[AArch64] Add patterns for sub from add negative immediates #156024
Merged
+67
−56
Conversation
This file contains hidden or 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 Author: David Green (davemgreen) Changes
The alternative is to add a isel combine, which seemed to work but created problems for mad and index patterns. This version does still need to add a lower-than-default Complexity to the ComplexPatterns to ensure that index was selected over sub-imm + index, as the default Complexity on ComplexPatterns is quite high. Fixes #155928 Full diff: https://github.com/llvm/llvm-project/pull/156024.diff 8 Files Affected:
diff --git a/llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp b/llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp
index bc786f415b554..42d1c1dd594f1 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp
@@ -246,9 +246,9 @@ class AArch64DAGToDAGISel : public SelectionDAGISel {
return false;
}
- template<MVT::SimpleValueType VT>
+ template<MVT::SimpleValueType VT, bool Negate>
bool SelectSVEAddSubImm(SDValue N, SDValue &Imm, SDValue &Shift) {
- return SelectSVEAddSubImm(N, VT, Imm, Shift);
+ return SelectSVEAddSubImm(N, VT, Imm, Shift, Negate);
}
template <MVT::SimpleValueType VT, bool Negate>
@@ -489,7 +489,7 @@ class AArch64DAGToDAGISel : public SelectionDAGISel {
bool SelectCMP_SWAP(SDNode *N);
- bool SelectSVEAddSubImm(SDValue N, MVT VT, SDValue &Imm, SDValue &Shift);
+ bool SelectSVEAddSubImm(SDValue N, MVT VT, SDValue &Imm, SDValue &Shift, bool Negate);
bool SelectSVEAddSubSSatImm(SDValue N, MVT VT, SDValue &Imm, SDValue &Shift,
bool Negate);
bool SelectSVECpyDupImm(SDValue N, MVT VT, SDValue &Imm, SDValue &Shift);
@@ -4227,35 +4227,36 @@ bool AArch64DAGToDAGISel::SelectCMP_SWAP(SDNode *N) {
}
bool AArch64DAGToDAGISel::SelectSVEAddSubImm(SDValue N, MVT VT, SDValue &Imm,
- SDValue &Shift) {
+ SDValue &Shift, bool Negate) {
if (!isa<ConstantSDNode>(N))
return false;
SDLoc DL(N);
- uint64_t Val = cast<ConstantSDNode>(N)
- ->getAPIntValue()
- .trunc(VT.getFixedSizeInBits())
- .getZExtValue();
+ APInt Val =
+ cast<ConstantSDNode>(N)->getAPIntValue().trunc(VT.getFixedSizeInBits());
+
+ if (Negate)
+ Val = -Val;
switch (VT.SimpleTy) {
case MVT::i8:
// All immediates are supported.
Shift = CurDAG->getTargetConstant(0, DL, MVT::i32);
- Imm = CurDAG->getTargetConstant(Val, DL, MVT::i32);
+ Imm = CurDAG->getTargetConstant(Val.getZExtValue(), DL, MVT::i32);
return true;
case MVT::i16:
case MVT::i32:
case MVT::i64:
// Support 8bit unsigned immediates.
- if (Val <= 255) {
+ if ((Val & ~0xff) == 0) {
Shift = CurDAG->getTargetConstant(0, DL, MVT::i32);
- Imm = CurDAG->getTargetConstant(Val, DL, MVT::i32);
+ Imm = CurDAG->getTargetConstant(Val.getZExtValue(), DL, MVT::i32);
return true;
}
// Support 16bit unsigned immediates that are a multiple of 256.
- if (Val <= 65280 && Val % 256 == 0) {
+ if ((Val & ~0xff00) == 0) {
Shift = CurDAG->getTargetConstant(8, DL, MVT::i32);
- Imm = CurDAG->getTargetConstant(Val >> 8, DL, MVT::i32);
+ Imm = CurDAG->getTargetConstant(Val.lshr(8).getZExtValue(), DL, MVT::i32);
return true;
}
break;
diff --git a/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td b/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
index eeb47b4d99750..58d8dcff4f34e 100644
--- a/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
+++ b/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
@@ -657,6 +657,12 @@ let Predicates = [HasSVE_or_SME] in {
defm SQSUB_ZI : sve_int_arith_imm0_ssat<0b110, "sqsub", ssubsat, saddsat>;
defm UQSUB_ZI : sve_int_arith_imm0<0b111, "uqsub", usubsat>;
+ // Extra patterns for add(x, splat(-ve)) -> sub(x, +ve). There is no i8
+ // pattern as all i8 constants can be handled by an add.
+ def : SVE_1_Op_Imm_OptLsl_Pat<nxv8i16, add, ZPR16, i32, SVEAddSubNegImm16Pat, SUB_ZI_H>;
+ def : SVE_1_Op_Imm_OptLsl_Pat<nxv4i32, add, ZPR32, i32, SVEAddSubNegImm32Pat, SUB_ZI_S>;
+ def : SVE_1_Op_Imm_OptLsl_Pat<nxv2i64, add, ZPR64, i64, SVEAddSubNegImm64Pat, SUB_ZI_D>;
+
defm MAD_ZPmZZ : sve_int_mladdsub_vvv_pred<0b0, "mad", AArch64mad_m1, "MLA_ZPmZZ", /*isReverseInstr*/ 1>;
defm MSB_ZPmZZ : sve_int_mladdsub_vvv_pred<0b1, "msb", AArch64msb_m1, "MLS_ZPmZZ", /*isReverseInstr*/ 1>;
defm MLA_ZPmZZ : sve_int_mlas_vvv_pred<0b0, "mla", AArch64mla_m1, "MLA_ZPZZZ", "MAD_ZPmZZ">;
diff --git a/llvm/lib/Target/AArch64/SVEInstrFormats.td b/llvm/lib/Target/AArch64/SVEInstrFormats.td
index a3a7d0f74e1bc..7389f90457234 100644
--- a/llvm/lib/Target/AArch64/SVEInstrFormats.td
+++ b/llvm/lib/Target/AArch64/SVEInstrFormats.td
@@ -315,10 +315,16 @@ def addsub_imm8_opt_lsl_i16 : imm8_opt_lsl<16, "uint16_t", SVEAddSubImmOperand16
def addsub_imm8_opt_lsl_i32 : imm8_opt_lsl<32, "uint32_t", SVEAddSubImmOperand32>;
def addsub_imm8_opt_lsl_i64 : imm8_opt_lsl<64, "uint64_t", SVEAddSubImmOperand64>;
-def SVEAddSubImm8Pat : ComplexPattern<i32, 2, "SelectSVEAddSubImm<MVT::i8>", []>;
-def SVEAddSubImm16Pat : ComplexPattern<i32, 2, "SelectSVEAddSubImm<MVT::i16>", []>;
-def SVEAddSubImm32Pat : ComplexPattern<i32, 2, "SelectSVEAddSubImm<MVT::i32>", []>;
-def SVEAddSubImm64Pat : ComplexPattern<i64, 2, "SelectSVEAddSubImm<MVT::i64>", []>;
+let Complexity = 1 in {
+def SVEAddSubImm8Pat : ComplexPattern<i32, 2, "SelectSVEAddSubImm<MVT::i8, false>", []>;
+def SVEAddSubImm16Pat : ComplexPattern<i32, 2, "SelectSVEAddSubImm<MVT::i16, false>", []>;
+def SVEAddSubImm32Pat : ComplexPattern<i32, 2, "SelectSVEAddSubImm<MVT::i32, false>", []>;
+def SVEAddSubImm64Pat : ComplexPattern<i64, 2, "SelectSVEAddSubImm<MVT::i64, false>", []>;
+
+def SVEAddSubNegImm8Pat : ComplexPattern<i32, 2, "SelectSVEAddSubImm<MVT::i8, true>", []>;
+def SVEAddSubNegImm16Pat : ComplexPattern<i32, 2, "SelectSVEAddSubImm<MVT::i16, true>", []>;
+def SVEAddSubNegImm32Pat : ComplexPattern<i32, 2, "SelectSVEAddSubImm<MVT::i32, true>", []>;
+def SVEAddSubNegImm64Pat : ComplexPattern<i64, 2, "SelectSVEAddSubImm<MVT::i64, true>", []>;
def SVEAddSubSSatNegImm8Pat : ComplexPattern<i32, 2, "SelectSVEAddSubSSatImm<MVT::i8, true>", []>;
def SVEAddSubSSatNegImm16Pat : ComplexPattern<i32, 2, "SelectSVEAddSubSSatImm<MVT::i16, true>", []>;
@@ -329,6 +335,7 @@ def SVEAddSubSSatPosImm8Pat : ComplexPattern<i32, 2, "SelectSVEAddSubSSatImm<MV
def SVEAddSubSSatPosImm16Pat : ComplexPattern<i32, 2, "SelectSVEAddSubSSatImm<MVT::i16, false>", []>;
def SVEAddSubSSatPosImm32Pat : ComplexPattern<i32, 2, "SelectSVEAddSubSSatImm<MVT::i32, false>", []>;
def SVEAddSubSSatPosImm64Pat : ComplexPattern<i64, 2, "SelectSVEAddSubSSatImm<MVT::i64, false>", []>;
+} // Complexity = 1
def SVECpyDupImm8Pat : ComplexPattern<i32, 2, "SelectSVECpyDupImm<MVT::i8>", []>;
def SVECpyDupImm16Pat : ComplexPattern<i32, 2, "SelectSVECpyDupImm<MVT::i16>", []>;
diff --git a/llvm/test/CodeGen/AArch64/sve-fixed-length-bitselect.ll b/llvm/test/CodeGen/AArch64/sve-fixed-length-bitselect.ll
index fb494afa11de2..258e399018ba8 100644
--- a/llvm/test/CodeGen/AArch64/sve-fixed-length-bitselect.ll
+++ b/llvm/test/CodeGen/AArch64/sve-fixed-length-bitselect.ll
@@ -13,15 +13,15 @@ define void @fixed_bitselect_v8i32(ptr %pre_cond_ptr, ptr %left_ptr, ptr %right_
; CHECK-LABEL: fixed_bitselect_v8i32:
; CHECK: // %bb.0:
; CHECK-NEXT: ptrue p0.s, vl8
-; CHECK-NEXT: mov z1.s, #-1 // =0xffffffffffffffff
; CHECK-NEXT: ld1w { z0.s }, p0/z, [x0]
-; CHECK-NEXT: ld1w { z2.s }, p0/z, [x1]
+; CHECK-NEXT: ld1w { z1.s }, p0/z, [x1]
; CHECK-NEXT: ld1w { z3.s }, p0/z, [x2]
-; CHECK-NEXT: add z1.s, z0.s, z1.s
-; CHECK-NEXT: subr z0.s, z0.s, #0 // =0x0
-; CHECK-NEXT: and z0.d, z0.d, z2.d
-; CHECK-NEXT: and z1.d, z1.d, z3.d
-; CHECK-NEXT: orr z0.d, z1.d, z0.d
+; CHECK-NEXT: mov z2.d, z0.d
+; CHECK-NEXT: sub z0.s, z0.s, #1 // =0x1
+; CHECK-NEXT: subr z2.s, z2.s, #0 // =0x0
+; CHECK-NEXT: and z0.d, z0.d, z3.d
+; CHECK-NEXT: and z1.d, z2.d, z1.d
+; CHECK-NEXT: orr z0.d, z0.d, z1.d
; CHECK-NEXT: st1w { z0.s }, p0, [x3]
; CHECK-NEXT: ret
%pre_cond = load <8 x i32>, ptr %pre_cond_ptr
diff --git a/llvm/test/CodeGen/AArch64/sve-index-const-step-vector.ll b/llvm/test/CodeGen/AArch64/sve-index-const-step-vector.ll
index 433ddbd4a261b..cf2ae02c14b18 100644
--- a/llvm/test/CodeGen/AArch64/sve-index-const-step-vector.ll
+++ b/llvm/test/CodeGen/AArch64/sve-index-const-step-vector.ll
@@ -94,8 +94,8 @@ define <4 x i32> @v4i32_neg_immediates() #0 {
define <4 x i32> @v4i32_out_range_start() #0 {
; CHECK-LABEL: v4i32_out_range_start:
; CHECK: // %bb.0:
-; CHECK-NEXT: index z0.s, #0, #1
-; CHECK-NEXT: add z0.s, z0.s, #16 // =0x10
+; CHECK-NEXT: mov w8, #16 // =0x10
+; CHECK-NEXT: index z0.s, w8, #1
; CHECK-NEXT: // kill: def $q0 killed $q0 killed $z0
; CHECK-NEXT: ret
ret <4 x i32> <i32 16, i32 17, i32 18, i32 19>
diff --git a/llvm/test/CodeGen/AArch64/sve-int-imm.ll b/llvm/test/CodeGen/AArch64/sve-int-imm.ll
index e34f4840f517c..985b7b9597705 100644
--- a/llvm/test/CodeGen/AArch64/sve-int-imm.ll
+++ b/llvm/test/CodeGen/AArch64/sve-int-imm.ll
@@ -229,8 +229,7 @@ define <vscale x 16 x i8> @addnve_i8_low(<vscale x 16 x i8> %a) {
define <vscale x 8 x i16> @addnve_i16_low(<vscale x 8 x i16> %a) {
; CHECK-LABEL: addnve_i16_low:
; CHECK: // %bb.0:
-; CHECK-NEXT: mov z1.h, #-30 // =0xffffffffffffffe2
-; CHECK-NEXT: add z0.h, z0.h, z1.h
+; CHECK-NEXT: sub z0.h, z0.h, #30 // =0x1e
; CHECK-NEXT: ret
%res = add <vscale x 8 x i16> %a, splat(i16 -30)
ret <vscale x 8 x i16> %res
@@ -248,8 +247,7 @@ define <vscale x 8 x i16> @addnve_i16_high(<vscale x 8 x i16> %a) {
define <vscale x 4 x i32> @addnve_i32_low(<vscale x 4 x i32> %a) {
; CHECK-LABEL: addnve_i32_low:
; CHECK: // %bb.0:
-; CHECK-NEXT: mov z1.s, #-30 // =0xffffffffffffffe2
-; CHECK-NEXT: add z0.s, z0.s, z1.s
+; CHECK-NEXT: sub z0.s, z0.s, #30 // =0x1e
; CHECK-NEXT: ret
%res = add <vscale x 4 x i32> %a, splat(i32 -30)
ret <vscale x 4 x i32> %res
@@ -258,8 +256,7 @@ define <vscale x 4 x i32> @addnve_i32_low(<vscale x 4 x i32> %a) {
define <vscale x 4 x i32> @addnve_i32_high(<vscale x 4 x i32> %a) {
; CHECK-LABEL: addnve_i32_high:
; CHECK: // %bb.0:
-; CHECK-NEXT: mov z1.s, #-1024 // =0xfffffffffffffc00
-; CHECK-NEXT: add z0.s, z0.s, z1.s
+; CHECK-NEXT: sub z0.s, z0.s, #1024 // =0x400
; CHECK-NEXT: ret
%res = add <vscale x 4 x i32> %a, splat(i32 -1024)
ret <vscale x 4 x i32> %res
@@ -268,8 +265,7 @@ define <vscale x 4 x i32> @addnve_i32_high(<vscale x 4 x i32> %a) {
define <vscale x 2 x i64> @addnve_i64_low(<vscale x 2 x i64> %a) {
; CHECK-LABEL: addnve_i64_low:
; CHECK: // %bb.0:
-; CHECK-NEXT: mov z1.d, #-30 // =0xffffffffffffffe2
-; CHECK-NEXT: add z0.d, z0.d, z1.d
+; CHECK-NEXT: sub z0.d, z0.d, #30 // =0x1e
; CHECK-NEXT: ret
%res = add <vscale x 2 x i64> %a, splat(i64 -30)
ret <vscale x 2 x i64> %res
@@ -278,8 +274,7 @@ define <vscale x 2 x i64> @addnve_i64_low(<vscale x 2 x i64> %a) {
define <vscale x 2 x i64> @addnve_i64_high(<vscale x 2 x i64> %a) {
; CHECK-LABEL: addnve_i64_high:
; CHECK: // %bb.0:
-; CHECK-NEXT: mov z1.d, #-1024 // =0xfffffffffffffc00
-; CHECK-NEXT: add z0.d, z0.d, z1.d
+; CHECK-NEXT: sub z0.d, z0.d, #1024 // =0x400
; CHECK-NEXT: ret
%res = add <vscale x 2 x i64> %a, splat(i64 -1024)
ret <vscale x 2 x i64> %res
diff --git a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-bitselect.ll b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-bitselect.ll
index d29e43509dfe9..71396da004002 100644
--- a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-bitselect.ll
+++ b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-bitselect.ll
@@ -14,20 +14,21 @@ target triple = "aarch64"
define <8 x i32> @fixed_bitselect_v8i32(ptr %pre_cond_ptr, ptr %left_ptr, ptr %right_ptr) {
; CHECK-LABEL: fixed_bitselect_v8i32:
; CHECK: // %bb.0:
-; CHECK-NEXT: mov z0.s, #-1 // =0xffffffffffffffff
-; CHECK-NEXT: ldp q2, q1, [x0]
+; CHECK-NEXT: ldp q1, q0, [x0]
; CHECK-NEXT: ldp q5, q4, [x1]
; CHECK-NEXT: ldp q6, q7, [x2]
-; CHECK-NEXT: add z3.s, z1.s, z0.s
-; CHECK-NEXT: subr z1.s, z1.s, #0 // =0x0
-; CHECK-NEXT: add z0.s, z2.s, z0.s
+; CHECK-NEXT: mov z2.d, z0.d
+; CHECK-NEXT: mov z3.d, z1.d
+; CHECK-NEXT: sub z1.s, z1.s, #1 // =0x1
+; CHECK-NEXT: sub z0.s, z0.s, #1 // =0x1
; CHECK-NEXT: subr z2.s, z2.s, #0 // =0x0
-; CHECK-NEXT: and z1.d, z1.d, z4.d
-; CHECK-NEXT: and z3.d, z3.d, z7.d
-; CHECK-NEXT: and z0.d, z0.d, z6.d
-; CHECK-NEXT: and z2.d, z2.d, z5.d
-; CHECK-NEXT: orr z1.d, z3.d, z1.d
-; CHECK-NEXT: orr z0.d, z0.d, z2.d
+; CHECK-NEXT: subr z3.s, z3.s, #0 // =0x0
+; CHECK-NEXT: and z2.d, z2.d, z4.d
+; CHECK-NEXT: and z3.d, z3.d, z5.d
+; CHECK-NEXT: and z4.d, z0.d, z7.d
+; CHECK-NEXT: and z0.d, z1.d, z6.d
+; CHECK-NEXT: orr z1.d, z4.d, z2.d
+; CHECK-NEXT: orr z0.d, z0.d, z3.d
; CHECK-NEXT: // kill: def $q1 killed $q1 killed $z1
; CHECK-NEXT: // kill: def $q0 killed $q0 killed $z0
; CHECK-NEXT: ret
diff --git a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-build-vector.ll b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-build-vector.ll
index 3a6445dd1d99b..d226fc89c3381 100644
--- a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-build-vector.ll
+++ b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-build-vector.ll
@@ -69,9 +69,9 @@ define void @build_vector_0_dec3_v8i32(ptr %a) {
; CHECK-LABEL: build_vector_0_dec3_v8i32:
; CHECK: // %bb.0:
; CHECK-NEXT: index z0.s, #0, #-3
-; CHECK-NEXT: mov z1.s, #-12 // =0xfffffffffffffff4
-; CHECK-NEXT: add z1.s, z0.s, z1.s
-; CHECK-NEXT: stp q0, q1, [x0]
+; CHECK-NEXT: str q0, [x0]
+; CHECK-NEXT: sub z0.s, z0.s, #12 // =0xc
+; CHECK-NEXT: str q0, [x0, #16]
; CHECK-NEXT: ret
;
; NONEON-NOSVE-LABEL: build_vector_0_dec3_v8i32:
@@ -91,11 +91,10 @@ define void @build_vector_minus2_dec32_v4i64(ptr %a) {
; CHECK-LABEL: build_vector_minus2_dec32_v4i64:
; CHECK: // %bb.0:
; CHECK-NEXT: mov x8, #-32 // =0xffffffffffffffe0
-; CHECK-NEXT: mov z1.d, #-66 // =0xffffffffffffffbe
-; CHECK-NEXT: mov z2.d, #-2 // =0xfffffffffffffffe
; CHECK-NEXT: index z0.d, #0, x8
-; CHECK-NEXT: add z1.d, z0.d, z1.d
-; CHECK-NEXT: add z0.d, z0.d, z2.d
+; CHECK-NEXT: mov z1.d, z0.d
+; CHECK-NEXT: sub z0.d, z0.d, #2 // =0x2
+; CHECK-NEXT: sub z1.d, z1.d, #66 // =0x42
; CHECK-NEXT: stp q0, q1, [x0]
; CHECK-NEXT: ret
;
|
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
4a11da9 to
f3bddf8
Compare
paulwalker-arm
approved these changes
Aug 29, 2025
Comment on lines
662
to
664
| def : SVE_1_Op_Imm_OptLsl_Pat<nxv8i16, add, ZPR16, i32, SVEAddSubNegImm16Pat, SUB_ZI_H>; | ||
| def : SVE_1_Op_Imm_OptLsl_Pat<nxv4i32, add, ZPR32, i32, SVEAddSubNegImm32Pat, SUB_ZI_S>; | ||
| def : SVE_1_Op_Imm_OptLsl_Pat<nxv2i64, add, ZPR64, i64, SVEAddSubNegImm64Pat, SUB_ZI_D>; |
There was a problem hiding this comment.
Any objections to moving these into sve_int_arith_imm0 similar to sve_int_arith_imm0_ssat, albeit this time with inv_op defaulting to null_frag?
f3bddf8 to
289ae65
Compare
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.
sub 3will be canonicalized in llvm toadd -3. This adds some tablegen patterns for add from a negative immediate so that we can still generate sub imm SVE instructions.The alternative is to add a isel combine, which seemed to work but created problems for mad and index patterns. This version does still need to add a lower-than-default Complexity to the ComplexPatterns to ensure that index was selected over sub-imm + index, as the default Complexity on ComplexPatterns is quite high.
Fixes #155928