[llvm-branch-commits] [llvm] release/23.x: [LLVM][CodeGen][SME] Don't use non-streaming zeroing instruction when in streaming mode. (#222353) (PR #222588)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Thu Sep 10 03:25:31 PDT 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-aarch64
Author: llvmbot
<details>
<summary>Changes</summary>
Backport 9679fbfea993cb2e815fe0fa5eefde310f917349
Requested by: @<!-- -->paulwalker-arm
---
Patch is 95.16 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/222588.diff
7 Files Affected:
- (modified) llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td (+11-2)
- (modified) llvm/test/CodeGen/AArch64/sve-bf16-arith.ll (+265-130)
- (modified) llvm/test/CodeGen/AArch64/sve-bf16-compares.ll (+538-260)
- (modified) llvm/test/CodeGen/AArch64/sve-bf16-converts.ll (+25-9)
- (modified) llvm/test/CodeGen/AArch64/sve-bf16-int-converts.ll (+176-84)
- (modified) llvm/test/CodeGen/AArch64/sve-bf16-reductions.ll (+105-50)
- (modified) llvm/test/CodeGen/AArch64/sve-bf16-rounding.ll (+177-86)
``````````diff
diff --git a/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td b/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
index 26b0a38f6bae0..0d8fac0d95706 100644
--- a/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
+++ b/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
@@ -2653,9 +2653,9 @@ let Predicates = [HasSVE_or_SME] in {
// A common sequence that occurs when promoting nxv8bf16 operations.
def : Pat<(nxv4f32 (AArch64fcvte_mt (SVEAnyPredicate), (extract_subvector nxv8bf16:$Zs, (i64 0)), undef)),
- (ZIP1_ZZZ_H (SUBREG_TO_REG (MOVIv2d_ns (i32 0)), zsub), $Zs)>;
+ (ZIP1_ZZZ_H (DUP_ZI_H 0, 0), $Zs)>;
def : Pat<(nxv4f32 (AArch64fcvte_mt (SVEAnyPredicate), (extract_subvector nxv8bf16:$Zs, (i64 4)), undef)),
- (ZIP2_ZZZ_H (SUBREG_TO_REG (MOVIv2d_ns (i32 0)), zsub), $Zs)>;
+ (ZIP2_ZZZ_H (DUP_ZI_H 0, 0), $Zs)>;
// Signed integer -> Floating-point
def : Pat<(nxv2f16 (AArch64scvtf_mt (nxv2i1 (SVEAllActive):$Pg),
@@ -2715,6 +2715,15 @@ let Predicates = [HasSVE_or_SME] in {
defm FSQRT_ZPmZ : sve_fp_2op_p_zd_HSD<0b01101, "fsqrt", AArch64fsqrt_mt>;
} // End HasSVE_or_SME
+let Predicates = [HasSVE, HasNEON], AddedComplexity = 1 in {
+ // A common sequence that occurs when promoting nxv8bf16 operations.
+ // NOTE: MOVIv2d_ns is not available in streaming-mode.
+ def : Pat<(nxv4f32 (AArch64fcvte_mt (SVEAnyPredicate), (extract_subvector nxv8bf16:$Zs, (i64 0)), undef)),
+ (ZIP1_ZZZ_H (SUBREG_TO_REG (MOVIv2d_ns (i32 0)), zsub), $Zs)>;
+ def : Pat<(nxv4f32 (AArch64fcvte_mt (SVEAnyPredicate), (extract_subvector nxv8bf16:$Zs, (i64 4)), undef)),
+ (ZIP2_ZZZ_H (SUBREG_TO_REG (MOVIv2d_ns (i32 0)), zsub), $Zs)>;
+}
+
let Predicates = [HasBF16, HasSVE_or_SME] in {
defm BFDOT_ZZZ : sve_float_dot<0b1, 0b0, ZPR32, ZPR16, "bfdot", nxv8bf16, int_aarch64_sve_bfdot>;
defm BFDOT_ZZI : sve_float_dot_indexed<0b1, 0b00, ZPR16, ZPR3b16, "bfdot", nxv8bf16, int_aarch64_sve_bfdot_lane_v2>;
diff --git a/llvm/test/CodeGen/AArch64/sve-bf16-arith.ll b/llvm/test/CodeGen/AArch64/sve-bf16-arith.ll
index c916f81edf669..6337fa8ee7fe8 100644
--- a/llvm/test/CodeGen/AArch64/sve-bf16-arith.ll
+++ b/llvm/test/CodeGen/AArch64/sve-bf16-arith.ll
@@ -1,8 +1,8 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-; RUN: llc -mattr=+sve,+bf16 < %s | FileCheck %s --check-prefixes=CHECK,NOB16B16,NOB16B16-NONSTREAMING
-; RUN: llc -mattr=+sve,+bf16,+sve-b16b16 < %s | FileCheck %s --check-prefixes=CHECK,B16B16
-; RUN: llc -mattr=+sme,+sve-b16b16 -force-streaming < %s | FileCheck %s --check-prefixes=CHECK,NOB16B16,NOB16B16-STREAMING
-; RUN: llc -mattr=+sme2,+sve-b16b16 -force-streaming < %s | FileCheck %s --check-prefixes=CHECK,B16B16
+; RUN: llc -mattr=+sve,+bf16 < %s | FileCheck %s --check-prefixes=CHECK,NONSTREAMING,NOB16B16,NOB16B16-NONSTREAMING
+; RUN: llc -mattr=+sve,+bf16,+sve-b16b16 < %s | FileCheck %s --check-prefixes=CHECK,NONSTREAMING,B16B16
+; RUN: llc -mattr=+sme,+sve-b16b16 -force-streaming < %s | FileCheck %s --check-prefixes=CHECK,STREAMING,NOB16B16,NOB16B16-STREAMING
+; RUN: llc -mattr=+sme2,+sve-b16b16 -force-streaming < %s | FileCheck %s --check-prefixes=CHECK,STREAMING,B16B16
target triple = "aarch64-unknown-linux-gnu"
@@ -83,25 +83,40 @@ define <vscale x 4 x bfloat> @fadd_nxv4bf16(<vscale x 4 x bfloat> %a, <vscale x
}
define <vscale x 8 x bfloat> @fadd_nxv8bf16(<vscale x 8 x bfloat> %a, <vscale x 8 x bfloat> %b) {
-; NOB16B16-LABEL: fadd_nxv8bf16:
-; NOB16B16: // %bb.0:
-; NOB16B16-NEXT: movi v2.2d, #0000000000000000
-; NOB16B16-NEXT: ptrue p0.s
-; NOB16B16-NEXT: zip2 z3.h, z2.h, z1.h
-; NOB16B16-NEXT: zip2 z4.h, z2.h, z0.h
-; NOB16B16-NEXT: zip1 z1.h, z2.h, z1.h
-; NOB16B16-NEXT: zip1 z0.h, z2.h, z0.h
-; NOB16B16-NEXT: fadd z2.s, z4.s, z3.s
-; NOB16B16-NEXT: fadd z0.s, z0.s, z1.s
-; NOB16B16-NEXT: bfcvt z1.h, p0/m, z2.s
-; NOB16B16-NEXT: bfcvt z0.h, p0/m, z0.s
-; NOB16B16-NEXT: uzp1 z0.h, z0.h, z1.h
-; NOB16B16-NEXT: ret
+; NOB16B16-NONSTREAMING-LABEL: fadd_nxv8bf16:
+; NOB16B16-NONSTREAMING: // %bb.0:
+; NOB16B16-NONSTREAMING-NEXT: movi v2.2d, #0000000000000000
+; NOB16B16-NONSTREAMING-NEXT: ptrue p0.s
+; NOB16B16-NONSTREAMING-NEXT: zip2 z3.h, z2.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: zip2 z4.h, z2.h, z0.h
+; NOB16B16-NONSTREAMING-NEXT: zip1 z1.h, z2.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: zip1 z0.h, z2.h, z0.h
+; NOB16B16-NONSTREAMING-NEXT: fadd z2.s, z4.s, z3.s
+; NOB16B16-NONSTREAMING-NEXT: fadd z0.s, z0.s, z1.s
+; NOB16B16-NONSTREAMING-NEXT: bfcvt z1.h, p0/m, z2.s
+; NOB16B16-NONSTREAMING-NEXT: bfcvt z0.h, p0/m, z0.s
+; NOB16B16-NONSTREAMING-NEXT: uzp1 z0.h, z0.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: ret
;
; B16B16-LABEL: fadd_nxv8bf16:
; B16B16: // %bb.0:
; B16B16-NEXT: bfadd z0.h, z0.h, z1.h
; B16B16-NEXT: ret
+;
+; NOB16B16-STREAMING-LABEL: fadd_nxv8bf16:
+; NOB16B16-STREAMING: // %bb.0:
+; NOB16B16-STREAMING-NEXT: mov z2.h, #0 // =0x0
+; NOB16B16-STREAMING-NEXT: ptrue p0.s
+; NOB16B16-STREAMING-NEXT: zip2 z3.h, z2.h, z1.h
+; NOB16B16-STREAMING-NEXT: zip2 z4.h, z2.h, z0.h
+; NOB16B16-STREAMING-NEXT: zip1 z1.h, z2.h, z1.h
+; NOB16B16-STREAMING-NEXT: zip1 z0.h, z2.h, z0.h
+; NOB16B16-STREAMING-NEXT: fadd z2.s, z4.s, z3.s
+; NOB16B16-STREAMING-NEXT: fadd z0.s, z0.s, z1.s
+; NOB16B16-STREAMING-NEXT: bfcvt z1.h, p0/m, z2.s
+; NOB16B16-STREAMING-NEXT: bfcvt z0.h, p0/m, z0.s
+; NOB16B16-STREAMING-NEXT: uzp1 z0.h, z0.h, z1.h
+; NOB16B16-STREAMING-NEXT: ret
%res = fadd <vscale x 8 x bfloat> %a, %b
ret <vscale x 8 x bfloat> %res
}
@@ -137,20 +152,35 @@ define <vscale x 4 x bfloat> @fdiv_nxv4bf16(<vscale x 4 x bfloat> %a, <vscale x
}
define <vscale x 8 x bfloat> @fdiv_nxv8bf16(<vscale x 8 x bfloat> %a, <vscale x 8 x bfloat> %b) {
-; CHECK-LABEL: fdiv_nxv8bf16:
-; CHECK: // %bb.0:
-; CHECK-NEXT: movi v2.2d, #0000000000000000
-; CHECK-NEXT: ptrue p0.s
-; CHECK-NEXT: zip2 z3.h, z2.h, z1.h
-; CHECK-NEXT: zip2 z4.h, z2.h, z0.h
-; CHECK-NEXT: zip1 z1.h, z2.h, z1.h
-; CHECK-NEXT: zip1 z0.h, z2.h, z0.h
-; CHECK-NEXT: fdivr z3.s, p0/m, z3.s, z4.s
-; CHECK-NEXT: fdiv z0.s, p0/m, z0.s, z1.s
-; CHECK-NEXT: bfcvt z1.h, p0/m, z3.s
-; CHECK-NEXT: bfcvt z0.h, p0/m, z0.s
-; CHECK-NEXT: uzp1 z0.h, z0.h, z1.h
-; CHECK-NEXT: ret
+; NONSTREAMING-LABEL: fdiv_nxv8bf16:
+; NONSTREAMING: // %bb.0:
+; NONSTREAMING-NEXT: movi v2.2d, #0000000000000000
+; NONSTREAMING-NEXT: ptrue p0.s
+; NONSTREAMING-NEXT: zip2 z3.h, z2.h, z1.h
+; NONSTREAMING-NEXT: zip2 z4.h, z2.h, z0.h
+; NONSTREAMING-NEXT: zip1 z1.h, z2.h, z1.h
+; NONSTREAMING-NEXT: zip1 z0.h, z2.h, z0.h
+; NONSTREAMING-NEXT: fdivr z3.s, p0/m, z3.s, z4.s
+; NONSTREAMING-NEXT: fdiv z0.s, p0/m, z0.s, z1.s
+; NONSTREAMING-NEXT: bfcvt z1.h, p0/m, z3.s
+; NONSTREAMING-NEXT: bfcvt z0.h, p0/m, z0.s
+; NONSTREAMING-NEXT: uzp1 z0.h, z0.h, z1.h
+; NONSTREAMING-NEXT: ret
+;
+; STREAMING-LABEL: fdiv_nxv8bf16:
+; STREAMING: // %bb.0:
+; STREAMING-NEXT: mov z2.h, #0 // =0x0
+; STREAMING-NEXT: ptrue p0.s
+; STREAMING-NEXT: zip2 z3.h, z2.h, z1.h
+; STREAMING-NEXT: zip2 z4.h, z2.h, z0.h
+; STREAMING-NEXT: zip1 z1.h, z2.h, z1.h
+; STREAMING-NEXT: zip1 z0.h, z2.h, z0.h
+; STREAMING-NEXT: fdivr z3.s, p0/m, z3.s, z4.s
+; STREAMING-NEXT: fdiv z0.s, p0/m, z0.s, z1.s
+; STREAMING-NEXT: bfcvt z1.h, p0/m, z3.s
+; STREAMING-NEXT: bfcvt z0.h, p0/m, z0.s
+; STREAMING-NEXT: uzp1 z0.h, z0.h, z1.h
+; STREAMING-NEXT: ret
%res = fdiv <vscale x 8 x bfloat> %a, %b
ret <vscale x 8 x bfloat> %res
}
@@ -198,26 +228,41 @@ define <vscale x 4 x bfloat> @fmax_nxv4bf16(<vscale x 4 x bfloat> %a, <vscale x
}
define <vscale x 8 x bfloat> @fmax_nxv8bf16(<vscale x 8 x bfloat> %a, <vscale x 8 x bfloat> %b) {
-; NOB16B16-LABEL: fmax_nxv8bf16:
-; NOB16B16: // %bb.0:
-; NOB16B16-NEXT: movi v2.2d, #0000000000000000
-; NOB16B16-NEXT: ptrue p0.s
-; NOB16B16-NEXT: zip2 z3.h, z2.h, z1.h
-; NOB16B16-NEXT: zip2 z4.h, z2.h, z0.h
-; NOB16B16-NEXT: zip1 z1.h, z2.h, z1.h
-; NOB16B16-NEXT: zip1 z0.h, z2.h, z0.h
-; NOB16B16-NEXT: fmax z3.s, p0/m, z3.s, z4.s
-; NOB16B16-NEXT: fmax z0.s, p0/m, z0.s, z1.s
-; NOB16B16-NEXT: bfcvt z1.h, p0/m, z3.s
-; NOB16B16-NEXT: bfcvt z0.h, p0/m, z0.s
-; NOB16B16-NEXT: uzp1 z0.h, z0.h, z1.h
-; NOB16B16-NEXT: ret
+; NOB16B16-NONSTREAMING-LABEL: fmax_nxv8bf16:
+; NOB16B16-NONSTREAMING: // %bb.0:
+; NOB16B16-NONSTREAMING-NEXT: movi v2.2d, #0000000000000000
+; NOB16B16-NONSTREAMING-NEXT: ptrue p0.s
+; NOB16B16-NONSTREAMING-NEXT: zip2 z3.h, z2.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: zip2 z4.h, z2.h, z0.h
+; NOB16B16-NONSTREAMING-NEXT: zip1 z1.h, z2.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: zip1 z0.h, z2.h, z0.h
+; NOB16B16-NONSTREAMING-NEXT: fmax z3.s, p0/m, z3.s, z4.s
+; NOB16B16-NONSTREAMING-NEXT: fmax z0.s, p0/m, z0.s, z1.s
+; NOB16B16-NONSTREAMING-NEXT: bfcvt z1.h, p0/m, z3.s
+; NOB16B16-NONSTREAMING-NEXT: bfcvt z0.h, p0/m, z0.s
+; NOB16B16-NONSTREAMING-NEXT: uzp1 z0.h, z0.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: ret
;
; B16B16-LABEL: fmax_nxv8bf16:
; B16B16: // %bb.0:
; B16B16-NEXT: ptrue p0.h
; B16B16-NEXT: bfmax z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT: ret
+;
+; NOB16B16-STREAMING-LABEL: fmax_nxv8bf16:
+; NOB16B16-STREAMING: // %bb.0:
+; NOB16B16-STREAMING-NEXT: mov z2.h, #0 // =0x0
+; NOB16B16-STREAMING-NEXT: ptrue p0.s
+; NOB16B16-STREAMING-NEXT: zip2 z3.h, z2.h, z1.h
+; NOB16B16-STREAMING-NEXT: zip2 z4.h, z2.h, z0.h
+; NOB16B16-STREAMING-NEXT: zip1 z1.h, z2.h, z1.h
+; NOB16B16-STREAMING-NEXT: zip1 z0.h, z2.h, z0.h
+; NOB16B16-STREAMING-NEXT: fmax z3.s, p0/m, z3.s, z4.s
+; NOB16B16-STREAMING-NEXT: fmax z0.s, p0/m, z0.s, z1.s
+; NOB16B16-STREAMING-NEXT: bfcvt z1.h, p0/m, z3.s
+; NOB16B16-STREAMING-NEXT: bfcvt z0.h, p0/m, z0.s
+; NOB16B16-STREAMING-NEXT: uzp1 z0.h, z0.h, z1.h
+; NOB16B16-STREAMING-NEXT: ret
%res = call <vscale x 8 x bfloat> @llvm.maximum.nxv8bf16(<vscale x 8 x bfloat> %a, <vscale x 8 x bfloat> %b)
ret <vscale x 8 x bfloat> %res
}
@@ -265,26 +310,41 @@ define <vscale x 4 x bfloat> @fmaxnm_nxv4bf16(<vscale x 4 x bfloat> %a, <vscale
}
define <vscale x 8 x bfloat> @fmaxnm_nxv8bf16(<vscale x 8 x bfloat> %a, <vscale x 8 x bfloat> %b) {
-; NOB16B16-LABEL: fmaxnm_nxv8bf16:
-; NOB16B16: // %bb.0:
-; NOB16B16-NEXT: movi v2.2d, #0000000000000000
-; NOB16B16-NEXT: ptrue p0.s
-; NOB16B16-NEXT: zip2 z3.h, z2.h, z1.h
-; NOB16B16-NEXT: zip2 z4.h, z2.h, z0.h
-; NOB16B16-NEXT: zip1 z1.h, z2.h, z1.h
-; NOB16B16-NEXT: zip1 z0.h, z2.h, z0.h
-; NOB16B16-NEXT: fmaxnm z3.s, p0/m, z3.s, z4.s
-; NOB16B16-NEXT: fmaxnm z0.s, p0/m, z0.s, z1.s
-; NOB16B16-NEXT: bfcvt z1.h, p0/m, z3.s
-; NOB16B16-NEXT: bfcvt z0.h, p0/m, z0.s
-; NOB16B16-NEXT: uzp1 z0.h, z0.h, z1.h
-; NOB16B16-NEXT: ret
+; NOB16B16-NONSTREAMING-LABEL: fmaxnm_nxv8bf16:
+; NOB16B16-NONSTREAMING: // %bb.0:
+; NOB16B16-NONSTREAMING-NEXT: movi v2.2d, #0000000000000000
+; NOB16B16-NONSTREAMING-NEXT: ptrue p0.s
+; NOB16B16-NONSTREAMING-NEXT: zip2 z3.h, z2.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: zip2 z4.h, z2.h, z0.h
+; NOB16B16-NONSTREAMING-NEXT: zip1 z1.h, z2.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: zip1 z0.h, z2.h, z0.h
+; NOB16B16-NONSTREAMING-NEXT: fmaxnm z3.s, p0/m, z3.s, z4.s
+; NOB16B16-NONSTREAMING-NEXT: fmaxnm z0.s, p0/m, z0.s, z1.s
+; NOB16B16-NONSTREAMING-NEXT: bfcvt z1.h, p0/m, z3.s
+; NOB16B16-NONSTREAMING-NEXT: bfcvt z0.h, p0/m, z0.s
+; NOB16B16-NONSTREAMING-NEXT: uzp1 z0.h, z0.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: ret
;
; B16B16-LABEL: fmaxnm_nxv8bf16:
; B16B16: // %bb.0:
; B16B16-NEXT: ptrue p0.h
; B16B16-NEXT: bfmaxnm z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT: ret
+;
+; NOB16B16-STREAMING-LABEL: fmaxnm_nxv8bf16:
+; NOB16B16-STREAMING: // %bb.0:
+; NOB16B16-STREAMING-NEXT: mov z2.h, #0 // =0x0
+; NOB16B16-STREAMING-NEXT: ptrue p0.s
+; NOB16B16-STREAMING-NEXT: zip2 z3.h, z2.h, z1.h
+; NOB16B16-STREAMING-NEXT: zip2 z4.h, z2.h, z0.h
+; NOB16B16-STREAMING-NEXT: zip1 z1.h, z2.h, z1.h
+; NOB16B16-STREAMING-NEXT: zip1 z0.h, z2.h, z0.h
+; NOB16B16-STREAMING-NEXT: fmaxnm z3.s, p0/m, z3.s, z4.s
+; NOB16B16-STREAMING-NEXT: fmaxnm z0.s, p0/m, z0.s, z1.s
+; NOB16B16-STREAMING-NEXT: bfcvt z1.h, p0/m, z3.s
+; NOB16B16-STREAMING-NEXT: bfcvt z0.h, p0/m, z0.s
+; NOB16B16-STREAMING-NEXT: uzp1 z0.h, z0.h, z1.h
+; NOB16B16-STREAMING-NEXT: ret
%res = call <vscale x 8 x bfloat> @llvm.maxnum.nxv8bf16(<vscale x 8 x bfloat> %a, <vscale x 8 x bfloat> %b)
ret <vscale x 8 x bfloat> %res
}
@@ -332,26 +392,41 @@ define <vscale x 4 x bfloat> @fmin_nxv4bf16(<vscale x 4 x bfloat> %a, <vscale x
}
define <vscale x 8 x bfloat> @fmin_nxv8bf16(<vscale x 8 x bfloat> %a, <vscale x 8 x bfloat> %b) {
-; NOB16B16-LABEL: fmin_nxv8bf16:
-; NOB16B16: // %bb.0:
-; NOB16B16-NEXT: movi v2.2d, #0000000000000000
-; NOB16B16-NEXT: ptrue p0.s
-; NOB16B16-NEXT: zip2 z3.h, z2.h, z1.h
-; NOB16B16-NEXT: zip2 z4.h, z2.h, z0.h
-; NOB16B16-NEXT: zip1 z1.h, z2.h, z1.h
-; NOB16B16-NEXT: zip1 z0.h, z2.h, z0.h
-; NOB16B16-NEXT: fmin z3.s, p0/m, z3.s, z4.s
-; NOB16B16-NEXT: fmin z0.s, p0/m, z0.s, z1.s
-; NOB16B16-NEXT: bfcvt z1.h, p0/m, z3.s
-; NOB16B16-NEXT: bfcvt z0.h, p0/m, z0.s
-; NOB16B16-NEXT: uzp1 z0.h, z0.h, z1.h
-; NOB16B16-NEXT: ret
+; NOB16B16-NONSTREAMING-LABEL: fmin_nxv8bf16:
+; NOB16B16-NONSTREAMING: // %bb.0:
+; NOB16B16-NONSTREAMING-NEXT: movi v2.2d, #0000000000000000
+; NOB16B16-NONSTREAMING-NEXT: ptrue p0.s
+; NOB16B16-NONSTREAMING-NEXT: zip2 z3.h, z2.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: zip2 z4.h, z2.h, z0.h
+; NOB16B16-NONSTREAMING-NEXT: zip1 z1.h, z2.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: zip1 z0.h, z2.h, z0.h
+; NOB16B16-NONSTREAMING-NEXT: fmin z3.s, p0/m, z3.s, z4.s
+; NOB16B16-NONSTREAMING-NEXT: fmin z0.s, p0/m, z0.s, z1.s
+; NOB16B16-NONSTREAMING-NEXT: bfcvt z1.h, p0/m, z3.s
+; NOB16B16-NONSTREAMING-NEXT: bfcvt z0.h, p0/m, z0.s
+; NOB16B16-NONSTREAMING-NEXT: uzp1 z0.h, z0.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: ret
;
; B16B16-LABEL: fmin_nxv8bf16:
; B16B16: // %bb.0:
; B16B16-NEXT: ptrue p0.h
; B16B16-NEXT: bfmin z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT: ret
+;
+; NOB16B16-STREAMING-LABEL: fmin_nxv8bf16:
+; NOB16B16-STREAMING: // %bb.0:
+; NOB16B16-STREAMING-NEXT: mov z2.h, #0 // =0x0
+; NOB16B16-STREAMING-NEXT: ptrue p0.s
+; NOB16B16-STREAMING-NEXT: zip2 z3.h, z2.h, z1.h
+; NOB16B16-STREAMING-NEXT: zip2 z4.h, z2.h, z0.h
+; NOB16B16-STREAMING-NEXT: zip1 z1.h, z2.h, z1.h
+; NOB16B16-STREAMING-NEXT: zip1 z0.h, z2.h, z0.h
+; NOB16B16-STREAMING-NEXT: fmin z3.s, p0/m, z3.s, z4.s
+; NOB16B16-STREAMING-NEXT: fmin z0.s, p0/m, z0.s, z1.s
+; NOB16B16-STREAMING-NEXT: bfcvt z1.h, p0/m, z3.s
+; NOB16B16-STREAMING-NEXT: bfcvt z0.h, p0/m, z0.s
+; NOB16B16-STREAMING-NEXT: uzp1 z0.h, z0.h, z1.h
+; NOB16B16-STREAMING-NEXT: ret
%res = call <vscale x 8 x bfloat> @llvm.minimum.nxv8bf16(<vscale x 8 x bfloat> %a, <vscale x 8 x bfloat> %b)
ret <vscale x 8 x bfloat> %res
}
@@ -399,26 +474,41 @@ define <vscale x 4 x bfloat> @fminnm_nxv4bf16(<vscale x 4 x bfloat> %a, <vscale
}
define <vscale x 8 x bfloat> @fminnm_nxv8bf16(<vscale x 8 x bfloat> %a, <vscale x 8 x bfloat> %b) {
-; NOB16B16-LABEL: fminnm_nxv8bf16:
-; NOB16B16: // %bb.0:
-; NOB16B16-NEXT: movi v2.2d, #0000000000000000
-; NOB16B16-NEXT: ptrue p0.s
-; NOB16B16-NEXT: zip2 z3.h, z2.h, z1.h
-; NOB16B16-NEXT: zip2 z4.h, z2.h, z0.h
-; NOB16B16-NEXT: zip1 z1.h, z2.h, z1.h
-; NOB16B16-NEXT: zip1 z0.h, z2.h, z0.h
-; NOB16B16-NEXT: fminnm z3.s, p0/m, z3.s, z4.s
-; NOB16B16-NEXT: fminnm z0.s, p0/m, z0.s, z1.s
-; NOB16B16-NEXT: bfcvt z1.h, p0/m, z3.s
-; NOB16B16-NEXT: bfcvt z0.h, p0/m, z0.s
-; NOB16B16-NEXT: uzp1 z0.h, z0.h, z1.h
-; NOB16B16-NEXT: ret
+; NOB16B16-NONSTREAMING-LABEL: fminnm_nxv8bf16:
+; NOB16B16-NONSTREAMING: // %bb.0:
+; NOB16B16-NONSTREAMING-NEXT: movi v2.2d, #0000000000000000
+; NOB16B16-NONSTREAMING-NEXT: ptrue p0.s
+; NOB16B16-NONSTREAMING-NEXT: zip2 z3.h, z2.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: zip2 z4.h, z2.h, z0.h
+; NOB16B16-NONSTREAMING-NEXT: zip1 z1.h, z2.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: zip1 z0.h, z2.h, z0.h
+; NOB16B16-NONSTREAMING-NEXT: fminnm z3.s, p0/m, z3.s, z4.s
+; NOB16B16-NONSTREAMING-NEXT: fminnm z0.s, p0/m, z0.s, z1.s
+; NOB16B16-NONSTREAMING-NEXT: bfcvt z1.h, p0/m, z3.s
+; NOB16B16-NONSTREAMING-NEXT: bfcvt z0.h, p0/m, z0.s
+; NOB16B16-NONSTREAMING-NEXT: uzp1 z0.h, z0.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: ret
;
; B16B16-LABEL: fminnm_nxv8bf16:
; B16B16: // %bb.0:
; B16B16-NEXT: ptrue p0.h
; B16B16-NEXT: bfminnm z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT: ret
+;
+; NOB16B16-STREAMING-LABEL: fminnm_nxv8bf16:
+; NOB16B16-STREAMING: // %bb.0:
+; NOB16B16-STREAMING-NEXT: mov z2.h, #0 // =0x0
+; NOB16B16-STREAMING-NEXT: ptrue p0.s
+; NOB16B16-STREAMING-NEXT: zip2 z3.h, z2.h, z1.h
+; NOB16B16-STREAMING-NEXT: zip2 z4.h, z2.h, z0.h
+; NOB16B16-STREAMING-NEXT: zip1 z1.h, z2.h, z1.h
+; NOB16B16-STREAMING-NEXT: zip1 z0.h, z2.h, z0.h
+; NOB16B16-STREAMING-NEXT: fminnm z3.s, p0/m, z3.s, z4.s
+; NOB16B16-STREAMING-NEXT: fminnm z0.s, p0/m, z0.s, z1.s
+; NOB16B16-STREAMING-NEXT: bfcvt z1.h, p0/m, z3.s
+; NOB16B16-STREAMING-NEXT: bfcvt z0.h, p0/m, z0.s
+; NOB16B16-STREAMING-NEXT: uzp1 z0.h, z0.h, z1.h
+; NOB16B16-STREAMING-NEXT: ret
%res = call <vscale x 8 x bfloat> @llvm.minnum.nxv8bf16(<vscale x 8 x bfloat> %a, <vscale x 8 x bfloat> %b)
ret <vscale x 8 x bfloat> %res
}
@@ -466,28 +556,45 @@ define <vscale x 4 x bfloat> @fmla_nxv4bf16(<vscale x 4 x bfloat> %a, <vscale x
}
define <vscale x 8 x bfloat> @fmla_nxv8bf16(<vscale x 8 x bfloat> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) {
-; NOB16B16-LABEL: fmla_nxv8bf16:
-; NOB16B16: // %bb.0:
-; NOB16B16-NEXT: movi v3.2d, #0000000000000000
-; NOB16B16-NEXT: uunpkhi z4.s, z1.h
-; NOB16B16-NEXT: uunpkhi z5.s, z0.h
-; NOB16B16-NEXT: uunpklo z1.s, z1.h
-; NOB16B16-NEXT: uunpklo z0.s, z0.h
-; NOB16B16-NEXT: ptrue p0.s
-; NOB16B16-NEXT: zip2 z6.h, z3.h, z2.h
-; NOB16B16-NEXT: zip1 z2.h, z3.h, z2.h
-; NOB16B16-NEXT: bfmlalb z6.s, z5.h, z4.h
-; NOB16B16-NEXT: bfmlalb z2.s, z0.h, z1.h
-; NOB16B16-NEXT: bfcvt z0.h, p0/m, z6.s
-; NOB16B16-NEXT: bfcvt z1.h, p0/m, z2.s
-; NOB16B16-NEXT: uzp1 z0.h, z1.h, z0.h
-; NOB16B16-NEXT: ret
+; NOB16B16-NONSTREAMING-LABEL: fmla_nxv8bf16:
+; NOB16B16-NONSTREAMING: // %bb.0:
+; NOB16B16-NONSTREAMING-NEXT: movi v3.2d, #0000000000000000
+; NOB16B16-NONSTREAMING-NEXT: uunpkhi z4.s, z1.h
+; NOB16B16-NONSTREAMING-NEXT: uunpkhi z5.s, z0.h
+; NOB16B16-NONSTREAMING-NEXT: uunpklo z1.s, z1.h
+; NOB16B16-NONSTREAMING-NEXT: uunpklo z0.s, z0.h
+; NOB16B16-NONSTREAMING-NEXT: ptrue p0.s
+; NOB16B16-NONSTREAMING-NEXT: zip2 z6.h, z3.h, z2.h
+; NOB16B16-NONSTREAMING-NEXT: zip1 z2.h, z3.h, z2.h
+; NOB16B16-NONSTREAMING-NEXT: bfmlalb z6.s, z5.h, z4.h
+; NOB16B16-NONSTREAMING-NEXT: bfmlalb z2.s, z0.h, z1.h
+; NOB16B16-NONSTREAMING-NEXT: bfcvt z0.h, p0/m, z6.s
+; NOB16B16-NONSTREAMING-NEXT: bfcvt z1.h, p0/m, z2.s
+; NOB16B16-NONSTREAMING-NEXT: uzp1 z0.h, z1.h, z0.h
+; NOB16B16-NONSTREAMING-NEXT: ret
;
; B16B16-LABEL: fmla_nxv8bf16:
; B16B16: // %bb.0:
; B16B16-NEXT: ptrue p0.h
; B16B16-NEXT: bfmla z0.h, p0/m, z1.h, z2.h
; B16B16-NEXT: ret
+;
+; NOB16B16-STREAMING-LABEL: fmla_nxv8bf16:
+; NOB16B16-STREAMING: // %bb.0:
+; NOB16B16-STREAMING-NEXT: mov z3.h, #0 // =0x0
+; NOB16B16-STREAMING-NEXT: uunpkhi z4.s, z1.h
+; NOB16B16-STREAMING-NEXT: uunpkhi z5.s, z0.h
+; NOB16B16-STREAMING-NEXT: uunpklo z1.s, z1.h
+; NOB16B16-STREAMING-NEXT: uunpklo z0.s, z0.h
+; NOB16B16-STREAMING-NEXT: ptrue p0.s
+; NOB16B16-STREAMING-NEXT: zip2 z6.h, z3.h, z2.h
+; NOB16B16-STREAMING-NEXT: zip1 z2.h, z3.h, z2.h
+; NOB16B16-STREAMING-NEXT: bfmlalb z6.s, z5.h, z4.h
+; NOB16B16-S...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/222588
More information about the llvm-branch-commits
mailing list