL-NEXT: s_bitset0_b64 s[4:5], s7
+; GFX8GISEL-NEXT: s_and_b32 s6, s6, s8
+; GFX8GISEL-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX8GISEL-NEXT: s_cbranch_scc1 .LBB1_1
+; GFX8GISEL-NEXT: ; %bb.2:
+; GFX8GISEL-NEXT: v_mov_b32_e32 v2, s6
+; GFX8GISEL-NEXT: flat_store_short v[0:1], v2
+; GFX8GISEL-NEXT: s_waitcnt vmcnt(0)
+; GFX8GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9DAGISEL-LABEL: divergnet_value_i16:
+; GFX9DAGISEL: ; %bb.0: ; %entry
+; GFX9DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9DAGISEL-NEXT: s_mov_b64 s[4:5], exec
+; GFX9DAGISEL-NEXT: s_mov_b32 s6, -1
+; GFX9DAGISEL-NEXT: .LBB1_1: ; =>This Inner Loop Header: Depth=1
+; GFX9DAGISEL-NEXT: s_ff1_i32_b64 s7, s[4:5]
+; GFX9DAGISEL-NEXT: v_readlane_b32 s8, v2, s7
+; GFX9DAGISEL-NEXT: s_bitset0_b64 s[4:5], s7
+; GFX9DAGISEL-NEXT: s_and_b32 s6, s6, s8
+; GFX9DAGISEL-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX9DAGISEL-NEXT: s_cbranch_scc1 .LBB1_1
+; GFX9DAGISEL-NEXT: ; %bb.2:
+; GFX9DAGISEL-NEXT: v_mov_b32_e32 v2, s6
+; GFX9DAGISEL-NEXT: global_store_short v[0:1], v2, off
+; GFX9DAGISEL-NEXT: s_waitcnt vmcnt(0)
+; GFX9DAGISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9GISEL-LABEL: divergnet_value_i16:
+; GFX9GISEL: ; %bb.0: ; %entry
+; GFX9GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9GISEL-NEXT: s_mov_b64 s[4:5], exec
+; GFX9GISEL-NEXT: s_mov_b32 s6, -1
+; GFX9GISEL-NEXT: .LBB1_1: ; =>This Inner Loop Header: Depth=1
+; GFX9GISEL-NEXT: s_ff1_i32_b64 s7, s[4:5]
+; GFX9GISEL-NEXT: v_readlane_b32 s8, v2, s7
+; GFX9GISEL-NEXT: s_bitset0_b64 s[4:5], s7
+; GFX9GISEL-NEXT: s_and_b32 s6, s6, s8
+; GFX9GISEL-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX9GISEL-NEXT: s_cbranch_scc1 .LBB1_1
+; GFX9GISEL-NEXT: ; %bb.2:
+; GFX9GISEL-NEXT: v_mov_b32_e32 v2, s6
+; GFX9GISEL-NEXT: global_store_short v[0:1], v2, off
+; GFX9GISEL-NEXT: s_waitcnt vmcnt(0)
+; GFX9GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1064DAGISEL-LABEL: divergnet_value_i16:
+; GFX1064DAGISEL: ; %bb.0: ; %entry
+; GFX1064DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX1064DAGISEL-NEXT: s_mov_b64 s[4:5], exec
+; GFX1064DAGISEL-NEXT: s_mov_b32 s6, -1
+; GFX1064DAGISEL-NEXT: .LBB1_1: ; =>This Inner Loop Header: Depth=1
+; GFX1064DAGISEL-NEXT: s_ff1_i32_b64 s7, s[4:5]
+; GFX1064DAGISEL-NEXT: v_readlane_b32 s8, v2, s7
+; GFX1064DAGISEL-NEXT: s_bitset0_b64 s[4:5], s7
+; GFX1064DAGISEL-NEXT: s_and_b32 s6, s6, s8
+; GFX1064DAGISEL-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX1064DAGISEL-NEXT: s_cbranch_scc1 .LBB1_1
+; GFX1064DAGISEL-NEXT: ; %bb.2:
+; GFX1064DAGISEL-NEXT: v_mov_b32_e32 v2, s6
+; GFX1064DAGISEL-NEXT: global_store_short v[0:1], v2, off
+; GFX1064DAGISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1064GISEL-LABEL: divergnet_value_i16:
+; GFX1064GISEL: ; %bb.0: ; %entry
+; GFX1064GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX1064GISEL-NEXT: s_mov_b64 s[4:5], exec
+; GFX1064GISEL-NEXT: s_mov_b32 s6, -1
+; GFX1064GISEL-NEXT: .LBB1_1: ; =>This Inner Loop Header: Depth=1
+; GFX1064GISEL-NEXT: s_ff1_i32_b64 s7, s[4:5]
+; GFX1064GISEL-NEXT: v_readlane_b32 s8, v2, s7
+; GFX1064GISEL-NEXT: s_bitset0_b64 s[4:5], s7
+; GFX1064GISEL-NEXT: s_and_b32 s6, s6, s8
+; GFX1064GISEL-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX1064GISEL-NEXT: s_cbranch_scc1 .LBB1_1
+; GFX1064GISEL-NEXT: ; %bb.2:
+; GFX1064GISEL-NEXT: v_mov_b32_e32 v2, s6
+; GFX1064GISEL-NEXT: global_store_short v[0:1], v2, off
+; GFX1064GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1032DAGISEL-LABEL: divergnet_value_i16:
+; GFX1032DAGISEL: ; %bb.0: ; %entry
+; GFX1032DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX1032DAGISEL-NEXT: s_mov_b32 s5, exec_lo
+; GFX1032DAGISEL-NEXT: s_mov_b32 s4, -1
+; GFX1032DAGISEL-NEXT: .LBB1_1: ; =>This Inner Loop Header: Depth=1
+; GFX1032DAGISEL-NEXT: s_ff1_i32_b32 s6, s5
+; GFX1032DAGISEL-NEXT: v_readlane_b32 s7, v2, s6
+; GFX1032DAGISEL-NEXT: s_bitset0_b32 s5, s6
+; GFX1032DAGISEL-NEXT: s_and_b32 s4, s4, s7
+; GFX1032DAGISEL-NEXT: s_cmp_lg_u32 s5, 0
+; GFX1032DAGISEL-NEXT: s_cbranch_scc1 .LBB1_1
+; GFX1032DAGISEL-NEXT: ; %bb.2:
+; GFX1032DAGISEL-NEXT: v_mov_b32_e32 v2, s4
+; GFX1032DAGISEL-NEXT: global_store_short v[0:1], v2, off
+; GFX1032DAGISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1032GISEL-LABEL: divergnet_value_i16:
+; GFX1032GISEL: ; %bb.0: ; %entry
+; GFX1032GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX1032GISEL-NEXT: s_mov_b32 s5, exec_lo
+; GFX1032GISEL-NEXT: s_mov_b32 s4, -1
+; GFX1032GISEL-NEXT: .LBB1_1: ; =>This Inner Loop Header: Depth=1
+; GFX1032GISEL-NEXT: s_ff1_i32_b32 s6, s5
+; GFX1032GISEL-NEXT: v_readlane_b32 s7, v2, s6
+; GFX1032GISEL-NEXT: s_bitset0_b32 s5, s6
+; GFX1032GISEL-NEXT: s_and_b32 s4, s4, s7
+; GFX1032GISEL-NEXT: s_cmp_lg_u32 s5, 0
+; GFX1032GISEL-NEXT: s_cbranch_scc1 .LBB1_1
+; GFX1032GISEL-NEXT: ; %bb.2:
+; GFX1032GISEL-NEXT: v_mov_b32_e32 v2, s4
+; GFX1032GISEL-NEXT: global_store_short v[0:1], v2, off
+; GFX1032GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1164DAGISEL-LABEL: divergnet_value_i16:
+; GFX1164DAGISEL: ; %bb.0: ; %entry
+; GFX1164DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX1164DAGISEL-NEXT: s_mov_b64 s[0:1], exec
+; GFX1164DAGISEL-NEXT: s_mov_b32 s2, 0xffff
+; GFX1164DAGISEL-NEXT: .LBB1_1: ; =>This Inner Loop Header: Depth=1
+; GFX1164DAGISEL-NEXT: s_ctz_i32_b64 s3, s[0:1]
+; GFX1164DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX1164DAGISEL-NEXT: v_readlane_b32 s4, v2, s3
+; GFX1164DAGISEL-NEXT: s_bitset0_b64 s[0:1], s3
+; GFX1164DAGISEL-NEXT: s_cmp_lg_u64 s[0:1], 0
+; GFX1164DAGISEL-NEXT: v_and_b16 v3, s2, s4
+; GFX1164DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1164DAGISEL-NEXT: v_readfirstlane_b32 s2, v3
+; GFX1164DAGISEL-NEXT: s_cbranch_scc1 .LBB1_1
+; GFX1164DAGISEL-NEXT: ; %bb.2:
+; GFX1164DAGISEL-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164DAGISEL-NEXT: global_store_b16 v[0:1], v2, off
+; GFX1164DAGISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1164GISEL-LABEL: divergnet_value_i16:
+; GFX1164GISEL: ; %bb.0: ; %entry
+; GFX1164GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX1164GISEL-NEXT: s_mov_b64 s[0:1], exec
+; GFX1164GISEL-NEXT: s_mov_b32 s2, 0xffff
+; GFX1164GISEL-NEXT: .LBB1_1: ; =>This Inner Loop Header: Depth=1
+; GFX1164GISEL-NEXT: s_ctz_i32_b64 s3, s[0:1...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/194812
More information about the llvm-branch-commits
mailing list