[CombinerHelper]: Use undef for handling divisors of one #91037
Closed
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-backend-aarch64 @llvm/pr-subscribers-llvm-globalisel Author: AtariDreams (AtariDreams) ChangesFull diff: https://github.com/llvm/llvm-project/pull/91037.diff 3 Files Affected:
diff --git a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
index 22eb4a3e0d7cb..644f56b6bf276 100644
--- a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
@@ -5099,6 +5099,8 @@ MachineInstr *CombinerHelper::buildUDivUsingMul(MachineInstr &MI) {
auto &MIB = Builder;
bool UseNPQ = false;
+ bool UsePreShift = false;
+ bool UsePostShift = false;
SmallVector<Register, 16> PreShifts, PostShifts, MagicFactors, NPQFactors;
auto BuildUDIVPattern = [&](const Constant *C) {
@@ -5111,27 +5113,30 @@ MachineInstr *CombinerHelper::buildUDivUsingMul(MachineInstr &MI) {
// Magic algorithm doesn't work for division by 1. We need to emit a select
// at the end.
- // TODO: Use undef values for divisor of 1.
- if (!Divisor.isOne()) {
-
- // UnsignedDivisionByConstantInfo doesn't work correctly if leading zeros
- // in the dividend exceeds the leading zeros for the divisor.
- UnsignedDivisionByConstantInfo magics =
- UnsignedDivisionByConstantInfo::get(
- Divisor, std::min(KnownLeadingZeros, Divisor.countl_zero()));
-
- Magic = std::move(magics.Magic);
-
- assert(magics.PreShift < Divisor.getBitWidth() &&
- "We shouldn't generate an undefined shift!");
- assert(magics.PostShift < Divisor.getBitWidth() &&
- "We shouldn't generate an undefined shift!");
- assert((!magics.IsAdd || magics.PreShift == 0) && "Unexpected pre-shift");
- PreShift = magics.PreShift;
- PostShift = magics.PostShift;
- SelNPQ = magics.IsAdd;
+ if (Divisor.isOne()) {
+ PreShifts.push_back(MIB.buildUndef(ScalarShiftAmtTy).getReg(0));
+ MagicFactors.push_back(MIB.buildUndef(ScalarTy).getReg(0));
+ NPQFactors.push_back(MIB.buildUndef(ScalarTy).getReg(0));
+ PostShifts.push_back(MIB.buildUndef(ScalarShiftAmtTy).getReg(0));
+ return true;
}
+ // UnsignedDivisionByConstantInfo doesn't work correctly if leading zeros
+ // in the dividend exceeds the leading zeros for the divisor.
+ UnsignedDivisionByConstantInfo magics = UnsignedDivisionByConstantInfo::get(
+ Divisor, std::min(KnownLeadingZeros, Divisor.countl_zero()));
+
+ Magic = std::move(magics.Magic);
+
+ assert(magics.PreShift < Divisor.getBitWidth() &&
+ "We shouldn't generate an undefined shift!");
+ assert(magics.PostShift < Divisor.getBitWidth() &&
+ "We shouldn't generate an undefined shift!");
+ assert((!magics.IsAdd || magics.PreShift == 0) && "Unexpected pre-shift");
+ PreShift = magics.PreShift;
+ PostShift = magics.PostShift;
+ SelNPQ = magics.IsAdd;
+
PreShifts.push_back(
MIB.buildConstant(ScalarShiftAmtTy, PreShift).getReg(0));
MagicFactors.push_back(MIB.buildConstant(ScalarTy, Magic).getReg(0));
@@ -5143,6 +5148,8 @@ MachineInstr *CombinerHelper::buildUDivUsingMul(MachineInstr &MI) {
PostShifts.push_back(
MIB.buildConstant(ScalarShiftAmtTy, PostShift).getReg(0));
UseNPQ |= SelNPQ;
+ UsePreShift |= PreShift != 0;
+ UsePostShift |= magics.PostShift != 0;
return true;
};
@@ -5167,7 +5174,9 @@ MachineInstr *CombinerHelper::buildUDivUsingMul(MachineInstr &MI) {
}
Register Q = LHS;
- Q = MIB.buildLShr(Ty, Q, PreShift).getReg(0);
+
+ if (UsePreShift)
+ Q = MIB.buildLShr(Ty, Q, PreShift).getReg(0);
// Multiply the numerator (operand 0) by the magic value.
Q = MIB.buildUMulH(Ty, Q, MagicFactor).getReg(0);
@@ -5185,7 +5194,8 @@ MachineInstr *CombinerHelper::buildUDivUsingMul(MachineInstr &MI) {
Q = MIB.buildAdd(Ty, NPQ, Q).getReg(0);
}
- Q = MIB.buildLShr(Ty, Q, PostShift).getReg(0);
+ if (UsePostShift)
+ Q = MIB.buildLShr(Ty, Q, PostShift).getReg(0);
auto One = MIB.buildConstant(Ty, 1);
auto IsOne = MIB.buildICmp(
CmpInst::Predicate::ICMP_EQ,
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.ll b/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.ll
index c97a00ccdd455..ab2f2cb5f12a9 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.ll
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.ll
@@ -169,17 +169,17 @@ define <16 x i8> @combine_vec_udiv_nonuniform4(<16 x i8> %x) {
;
; GISEL-LABEL: combine_vec_udiv_nonuniform4:
; GISEL: // %bb.0:
-; GISEL-NEXT: adrp x8, .LCPI4_2
-; GISEL-NEXT: adrp x9, .LCPI4_0
-; GISEL-NEXT: ldr q1, [x8, :lo12:.LCPI4_2]
-; GISEL-NEXT: adrp x8, .LCPI4_1
-; GISEL-NEXT: ldr q4, [x9, :lo12:.LCPI4_0]
-; GISEL-NEXT: ldr q3, [x8, :lo12:.LCPI4_1]
+; GISEL-NEXT: mov w8, #171 // =0xab
+; GISEL-NEXT: fmov s1, w8
+; GISEL-NEXT: adrp x8, .LCPI4_0
+; GISEL-NEXT: ldr q3, [x8, :lo12:.LCPI4_0]
+; GISEL-NEXT: mov w8, #7 // =0x7
; GISEL-NEXT: umull2 v2.8h, v0.16b, v1.16b
; GISEL-NEXT: umull v1.8h, v0.8b, v1.8b
+; GISEL-NEXT: shl v3.16b, v3.16b, #7
; GISEL-NEXT: uzp2 v1.16b, v1.16b, v2.16b
-; GISEL-NEXT: neg v2.16b, v3.16b
-; GISEL-NEXT: shl v3.16b, v4.16b, #7
+; GISEL-NEXT: fmov s2, w8
+; GISEL-NEXT: neg v2.16b, v2.16b
; GISEL-NEXT: ushl v1.16b, v1.16b, v2.16b
; GISEL-NEXT: sshr v2.16b, v3.16b, #7
; GISEL-NEXT: bif v0.16b, v1.16b, v2.16b
@@ -217,25 +217,66 @@ define <8 x i16> @pr38477(<8 x i16> %a0) {
;
; GISEL-LABEL: pr38477:
; GISEL: // %bb.0:
-; GISEL-NEXT: adrp x8, .LCPI5_3
-; GISEL-NEXT: ldr q1, [x8, :lo12:.LCPI5_3]
-; GISEL-NEXT: adrp x8, .LCPI5_2
-; GISEL-NEXT: ldr q3, [x8, :lo12:.LCPI5_2]
-; GISEL-NEXT: adrp x8, .LCPI5_0
-; GISEL-NEXT: umull2 v2.4s, v0.8h, v1.8h
+; GISEL-NEXT: mov w8, #4957 // =0x135d
+; GISEL-NEXT: mov w9, #16385 // =0x4001
+; GISEL-NEXT: fmov s1, w8
+; GISEL-NEXT: mov w8, #57457 // =0xe071
+; GISEL-NEXT: fmov s4, w9
+; GISEL-NEXT: fmov s2, w8
+; GISEL-NEXT: mov w8, #4103 // =0x1007
+; GISEL-NEXT: mov w9, #35545 // =0x8ad9
+; GISEL-NEXT: fmov s5, w9
+; GISEL-NEXT: mov w9, #2048 // =0x800
+; GISEL-NEXT: mov v1.h[1], v1.h[0]
+; GISEL-NEXT: fmov s6, w9
+; GISEL-NEXT: adrp x9, .LCPI5_0
+; GISEL-NEXT: mov v1.h[2], v2.h[0]
+; GISEL-NEXT: fmov s2, w8
+; GISEL-NEXT: mov w8, #32768 // =0x8000
+; GISEL-NEXT: fmov s3, w8
+; GISEL-NEXT: mov w8, #0 // =0x0
+; GISEL-NEXT: mov v1.h[3], v2.h[0]
+; GISEL-NEXT: mov v2.h[1], v3.h[0]
+; GISEL-NEXT: mov v1.h[4], v4.h[0]
+; GISEL-NEXT: fmov s4, w8
+; GISEL-NEXT: mov w8, #6 // =0x6
+; GISEL-NEXT: mov v2.h[2], v4.h[0]
+; GISEL-NEXT: mov v1.h[5], v5.h[0]
+; GISEL-NEXT: fmov s5, w8
+; GISEL-NEXT: mov w8, #2115 // =0x843
+; GISEL-NEXT: mov v2.h[3], v4.h[0]
+; GISEL-NEXT: mov v7.h[1], v5.h[0]
+; GISEL-NEXT: mov v1.h[6], v6.h[0]
+; GISEL-NEXT: fmov s6, w8
+; GISEL-NEXT: mov w8, #12 // =0xc
+; GISEL-NEXT: mov v2.h[4], v4.h[0]
+; GISEL-NEXT: mov v7.h[2], v5.h[0]
+; GISEL-NEXT: mov v1.h[7], v6.h[0]
+; GISEL-NEXT: fmov s6, w8
+; GISEL-NEXT: mov w8, #14 // =0xe
+; GISEL-NEXT: fmov s16, w8
+; GISEL-NEXT: mov w8, #4 // =0x4
+; GISEL-NEXT: mov v2.h[5], v4.h[0]
+; GISEL-NEXT: mov v7.h[3], v6.h[0]
+; GISEL-NEXT: umull2 v6.4s, v0.8h, v1.8h
; GISEL-NEXT: umull v1.4s, v0.4h, v1.4h
-; GISEL-NEXT: uzp2 v1.8h, v1.8h, v2.8h
-; GISEL-NEXT: sub v2.8h, v0.8h, v1.8h
-; GISEL-NEXT: umull2 v4.4s, v2.8h, v3.8h
-; GISEL-NEXT: umull v2.4s, v2.4h, v3.4h
-; GISEL-NEXT: ldr d3, [x8, :lo12:.LCPI5_0]
-; GISEL-NEXT: adrp x8, .LCPI5_1
-; GISEL-NEXT: ushll v3.8h, v3.8b, #0
+; GISEL-NEXT: mov v2.h[6], v4.h[0]
+; GISEL-NEXT: mov v7.h[4], v16.h[0]
+; GISEL-NEXT: uzp2 v1.8h, v1.8h, v6.8h
+; GISEL-NEXT: mov v2.h[7], v3.h[0]
+; GISEL-NEXT: mov v7.h[5], v5.h[0]
+; GISEL-NEXT: ldr d5, [x9, :lo12:.LCPI5_0]
+; GISEL-NEXT: sub v3.8h, v0.8h, v1.8h
+; GISEL-NEXT: mov v7.h[6], v4.h[0]
+; GISEL-NEXT: umull2 v4.4s, v3.8h, v2.8h
+; GISEL-NEXT: umull v2.4s, v3.4h, v2.4h
+; GISEL-NEXT: fmov s3, w8
+; GISEL-NEXT: mov v7.h[7], v3.h[0]
; GISEL-NEXT: uzp2 v2.8h, v2.8h, v4.8h
-; GISEL-NEXT: ldr q4, [x8, :lo12:.LCPI5_1]
+; GISEL-NEXT: ushll v3.8h, v5.8b, #0
; GISEL-NEXT: shl v3.8h, v3.8h, #15
; GISEL-NEXT: add v1.8h, v2.8h, v1.8h
-; GISEL-NEXT: neg v2.8h, v4.8h
+; GISEL-NEXT: neg v2.8h, v7.8h
; GISEL-NEXT: ushl v1.8h, v1.8h, v2.8h
; GISEL-NEXT: sshr v2.8h, v3.8h, #15
; GISEL-NEXT: bif v0.16b, v1.16b, v2.16b
@@ -257,10 +298,10 @@ define i32 @udiv_div_by_180(i32 %x)
;
; GISEL-LABEL: udiv_div_by_180:
; GISEL: // %bb.0:
-; GISEL-NEXT: uxtb w8, w0
-; GISEL-NEXT: mov w9, #5826 // =0x16c2
-; GISEL-NEXT: movk w9, #364, lsl #16
-; GISEL-NEXT: umull x8, w8, w9
+; GISEL-NEXT: mov w8, #5826 // =0x16c2
+; GISEL-NEXT: and w9, w0, #0xff
+; GISEL-NEXT: movk w8, #364, lsl #16
+; GISEL-NEXT: umull x8, w9, w8
; GISEL-NEXT: lsr x0, x8, #32
; GISEL-NEXT: // kill: def $w0 killed $w0 killed $x0
; GISEL-NEXT: ret
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.mir b/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.mir
index 02233b9f498bd..11203925d1fe0 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.mir
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.mir
@@ -228,16 +228,16 @@ body: |
; CHECK: liveins: $q0
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: [[COPY:%[0-9]+]]:_(<16 x s8>) = COPY $q0
- ; CHECK-NEXT: [[C:%[0-9]+]]:_(s8) = G_CONSTANT i8 0
- ; CHECK-NEXT: [[C1:%[0-9]+]]:_(s8) = G_CONSTANT i8 -85
- ; CHECK-NEXT: [[C2:%[0-9]+]]:_(s8) = G_CONSTANT i8 7
- ; CHECK-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<16 x s8>) = G_BUILD_VECTOR [[C1]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8)
- ; CHECK-NEXT: [[BUILD_VECTOR1:%[0-9]+]]:_(<16 x s8>) = G_BUILD_VECTOR [[C2]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8), [[C]](s8)
+ ; CHECK-NEXT: [[C:%[0-9]+]]:_(s8) = G_CONSTANT i8 -85
+ ; CHECK-NEXT: [[C1:%[0-9]+]]:_(s8) = G_CONSTANT i8 7
+ ; CHECK-NEXT: [[DEF:%[0-9]+]]:_(s8) = G_IMPLICIT_DEF
+ ; CHECK-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<16 x s8>) = G_BUILD_VECTOR [[C]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8)
+ ; CHECK-NEXT: [[BUILD_VECTOR1:%[0-9]+]]:_(<16 x s8>) = G_BUILD_VECTOR [[C1]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8), [[DEF]](s8)
; CHECK-NEXT: [[UMULH:%[0-9]+]]:_(<16 x s8>) = G_UMULH [[COPY]], [[BUILD_VECTOR]]
; CHECK-NEXT: [[LSHR:%[0-9]+]]:_(<16 x s8>) = G_LSHR [[UMULH]], [[BUILD_VECTOR1]](<16 x s8>)
- ; CHECK-NEXT: [[C3:%[0-9]+]]:_(s1) = G_CONSTANT i1 false
- ; CHECK-NEXT: [[C4:%[0-9]+]]:_(s1) = G_CONSTANT i1 true
- ; CHECK-NEXT: [[BUILD_VECTOR2:%[0-9]+]]:_(<16 x s1>) = G_BUILD_VECTOR [[C3]](s1), [[C4]](s1), [[C4]](s1), [[C4]](s1), [[C4]](s1), [[C4]](s1), [[C4]](s1), [[C4]](s1), [[C4]](s1), [[C4]](s1), [[C4]](s1), [[C4]](s1), [[C4]](s1), [[C4]](s1), [[C4]](s1), [[C4]](s1)
+ ; CHECK-NEXT: [[C2:%[0-9]+]]:_(s1) = G_CONSTANT i1 false
+ ; CHECK-NEXT: [[C3:%[0-9]+]]:_(s1) = G_CONSTANT i1 true
+ ; CHECK-NEXT: [[BUILD_VECTOR2:%[0-9]+]]:_(<16 x s1>) = G_BUILD_VECTOR [[C2]](s1), [[C3]](s1), [[C3]](s1), [[C3]](s1), [[C3]](s1), [[C3]](s1), [[C3]](s1), [[C3]](s1), [[C3]](s1), [[C3]](s1), [[C3]](s1), [[C3]](s1), [[C3]](s1), [[C3]](s1), [[C3]](s1), [[C3]](s1)
; CHECK-NEXT: [[SELECT:%[0-9]+]]:_(<16 x s8>) = G_SELECT [[BUILD_VECTOR2]](<16 x s1>), [[COPY]], [[LSHR]]
; CHECK-NEXT: $q0 = COPY [[SELECT]](<16 x s8>)
; CHECK-NEXT: RET_ReallyLR implicit $q0
@@ -264,6 +264,7 @@ body: |
; CHECK: liveins: $q0
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: [[COPY:%[0-9]+]]:_(<8 x s16>) = COPY $q0
+ ; CHECK-NEXT: [[DEF:%[0-9]+]]:_(s16) = G_IMPLICIT_DEF
; CHECK-NEXT: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 0
; CHECK-NEXT: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 4957
; CHECK-NEXT: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 -32768
@@ -277,9 +278,9 @@ body: |
; CHECK-NEXT: [[C10:%[0-9]+]]:_(s16) = G_CONSTANT i16 2048
; CHECK-NEXT: [[C11:%[0-9]+]]:_(s16) = G_CONSTANT i16 2115
; CHECK-NEXT: [[C12:%[0-9]+]]:_(s16) = G_CONSTANT i16 4
- ; CHECK-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s16>) = G_BUILD_VECTOR [[C]](s16), [[C1]](s16), [[C4]](s16), [[C5]](s16), [[C7]](s16), [[C9]](s16), [[C10]](s16), [[C11]](s16)
- ; CHECK-NEXT: [[BUILD_VECTOR1:%[0-9]+]]:_(<8 x s16>) = G_BUILD_VECTOR [[C]](s16), [[C2]](s16), [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16), [[C2]](s16)
- ; CHECK-NEXT: [[BUILD_VECTOR2:%[0-9]+]]:_(<8 x s16>) = G_BUILD_VECTOR [[C]](s16), [[C3]](s16), [[C3]](s16), [[C6]](s16), [[C8]](s16), [[C3]](s16), [[C]](s16), [[C12]](s16)
+ ; CHECK-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s16>) = G_BUILD_VECTOR [[DEF]](s16), [[C1]](s16), [[C4]](s16), [[C5]](s16), [[C7]](s16), [[C9]](s16), [[C10]](s16), [[C11]](s16)
+ ; CHECK-NEXT: [[BUILD_VECTOR1:%[0-9]+]]:_(<8 x s16>) = G_BUILD_VECTOR [[DEF]](s16), [[C2]](s16), [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16), [[C2]](s16)
+ ; CHECK-NEXT: [[BUILD_VECTOR2:%[0-9]+]]:_(<8 x s16>) = G_BUILD_VECTOR [[DEF]](s16), [[C3]](s16), [[C3]](s16), [[C6]](s16), [[C8]](s16), [[C3]](s16), [[C]](s16), [[C12]](s16)
; CHECK-NEXT: [[UMULH:%[0-9]+]]:_(<8 x s16>) = G_UMULH [[COPY]], [[BUILD_VECTOR]]
; CHECK-NEXT: [[SUB:%[0-9]+]]:_(<8 x s16>) = G_SUB [[COPY]], [[UMULH]]
; CHECK-NEXT: [[UMULH1:%[0-9]+]]:_(<8 x s16>) = G_UMULH [[SUB]], [[BUILD_VECTOR1]]
|
arsenm
reviewed
May 20, 2024
Comment on lines
+5117
to
+5120
| PreShifts.push_back(MIB.buildUndef(ScalarShiftAmtTy).getReg(0)); | ||
| MagicFactors.push_back(MIB.buildUndef(ScalarTy).getReg(0)); | ||
| NPQFactors.push_back(MIB.buildUndef(ScalarTy).getReg(0)); | ||
| PostShifts.push_back(MIB.buildUndef(ScalarShiftAmtTy).getReg(0)); |
There was a problem hiding this comment.
call buildUndef once per type
| ; GISEL-NEXT: fmov s5, w9 | ||
| ; GISEL-NEXT: mov w9, #2048 // =0x800 | ||
| ; GISEL-NEXT: mov v1.h[1], v1.h[0] | ||
| ; GISEL-NEXT: fmov s6, w9 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.