[llvm] [TaiDup] Allow large number of predecessors/successors without phis. (PR #116072)
via llvm-commits
llvm-commits at lists.llvm.org
Wed Nov 13 08:02:54 PST 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-x86
Author: Florian Hahn (fhahn)
<details>
<summary>Changes</summary>
This adjusts the threshold logic added in #<!-- -->78582 to only trigger for cases where there are actually phis to duplicate in either TailBB or in one of the successors.
In cases there are no phis, we only have to pay the cost of extra edges, but have no explosion in PHI related instructions.
This improves performance of Python on some inputs by 2-3% on Apple Silicon CPUs.
---
Full diff: https://github.com/llvm/llvm-project/pull/116072.diff
2 Files Affected:
- (modified) llvm/lib/CodeGen/TailDuplicator.cpp (+15-8)
- (modified) llvm/test/CodeGen/X86/tail-dup-pred-succ-size.mir (+32-28)
``````````diff
diff --git a/llvm/lib/CodeGen/TailDuplicator.cpp b/llvm/lib/CodeGen/TailDuplicator.cpp
index d0013923838b35..a333122b2ba035 100644
--- a/llvm/lib/CodeGen/TailDuplicator.cpp
+++ b/llvm/lib/CodeGen/TailDuplicator.cpp
@@ -573,14 +573,6 @@ bool TailDuplicator::shouldTailDuplicate(bool IsSimple,
if (TailBB.isSuccessor(&TailBB))
return false;
- // Duplicating a BB which has both multiple predecessors and successors will
- // result in a complex CFG and also may cause huge amount of PHI nodes. If we
- // want to remove this limitation, we have to address
- // https://github.com/llvm/llvm-project/issues/78578.
- if (TailBB.pred_size() > TailDupPredSize &&
- TailBB.succ_size() > TailDupSuccSize)
- return false;
-
// Set the limit on the cost to duplicate. When optimizing for size,
// duplicate only one, because one branch instruction can be eliminated to
// compensate for the duplication.
@@ -618,6 +610,7 @@ bool TailDuplicator::shouldTailDuplicate(bool IsSimple,
// Check the instructions in the block to determine whether tail-duplication
// is invalid or unlikely to be profitable.
unsigned InstrCount = 0;
+ unsigned NumPhis = 0;
for (MachineInstr &MI : TailBB) {
// Non-duplicable things shouldn't be tail-duplicated.
// CFI instructions are marked as non-duplicable, because Darwin compact
@@ -661,6 +654,20 @@ bool TailDuplicator::shouldTailDuplicate(bool IsSimple,
if (InstrCount > MaxDuplicateCount)
return false;
+ NumPhis += MI.isPHI();
+ }
+
+ // Duplicating a BB which has both multiple predecessors and successors will
+ // may cause huge amount of PHI nodes. If we want to remove this limitation,
+ // we have to address https://github.com/llvm/llvm-project/issues/78578.
+ if (TailBB.pred_size() > TailDupPredSize &&
+ TailBB.succ_size() > TailDupSuccSize) {
+ // If TailBB or any of its successors contains a phi, we may have to add a
+ // large number of additional phis with additional incoming values.
+ if (NumPhis != 0 || any_of(TailBB.successors(), [](MachineBasicBlock *MBB) {
+ return any_of(*MBB, [](MachineInstr &MI) { return MI.isPHI(); });
+ }))
+ return false;
}
// Check if any of the successors of TailBB has a PHI node in which the
diff --git a/llvm/test/CodeGen/X86/tail-dup-pred-succ-size.mir b/llvm/test/CodeGen/X86/tail-dup-pred-succ-size.mir
index 1d17672e2c6bd0..2f1ff76fda76c9 100644
--- a/llvm/test/CodeGen/X86/tail-dup-pred-succ-size.mir
+++ b/llvm/test/CodeGen/X86/tail-dup-pred-succ-size.mir
@@ -538,43 +538,47 @@ body: |
; LIMIT-NEXT: JMP64m $noreg, 8, [[SUBREG_TO_REG]], %jump-table.0, $noreg
; LIMIT-NEXT: {{ $}}
; LIMIT-NEXT: bb.2:
- ; LIMIT-NEXT: successors: %bb.7(0x80000000)
+ ; LIMIT-NEXT: successors: %bb.9(0x20000000), %bb.10(0x20000000), %bb.11(0x20000000), %bb.12(0x20000000)
; LIMIT-NEXT: {{ $}}
; LIMIT-NEXT: [[MOV32rm:%[0-9]+]]:gr32 = MOV32rm [[COPY1]], 1, $noreg, 0, $noreg
- ; LIMIT-NEXT: JMP_1 %bb.7
+ ; LIMIT-NEXT: [[SHR32ri1:%[0-9]+]]:gr32 = SHR32ri [[COPY]], 2, implicit-def dead $eflags
+ ; LIMIT-NEXT: [[AND32ri1:%[0-9]+]]:gr32 = AND32ri [[SHR32ri1]], 7, implicit-def dead $eflags
+ ; LIMIT-NEXT: [[SUBREG_TO_REG1:%[0-9]+]]:gr64_nosp = SUBREG_TO_REG 0, [[AND32ri1]], %subreg.sub_32bit
+ ; LIMIT-NEXT: JMP64m $noreg, 8, [[SUBREG_TO_REG1]], %jump-table.1, $noreg
; LIMIT-NEXT: {{ $}}
; LIMIT-NEXT: bb.3:
- ; LIMIT-NEXT: successors: %bb.7(0x80000000)
+ ; LIMIT-NEXT: successors: %bb.9(0x20000000), %bb.10(0x20000000), %bb.11(0x20000000), %bb.12(0x20000000)
; LIMIT-NEXT: {{ $}}
; LIMIT-NEXT: [[MOV32rm1:%[0-9]+]]:gr32 = MOV32rm [[COPY1]], 1, $noreg, 0, $noreg
- ; LIMIT-NEXT: [[SHR32ri1:%[0-9]+]]:gr32 = SHR32ri [[MOV32rm1]], 1, implicit-def dead $eflags
- ; LIMIT-NEXT: JMP_1 %bb.7
+ ; LIMIT-NEXT: [[SHR32ri2:%[0-9]+]]:gr32 = SHR32ri [[MOV32rm1]], 1, implicit-def dead $eflags
+ ; LIMIT-NEXT: [[SHR32ri3:%[0-9]+]]:gr32 = SHR32ri [[COPY]], 2, implicit-def dead $eflags
+ ; LIMIT-NEXT: [[AND32ri2:%[0-9]+]]:gr32 = AND32ri [[SHR32ri3]], 7, implicit-def dead $eflags
+ ; LIMIT-NEXT: [[SUBREG_TO_REG2:%[0-9]+]]:gr64_nosp = SUBREG_TO_REG 0, [[AND32ri2]], %subreg.sub_32bit
+ ; LIMIT-NEXT: JMP64m $noreg, 8, [[SUBREG_TO_REG2]], %jump-table.1, $noreg
; LIMIT-NEXT: {{ $}}
; LIMIT-NEXT: bb.4:
- ; LIMIT-NEXT: successors: %bb.7(0x80000000)
+ ; LIMIT-NEXT: successors: %bb.9(0x20000000), %bb.10(0x20000000), %bb.11(0x20000000), %bb.12(0x20000000)
; LIMIT-NEXT: {{ $}}
; LIMIT-NEXT: [[MOV32rm2:%[0-9]+]]:gr32 = MOV32rm [[COPY1]], 1, $noreg, 0, $noreg
- ; LIMIT-NEXT: [[SHR32ri2:%[0-9]+]]:gr32 = SHR32ri [[MOV32rm2]], 2, implicit-def dead $eflags
- ; LIMIT-NEXT: JMP_1 %bb.7
+ ; LIMIT-NEXT: [[SHR32ri4:%[0-9]+]]:gr32 = SHR32ri [[MOV32rm2]], 2, implicit-def dead $eflags
+ ; LIMIT-NEXT: [[SHR32ri5:%[0-9]+]]:gr32 = SHR32ri [[COPY]], 2, implicit-def dead $eflags
+ ; LIMIT-NEXT: [[AND32ri3:%[0-9]+]]:gr32 = AND32ri [[SHR32ri5]], 7, implicit-def dead $eflags
+ ; LIMIT-NEXT: [[SUBREG_TO_REG3:%[0-9]+]]:gr64_nosp = SUBREG_TO_REG 0, [[AND32ri3]], %subreg.sub_32bit
+ ; LIMIT-NEXT: JMP64m $noreg, 8, [[SUBREG_TO_REG3]], %jump-table.1, $noreg
; LIMIT-NEXT: {{ $}}
; LIMIT-NEXT: bb.5:
- ; LIMIT-NEXT: successors: %bb.7(0x80000000)
+ ; LIMIT-NEXT: successors: %bb.9(0x20000000), %bb.10(0x20000000), %bb.11(0x20000000), %bb.12(0x20000000)
; LIMIT-NEXT: {{ $}}
; LIMIT-NEXT: [[MOV32rm3:%[0-9]+]]:gr32 = MOV32rm [[COPY1]], 1, $noreg, 0, $noreg
- ; LIMIT-NEXT: [[SHR32ri3:%[0-9]+]]:gr32 = SHR32ri [[MOV32rm3]], 3, implicit-def dead $eflags
- ; LIMIT-NEXT: JMP_1 %bb.7
+ ; LIMIT-NEXT: [[SHR32ri6:%[0-9]+]]:gr32 = SHR32ri [[MOV32rm3]], 3, implicit-def dead $eflags
+ ; LIMIT-NEXT: [[SHR32ri7:%[0-9]+]]:gr32 = SHR32ri [[COPY]], 2, implicit-def dead $eflags
+ ; LIMIT-NEXT: [[AND32ri4:%[0-9]+]]:gr32 = AND32ri [[SHR32ri7]], 7, implicit-def dead $eflags
+ ; LIMIT-NEXT: [[SUBREG_TO_REG4:%[0-9]+]]:gr64_nosp = SUBREG_TO_REG 0, [[AND32ri4]], %subreg.sub_32bit
+ ; LIMIT-NEXT: JMP64m $noreg, 8, [[SUBREG_TO_REG4]], %jump-table.1, $noreg
; LIMIT-NEXT: {{ $}}
; LIMIT-NEXT: bb.6:
; LIMIT-NEXT: successors:
; LIMIT-NEXT: {{ $}}
- ; LIMIT-NEXT: bb.7:
- ; LIMIT-NEXT: successors: %bb.9(0x20000000), %bb.10(0x20000000), %bb.11(0x20000000), %bb.12(0x20000000)
- ; LIMIT-NEXT: {{ $}}
- ; LIMIT-NEXT: [[SHR32ri4:%[0-9]+]]:gr32 = SHR32ri [[COPY]], 2, implicit-def dead $eflags
- ; LIMIT-NEXT: [[AND32ri1:%[0-9]+]]:gr32 = AND32ri [[SHR32ri4]], 7, implicit-def dead $eflags
- ; LIMIT-NEXT: [[SUBREG_TO_REG1:%[0-9]+]]:gr64_nosp = SUBREG_TO_REG 0, killed [[AND32ri1]], %subreg.sub_32bit
- ; LIMIT-NEXT: JMP64m $noreg, 8, [[SUBREG_TO_REG1]], %jump-table.1, $noreg
- ; LIMIT-NEXT: {{ $}}
; LIMIT-NEXT: bb.9:
; LIMIT-NEXT: [[MOV32rm4:%[0-9]+]]:gr32 = MOV32rm [[COPY1]], 1, $noreg, 0, $noreg
; LIMIT-NEXT: MOV32mr [[COPY1]], 1, $noreg, 0, $noreg, [[MOV32rm4]] :: (store (s32))
@@ -583,23 +587,23 @@ body: |
; LIMIT-NEXT: {{ $}}
; LIMIT-NEXT: bb.10:
; LIMIT-NEXT: [[MOV32rm5:%[0-9]+]]:gr32 = MOV32rm [[COPY1]], 1, $noreg, 0, $noreg
- ; LIMIT-NEXT: [[SHR32ri5:%[0-9]+]]:gr32 = SHR32ri [[MOV32rm5]], 1, implicit-def dead $eflags
- ; LIMIT-NEXT: MOV32mr [[COPY1]], 1, $noreg, 0, $noreg, [[SHR32ri5]] :: (store (s32))
- ; LIMIT-NEXT: $eax = COPY [[SHR32ri5]]
+ ; LIMIT-NEXT: [[SHR32ri8:%[0-9]+]]:gr32 = SHR32ri [[MOV32rm5]], 1, implicit-def dead $eflags
+ ; LIMIT-NEXT: MOV32mr [[COPY1]], 1, $noreg, 0, $noreg, [[SHR32ri8]] :: (store (s32))
+ ; LIMIT-NEXT: $eax = COPY [[SHR32ri8]]
; LIMIT-NEXT: RET 0, $eax
; LIMIT-NEXT: {{ $}}
; LIMIT-NEXT: bb.11:
; LIMIT-NEXT: [[MOV32rm6:%[0-9]+]]:gr32 = MOV32rm [[COPY1]], 1, $noreg, 0, $noreg
- ; LIMIT-NEXT: [[SHR32ri6:%[0-9]+]]:gr32 = SHR32ri [[MOV32rm6]], 2, implicit-def dead $eflags
- ; LIMIT-NEXT: MOV32mr [[COPY1]], 1, $noreg, 0, $noreg, [[SHR32ri6]] :: (store (s32))
- ; LIMIT-NEXT: $eax = COPY [[SHR32ri6]]
+ ; LIMIT-NEXT: [[SHR32ri9:%[0-9]+]]:gr32 = SHR32ri [[MOV32rm6]], 2, implicit-def dead $eflags
+ ; LIMIT-NEXT: MOV32mr [[COPY1]], 1, $noreg, 0, $noreg, [[SHR32ri9]] :: (store (s32))
+ ; LIMIT-NEXT: $eax = COPY [[SHR32ri9]]
; LIMIT-NEXT: RET 0, $eax
; LIMIT-NEXT: {{ $}}
; LIMIT-NEXT: bb.12:
; LIMIT-NEXT: [[MOV32rm7:%[0-9]+]]:gr32 = MOV32rm [[COPY1]], 1, $noreg, 0, $noreg
- ; LIMIT-NEXT: [[SHR32ri7:%[0-9]+]]:gr32 = SHR32ri [[MOV32rm7]], 6, implicit-def dead $eflags
- ; LIMIT-NEXT: MOV32mr [[COPY1]], 1, $noreg, 0, $noreg, [[SHR32ri7]] :: (store (s32))
- ; LIMIT-NEXT: $eax = COPY [[SHR32ri7]]
+ ; LIMIT-NEXT: [[SHR32ri10:%[0-9]+]]:gr32 = SHR32ri [[MOV32rm7]], 6, implicit-def dead $eflags
+ ; LIMIT-NEXT: MOV32mr [[COPY1]], 1, $noreg, 0, $noreg, [[SHR32ri10]] :: (store (s32))
+ ; LIMIT-NEXT: $eax = COPY [[SHR32ri10]]
; LIMIT-NEXT: RET 0, $eax
;
; NOLIMIT-LABEL: name: foo_no_phis
``````````
</details>
https://github.com/llvm/llvm-project/pull/116072
More information about the llvm-commits
mailing list