[llvm] [AMDGPU] Delete redundant s_cmp_or32 (PR #165261)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Oct 27 08:06:27 PDT 2025
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: None (LU-JOHN)
<details>
<summary>Changes</summary>
Transform sequences like:
```
s_cselect_b64 s12, -1, 0
s_or_b32 s6, s12, s13
```
where s6 is dead to:
` s_cselect_b64 s12, -1, 0`
---
Patch is 135.87 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/165261.diff
11 Files Affected:
- (modified) llvm/lib/Target/AMDGPU/SIInstrInfo.cpp (+50-32)
- (modified) llvm/lib/Target/AMDGPU/SIInstrInfo.h (+1-1)
- (modified) llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll (+448-493)
- (modified) llvm/test/CodeGen/AMDGPU/carryout-selection.ll (-4)
- (modified) llvm/test/CodeGen/AMDGPU/expand-scalar-carry-out-select-user.ll (+3-7)
- (modified) llvm/test/CodeGen/AMDGPU/sdiv64.ll (+173-195)
- (modified) llvm/test/CodeGen/AMDGPU/srem64.ll (+187-223)
- (modified) llvm/test/CodeGen/AMDGPU/uaddo.ll (+1-5)
- (modified) llvm/test/CodeGen/AMDGPU/udiv64.ll (+90-109)
- (modified) llvm/test/CodeGen/AMDGPU/urem64.ll (+134-162)
- (modified) llvm/test/CodeGen/AMDGPU/usubo.ll (+1-5)
``````````diff
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index d930a21c2d7f5..2172e733ef732 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -10160,7 +10160,7 @@ static bool followSubRegDef(MachineInstr &MI,
}
MachineInstr *llvm::getVRegSubRegDef(const TargetInstrInfo::RegSubRegPair &P,
- MachineRegisterInfo &MRI) {
+ const MachineRegisterInfo &MRI) {
assert(MRI.isSSA());
if (!P.Reg.isVirtual())
return nullptr;
@@ -10628,7 +10628,31 @@ bool SIInstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg,
if (SrcReg2 && !getFoldableImm(SrcReg2, *MRI, CmpValue))
return false;
- const auto optimizeCmpSelect = [&CmpInstr, SrcReg, CmpValue, MRI,
+ // SCC is already valid after SCCValid.
+ // SCCRedefine will redefine SCC to the same value already available after
+ // SCCValid. If there are no intervening SCC conflicts delete SCCRedefine and
+ // update kill/dead flags if necessary.
+ const auto optimizeSCC = [this](MachineInstr *SCCValid,
+ MachineInstr *SCCRedefine) -> bool {
+ MachineInstr *KillsSCC = nullptr;
+ for (MachineInstr &MI : make_range(std::next(SCCValid->getIterator()),
+ SCCRedefine->getIterator())) {
+ if (MI.modifiesRegister(AMDGPU::SCC, &RI))
+ return false;
+ if (MI.killsRegister(AMDGPU::SCC, &RI))
+ KillsSCC = &MI;
+ }
+ if (MachineOperand *SccDef =
+ SCCValid->findRegisterDefOperand(AMDGPU::SCC, /*TRI=*/nullptr))
+ SccDef->setIsDead(false);
+ if (KillsSCC)
+ KillsSCC->clearRegisterKills(AMDGPU::SCC, /*TRI=*/nullptr);
+ SCCRedefine->eraseFromParent();
+
+ return true;
+ };
+
+ const auto optimizeCmpSelect = [&CmpInstr, SrcReg, CmpValue, MRI, optimizeSCC,
this]() -> bool {
if (CmpValue != 0)
return false;
@@ -10663,25 +10687,32 @@ bool SIInstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg,
if (!setsSCCifResultIsNonZero(*Def) && !foldableSelect(Def))
return false;
- MachineInstr *KillsSCC = nullptr;
- for (MachineInstr &MI :
- make_range(std::next(Def->getIterator()), CmpInstr.getIterator())) {
- if (MI.modifiesRegister(AMDGPU::SCC, &RI))
- return false;
- if (MI.killsRegister(AMDGPU::SCC, &RI))
- KillsSCC = &MI;
- }
+ if (!optimizeSCC(Def, &CmpInstr))
+ return false;
- if (MachineOperand *SccDef =
- Def->findRegisterDefOperand(AMDGPU::SCC, /*TRI=*/nullptr))
- SccDef->setIsDead(false);
- if (KillsSCC)
- KillsSCC->clearRegisterKills(AMDGPU::SCC, /*TRI=*/nullptr);
- CmpInstr.eraseFromParent();
+ // If s_or_32 result is unused (i.e. it is effectively a 64-bit s_cmp_lg of
+ // a register pair) and the input is a 64-bit foldableSelect then transform:
+ //
+ // (s_or_b32 (S_CSELECT_B64 (non-zero imm), 0), 0 => (S_CSELECT_B64
+ // (non-zero
+ // imm), 0)
+ if (Def->getOpcode() == AMDGPU::S_OR_B32 &&
+ MRI->use_nodbg_empty(Def->getOperand(0).getReg())) {
+ MachineOperand OrOpnd1 = Def->getOperand(1);
+ MachineOperand OrOpnd2 = Def->getOperand(2);
+
+ if (OrOpnd1.isReg() && OrOpnd2.isReg() &&
+ OrOpnd1.getReg() != OrOpnd2.getReg()) {
+ auto *Def1 = getVRegSubRegDef(getRegSubRegPair(OrOpnd1), *MRI);
+ auto *Def2 = getVRegSubRegDef(getRegSubRegPair(OrOpnd2), *MRI);
+ if (Def1 == Def2 && foldableSelect(Def1))
+ optimizeSCC(Def1, Def);
+ }
+ }
return true;
};
- const auto optimizeCmpAnd = [&CmpInstr, SrcReg, CmpValue, MRI,
+ const auto optimizeCmpAnd = [&CmpInstr, SrcReg, CmpValue, MRI, optimizeSCC,
this](int64_t ExpectedValue, unsigned SrcSize,
bool IsReversible, bool IsSigned) -> bool {
// s_cmp_eq_u32 (s_and_b32 $src, 1 << n), 1 << n => s_and_b32 $src, 1 << n
@@ -10755,21 +10786,8 @@ bool SIInstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg,
if (IsReversedCC && !MRI->hasOneNonDBGUse(DefReg))
return false;
- MachineInstr *KillsSCC = nullptr;
- for (MachineInstr &MI :
- make_range(std::next(Def->getIterator()), CmpInstr.getIterator())) {
- if (MI.modifiesRegister(AMDGPU::SCC, &RI))
- return false;
- if (MI.killsRegister(AMDGPU::SCC, &RI))
- KillsSCC = &MI;
- }
-
- MachineOperand *SccDef =
- Def->findRegisterDefOperand(AMDGPU::SCC, /*TRI=*/nullptr);
- SccDef->setIsDead(false);
- if (KillsSCC)
- KillsSCC->clearRegisterKills(AMDGPU::SCC, /*TRI=*/nullptr);
- CmpInstr.eraseFromParent();
+ if (!optimizeSCC(Def, &CmpInstr))
+ return false;
if (!MRI->use_nodbg_empty(DefReg)) {
assert(!IsReversedCC);
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.h b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
index 5fdeddaf3f736..f7caae9e257bc 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.h
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
@@ -1687,7 +1687,7 @@ TargetInstrInfo::RegSubRegPair getRegSequenceSubReg(MachineInstr &MI,
/// skipping copy like instructions and subreg-manipulation pseudos.
/// Following another subreg of a reg:subreg isn't supported.
MachineInstr *getVRegSubRegDef(const TargetInstrInfo::RegSubRegPair &P,
- MachineRegisterInfo &MRI);
+ const MachineRegisterInfo &MRI);
/// \brief Return false if EXEC is not changed between the def of \p VReg at \p
/// DefMI and the use at \p UseMI. Should be run on SSA. Currently does not
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll
index 51df8c34cc55e..54b1554ae5d04 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll
@@ -7772,7 +7772,6 @@ define amdgpu_kernel void @sdiv_i64_pow2_shl_denom(ptr addrspace(1) %out, i64 %x
; GFX6-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0xd
; GFX6-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9
; GFX6-NEXT: s_mov_b32 s3, 0xf000
-; GFX6-NEXT: s_mov_b32 s2, -1
; GFX6-NEXT: s_waitcnt lgkmcnt(0)
; GFX6-NEXT: s_lshl_b64 s[0:1], 0x1000, s0
; GFX6-NEXT: s_ashr_i32 s8, s1, 31
@@ -7782,8 +7781,8 @@ define amdgpu_kernel void @sdiv_i64_pow2_shl_denom(ptr addrspace(1) %out, i64 %x
; GFX6-NEXT: s_xor_b64 s[10:11], s[0:1], s[8:9]
; GFX6-NEXT: v_cvt_f32_u32_e32 v0, s10
; GFX6-NEXT: v_cvt_f32_u32_e32 v1, s11
-; GFX6-NEXT: s_sub_u32 s12, 0, s10
-; GFX6-NEXT: s_subb_u32 s13, 0, s11
+; GFX6-NEXT: s_sub_u32 s0, 0, s10
+; GFX6-NEXT: s_subb_u32 s1, 0, s11
; GFX6-NEXT: v_madmk_f32 v0, v1, 0x4f800000, v0
; GFX6-NEXT: v_rcp_f32_e32 v0, v0
; GFX6-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0
@@ -7792,128 +7791,121 @@ define amdgpu_kernel void @sdiv_i64_pow2_shl_denom(ptr addrspace(1) %out, i64 %x
; GFX6-NEXT: v_madmk_f32 v0, v1, 0xcf800000, v0
; GFX6-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX6-NEXT: v_cvt_u32_f32_e32 v1, v1
-; GFX6-NEXT: v_mul_hi_u32 v2, s12, v0
-; GFX6-NEXT: v_readfirstlane_b32 s14, v1
-; GFX6-NEXT: v_readfirstlane_b32 s0, v0
-; GFX6-NEXT: s_mul_i32 s1, s12, s14
-; GFX6-NEXT: v_readfirstlane_b32 s17, v2
-; GFX6-NEXT: s_mul_i32 s15, s13, s0
-; GFX6-NEXT: s_mul_i32 s16, s12, s0
-; GFX6-NEXT: s_add_i32 s1, s17, s1
-; GFX6-NEXT: v_mul_hi_u32 v3, v0, s16
-; GFX6-NEXT: s_add_i32 s1, s1, s15
-; GFX6-NEXT: v_mul_hi_u32 v0, v0, s1
-; GFX6-NEXT: v_mul_hi_u32 v4, v1, s16
-; GFX6-NEXT: v_readfirstlane_b32 s15, v3
-; GFX6-NEXT: s_mul_i32 s17, s0, s1
-; GFX6-NEXT: v_mul_hi_u32 v1, v1, s1
-; GFX6-NEXT: s_add_u32 s15, s15, s17
-; GFX6-NEXT: v_readfirstlane_b32 s17, v0
-; GFX6-NEXT: s_addc_u32 s17, 0, s17
-; GFX6-NEXT: s_mul_i32 s16, s14, s16
-; GFX6-NEXT: v_readfirstlane_b32 s18, v4
-; GFX6-NEXT: s_add_u32 s15, s15, s16
-; GFX6-NEXT: s_addc_u32 s15, s17, s18
-; GFX6-NEXT: v_readfirstlane_b32 s16, v1
-; GFX6-NEXT: s_addc_u32 s16, s16, 0
-; GFX6-NEXT: s_mul_i32 s1, s14, s1
-; GFX6-NEXT: s_add_u32 s1, s15, s1
-; GFX6-NEXT: s_addc_u32 s15, 0, s16
-; GFX6-NEXT: s_add_u32 s16, s0, s1
-; GFX6-NEXT: v_mov_b32_e32 v0, s16
-; GFX6-NEXT: v_mul_hi_u32 v0, s12, v0
-; GFX6-NEXT: s_cselect_b64 s[0:1], -1, 0
-; GFX6-NEXT: s_or_b32 s0, s0, s1
-; GFX6-NEXT: s_addc_u32 s14, s14, s15
-; GFX6-NEXT: s_mul_i32 s0, s12, s14
-; GFX6-NEXT: v_readfirstlane_b32 s1, v0
-; GFX6-NEXT: s_add_i32 s0, s1, s0
-; GFX6-NEXT: s_mul_i32 s13, s13, s16
-; GFX6-NEXT: s_mul_i32 s1, s12, s16
-; GFX6-NEXT: s_add_i32 s0, s0, s13
-; GFX6-NEXT: v_mov_b32_e32 v2, s1
-; GFX6-NEXT: v_mov_b32_e32 v0, s0
-; GFX6-NEXT: v_mul_hi_u32 v3, s14, v2
-; GFX6-NEXT: v_mul_hi_u32 v2, s16, v2
-; GFX6-NEXT: v_mul_hi_u32 v1, s14, v0
-; GFX6-NEXT: v_mul_hi_u32 v0, s16, v0
-; GFX6-NEXT: s_mul_i32 s13, s16, s0
-; GFX6-NEXT: v_readfirstlane_b32 s17, v2
-; GFX6-NEXT: s_add_u32 s13, s17, s13
-; GFX6-NEXT: v_readfirstlane_b32 s15, v0
-; GFX6-NEXT: s_mul_i32 s1, s14, s1
-; GFX6-NEXT: s_addc_u32 s15, 0, s15
-; GFX6-NEXT: v_readfirstlane_b32 s12, v3
-; GFX6-NEXT: s_add_u32 s1, s13, s1
-; GFX6-NEXT: s_addc_u32 s1, s15, s12
+; GFX6-NEXT: v_mul_hi_u32 v2, s0, v0
; GFX6-NEXT: v_readfirstlane_b32 s12, v1
-; GFX6-NEXT: s_addc_u32 s12, s12, 0
-; GFX6-NEXT: s_mul_i32 s0, s14, s0
-; GFX6-NEXT: s_add_u32 s0, s1, s0
-; GFX6-NEXT: s_addc_u32 s12, 0, s12
-; GFX6-NEXT: s_add_u32 s15, s16, s0
-; GFX6-NEXT: s_cselect_b64 s[0:1], -1, 0
-; GFX6-NEXT: s_or_b32 s0, s0, s1
-; GFX6-NEXT: s_addc_u32 s14, s14, s12
+; GFX6-NEXT: v_readfirstlane_b32 s2, v0
+; GFX6-NEXT: s_mul_i32 s13, s0, s12
+; GFX6-NEXT: v_readfirstlane_b32 s16, v2
+; GFX6-NEXT: s_mul_i32 s14, s1, s2
+; GFX6-NEXT: s_mul_i32 s15, s0, s2
+; GFX6-NEXT: s_add_i32 s13, s16, s13
+; GFX6-NEXT: v_mul_hi_u32 v3, v0, s15
+; GFX6-NEXT: s_add_i32 s13, s13, s14
+; GFX6-NEXT: v_mul_hi_u32 v0, v0, s13
+; GFX6-NEXT: v_mul_hi_u32 v4, v1, s15
+; GFX6-NEXT: v_readfirstlane_b32 s14, v3
+; GFX6-NEXT: s_mul_i32 s16, s2, s13
+; GFX6-NEXT: v_mul_hi_u32 v1, v1, s13
+; GFX6-NEXT: s_add_u32 s14, s14, s16
+; GFX6-NEXT: v_readfirstlane_b32 s16, v0
+; GFX6-NEXT: s_mul_i32 s15, s12, s15
+; GFX6-NEXT: s_addc_u32 s16, 0, s16
+; GFX6-NEXT: v_readfirstlane_b32 s17, v4
+; GFX6-NEXT: s_add_u32 s14, s14, s15
+; GFX6-NEXT: s_addc_u32 s14, s16, s17
+; GFX6-NEXT: v_readfirstlane_b32 s15, v1
+; GFX6-NEXT: s_addc_u32 s15, s15, 0
+; GFX6-NEXT: s_mul_i32 s13, s12, s13
+; GFX6-NEXT: s_add_u32 s13, s14, s13
+; GFX6-NEXT: s_addc_u32 s14, 0, s15
+; GFX6-NEXT: s_add_u32 s13, s2, s13
+; GFX6-NEXT: v_mov_b32_e32 v0, s13
+; GFX6-NEXT: v_mul_hi_u32 v0, s0, v0
+; GFX6-NEXT: s_addc_u32 s12, s12, s14
+; GFX6-NEXT: s_mul_i32 s14, s0, s12
+; GFX6-NEXT: s_mul_i32 s1, s1, s13
+; GFX6-NEXT: v_readfirstlane_b32 s15, v0
+; GFX6-NEXT: s_add_i32 s14, s15, s14
+; GFX6-NEXT: s_mul_i32 s0, s0, s13
+; GFX6-NEXT: s_add_i32 s1, s14, s1
+; GFX6-NEXT: v_mov_b32_e32 v2, s0
+; GFX6-NEXT: v_mov_b32_e32 v0, s1
+; GFX6-NEXT: v_mul_hi_u32 v3, s12, v2
+; GFX6-NEXT: v_mul_hi_u32 v2, s13, v2
+; GFX6-NEXT: v_mul_hi_u32 v1, s12, v0
+; GFX6-NEXT: v_mul_hi_u32 v0, s13, v0
+; GFX6-NEXT: s_mul_i32 s15, s13, s1
+; GFX6-NEXT: v_readfirstlane_b32 s17, v2
+; GFX6-NEXT: s_add_u32 s15, s17, s15
+; GFX6-NEXT: v_readfirstlane_b32 s16, v0
+; GFX6-NEXT: s_mul_i32 s0, s12, s0
+; GFX6-NEXT: s_addc_u32 s16, 0, s16
+; GFX6-NEXT: v_readfirstlane_b32 s14, v3
+; GFX6-NEXT: s_add_u32 s0, s15, s0
+; GFX6-NEXT: s_addc_u32 s0, s16, s14
+; GFX6-NEXT: v_readfirstlane_b32 s14, v1
+; GFX6-NEXT: s_addc_u32 s14, s14, 0
+; GFX6-NEXT: s_mul_i32 s1, s12, s1
+; GFX6-NEXT: s_add_u32 s0, s0, s1
+; GFX6-NEXT: s_addc_u32 s1, 0, s14
+; GFX6-NEXT: s_add_u32 s14, s13, s0
+; GFX6-NEXT: s_addc_u32 s15, s12, s1
; GFX6-NEXT: s_ashr_i32 s12, s7, 31
; GFX6-NEXT: s_add_u32 s0, s6, s12
; GFX6-NEXT: s_mov_b32 s13, s12
; GFX6-NEXT: s_addc_u32 s1, s7, s12
; GFX6-NEXT: s_xor_b64 s[6:7], s[0:1], s[12:13]
-; GFX6-NEXT: v_mov_b32_e32 v0, s14
+; GFX6-NEXT: v_mov_b32_e32 v0, s15
; GFX6-NEXT: v_mul_hi_u32 v1, s6, v0
-; GFX6-NEXT: v_mov_b32_e32 v2, s15
+; GFX6-NEXT: v_mov_b32_e32 v2, s14
; GFX6-NEXT: v_mul_hi_u32 v3, s6, v2
; GFX6-NEXT: s_mov_b32 s0, s4
; GFX6-NEXT: v_readfirstlane_b32 s4, v1
; GFX6-NEXT: v_mul_hi_u32 v1, s7, v2
-; GFX6-NEXT: s_mul_i32 s1, s6, s14
+; GFX6-NEXT: s_mul_i32 s1, s6, s15
; GFX6-NEXT: v_readfirstlane_b32 s16, v3
; GFX6-NEXT: v_mul_hi_u32 v0, s7, v0
; GFX6-NEXT: s_add_u32 s1, s16, s1
; GFX6-NEXT: s_addc_u32 s4, 0, s4
-; GFX6-NEXT: s_mul_i32 s15, s7, s15
+; GFX6-NEXT: s_mul_i32 s14, s7, s14
; GFX6-NEXT: v_readfirstlane_b32 s16, v1
-; GFX6-NEXT: s_add_u32 s1, s1, s15
+; GFX6-NEXT: s_add_u32 s1, s1, s14
; GFX6-NEXT: s_addc_u32 s1, s4, s16
; GFX6-NEXT: v_readfirstlane_b32 s4, v0
; GFX6-NEXT: s_addc_u32 s4, s4, 0
-; GFX6-NEXT: s_mul_i32 s14, s7, s14
-; GFX6-NEXT: s_add_u32 s16, s1, s14
-; GFX6-NEXT: v_mov_b32_e32 v0, s16
+; GFX6-NEXT: s_mul_i32 s14, s7, s15
+; GFX6-NEXT: s_add_u32 s14, s1, s14
+; GFX6-NEXT: v_mov_b32_e32 v0, s14
; GFX6-NEXT: v_mul_hi_u32 v0, s10, v0
-; GFX6-NEXT: s_addc_u32 s17, 0, s4
+; GFX6-NEXT: s_addc_u32 s15, 0, s4
; GFX6-NEXT: s_mov_b32 s1, s5
-; GFX6-NEXT: s_mul_i32 s4, s10, s17
+; GFX6-NEXT: s_mul_i32 s4, s10, s15
; GFX6-NEXT: v_readfirstlane_b32 s5, v0
; GFX6-NEXT: s_add_i32 s4, s5, s4
-; GFX6-NEXT: s_mul_i32 s5, s11, s16
-; GFX6-NEXT: s_add_i32 s18, s4, s5
-; GFX6-NEXT: s_sub_i32 s14, s7, s18
-; GFX6-NEXT: s_mul_i32 s4, s10, s16
+; GFX6-NEXT: s_mul_i32 s5, s11, s14
+; GFX6-NEXT: s_add_i32 s16, s4, s5
+; GFX6-NEXT: s_sub_i32 s17, s7, s16
+; GFX6-NEXT: s_mul_i32 s4, s10, s14
; GFX6-NEXT: s_sub_u32 s6, s6, s4
; GFX6-NEXT: s_cselect_b64 s[4:5], -1, 0
-; GFX6-NEXT: s_or_b32 s15, s4, s5
-; GFX6-NEXT: s_subb_u32 s19, s14, s11
-; GFX6-NEXT: s_sub_u32 s20, s6, s10
-; GFX6-NEXT: s_cselect_b64 s[14:15], -1, 0
-; GFX6-NEXT: s_or_b32 s14, s14, s15
-; GFX6-NEXT: s_subb_u32 s14, s19, 0
-; GFX6-NEXT: s_cmp_ge_u32 s14, s11
-; GFX6-NEXT: s_cselect_b32 s15, -1, 0
-; GFX6-NEXT: s_cmp_ge_u32 s20, s10
+; GFX6-NEXT: s_subb_u32 s17, s17, s11
+; GFX6-NEXT: s_sub_u32 s18, s6, s10
+; GFX6-NEXT: s_subb_u32 s17, s17, 0
+; GFX6-NEXT: s_cmp_ge_u32 s17, s11
; GFX6-NEXT: s_cselect_b32 s19, -1, 0
-; GFX6-NEXT: s_cmp_eq_u32 s14, s11
-; GFX6-NEXT: s_cselect_b32 s14, s19, s15
-; GFX6-NEXT: s_add_u32 s15, s16, 1
-; GFX6-NEXT: s_addc_u32 s19, s17, 0
-; GFX6-NEXT: s_add_u32 s20, s16, 2
-; GFX6-NEXT: s_addc_u32 s21, s17, 0
-; GFX6-NEXT: s_cmp_lg_u32 s14, 0
-; GFX6-NEXT: s_cselect_b32 s14, s20, s15
-; GFX6-NEXT: s_cselect_b32 s15, s21, s19
+; GFX6-NEXT: s_cmp_ge_u32 s18, s10
+; GFX6-NEXT: s_cselect_b32 s18, -1, 0
+; GFX6-NEXT: s_cmp_eq_u32 s17, s11
+; GFX6-NEXT: s_cselect_b32 s17, s18, s19
+; GFX6-NEXT: s_add_u32 s18, s14, 1
+; GFX6-NEXT: s_addc_u32 s19, s15, 0
+; GFX6-NEXT: s_add_u32 s20, s14, 2
+; GFX6-NEXT: s_addc_u32 s21, s15, 0
+; GFX6-NEXT: s_cmp_lg_u32 s17, 0
+; GFX6-NEXT: s_cselect_b32 s17, s20, s18
+; GFX6-NEXT: s_cselect_b32 s18, s21, s19
; GFX6-NEXT: s_or_b32 s4, s4, s5
-; GFX6-NEXT: s_subb_u32 s4, s7, s18
+; GFX6-NEXT: s_subb_u32 s4, s7, s16
; GFX6-NEXT: s_cmp_ge_u32 s4, s11
; GFX6-NEXT: s_cselect_b32 s5, -1, 0
; GFX6-NEXT: s_cmp_ge_u32 s6, s10
@@ -7921,13 +7913,14 @@ define amdgpu_kernel void @sdiv_i64_pow2_shl_denom(ptr addrspace(1) %out, i64 %x
; GFX6-NEXT: s_cmp_eq_u32 s4, s11
; GFX6-NEXT: s_cselect_b32 s4, s6, s5
; GFX6-NEXT: s_cmp_lg_u32 s4, 0
-; GFX6-NEXT: s_cselect_b32 s5, s15, s17
-; GFX6-NEXT: s_cselect_b32 s4, s14, s16
+; GFX6-NEXT: s_cselect_b32 s5, s18, s15
+; GFX6-NEXT: s_cselect_b32 s4, s17, s14
; GFX6-NEXT: s_xor_b64 s[6:7], s[12:13], s[8:9]
; GFX6-NEXT: s_xor_b64 s[4:5], s[4:5], s[6:7]
; GFX6-NEXT: s_sub_u32 s4, s4, s6
; GFX6-NEXT: s_subb_u32 s5, s5, s7
; GFX6-NEXT: v_mov_b32_e32 v0, s4
+; GFX6-NEXT: s_mov_b32 s2, -1
; GFX6-NEXT: v_mov_b32_e32 v1, s5
; GFX6-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; GFX6-NEXT: s_endpgm
@@ -8278,8 +8271,8 @@ define amdgpu_kernel void @sdiv_v2i64_pow2_shl_denom(ptr addrspace(1) %out, <2 x
; GFX6-NEXT: s_xor_b64 s[6:7], s[6:7], s[2:3]
; GFX6-NEXT: v_cvt_f32_u32_e32 v0, s6
; GFX6-NEXT: v_cvt_f32_u32_e32 v1, s7
-; GFX6-NEXT: s_sub_u32 s14, 0, s6
-; GFX6-NEXT: s_subb_u32 s15, 0, s7
+; GFX6-NEXT: s_sub_u32 s12, 0, s6
+; GFX6-NEXT: s_subb_u32 s13, 0, s7
; GFX6-NEXT: v_mac_f32_e32 v0, 0x4f800000, v1
; GFX6-NEXT: v_rcp_f32_e32 v0, v0
; GFX6-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0
@@ -8288,69 +8281,65 @@ define amdgpu_kernel void @sdiv_v2i64_pow2_shl_denom(ptr addrspace(1) %out, <2 x
; GFX6-NEXT: v_mac_f32_e32 v0, 0xcf800000, v1
; GFX6-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX6-NEXT: v_cvt_u32_f32_e32 v1, v1
-; GFX6-NEXT: v_mul_hi_u32 v2, s14, v0
-; GFX6-NEXT: v_readfirstlane_b32 s16, v1
-; GFX6-NEXT: v_readfirstlane_b32 s12, v0
-; GFX6-NEXT: s_mul_i32 s13, s14, s16
+; GFX6-NEXT: v_mul_hi_u32 v2, s12, v0
+; GFX6-NEXT: v_readfirstlane_b32 s14, v1
+; GFX6-NEXT: v_readfirstlane_b32 s15, v0
+; GFX6-NEXT: s_mul_i32 s16, s12, s14
; GFX6-NEXT: v_readfirstlane_b32 s19, v2
-; GFX6-NEXT: s_mul_i32 s17, s15, s12
-; GFX6-NEXT: s_mul_i32 s18, s14, s12
-; GFX6-NEXT: s_add_i32 s13, s19, s13
+; GFX6-NEXT: s_mul_i32 s17, s13, s15
+; GFX6-NEXT: s_mul_i32 s18, s12, s15
+; GFX6-NEXT: s_add_i32 s16, s19, s16
; GFX6-NEXT: v_mul_hi_u32 v3, v0, s18
-; GFX6-NEXT: s_add_i32 s13, s13, s17
-; GFX6-NEXT: v_mul_hi_u32 v0, v0, s13
+; GFX6-NEXT: s_add_i32 s16, s16, s17
+; GFX6-NEXT: v_mul_hi_u32 v0, v0, s16
; GFX6-NEXT: v_mul_hi_u32 v4, v1, s18
; GFX6-NEXT: v_readfirstlane_b32 s17, v3
-; GFX6-NEXT: s_mul_i32 s20, s12, s13
-; GFX6-NEXT: v_mul_hi_u32 v1, v1, s13
+; GFX6-NEXT: s_mul_i32 s20, s15, s16
+; GFX6-NEXT: v_mul_hi_u32 v1, v1, s16
; GFX6-NEXT: s_add_u32 s17, s17, s20
; GFX6-NEXT: v_readfirstlane_b32 s20, v0
-; GFX6-NEXT: s_mul_i32 s18, s16, s18
+; GFX6-NEXT: s_mul_i32 s18, s14, s18
; GFX6-NEXT: s_addc_u32 s20, 0, s20
; GFX6-NEXT: v_readfirstlane_b32 s19, v4
; GFX6-NEXT: s_add_u32 s17, s17, s18
; GFX6-NEXT: s_addc_u32 s17, s20, s19
; GFX6-NEXT: v_readfirstlane_b32 s18, v1
; GFX6-NEXT: s_addc_u32 s18, s18, 0
-; GFX6-NEXT: s_mul_i32 s13, s16, s13
-; GFX6-NEXT: s_add_u32 s13, s17, s13
+; GFX6-NEXT: s_mul_i32 s16, s14, s16
+; GFX6-NEXT: s_add_u32 s16, s17, s16
; GFX6-NEXT: s_addc_u32 s17, 0, s18
-; GFX6-NEXT: s_add_u32 s18, s12, s13
-; GFX6-NEXT: v_mov_b32_e32 v0, s18
-; GFX6-NEXT: v_mul_hi_u32 v0, s14, v0
-; GFX6-NEXT: s_cselect_b64 s[12:13], -1, 0
-; GFX6-NEXT: s_or_b32 s12, s12, s13
-; GFX6-NEXT: s_addc_u32 s16, s16, s17
-; GFX6-NEXT: s_mul_i32 s12, s14, s16
-; GFX6-NEXT: v_readfirstlane_b32 s13, v0
-; GFX6-NEXT: s_add_i32 s12, s13, s12
-; GFX6-NEXT: s_mul_i32 s15, s15, s18
-; GFX6-NEXT: s_mul_i32 s13, s14, s18
-; GFX6-NEXT: s_add_i32 s12, s12, s15
-; GFX6-NEXT: v_mov_b32_e32 v2, s13
-; GFX6-NEXT: v_mov_b32_e32 v0, s12
-; GFX6-NEXT: v_mul_hi_u32 v3, s16, v2
-; GFX6-NEXT: v_mul_hi_u32 v2, s18, v2
-; GFX6-NEXT: v_mul_hi_u32 v1, s16, v0
-; GFX6-NEXT: v_mul_hi_u32 v0, s18, v0
-; GFX6-NEXT: s_mul_i32 s15, s18, s12
-; GFX6-NEXT: v_readfirstlane_b32 s19, v2
-; GFX6-NEXT: s_add_u32 s15, s19, s15
+; GFX6-NEXT: s_add_u32 s15, s15, s16
+; GFX6-NEXT: v_mov_b32_e32 v0, s15
+; GFX6-NEXT: v_mul_hi_u32 v0, s12, v0
+; GFX6-NEXT: s_addc_u32 s14, s14, s17
+; GFX6-NEXT: s_mul_i32 s16, s12, s14
+; GFX6-NEXT: s_mul_i32 s13, s13, s15
; GFX6-NEXT: v_readfirstlane_b32 s17, v0
-; GFX6-NEXT: s_mul_i32 s13, s16, s13
-; GFX6-NEXT: s_addc_u32 s17, 0, s17
-; GFX6-NEXT: v_readfirstl...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/165261
More information about the llvm-commits
mailing list