[llvm] [AMDGPU][ISEL] Use target specific hook to block narrowing insert-extract sub-vector (PR #201056)
via llvm-commits
llvm-commits at lists.llvm.org
Tue Jun 2 01:26:54 PDT 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: Shoreshen
<details>
<summary>Changes</summary>
Function `narrowInsertExtractVectorBinOp` some time will create duplicate binOp. For example:
```llvm
define <4 x i32> @<!-- -->insert_extract_and_inreg_v4i32(ptr addrspace(1) inreg %out_sub, <2 x i32> inreg %a, <2 x i32> inreg %b) {
entry:
%v0 = call <4 x i32> @<!-- -->llvm.vector.insert.v4i32.v2i32(<4 x i32> zeroinitializer, <2 x i32> %a, i64 2)
%v1 = call <4 x i32> @<!-- -->llvm.vector.insert.v4i32.v2i32(<4 x i32> zeroinitializer, <2 x i32> %b, i64 2)
%r = and <4 x i32> %v0, %v1
%sub = call <2 x i32> @<!-- -->llvm.vector.extract.v2i32.v4i32(<4 x i32> %r, i64 2)
store <2 x i32> %sub, ptr addrspace(1) %out_sub, align 8
ret <4 x i32> %r
}
```
Will result:
```asm
insert_extract_and_inreg_v4i32: ; @<!-- -->insert_extract_and_inreg_v4i32
; %bb.0: ; %entry
s_wait_loadcnt_dscnt 0x0
s_wait_kmcnt 0x0
s_and_b64 s[4:5], s[2:3], s[16:17]
v_mov_b32_e32 v2, 0
v_mov_b64_e32 v[0:1], s[4:5]
s_and_b32 s2, s2, s16
s_and_b32 s3, s3, s17
s_delay_alu instid0(SALU_CYCLE_1)
v_mov_b32_e32 v3, s3
global_store_b64 v2, v[0:1], s[0:1]
s_wait_xcnt 0x0
v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0
v_mov_b32_e32 v2, s2
s_set_pc_i64 s[30:31]
```
The reason is `%r` is used for return and `%sub` is used for store, while `narrowInsertExtractVectorBinOp` created a `and` with type of `v2i32`, the original `and` of `v4i32` cannot be erased because of used by return.
This PR add a target specific hook for this `narrowInsertExtractVectorBinOp`. And implement special handling for AMDGPU backend based on the logic that if the insert & extract vector can be reference by register, then they are free and no need for `narrowInsertExtractVectorBinOp`
---
Patch is 26.11 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/201056.diff
14 Files Affected:
- (modified) llvm/include/llvm/CodeGen/TargetLowering.h (+7)
- (modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+2)
- (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+14-1)
- (modified) llvm/lib/Target/AMDGPU/SIISelLowering.h (+2-1)
- (modified) llvm/test/CodeGen/AMDGPU/fmax_legacy.f16.ll (+1-1)
- (modified) llvm/test/CodeGen/AMDGPU/fmin_legacy.f16.ll (+1-1)
- (modified) llvm/test/CodeGen/AMDGPU/llvm.maximum.f16.ll (+4-4)
- (modified) llvm/test/CodeGen/AMDGPU/llvm.minimum.f16.ll (+4-4)
- (modified) llvm/test/CodeGen/AMDGPU/maximumnum.ll (+41-43)
- (modified) llvm/test/CodeGen/AMDGPU/minimumnum.ll (+41-43)
- (modified) llvm/test/CodeGen/AMDGPU/saddsat.ll (+1-1)
- (modified) llvm/test/CodeGen/AMDGPU/ssubsat.ll (+1-1)
- (modified) llvm/test/CodeGen/AMDGPU/uaddsat.ll (+1-1)
- (modified) llvm/test/CodeGen/AMDGPU/usubsat.ll (+1-1)
``````````diff
diff --git a/llvm/include/llvm/CodeGen/TargetLowering.h b/llvm/include/llvm/CodeGen/TargetLowering.h
index 82c47cce0f522..a57fd2ede44ea 100644
--- a/llvm/include/llvm/CodeGen/TargetLowering.h
+++ b/llvm/include/llvm/CodeGen/TargetLowering.h
@@ -3527,6 +3527,13 @@ class LLVM_ABI TargetLoweringBase {
return false;
}
+ /// Return true if it is profitable to narrow a vector binop feeding an
+ /// extract_subvector into a binop of the extracted vector type.
+ virtual bool shouldNarrowExtractedVectorBinOp(EVT SrcVT, EVT SubVT,
+ unsigned Index) const {
+ return true;
+ }
+
/// Return true if extraction of a scalar element from the given vector type
/// at the given index is cheap. For example, if scalar operations occur on
/// the same register file as vector operations, then an extract element may
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 0c9820fb64de9..c4c12fbf5dde6 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -27192,6 +27192,8 @@ static SDValue narrowInsertExtractVectorBinOp(EVT SubVT, SDValue BinOp,
return SDValue();
EVT VecVT = BinOp.getValueType();
+ if (!TLI.shouldNarrowExtractedVectorBinOp(VecVT, SubVT, Index))
+ return SDValue();
SDValue Bop0 = BinOp.getOperand(0), Bop1 = BinOp.getOperand(1);
if (VecVT != Bop0.getValueType() || VecVT != Bop1.getValueType())
return SDValue();
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 86f2479490c29..b735d00be58e7 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -892,7 +892,7 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
for (MVT VT : {MVT::v4bf16, MVT::v8bf16, MVT::v16bf16, MVT::v32bf16})
// Split vector operations.
setOperationAction({ISD::FADD, ISD::FMUL, ISD::FMA, ISD::FCANONICALIZE,
- ISD::FNEG, ISD::FABS},
+ ISD::FNEG, ISD::FABS, ISD::FMAXNUM, ISD::FMINNUM},
VT, Custom);
}
@@ -2354,6 +2354,19 @@ bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
return true;
}
+bool SITargetLowering::shouldNarrowExtractedVectorBinOp(EVT SrcVT, EVT SubVT,
+ unsigned Index) const {
+ unsigned ScalarBits = SubVT.getScalarSizeInBits();
+ unsigned OffsetBits = Index * ScalarBits;
+ unsigned SubvectorBits = SubVT.getSizeInBits();
+
+ // If the extracted subvector starts on a dword boundary and has a dword
+ // sized payload, keeping the original wide result is preferable because the
+ // subvector can be referenced directly through subregisters. Otherwise,
+ // narrowing may avoid element extraction or reconstruction.
+ return OffsetBits % 32 != 0 || SubvectorBits % 32 != 0;
+}
+
bool SITargetLowering::isExtractSubvectorCheap(EVT ResVT, EVT SrcVT,
unsigned Index) const {
if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.h b/llvm/lib/Target/AMDGPU/SIISelLowering.h
index 95ff5bba7cfff..154faefb37ae1 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.h
@@ -399,7 +399,8 @@ class SITargetLowering final : public AMDGPUTargetLowering {
bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
Type *Ty) const override;
-
+ bool shouldNarrowExtractedVectorBinOp(EVT SrcVT, EVT SubVT,
+ unsigned Index) const override;
bool isExtractSubvectorCheap(EVT ResVT, EVT SrcVT,
unsigned Index) const override;
bool isExtractVecEltCheap(EVT VT, unsigned Index) const override;
diff --git a/llvm/test/CodeGen/AMDGPU/fmax_legacy.f16.ll b/llvm/test/CodeGen/AMDGPU/fmax_legacy.f16.ll
index f980288865b8b..8a0cb0d6de739 100644
--- a/llvm/test/CodeGen/AMDGPU/fmax_legacy.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fmax_legacy.f16.ll
@@ -289,8 +289,8 @@ define <3 x half> @test_fmax_legacy_ugt_v3f16_fast(<3 x half> %a, <3 x half> %b)
; GFX9-LABEL: test_fmax_legacy_ugt_v3f16_fast:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_pk_max_f16 v1, v1, v3
; GFX9-NEXT: v_pk_max_f16 v0, v0, v2
+; GFX9-NEXT: v_pk_max_f16 v1, v1, v3
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; VI-LABEL: test_fmax_legacy_ugt_v3f16_fast:
diff --git a/llvm/test/CodeGen/AMDGPU/fmin_legacy.f16.ll b/llvm/test/CodeGen/AMDGPU/fmin_legacy.f16.ll
index 41b7b7fe6b21a..04d9bfd4e9192 100644
--- a/llvm/test/CodeGen/AMDGPU/fmin_legacy.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fmin_legacy.f16.ll
@@ -290,8 +290,8 @@ define <3 x half> @test_fmin_legacy_ule_v3f16_fast(<3 x half> %a, <3 x half> %b)
; GFX9-LABEL: test_fmin_legacy_ule_v3f16_fast:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_pk_min_f16 v1, v1, v3
; GFX9-NEXT: v_pk_min_f16 v0, v0, v2
+; GFX9-NEXT: v_pk_min_f16 v1, v1, v3
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; VI-LABEL: test_fmin_legacy_ule_v3f16_fast:
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.maximum.f16.ll b/llvm/test/CodeGen/AMDGPU/llvm.maximum.f16.ll
index a60741905bdbd..ab5605d6d89db 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.maximum.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.maximum.f16.ll
@@ -1310,8 +1310,8 @@ define <3 x half> @v_maximum_v3f16(<3 x half> %src0, <3 x half> %src1) {
; GFX950-LABEL: v_maximum_v3f16:
; GFX950: ; %bb.0:
; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX950-NEXT: v_pk_maximum3_f16 v1, v1, v3, v3
; GFX950-NEXT: v_pk_maximum3_f16 v0, v0, v2, v2
+; GFX950-NEXT: v_pk_maximum3_f16 v1, v1, v3, v3
; GFX950-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximum_v3f16:
@@ -1419,8 +1419,8 @@ define <3 x half> @v_maximum_v3f16__nnan(<3 x half> %src0, <3 x half> %src1) {
; GFX9-LABEL: v_maximum_v3f16__nnan:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_pk_max_f16 v1, v1, v3
; GFX9-NEXT: v_pk_max_f16 v0, v0, v2
+; GFX9-NEXT: v_pk_max_f16 v1, v1, v3
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximum_v3f16__nnan:
@@ -1524,8 +1524,8 @@ define <3 x half> @v_maximum_v3f16__nsz(<3 x half> %src0, <3 x half> %src1) {
; GFX950-LABEL: v_maximum_v3f16__nsz:
; GFX950: ; %bb.0:
; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX950-NEXT: v_pk_maximum3_f16 v1, v1, v3, v3
; GFX950-NEXT: v_pk_maximum3_f16 v0, v0, v2, v2
+; GFX950-NEXT: v_pk_maximum3_f16 v1, v1, v3, v3
; GFX950-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximum_v3f16__nsz:
@@ -1633,8 +1633,8 @@ define <3 x half> @v_maximum_v3f16__nnan_nsz(<3 x half> %src0, <3 x half> %src1)
; GFX9-LABEL: v_maximum_v3f16__nnan_nsz:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_pk_max_f16 v1, v1, v3
; GFX9-NEXT: v_pk_max_f16 v0, v0, v2
+; GFX9-NEXT: v_pk_max_f16 v1, v1, v3
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximum_v3f16__nnan_nsz:
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.minimum.f16.ll b/llvm/test/CodeGen/AMDGPU/llvm.minimum.f16.ll
index cd01dfc2d3dc8..a262181921f15 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.minimum.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.minimum.f16.ll
@@ -1092,8 +1092,8 @@ define <3 x half> @v_minimum_v3f16(<3 x half> %src0, <3 x half> %src1) {
; GFX950-LABEL: v_minimum_v3f16:
; GFX950: ; %bb.0:
; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX950-NEXT: v_pk_minimum3_f16 v1, v1, v3, v3
; GFX950-NEXT: v_pk_minimum3_f16 v0, v0, v2, v2
+; GFX950-NEXT: v_pk_minimum3_f16 v1, v1, v3, v3
; GFX950-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_minimum_v3f16:
@@ -1180,8 +1180,8 @@ define <3 x half> @v_minimum_v3f16__nnan(<3 x half> %src0, <3 x half> %src1) {
; GFX9-LABEL: v_minimum_v3f16__nnan:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_pk_min_f16 v1, v1, v3
; GFX9-NEXT: v_pk_min_f16 v0, v0, v2
+; GFX9-NEXT: v_pk_min_f16 v1, v1, v3
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_minimum_v3f16__nnan:
@@ -1257,8 +1257,8 @@ define <3 x half> @v_minimum_v3f16__nsz(<3 x half> %src0, <3 x half> %src1) {
; GFX950-LABEL: v_minimum_v3f16__nsz:
; GFX950: ; %bb.0:
; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX950-NEXT: v_pk_minimum3_f16 v1, v1, v3, v3
; GFX950-NEXT: v_pk_minimum3_f16 v0, v0, v2, v2
+; GFX950-NEXT: v_pk_minimum3_f16 v1, v1, v3, v3
; GFX950-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_minimum_v3f16__nsz:
@@ -1345,8 +1345,8 @@ define <3 x half> @v_minimum_v3f16__nnan_nsz(<3 x half> %src0, <3 x half> %src1)
; GFX9-LABEL: v_minimum_v3f16__nnan_nsz:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_pk_min_f16 v1, v1, v3
; GFX9-NEXT: v_pk_min_f16 v0, v0, v2
+; GFX9-NEXT: v_pk_min_f16 v1, v1, v3
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_minimum_v3f16__nnan_nsz:
diff --git a/llvm/test/CodeGen/AMDGPU/maximumnum.ll b/llvm/test/CodeGen/AMDGPU/maximumnum.ll
index c0f80363e6850..4ba8ff42390ac 100644
--- a/llvm/test/CodeGen/AMDGPU/maximumnum.ll
+++ b/llvm/test/CodeGen/AMDGPU/maximumnum.ll
@@ -4301,16 +4301,16 @@ define <3 x half> @v_maximumnum_v3f16(<3 x half> %x, <3 x half> %y) #1 {
; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v4, v0
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
;
-; GFX9-SDAG-LABEL: v_maximumnum_v3f16:
-; GFX9-SDAG: ; %bb.0:
-; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-SDAG-NEXT: v_pk_max_f16 v3, v3, v3
-; GFX9-SDAG-NEXT: v_pk_max_f16 v1, v1, v1
-; GFX9-SDAG-NEXT: v_pk_max_f16 v2, v2, v2
-; GFX9-SDAG-NEXT: v_pk_max_f16 v0, v0, v0
-; GFX9-SDAG-NEXT: v_pk_max_f16 v1, v1, v3
-; GFX9-SDAG-NEXT: v_pk_max_f16 v0, v0, v2
-; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
+; GFX900-SDAG-LABEL: v_maximumnum_v3f16:
+; GFX900-SDAG: ; %bb.0:
+; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX900-SDAG-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX900-SDAG-NEXT: v_pk_max_f16 v0, v0, v0
+; GFX900-SDAG-NEXT: v_pk_max_f16 v0, v0, v2
+; GFX900-SDAG-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX900-SDAG-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX900-SDAG-NEXT: v_pk_max_f16 v1, v1, v2
+; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX900-GISEL-LABEL: v_maximumnum_v3f16:
; GFX900-GISEL: ; %bb.0:
@@ -4323,6 +4323,18 @@ define <3 x half> @v_maximumnum_v3f16(<3 x half> %x, <3 x half> %y) #1 {
; GFX900-GISEL-NEXT: v_pk_max_f16 v1, v1, v2
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
;
+; GFX950-SDAG-LABEL: v_maximumnum_v3f16:
+; GFX950-SDAG: ; %bb.0:
+; GFX950-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX950-SDAG-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX950-SDAG-NEXT: v_pk_max_f16 v0, v0, v0
+; GFX950-SDAG-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX950-SDAG-NEXT: v_pk_max_f16 v0, v0, v2
+; GFX950-SDAG-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX950-SDAG-NEXT: s_nop 0
+; GFX950-SDAG-NEXT: v_pk_max_f16 v1, v1, v2
+; GFX950-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
; GFX950-GISEL-LABEL: v_maximumnum_v3f16:
; GFX950-GISEL: ; %bb.0:
; GFX950-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
@@ -4338,9 +4350,9 @@ define <3 x half> @v_maximumnum_v3f16(<3 x half> %x, <3 x half> %y) #1 {
; GFX10-SDAG-LABEL: v_maximumnum_v3f16:
; GFX10-SDAG: ; %bb.0:
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-SDAG-NEXT: v_pk_max_f16 v3, v3, v3
; GFX10-SDAG-NEXT: v_pk_max_f16 v2, v2, v2
; GFX10-SDAG-NEXT: v_pk_max_f16 v0, v0, v0
+; GFX10-SDAG-NEXT: v_pk_max_f16 v3, v3, v3
; GFX10-SDAG-NEXT: v_pk_max_f16 v1, v1, v1
; GFX10-SDAG-NEXT: v_pk_max_f16 v0, v0, v2
; GFX10-SDAG-NEXT: v_pk_max_f16 v1, v1, v3
@@ -4360,11 +4372,11 @@ define <3 x half> @v_maximumnum_v3f16(<3 x half> %x, <3 x half> %y) #1 {
; GFX11-SDAG-LABEL: v_maximumnum_v3f16:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-SDAG-NEXT: v_pk_max_f16 v3, v3, v3
; GFX11-SDAG-NEXT: v_pk_max_f16 v2, v2, v2
; GFX11-SDAG-NEXT: v_pk_max_f16 v0, v0, v0
+; GFX11-SDAG-NEXT: v_pk_max_f16 v3, v3, v3
; GFX11-SDAG-NEXT: v_pk_max_f16 v1, v1, v1
-; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-SDAG-NEXT: v_pk_max_f16 v0, v0, v2
; GFX11-SDAG-NEXT: v_pk_max_f16 v1, v1, v3
; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
@@ -4384,11 +4396,11 @@ define <3 x half> @v_maximumnum_v3f16(<3 x half> %x, <3 x half> %y) #1 {
; GFX1170-SDAG-LABEL: v_maximumnum_v3f16:
; GFX1170-SDAG: ; %bb.0:
; GFX1170-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX1170-SDAG-NEXT: v_pk_max_num_f16 v3, v3, v3
; GFX1170-SDAG-NEXT: v_pk_max_num_f16 v2, v2, v2
; GFX1170-SDAG-NEXT: v_pk_max_num_f16 v0, v0, v0
+; GFX1170-SDAG-NEXT: v_pk_max_num_f16 v3, v3, v3
; GFX1170-SDAG-NEXT: v_pk_max_num_f16 v1, v1, v1
-; GFX1170-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1170-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX1170-SDAG-NEXT: v_pk_max_num_f16 v0, v0, v2
; GFX1170-SDAG-NEXT: v_pk_max_num_f16 v1, v1, v3
; GFX1170-SDAG-NEXT: s_setpc_b64 s[30:31]
@@ -4412,11 +4424,11 @@ define <3 x half> @v_maximumnum_v3f16(<3 x half> %x, <3 x half> %y) #1 {
; GFX12-SDAG-NEXT: s_wait_samplecnt 0x0
; GFX12-SDAG-NEXT: s_wait_bvhcnt 0x0
; GFX12-SDAG-NEXT: s_wait_kmcnt 0x0
-; GFX12-SDAG-NEXT: v_pk_max_num_f16 v3, v3, v3
; GFX12-SDAG-NEXT: v_pk_max_num_f16 v2, v2, v2
; GFX12-SDAG-NEXT: v_pk_max_num_f16 v0, v0, v0
+; GFX12-SDAG-NEXT: v_pk_max_num_f16 v3, v3, v3
; GFX12-SDAG-NEXT: v_pk_max_num_f16 v1, v1, v1
-; GFX12-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-SDAG-NEXT: v_pk_max_num_f16 v0, v0, v2
; GFX12-SDAG-NEXT: v_pk_max_num_f16 v1, v1, v3
; GFX12-SDAG-NEXT: s_setpc_b64 s[30:31]
@@ -4501,19 +4513,12 @@ define <3 x half> @v_maximumnum_v3f16_nnan(<3 x half> %x, <3 x half> %y) #1 {
; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v4, v0
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
;
-; GFX9-SDAG-LABEL: v_maximumnum_v3f16_nnan:
-; GFX9-SDAG: ; %bb.0:
-; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-SDAG-NEXT: v_pk_max_f16 v1, v1, v3
-; GFX9-SDAG-NEXT: v_pk_max_f16 v0, v0, v2
-; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-GISEL-LABEL: v_maximumnum_v3f16_nnan:
-; GFX9-GISEL: ; %bb.0:
-; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-GISEL-NEXT: v_pk_max_f16 v0, v0, v2
-; GFX9-GISEL-NEXT: v_pk_max_f16 v1, v1, v3
-; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
+; GFX9-LABEL: v_maximumnum_v3f16_nnan:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_pk_max_f16 v0, v0, v2
+; GFX9-NEXT: v_pk_max_f16 v1, v1, v3
+; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v3f16_nnan:
; GFX10: ; %bb.0:
@@ -10079,19 +10084,12 @@ define <3 x half> @v_maximumnum_v3f16_nnan_no_ieee(<3 x half> %x, <3 x half> %y)
; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v4, v0
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
;
-; GFX9-SDAG-LABEL: v_maximumnum_v3f16_nnan_no_ieee:
-; GFX9-SDAG: ; %bb.0:
-; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-SDAG-NEXT: v_pk_max_f16 v1, v1, v3
-; GFX9-SDAG-NEXT: v_pk_max_f16 v0, v0, v2
-; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-GISEL-LABEL: v_maximumnum_v3f16_nnan_no_ieee:
-; GFX9-GISEL: ; %bb.0:
-; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-GISEL-NEXT: v_pk_max_f16 v0, v0, v2
-; GFX9-GISEL-NEXT: v_pk_max_f16 v1, v1, v3
-; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
+; GFX9-LABEL: v_maximumnum_v3f16_nnan_no_ieee:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_pk_max_f16 v0, v0, v2
+; GFX9-NEXT: v_pk_max_f16 v1, v1, v3
+; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v3f16_nnan_no_ieee:
; GFX10: ; %bb.0:
diff --git a/llvm/test/CodeGen/AMDGPU/minimumnum.ll b/llvm/test/CodeGen/AMDGPU/minimumnum.ll
index 18ea5e2dd0e6c..187e8b26d4394 100644
--- a/llvm/test/CodeGen/AMDGPU/minimumnum.ll
+++ b/llvm/test/CodeGen/AMDGPU/minimumnum.ll
@@ -4091,16 +4091,16 @@ define <3 x half> @v_minimumnum_v3f16(<3 x half> %x, <3 x half> %y) #1 {
; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v4, v0
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
;
-; GFX9-SDAG-LABEL: v_minimumnum_v3f16:
-; GFX9-SDAG: ; %bb.0:
-; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-SDAG-NEXT: v_pk_max_f16 v3, v3, v3
-; GFX9-SDAG-NEXT: v_pk_max_f16 v1, v1, v1
-; GFX9-SDAG-NEXT: v_pk_max_f16 v2, v2, v2
-; GFX9-SDAG-NEXT: v_pk_max_f16 v0, v0, v0
-; GFX9-SDAG-NEXT: v_pk_min_f16 v1, v1, v3
-; GFX9-SDAG-NEXT: v_pk_min_f16 v0, v0, v2
-; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
+; GFX900-SDAG-LABEL: v_minimumnum_v3f16:
+; GFX900-SDAG: ; %bb.0:
+; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX900-SDAG-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX900-SDAG-NEXT: v_pk_max_f16 v0, v0, v0
+; GFX900-SDAG-NEXT: v_pk_min_f16 v0, v0, v2
+; GFX900-SDAG-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX900-SDAG-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX900-SDAG-NEXT: v_pk_min_f16 v1, v1, v2
+; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX900-GISEL-LABEL: v_minimumnum_v3f16:
; GFX900-GISEL: ; %bb.0:
@@ -4113,6 +4113,18 @@ define <3 x half> @v_minimumnum_v3f16(<3 x half> %x, <3 x half> %y) #1 {
; GFX900-GISEL-NEXT: v_pk_min_f16 v1, v1, v2
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
;
+; GFX950-SDAG-LABEL: v_minimumnum_v3f16:
+; GFX950-SDAG: ; %bb.0:
+; GFX950-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX950-SDAG-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX950-SDAG-NEXT: v_pk_max_f16 v0, v0, v0
+; GFX950-SDAG-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX950-SDAG-NEXT: v_pk_min_f16 v0, v0, v2
+; GFX950-SDAG-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX950-SDAG-NEXT: s_nop 0
+; GFX950-SDAG-NEXT: v_pk_min_f16 v1, v1, v2
+; GFX950-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
; GFX950-GISEL-LABEL: v_minimumnum_v3f16:
; GFX950-GISEL: ; %bb.0:
; GFX950-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
@@ -4128,9 +4140,9 @@ define <3 x half> @v_minimumnum_v3f16(<3 x half> %x, <3 x half> %y) #1 {
; GFX10-SDAG-LABEL: v_minimumnum_v3f16:
; GFX10-SDAG: ; %bb.0:
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-SDAG-NEXT: v_pk_max_f16 v3, v3, v3
; GFX10-SDAG-NEXT: v_pk_max_f16 v2, v2, v2
; GFX10-SDAG-NEXT: v_pk_max_f16 v0, v0, v0
+; GFX10-SDAG-NEXT: v_pk_max_f16 v3, v3, v3
; GFX10-SDAG-NEXT: v_pk_max_f16 v1, v1, v1
; GFX10-SDAG-NEXT: v_pk_min_f16 v0, v0, v2
; GFX10-SDAG-NEXT: v_pk_min_f16 v1, v1, v3
@@ -4150,11 +4162,11 @@ define <3 x half> @v_minimumnum_v3f16(<3 x half> %x, <3 x half> %y) #1 {
; GFX11-SDAG-LABEL: v_minimumnum_v3f16:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-SDAG-NEXT: v_pk_max_f16 v3, v3, v3
; GFX11-SDAG-NEXT: v_pk_max_f16 v2, v2, v2
; GFX11-SDAG-NEXT: v_pk_max_f16 v0, v0, v0
+; GFX11-SDAG-NEXT: v_pk_max_f16 v3, v3, v3
; GFX11-SDAG-NEXT: v_pk_max_f16 v1, v1, v1
-; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-SDAG-NEXT: v_pk_min_f16 v0, v0, v2
; GFX11-SDAG-NEXT: v_pk_min_f16 v1, v1, v3
; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
@@ -4174,11 +4186,11 @@ define <3 x half> @v_minimumnum_v3f16(<3 x half> %x, <3...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/201056
More information about the llvm-commits
mailing list