[llvm] dd8d0d2 - [InstCombine] canonicalize trunc + insert as bitcast + shuffle, part 2
Sanjay Patel via llvm-commits
llvm-commits at lists.llvm.org
Thu Dec 1 11:47:47 PST 2022
Author: Sanjay Patel
Date: 2022-12-01T14:47:37-05:00
New Revision: dd8d0d21ce6d0665ef5d426372096aaed85b479a
URL: https://github.com/llvm/llvm-project/commit/dd8d0d21ce6d0665ef5d426372096aaed85b479a
DIFF: https://github.com/llvm/llvm-project/commit/dd8d0d21ce6d0665ef5d426372096aaed85b479a.diff
LOG: [InstCombine] canonicalize trunc + insert as bitcast + shuffle, part 2
This enhances the base fold from part 1 to allow mapping a
right-shift to an insert index.
Example of translating a middle chunk of the scalar to vector
for either endian:
https://alive2.llvm.org/ce/z/fRXCOZ
This only allows creating an identity shuffle (with optional
shortening/lengthening) because that is considered the safe
baseline for any target (can be inverted if needed). If we
tried this fold with target-specific costs/legality, then we
could do the transform more generally.
Differential Revision: https://reviews.llvm.org/D138873
Added:
Modified:
llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
llvm/test/Transforms/InstCombine/insert-trunc.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
index 8b353abb94923..8a5ddbbdc3b27 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
@@ -1544,8 +1544,17 @@ static Instruction *foldTruncInsElt(InsertElementInst &InsElt, bool IsBigEndian,
// The insert is to the LSB end of the vector (depends on endian).
// That's all we need.
} else {
- // TODO: Look through a shift-right and translate the insert index.
- return nullptr;
+ // If not, we must match a right-shift to translate the insert index.
+ uint64_t ShiftC;
+ if (!match(T, m_OneUse(m_LShr(m_Value(X), m_ConstantInt(ShiftC)))))
+ return nullptr;
+
+ // Check the shift amount to see if this can be folded to an identity
+ // shuffle (assuming we are shuffling with an undef base vector).
+ // Big endian has MSB at vector index 0, so the insert index is flipped.
+ if (ShiftC != (IsBigEndian ? (NumEltsInScalar - 1 - IndexC) * VecEltWidth
+ : IndexC * VecEltWidth))
+ return nullptr;
}
// Bitcast the scalar to a vector type with the destination element type.
diff --git a/llvm/test/Transforms/InstCombine/insert-trunc.ll b/llvm/test/Transforms/InstCombine/insert-trunc.ll
index 20922f97648b9..07b810009228d 100644
--- a/llvm/test/Transforms/InstCombine/insert-trunc.ll
+++ b/llvm/test/Transforms/InstCombine/insert-trunc.ll
@@ -143,11 +143,16 @@ define <4 x i8> @low_index_shorter_length_poison_basevec_extra_use(i64 %x) {
}
define <4 x i16> @lshr_same_length_poison_basevec_le(i64 %x) {
-; ALL-LABEL: @lshr_same_length_poison_basevec_le(
-; ALL-NEXT: [[S:%.*]] = lshr i64 [[X:%.*]], 32
-; ALL-NEXT: [[T:%.*]] = trunc i64 [[S]] to i16
-; ALL-NEXT: [[R:%.*]] = insertelement <4 x i16> poison, i16 [[T]], i64 2
-; ALL-NEXT: ret <4 x i16> [[R]]
+; BE-LABEL: @lshr_same_length_poison_basevec_le(
+; BE-NEXT: [[S:%.*]] = lshr i64 [[X:%.*]], 32
+; BE-NEXT: [[T:%.*]] = trunc i64 [[S]] to i16
+; BE-NEXT: [[R:%.*]] = insertelement <4 x i16> poison, i16 [[T]], i64 2
+; BE-NEXT: ret <4 x i16> [[R]]
+;
+; LE-LABEL: @lshr_same_length_poison_basevec_le(
+; LE-NEXT: [[VEC_X:%.*]] = bitcast i64 [[X:%.*]] to <4 x i16>
+; LE-NEXT: [[R:%.*]] = shufflevector <4 x i16> [[VEC_X]], <4 x i16> poison, <4 x i32> <i32 undef, i32 undef, i32 2, i32 undef>
+; LE-NEXT: ret <4 x i16> [[R]]
;
%s = lshr i64 %x, 32
%t = trunc i64 %s to i16
@@ -156,11 +161,16 @@ define <4 x i16> @lshr_same_length_poison_basevec_le(i64 %x) {
}
define <4 x i16> @lshr_same_length_poison_basevec_be(i64 %x) {
-; ALL-LABEL: @lshr_same_length_poison_basevec_be(
-; ALL-NEXT: [[S:%.*]] = lshr i64 [[X:%.*]], 32
-; ALL-NEXT: [[T:%.*]] = trunc i64 [[S]] to i16
-; ALL-NEXT: [[R:%.*]] = insertelement <4 x i16> poison, i16 [[T]], i64 1
-; ALL-NEXT: ret <4 x i16> [[R]]
+; BE-LABEL: @lshr_same_length_poison_basevec_be(
+; BE-NEXT: [[VEC_X:%.*]] = bitcast i64 [[X:%.*]] to <4 x i16>
+; BE-NEXT: [[R:%.*]] = shufflevector <4 x i16> [[VEC_X]], <4 x i16> poison, <4 x i32> <i32 undef, i32 1, i32 undef, i32 undef>
+; BE-NEXT: ret <4 x i16> [[R]]
+;
+; LE-LABEL: @lshr_same_length_poison_basevec_be(
+; LE-NEXT: [[S:%.*]] = lshr i64 [[X:%.*]], 32
+; LE-NEXT: [[T:%.*]] = trunc i64 [[S]] to i16
+; LE-NEXT: [[R:%.*]] = insertelement <4 x i16> poison, i16 [[T]], i64 1
+; LE-NEXT: ret <4 x i16> [[R]]
;
%s = lshr i64 %x, 32
%t = trunc i64 %s to i16
@@ -170,9 +180,8 @@ define <4 x i16> @lshr_same_length_poison_basevec_be(i64 %x) {
define <4 x i16> @lshr_same_length_poison_basevec_both_endian(i64 %x) {
; BE-LABEL: @lshr_same_length_poison_basevec_both_endian(
-; BE-NEXT: [[S:%.*]] = lshr i64 [[X:%.*]], 48
-; BE-NEXT: [[T:%.*]] = trunc i64 [[S]] to i16
-; BE-NEXT: [[R:%.*]] = insertelement <4 x i16> poison, i16 [[T]], i64 0
+; BE-NEXT: [[VEC_X:%.*]] = bitcast i64 [[X:%.*]] to <4 x i16>
+; BE-NEXT: [[R:%.*]] = shufflevector <4 x i16> [[VEC_X]], <4 x i16> poison, <4 x i32> <i32 0, i32 undef, i32 undef, i32 undef>
; BE-NEXT: ret <4 x i16> [[R]]
;
; LE-LABEL: @lshr_same_length_poison_basevec_both_endian(
@@ -208,9 +217,8 @@ define <8 x i16> @lshr_longer_length_poison_basevec_le(i64 %x) {
; BE-NEXT: ret <8 x i16> [[R]]
;
; LE-LABEL: @lshr_longer_length_poison_basevec_le(
-; LE-NEXT: [[S:%.*]] = lshr i64 [[X:%.*]], 48
-; LE-NEXT: [[T:%.*]] = trunc i64 [[S]] to i16
-; LE-NEXT: [[R:%.*]] = insertelement <8 x i16> poison, i16 [[T]], i64 3
+; LE-NEXT: [[VEC_X:%.*]] = bitcast i64 [[X:%.*]] to <4 x i16>
+; LE-NEXT: [[R:%.*]] = shufflevector <4 x i16> [[VEC_X]], <4 x i16> poison, <8 x i32> <i32 undef, i32 undef, i32 undef, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
; LE-NEXT: ret <8 x i16> [[R]]
;
%s = lshr i64 %x, 48
@@ -220,11 +228,16 @@ define <8 x i16> @lshr_longer_length_poison_basevec_le(i64 %x) {
}
define <8 x i16> @lshr_longer_length_poison_basevec_be(i64 %x) {
-; ALL-LABEL: @lshr_longer_length_poison_basevec_be(
-; ALL-NEXT: [[S:%.*]] = lshr i64 [[X:%.*]], 32
-; ALL-NEXT: [[T:%.*]] = trunc i64 [[S]] to i16
-; ALL-NEXT: [[R:%.*]] = insertelement <8 x i16> poison, i16 [[T]], i64 1
-; ALL-NEXT: ret <8 x i16> [[R]]
+; BE-LABEL: @lshr_longer_length_poison_basevec_be(
+; BE-NEXT: [[VEC_X:%.*]] = bitcast i64 [[X:%.*]] to <4 x i16>
+; BE-NEXT: [[R:%.*]] = shufflevector <4 x i16> [[VEC_X]], <4 x i16> poison, <8 x i32> <i32 undef, i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+; BE-NEXT: ret <8 x i16> [[R]]
+;
+; LE-LABEL: @lshr_longer_length_poison_basevec_be(
+; LE-NEXT: [[S:%.*]] = lshr i64 [[X:%.*]], 32
+; LE-NEXT: [[T:%.*]] = trunc i64 [[S]] to i16
+; LE-NEXT: [[R:%.*]] = insertelement <8 x i16> poison, i16 [[T]], i64 1
+; LE-NEXT: ret <8 x i16> [[R]]
;
%s = lshr i64 %x, 32
%t = trunc i64 %s to i16
@@ -246,11 +259,16 @@ define <8 x i16> @lshr_wrong_index_longer_length_poison_basevec(i64 %x) {
}
define <2 x i16> @lshr_shorter_length_poison_basevec_le(i64 %x) {
-; ALL-LABEL: @lshr_shorter_length_poison_basevec_le(
-; ALL-NEXT: [[S:%.*]] = lshr i64 [[X:%.*]], 16
-; ALL-NEXT: [[T:%.*]] = trunc i64 [[S]] to i16
-; ALL-NEXT: [[R:%.*]] = insertelement <2 x i16> poison, i16 [[T]], i64 1
-; ALL-NEXT: ret <2 x i16> [[R]]
+; BE-LABEL: @lshr_shorter_length_poison_basevec_le(
+; BE-NEXT: [[S:%.*]] = lshr i64 [[X:%.*]], 16
+; BE-NEXT: [[T:%.*]] = trunc i64 [[S]] to i16
+; BE-NEXT: [[R:%.*]] = insertelement <2 x i16> poison, i16 [[T]], i64 1
+; BE-NEXT: ret <2 x i16> [[R]]
+;
+; LE-LABEL: @lshr_shorter_length_poison_basevec_le(
+; LE-NEXT: [[VEC_X:%.*]] = bitcast i64 [[X:%.*]] to <4 x i16>
+; LE-NEXT: [[R:%.*]] = shufflevector <4 x i16> [[VEC_X]], <4 x i16> poison, <2 x i32> <i32 undef, i32 1>
+; LE-NEXT: ret <2 x i16> [[R]]
;
%s = lshr i64 %x, 16
%t = trunc i64 %s to i16
@@ -259,11 +277,16 @@ define <2 x i16> @lshr_shorter_length_poison_basevec_le(i64 %x) {
}
define <4 x i8> @lshr_shorter_length_poison_basevec_be(i64 %x) {
-; ALL-LABEL: @lshr_shorter_length_poison_basevec_be(
-; ALL-NEXT: [[S:%.*]] = lshr i64 [[X:%.*]], 40
-; ALL-NEXT: [[T:%.*]] = trunc i64 [[S]] to i8
-; ALL-NEXT: [[R:%.*]] = insertelement <4 x i8> poison, i8 [[T]], i64 2
-; ALL-NEXT: ret <4 x i8> [[R]]
+; BE-LABEL: @lshr_shorter_length_poison_basevec_be(
+; BE-NEXT: [[VEC_X:%.*]] = bitcast i64 [[X:%.*]] to <8 x i8>
+; BE-NEXT: [[R:%.*]] = shufflevector <8 x i8> [[VEC_X]], <8 x i8> poison, <4 x i32> <i32 undef, i32 undef, i32 2, i32 undef>
+; BE-NEXT: ret <4 x i8> [[R]]
+;
+; LE-LABEL: @lshr_shorter_length_poison_basevec_be(
+; LE-NEXT: [[S:%.*]] = lshr i64 [[X:%.*]], 40
+; LE-NEXT: [[T:%.*]] = trunc i64 [[S]] to i8
+; LE-NEXT: [[R:%.*]] = insertelement <4 x i8> poison, i8 [[T]], i64 2
+; LE-NEXT: ret <4 x i8> [[R]]
;
%s = lshr i64 %x, 40
%t = trunc i64 %s to i8
More information about the llvm-commits
mailing list