[llvm] WIP: [AMDGPU] Remove `UnsafeFPMath` (PR #151079)
via llvm-commits
llvm-commits at lists.llvm.org
Tue Jul 29 02:22:17 PDT 2025
https://github.com/paperchalice updated https://github.com/llvm/llvm-project/pull/151079
>From 29001f0f4dd552ef154d66e29c04a425da5abdf5 Mon Sep 17 00:00:00 2001
From: PaperChalice <liujunchang97 at outlook.com>
Date: Tue, 29 Jul 2025 12:44:26 +0800
Subject: [PATCH 1/4] Remove `UnsafeFPMath` in `AMDGPUCodeGenPrepare`
---
.../Target/AMDGPU/AMDGPUCodeGenPrepare.cpp | 15 +---
llvm/test/CodeGen/AMDGPU/llvm.amdgcn.rcp.ll | 16 ++--
llvm/test/CodeGen/AMDGPU/rsq.f32.ll | 90 +++++++++----------
3 files changed, 57 insertions(+), 64 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
index 5f1983791cfae..7e2c67d22cf6e 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
@@ -89,10 +89,6 @@ static cl::opt<bool> DisableFDivExpand(
cl::ReallyHidden,
cl::init(false));
-static bool hasUnsafeFPMath(const Function &F) {
- return F.getFnAttribute("unsafe-fp-math").getValueAsBool();
-}
-
class AMDGPUCodeGenPrepareImpl
: public InstVisitor<AMDGPUCodeGenPrepareImpl, bool> {
public:
@@ -104,7 +100,6 @@ class AMDGPUCodeGenPrepareImpl
const DominatorTree *DT;
const UniformityInfo &UA;
const DataLayout &DL;
- const bool HasUnsafeFPMath;
const bool HasFP32DenormalFlush;
bool FlowChanged = false;
mutable Function *SqrtF32 = nullptr;
@@ -117,7 +112,6 @@ class AMDGPUCodeGenPrepareImpl
const DominatorTree *DT, const UniformityInfo &UA)
: F(F), ST(TM.getSubtarget<GCNSubtarget>(F)), TM(TM), TLI(TLI), AC(AC),
DT(DT), UA(UA), DL(F.getDataLayout()),
- HasUnsafeFPMath(hasUnsafeFPMath(F)),
HasFP32DenormalFlush(SIModeRegisterDefaults(F, ST).FP32Denormals ==
DenormalMode::getPreserveSign()) {}
@@ -637,8 +631,7 @@ bool AMDGPUCodeGenPrepareImpl::canOptimizeWithRsq(const FPMathOperator *SqrtOp,
return false;
// v_rsq_f32 gives 1ulp
- return SqrtFMF.approxFunc() || HasUnsafeFPMath ||
- SqrtOp->getFPAccuracy() >= 1.0f;
+ return SqrtFMF.approxFunc() || SqrtOp->getFPAccuracy() >= 1.0f;
}
Value *AMDGPUCodeGenPrepareImpl::optimizeWithRsq(
@@ -664,7 +657,7 @@ Value *AMDGPUCodeGenPrepareImpl::optimizeWithRsq(
IRBuilder<>::FastMathFlagGuard Guard(Builder);
Builder.setFastMathFlags(DivFMF | SqrtFMF);
- if ((DivFMF.approxFunc() && SqrtFMF.approxFunc()) || HasUnsafeFPMath ||
+ if ((DivFMF.approxFunc() && SqrtFMF.approxFunc()) ||
canIgnoreDenormalInput(Den, CtxI)) {
Value *Result = Builder.CreateUnaryIntrinsic(Intrinsic::amdgcn_rsq, Den);
// -1.0 / sqrt(x) -> fneg(rsq(x))
@@ -852,7 +845,7 @@ bool AMDGPUCodeGenPrepareImpl::visitFDiv(BinaryOperator &FDiv) {
// expansion of afn to codegen. The current interpretation is so aggressive we
// don't need any pre-consideration here when we have better information. A
// more conservative interpretation could use handling here.
- const bool AllowInaccurateRcp = HasUnsafeFPMath || DivFMF.approxFunc();
+ const bool AllowInaccurateRcp = DivFMF.approxFunc();
if (!RsqOp && AllowInaccurateRcp)
return false;
@@ -2026,7 +2019,7 @@ bool AMDGPUCodeGenPrepareImpl::visitSqrt(IntrinsicInst &Sqrt) {
// We're trying to handle the fast-but-not-that-fast case only. The lowering
// of fast llvm.sqrt will give the raw instruction anyway.
- if (SqrtFMF.approxFunc() || HasUnsafeFPMath)
+ if (SqrtFMF.approxFunc())
return false;
const float ReqdAccuracy = FPOp->getFPAccuracy();
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.rcp.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.rcp.ll
index 425a8530afa97..477f0a610feec 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.rcp.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.rcp.ll
@@ -51,7 +51,7 @@ define amdgpu_kernel void @safe_no_fp32_denormals_rcp_f32(ptr addrspace(1) %out,
; SI-NOT: [[RESULT]]
; SI: buffer_store_dword [[RESULT]]
define amdgpu_kernel void @safe_f32_denormals_rcp_pat_f32(ptr addrspace(1) %out, float %src) #4 {
- %rcp = fdiv float 1.0, %src, !fpmath !0
+ %rcp = fdiv afn float 1.0, %src, !fpmath !0
store float %rcp, ptr addrspace(1) %out, align 4
ret void
}
@@ -105,8 +105,8 @@ define amdgpu_kernel void @safe_rsq_rcp_pat_amdgcn_sqrt_f32_nocontract(ptr addrs
; SI: v_sqrt_f32_e32
; SI: v_rcp_f32_e32
define amdgpu_kernel void @unsafe_rsq_rcp_pat_f32(ptr addrspace(1) %out, float %src) #2 {
- %sqrt = call float @llvm.sqrt.f32(float %src)
- %rcp = call float @llvm.amdgcn.rcp.f32(float %sqrt)
+ %sqrt = call afn float @llvm.sqrt.f32(float %src)
+ %rcp = call afn float @llvm.amdgcn.rcp.f32(float %sqrt)
store float %rcp, ptr addrspace(1) %out, align 4
ret void
}
@@ -148,7 +148,7 @@ define amdgpu_kernel void @rcp_pat_f64(ptr addrspace(1) %out, double %src) #1 {
; SI: v_fma_f64
; SI: v_fma_f64
define amdgpu_kernel void @unsafe_rcp_pat_f64(ptr addrspace(1) %out, double %src) #2 {
- %rcp = fdiv double 1.0, %src
+ %rcp = fdiv afn double 1.0, %src
store double %rcp, ptr addrspace(1) %out, align 8
ret void
}
@@ -214,9 +214,9 @@ define amdgpu_kernel void @unsafe_amdgcn_sqrt_rsq_rcp_pat_f64(ptr addrspace(1) %
}
attributes #0 = { nounwind readnone }
-attributes #1 = { nounwind "unsafe-fp-math"="false" "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
-attributes #2 = { nounwind "unsafe-fp-math"="true" "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
-attributes #3 = { nounwind "unsafe-fp-math"="false" "denormal-fp-math-f32"="ieee,ieee" }
-attributes #4 = { nounwind "unsafe-fp-math"="true" "denormal-fp-math-f32"="ieee,ieee" }
+attributes #1 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
+attributes #2 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
+attributes #3 = { nounwind "denormal-fp-math-f32"="ieee,ieee" }
+attributes #4 = { nounwind "denormal-fp-math-f32"="ieee,ieee" }
!0 = !{float 2.500000e+00}
diff --git a/llvm/test/CodeGen/AMDGPU/rsq.f32.ll b/llvm/test/CodeGen/AMDGPU/rsq.f32.ll
index f7e0388561104..f967e951b27a4 100644
--- a/llvm/test/CodeGen/AMDGPU/rsq.f32.ll
+++ b/llvm/test/CodeGen/AMDGPU/rsq.f32.ll
@@ -1,10 +1,10 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
-; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=tahiti -denormal-fp-math-f32=preserve-sign -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=GCN-DAZ,GCN-DAZ-UNSAFE,SI-DAZ-UNSAFE %s
-; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=tahiti -denormal-fp-math-f32=ieee -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=GCN-IEEE,GCN-IEEE-UNSAFE,SI-IEEE-UNSAFE %s
+; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=tahiti -denormal-fp-math-f32=preserve-sign < %s | FileCheck -check-prefixes=GCN-DAZ,GCN-DAZ-UNSAFE,SI-DAZ-UNSAFE %s
+; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=tahiti -denormal-fp-math-f32=ieee < %s | FileCheck -check-prefixes=GCN-IEEE,GCN-IEEE-UNSAFE,SI-IEEE-UNSAFE %s
-; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=hawaii -denormal-fp-math-f32=preserve-sign -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=GCN-DAZ,GCN-DAZ-UNSAFE,CI-DAZ-UNSAFE %s
-; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=hawaii -denormal-fp-math-f32=ieee -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=GCN-IEEE,GCN-IEEE-UNSAFE,CI-IEEE-UNSAFE %s
+; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=hawaii -denormal-fp-math-f32=preserve-sign < %s | FileCheck -check-prefixes=GCN-DAZ,GCN-DAZ-UNSAFE,CI-DAZ-UNSAFE %s
+; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=hawaii -denormal-fp-math-f32=ieee < %s | FileCheck -check-prefixes=GCN-IEEE,GCN-IEEE-UNSAFE,CI-IEEE-UNSAFE %s
declare i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
@@ -65,8 +65,8 @@ define amdgpu_kernel void @rsq_f32(ptr addrspace(1) noalias %out, ptr addrspace(
; GCN-UNSAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-UNSAFE-NEXT: s_endpgm
%val = load float, ptr addrspace(1) %in, align 4
- %sqrt = call contract float @llvm.sqrt.f32(float %val) nounwind readnone
- %div = fdiv contract float 1.0, %sqrt, !fpmath !0
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val) nounwind readnone
+ %div = fdiv afn contract float 1.0, %sqrt, !fpmath !0
store float %div, ptr addrspace(1) %out, align 4
ret void
}
@@ -103,8 +103,8 @@ define amdgpu_kernel void @rsq_f32_sgpr(ptr addrspace(1) noalias %out, float %va
; GCN-UNSAFE-NEXT: s_mov_b32 s2, -1
; GCN-UNSAFE-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GCN-UNSAFE-NEXT: s_endpgm
- %sqrt = call contract float @llvm.sqrt.f32(float %val) nounwind readnone
- %div = fdiv contract float 1.0, %sqrt, !fpmath !0
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val) nounwind readnone
+ %div = fdiv afn contract float 1.0, %sqrt, !fpmath !0
store float %div, ptr addrspace(1) %out, align 4
ret void
}
@@ -196,7 +196,7 @@ define amdgpu_kernel void @rsqrt_fmul(ptr addrspace(1) %out, ptr addrspace(1) %i
%x = call contract float @llvm.sqrt.f32(float %a)
%y = fmul contract float %x, %b
- %z = fdiv arcp contract float %c, %y
+ %z = fdiv arcp afn contract float %c, %y
store float %z, ptr addrspace(1) %out.gep
ret void
}
@@ -258,8 +258,8 @@ define amdgpu_kernel void @neg_rsq_f32(ptr addrspace(1) noalias %out, ptr addrsp
; GCN-UNSAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-UNSAFE-NEXT: s_endpgm
%val = load float, ptr addrspace(1) %in, align 4
- %sqrt = call contract float @llvm.sqrt.f32(float %val)
- %div = fdiv contract float -1.0, %sqrt, !fpmath !0
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val)
+ %div = fdiv afn contract float -1.0, %sqrt, !fpmath !0
store float %div, ptr addrspace(1) %out, align 4
ret void
}
@@ -322,8 +322,8 @@ define amdgpu_kernel void @neg_rsq_neg_f32(ptr addrspace(1) noalias %out, ptr ad
; GCN-UNSAFE-NEXT: s_endpgm
%val = load float, ptr addrspace(1) %in, align 4
%val.fneg = fneg float %val
- %sqrt = call contract float @llvm.sqrt.f32(float %val.fneg)
- %div = fdiv contract float -1.0, %sqrt, !fpmath !0
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val.fneg)
+ %div = fdiv afn contract float -1.0, %sqrt, !fpmath !0
store float %div, ptr addrspace(1) %out, align 4
ret void
}
@@ -343,8 +343,8 @@ define float @v_neg_rsq_neg_f32(float %val) {
; GCN-IEEE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
; GCN-IEEE-NEXT: s_setpc_b64 s[30:31]
%val.fneg = fneg float %val
- %sqrt = call contract float @llvm.sqrt.f32(float %val.fneg)
- %div = fdiv contract float -1.0, %sqrt, !fpmath !0
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val.fneg)
+ %div = fdiv afn contract float -1.0, %sqrt, !fpmath !0
ret float %div
}
@@ -367,8 +367,8 @@ define <2 x float> @v_neg_rsq_neg_v2f32(<2 x float> %val) {
; GCN-IEEE-NEXT: v_xor_b32_e32 v1, 0x80000000, v1
; GCN-IEEE-NEXT: s_setpc_b64 s[30:31]
%val.fneg = fneg <2 x float> %val
- %sqrt = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> %val.fneg)
- %div = fdiv contract <2 x float> <float -1.0, float -1.0>, %sqrt, !fpmath !0
+ %sqrt = call afn contract <2 x float> @llvm.sqrt.v2f32(<2 x float> %val.fneg)
+ %div = fdiv afn contract <2 x float> <float -1.0, float -1.0>, %sqrt, !fpmath !0
ret <2 x float> %div
}
@@ -387,8 +387,8 @@ define float @v_neg_rsq_neg_f32_foldable_user(float %val0, float %val1) {
; GCN-IEEE-NEXT: v_mul_f32_e64 v0, -v0, v1
; GCN-IEEE-NEXT: s_setpc_b64 s[30:31]
%val0.neg = fneg float %val0
- %sqrt = call contract float @llvm.sqrt.f32(float %val0.neg)
- %div = fdiv contract float -1.0, %sqrt, !fpmath !0
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val0.neg)
+ %div = fdiv afn contract float -1.0, %sqrt, !fpmath !0
%user = fmul contract float %div, %val1
ret float %user
}
@@ -412,8 +412,8 @@ define <2 x float> @v_neg_rsq_neg_v2f32_foldable_user(<2 x float> %val0, <2 x fl
; GCN-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v3
; GCN-IEEE-NEXT: s_setpc_b64 s[30:31]
%val0.fneg = fneg <2 x float> %val0
- %sqrt = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> %val0.fneg)
- %div = fdiv contract <2 x float> <float -1.0, float -1.0>, %sqrt, !fpmath !0
+ %sqrt = call afn contract <2 x float> @llvm.sqrt.v2f32(<2 x float> %val0.fneg)
+ %div = fdiv afn contract <2 x float> <float -1.0, float -1.0>, %sqrt, !fpmath !0
%user = fmul contract <2 x float> %div, %val1
ret <2 x float> %user
}
@@ -432,8 +432,8 @@ define float @v_neg_rsq_f32(float %val) {
; GCN-IEEE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
; GCN-IEEE-NEXT: s_setpc_b64 s[30:31]
- %sqrt = call contract float @llvm.sqrt.f32(float %val)
- %div = fdiv contract float -1.0, %sqrt, !fpmath !0
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val)
+ %div = fdiv afn contract float -1.0, %sqrt, !fpmath !0
ret float %div
}
@@ -455,8 +455,8 @@ define <2 x float> @v_neg_rsq_v2f32(<2 x float> %val) {
; GCN-IEEE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
; GCN-IEEE-NEXT: v_xor_b32_e32 v1, 0x80000000, v1
; GCN-IEEE-NEXT: s_setpc_b64 s[30:31]
- %sqrt = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> %val)
- %div = fdiv contract <2 x float> <float -1.0, float -1.0>, %sqrt, !fpmath !0
+ %sqrt = call afn contract <2 x float> @llvm.sqrt.v2f32(<2 x float> %val)
+ %div = fdiv afn contract <2 x float> <float -1.0, float -1.0>, %sqrt, !fpmath !0
ret <2 x float> %div
}
@@ -474,8 +474,8 @@ define float @v_neg_rsq_f32_foldable_user(float %val0, float %val1) {
; GCN-IEEE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-NEXT: v_mul_f32_e64 v0, -v0, v1
; GCN-IEEE-NEXT: s_setpc_b64 s[30:31]
- %sqrt = call contract float @llvm.sqrt.f32(float %val0)
- %div = fdiv contract float -1.0, %sqrt, !fpmath !0
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val0)
+ %div = fdiv afn contract float -1.0, %sqrt, !fpmath !0
%user = fmul contract float %div, %val1
ret float %user
}
@@ -643,8 +643,8 @@ define <2 x float> @v_neg_rsq_v2f32_foldable_user(<2 x float> %val0, <2 x float>
; CI-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v2
; CI-IEEE-SAFE-NEXT: v_mul_f32_e32 v1, v1, v3
; CI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
- %sqrt = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> %val0)
- %div = fdiv contract <2 x float> <float -1.0, float -1.0>, %sqrt, !fpmath !0
+ %sqrt = call afn contract <2 x float> @llvm.sqrt.v2f32(<2 x float> %val0)
+ %div = fdiv afn contract <2 x float> <float -1.0, float -1.0>, %sqrt, !fpmath !0
%user = fmul contract <2 x float> %div, %val1
ret <2 x float> %user
}
@@ -672,8 +672,8 @@ define float @v_rsq_f32(float %val) {
; GCN-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v1, 0, 12, vcc
; GCN-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v0, v1
; GCN-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
- %sqrt = call contract float @llvm.sqrt.f32(float %val), !fpmath !1
- %div = fdiv contract float 1.0, %sqrt, !fpmath !1
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val), !fpmath !1
+ %div = fdiv afn contract float 1.0, %sqrt, !fpmath !1
ret float %div
}
@@ -756,9 +756,9 @@ define { float, float } @v_rsq_f32_multi_use(float %val) {
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v2, vcc, 0, v2
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v1, v1, v2
; CI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
- %sqrt = call contract float @llvm.sqrt.f32(float %val), !fpmath !1
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val), !fpmath !1
%insert.0 = insertvalue { float, float } poison, float %sqrt, 0
- %div = fdiv arcp contract float 1.0, %sqrt, !fpmath !1
+ %div = fdiv arcp afn contract float 1.0, %sqrt, !fpmath !1
%insert.1 = insertvalue { float, float } %insert.0, float %div, 1
ret { float, float } %insert.1
}
@@ -838,8 +838,8 @@ define float @v_rsq_f32_missing_contract0(float %val) {
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; CI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
- %sqrt = call float @llvm.sqrt.f32(float %val), !fpmath !1
- %div = fdiv arcp contract float 1.0, %sqrt, !fpmath !1
+ %sqrt = call afn float @llvm.sqrt.f32(float %val), !fpmath !1
+ %div = fdiv arcp afn contract float 1.0, %sqrt, !fpmath !1
ret float %div
}
@@ -855,8 +855,8 @@ define float @v_rsq_f32_missing_contract1(float %val) {
; GCN-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-NEXT: s_setpc_b64 s[30:31]
- %sqrt = call contract float @llvm.sqrt.f32(float %val), !fpmath !1
- %div = fdiv arcp float 1.0, %sqrt, !fpmath !1
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val), !fpmath !1
+ %div = fdiv arcp afn float 1.0, %sqrt, !fpmath !1
ret float %div
}
@@ -876,8 +876,8 @@ define float @v_rsq_f32_contractable_user(float %val0, float %val1) {
; GCN-IEEE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-NEXT: v_add_f32_e32 v0, v0, v1
; GCN-IEEE-NEXT: s_setpc_b64 s[30:31]
- %sqrt = call contract float @llvm.sqrt.f32(float %val0), !fpmath !1
- %div = fdiv contract float 1.0, %sqrt, !fpmath !1
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val0), !fpmath !1
+ %div = fdiv afn contract float 1.0, %sqrt, !fpmath !1
%add = fadd contract float %div, %val1
ret float %add
}
@@ -897,8 +897,8 @@ define float @v_rsq_f32_contractable_user_missing_contract0(float %val0, float %
; GCN-IEEE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-NEXT: v_add_f32_e32 v0, v0, v1
; GCN-IEEE-NEXT: s_setpc_b64 s[30:31]
- %sqrt = call contract float @llvm.sqrt.f32(float %val0), !fpmath !1
- %div = fdiv contract float 1.0, %sqrt, !fpmath !1
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val0), !fpmath !1
+ %div = fdiv afn contract float 1.0, %sqrt, !fpmath !1
%add = fadd contract float %div, %val1
ret float %add
}
@@ -918,8 +918,8 @@ define float @v_rsq_f32_contractable_user_missing_contract1(float %val0, float %
; GCN-IEEE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-NEXT: v_add_f32_e32 v0, v0, v1
; GCN-IEEE-NEXT: s_setpc_b64 s[30:31]
- %sqrt = call contract float @llvm.sqrt.f32(float %val0), !fpmath !1
- %div = fdiv contract float 1.0, %sqrt, !fpmath !1
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val0), !fpmath !1
+ %div = fdiv afn contract float 1.0, %sqrt, !fpmath !1
%add = fadd float %div, %val1
ret float %add
}
@@ -953,8 +953,8 @@ define float @v_rsq_f32_known_never_posdenormal(float nofpclass(psub) %val) {
; GCN-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-NEXT: s_setpc_b64 s[30:31]
- %sqrt = call contract float @llvm.sqrt.f32(float %val), !fpmath !1
- %div = fdiv contract float 1.0, %sqrt, !fpmath !1
+ %sqrt = call afn contract float @llvm.sqrt.f32(float %val), !fpmath !1
+ %div = fdiv afn contract float 1.0, %sqrt, !fpmath !1
ret float %div
}
>From f8bb469a294c7de4e989b0039c7f0f6f19851a98 Mon Sep 17 00:00:00 2001
From: PaperChalice <liujunchang97 at outlook.com>
Date: Tue, 29 Jul 2025 14:44:15 +0800
Subject: [PATCH 2/4] Remove `UnsafeFPMath` in `AMDGPUInstructions.td`
---
llvm/lib/Target/AMDGPU/AMDGPUInstructions.td | 1 -
1 file changed, 1 deletion(-)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
index 7a50923ffedc6..511fc6967da31 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
@@ -94,7 +94,6 @@ def NoFP32Denormals : Predicate<"MF->getInfo<SIMachineFunctionInfo>()->getMode()
def NoFP64Denormals : Predicate<"MF->getInfo<SIMachineFunctionInfo>()->getMode().FP64FP16Denormals == DenormalMode::getPreserveSign()">;
def IEEEModeEnabled : Predicate<"MF->getInfo<SIMachineFunctionInfo>()->getMode().IEEE">;
def IEEEModeDisabled : Predicate<"!MF->getInfo<SIMachineFunctionInfo>()->getMode().IEEE">;
-def UnsafeFPMath : Predicate<"TM.Options.UnsafeFPMath">;
}
def FMA : Predicate<"Subtarget->hasFMA()">;
>From 226f3c4fa548b78c6cccd5cf594334976755e9b7 Mon Sep 17 00:00:00 2001
From: PaperChalice <liujunchang97 at outlook.com>
Date: Tue, 29 Jul 2025 16:15:16 +0800
Subject: [PATCH 3/4] Remove `UnsafeFPMath` in `*ISelLowering.cpp`
---
llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp | 6 +-
llvm/lib/Target/AMDGPU/SIISelLowering.cpp | 14 +-
llvm/test/CodeGen/AMDGPU/fdiv.f16.ll | 4 +-
llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll | 1045 +++--------------
llvm/test/CodeGen/AMDGPU/fptrunc.ll | 633 ++++------
llvm/test/CodeGen/AMDGPU/fsqrt.f32.ll | 5 +-
llvm/test/CodeGen/AMDGPU/llvm.exp.ll | 4 +-
llvm/test/CodeGen/AMDGPU/llvm.exp10.ll | 4 +-
llvm/test/CodeGen/AMDGPU/llvm.log.ll | 4 +-
llvm/test/CodeGen/AMDGPU/llvm.log10.ll | 4 +-
llvm/test/CodeGen/AMDGPU/rcp-pattern.ll | 8 +-
llvm/test/CodeGen/AMDGPU/rsq.f64.ll | 13 +-
12 files changed, 368 insertions(+), 1376 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 61189337e5233..31c4f62d24dfe 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -2634,7 +2634,7 @@ bool AMDGPUTargetLowering::allowApproxFunc(const SelectionDAG &DAG,
if (Flags.hasApproximateFuncs())
return true;
auto &Options = DAG.getTarget().Options;
- return Options.UnsafeFPMath || Options.ApproxFuncFPMath;
+ return Options.ApproxFuncFPMath;
}
bool AMDGPUTargetLowering::needsDenormHandlingF32(const SelectionDAG &DAG,
@@ -2757,7 +2757,7 @@ SDValue AMDGPUTargetLowering::LowerFLOGCommon(SDValue Op,
const auto &Options = getTargetMachine().Options;
if (VT == MVT::f16 || Flags.hasApproximateFuncs() ||
- Options.ApproxFuncFPMath || Options.UnsafeFPMath) {
+ Options.ApproxFuncFPMath) {
if (VT == MVT::f16 && !Subtarget->has16BitInsts()) {
// Log and multiply in f32 is good enough for f16.
@@ -3585,7 +3585,7 @@ SDValue AMDGPUTargetLowering::LowerFP_TO_FP16(SDValue Op, SelectionDAG &DAG) con
if (N0.getValueType() == MVT::f32)
return DAG.getNode(AMDGPUISD::FP_TO_FP16, DL, Op.getValueType(), N0);
- if (getTargetMachine().Options.UnsafeFPMath) {
+ if (Op->getFlags().hasApproximateFuncs()) {
// There is a generic expand for FP_TO_FP16 with unsafe fast math.
return SDValue();
}
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 9017f4f26f835..1ae4a927ad9fe 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -7148,7 +7148,7 @@ SDValue SITargetLowering::lowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, FpToFp16);
return DAG.getNode(ISD::BITCAST, DL, MVT::f16, Trunc);
}
- if (getTargetMachine().Options.UnsafeFPMath) {
+ if (Op->getFlags().hasApproximateFuncs()) {
SDValue Flags = Op.getOperand(1);
SDValue Src32 = DAG.getNode(ISD::FP_ROUND, DL, MVT::f32, Src, Flags);
return DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, Src32, Flags);
@@ -11243,8 +11243,7 @@ SDValue SITargetLowering::lowerFastUnsafeFDIV(SDValue Op,
EVT VT = Op.getValueType();
const SDNodeFlags Flags = Op->getFlags();
- bool AllowInaccurateRcp =
- Flags.hasApproximateFuncs() || DAG.getTarget().Options.UnsafeFPMath;
+ bool AllowInaccurateRcp = Flags.hasApproximateFuncs();
if (const ConstantFPSDNode *CLHS = dyn_cast<ConstantFPSDNode>(LHS)) {
// Without !fpmath accuracy information, we can't do more because we don't
@@ -11263,7 +11262,7 @@ SDValue SITargetLowering::lowerFastUnsafeFDIV(SDValue Op,
// 1.0 / sqrt(x) -> rsq(x)
- // XXX - Is UnsafeFPMath sufficient to do this for f64? The maximum ULP
+ // XXX - Is afn sufficient to do this for f64? The maximum ULP
// error seems really high at 2^29 ULP.
// 1.0 / x -> rcp(x)
return DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
@@ -11297,8 +11296,7 @@ SDValue SITargetLowering::lowerFastUnsafeFDIV64(SDValue Op,
EVT VT = Op.getValueType();
const SDNodeFlags Flags = Op->getFlags();
- bool AllowInaccurateDiv =
- Flags.hasApproximateFuncs() || DAG.getTarget().Options.UnsafeFPMath;
+ bool AllowInaccurateDiv = Flags.hasApproximateFuncs();
if (!AllowInaccurateDiv)
return SDValue();
@@ -14550,7 +14548,7 @@ unsigned SITargetLowering::getFusedOpcode(const SelectionDAG &DAG,
return ISD::FMAD;
const TargetOptions &Options = DAG.getTarget().Options;
- if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath ||
+ if ((Options.AllowFPOpFusion == FPOpFusion::Fast ||
(N0->getFlags().hasAllowContract() &&
N1->getFlags().hasAllowContract())) &&
isFMAFasterThanFMulAndFAdd(DAG.getMachineFunction(), VT)) {
@@ -15675,7 +15673,7 @@ SDValue SITargetLowering::performFMACombine(SDNode *N,
// regardless of the denorm mode setting. Therefore,
// unsafe-fp-math/fp-contract is sufficient to allow generating fdot2.
const TargetOptions &Options = DAG.getTarget().Options;
- if (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath ||
+ if (Options.AllowFPOpFusion == FPOpFusion::Fast ||
(N->getFlags().hasAllowContract() &&
FMA->getFlags().hasAllowContract())) {
Op1 = Op1.getOperand(0);
diff --git a/llvm/test/CodeGen/AMDGPU/fdiv.f16.ll b/llvm/test/CodeGen/AMDGPU/fdiv.f16.ll
index 9ae9d1977bd11..210e09fd9169a 100644
--- a/llvm/test/CodeGen/AMDGPU/fdiv.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fdiv.f16.ll
@@ -1702,7 +1702,7 @@ entry:
%gep.r = getelementptr inbounds half, ptr addrspace(1) %r, i64 %tid.ext
%a.val = load volatile half, ptr addrspace(1) %gep.a
%b.val = load volatile half, ptr addrspace(1) %gep.b
- %r.val = fdiv half %a.val, %b.val
+ %r.val = fdiv afn half %a.val, %b.val
store half %r.val, ptr addrspace(1) %gep.r
ret void
}
@@ -2475,4 +2475,4 @@ declare <2 x half> @llvm.sqrt.v2f16(<2 x half>) #2
attributes #0 = { nounwind }
attributes #1 = { nounwind readnone }
-attributes #2 = { nounwind "unsafe-fp-math"="true" }
+attributes #2 = { nounwind }
diff --git a/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll b/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll
index 57b4857776246..1d33c26686528 100644
--- a/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll
@@ -1101,62 +1101,21 @@ entry:
define amdgpu_kernel void @fptrunc_f64_to_f16_afn(
; SI-SDAG-LABEL: fptrunc_f64_to_f16_afn:
; SI-SDAG: ; %bb.0: ; %entry
-; SI-SDAG-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9
-; SI-SDAG-NEXT: s_mov_b32 s3, 0xf000
-; SI-SDAG-NEXT: s_mov_b32 s2, -1
-; SI-SDAG-NEXT: s_mov_b32 s10, s2
-; SI-SDAG-NEXT: s_mov_b32 s11, s3
+; SI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
+; SI-SDAG-NEXT: s_mov_b32 s7, 0xf000
+; SI-SDAG-NEXT: s_mov_b32 s6, -1
+; SI-SDAG-NEXT: s_mov_b32 s10, s6
+; SI-SDAG-NEXT: s_mov_b32 s11, s7
; SI-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; SI-SDAG-NEXT: s_mov_b32 s8, s6
-; SI-SDAG-NEXT: s_mov_b32 s9, s7
+; SI-SDAG-NEXT: s_mov_b32 s8, s2
+; SI-SDAG-NEXT: s_mov_b32 s9, s3
; SI-SDAG-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
-; SI-SDAG-NEXT: s_movk_i32 s0, 0x7e00
+; SI-SDAG-NEXT: s_mov_b32 s4, s0
+; SI-SDAG-NEXT: s_mov_b32 s5, s1
; SI-SDAG-NEXT: s_waitcnt vmcnt(0)
-; SI-SDAG-NEXT: v_readfirstlane_b32 s1, v1
-; SI-SDAG-NEXT: s_and_b32 s6, s1, 0x1ff
-; SI-SDAG-NEXT: s_lshr_b32 s7, s1, 8
-; SI-SDAG-NEXT: s_bfe_u32 s8, s1, 0xb0014
-; SI-SDAG-NEXT: v_or_b32_e32 v0, s6, v0
-; SI-SDAG-NEXT: s_and_b32 s6, s7, 0xffe
-; SI-SDAG-NEXT: s_sub_i32 s7, 0x3f1, s8
-; SI-SDAG-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; SI-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
-; SI-SDAG-NEXT: v_med3_i32 v1, s7, 0, 13
-; SI-SDAG-NEXT: v_readfirstlane_b32 s7, v0
-; SI-SDAG-NEXT: v_readfirstlane_b32 s9, v1
-; SI-SDAG-NEXT: s_or_b32 s6, s6, s7
-; SI-SDAG-NEXT: s_or_b32 s7, s6, 0x1000
-; SI-SDAG-NEXT: s_lshr_b32 s10, s7, s9
-; SI-SDAG-NEXT: s_lshl_b32 s9, s10, s9
-; SI-SDAG-NEXT: s_cmp_lg_u32 s9, s7
-; SI-SDAG-NEXT: s_cselect_b32 s7, 1, 0
-; SI-SDAG-NEXT: s_addk_i32 s8, 0xfc10
-; SI-SDAG-NEXT: s_or_b32 s7, s10, s7
-; SI-SDAG-NEXT: s_lshl_b32 s9, s8, 12
-; SI-SDAG-NEXT: s_or_b32 s9, s6, s9
-; SI-SDAG-NEXT: s_cmp_lt_i32 s8, 1
-; SI-SDAG-NEXT: s_cselect_b32 s7, s7, s9
-; SI-SDAG-NEXT: s_and_b32 s9, s7, 7
-; SI-SDAG-NEXT: s_cmp_gt_i32 s9, 5
-; SI-SDAG-NEXT: s_cselect_b32 s10, 1, 0
-; SI-SDAG-NEXT: s_cmp_eq_u32 s9, 3
-; SI-SDAG-NEXT: s_cselect_b32 s9, 1, 0
-; SI-SDAG-NEXT: s_lshr_b32 s7, s7, 2
-; SI-SDAG-NEXT: s_or_b32 s9, s9, s10
-; SI-SDAG-NEXT: s_add_i32 s7, s7, s9
-; SI-SDAG-NEXT: s_cmp_lt_i32 s8, 31
-; SI-SDAG-NEXT: s_cselect_b32 s7, s7, 0x7c00
-; SI-SDAG-NEXT: s_cmp_lg_u32 s6, 0
-; SI-SDAG-NEXT: s_cselect_b32 s0, s0, 0x7c00
-; SI-SDAG-NEXT: s_cmpk_eq_i32 s8, 0x40f
-; SI-SDAG-NEXT: s_cselect_b32 s0, s0, s7
-; SI-SDAG-NEXT: s_lshr_b32 s1, s1, 16
-; SI-SDAG-NEXT: s_and_b32 s1, s1, 0x8000
-; SI-SDAG-NEXT: s_or_b32 s6, s1, s0
-; SI-SDAG-NEXT: s_mov_b32 s0, s4
-; SI-SDAG-NEXT: s_mov_b32 s1, s5
-; SI-SDAG-NEXT: v_mov_b32_e32 v0, s6
-; SI-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
+; SI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; SI-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0
; SI-SDAG-NEXT: s_endpgm
;
; SI-GISEL-LABEL: fptrunc_f64_to_f16_afn:
@@ -1174,62 +1133,21 @@ define amdgpu_kernel void @fptrunc_f64_to_f16_afn(
;
; VI-SDAG-LABEL: fptrunc_f64_to_f16_afn:
; VI-SDAG: ; %bb.0: ; %entry
-; VI-SDAG-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x24
-; VI-SDAG-NEXT: s_mov_b32 s3, 0xf000
-; VI-SDAG-NEXT: s_mov_b32 s2, -1
-; VI-SDAG-NEXT: s_mov_b32 s10, s2
-; VI-SDAG-NEXT: s_mov_b32 s11, s3
+; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000
+; VI-SDAG-NEXT: s_mov_b32 s6, -1
+; VI-SDAG-NEXT: s_mov_b32 s10, s6
+; VI-SDAG-NEXT: s_mov_b32 s11, s7
; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; VI-SDAG-NEXT: s_mov_b32 s8, s6
-; VI-SDAG-NEXT: s_mov_b32 s9, s7
+; VI-SDAG-NEXT: s_mov_b32 s8, s2
+; VI-SDAG-NEXT: s_mov_b32 s9, s3
; VI-SDAG-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
-; VI-SDAG-NEXT: s_mov_b32 s0, s4
-; VI-SDAG-NEXT: s_mov_b32 s1, s5
-; VI-SDAG-NEXT: s_movk_i32 s6, 0x7e00
+; VI-SDAG-NEXT: s_mov_b32 s4, s0
+; VI-SDAG-NEXT: s_mov_b32 s5, s1
; VI-SDAG-NEXT: s_waitcnt vmcnt(0)
-; VI-SDAG-NEXT: v_readfirstlane_b32 s4, v1
-; VI-SDAG-NEXT: s_and_b32 s5, s4, 0x1ff
-; VI-SDAG-NEXT: v_or_b32_e32 v0, s5, v0
-; VI-SDAG-NEXT: s_lshr_b32 s7, s4, 8
-; VI-SDAG-NEXT: s_bfe_u32 s8, s4, 0xb0014
-; VI-SDAG-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; VI-SDAG-NEXT: s_and_b32 s5, s7, 0xffe
-; VI-SDAG-NEXT: s_sub_i32 s7, 0x3f1, s8
-; VI-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
-; VI-SDAG-NEXT: v_med3_i32 v1, s7, 0, 13
-; VI-SDAG-NEXT: v_readfirstlane_b32 s7, v0
-; VI-SDAG-NEXT: s_or_b32 s5, s5, s7
-; VI-SDAG-NEXT: v_readfirstlane_b32 s9, v1
-; VI-SDAG-NEXT: s_or_b32 s7, s5, 0x1000
-; VI-SDAG-NEXT: s_lshr_b32 s10, s7, s9
-; VI-SDAG-NEXT: s_lshl_b32 s9, s10, s9
-; VI-SDAG-NEXT: s_cmp_lg_u32 s9, s7
-; VI-SDAG-NEXT: s_cselect_b32 s7, 1, 0
-; VI-SDAG-NEXT: s_addk_i32 s8, 0xfc10
-; VI-SDAG-NEXT: s_lshl_b32 s9, s8, 12
-; VI-SDAG-NEXT: s_or_b32 s7, s10, s7
-; VI-SDAG-NEXT: s_or_b32 s9, s5, s9
-; VI-SDAG-NEXT: s_cmp_lt_i32 s8, 1
-; VI-SDAG-NEXT: s_cselect_b32 s7, s7, s9
-; VI-SDAG-NEXT: s_and_b32 s9, s7, 7
-; VI-SDAG-NEXT: s_cmp_gt_i32 s9, 5
-; VI-SDAG-NEXT: s_cselect_b32 s10, 1, 0
-; VI-SDAG-NEXT: s_cmp_eq_u32 s9, 3
-; VI-SDAG-NEXT: s_cselect_b32 s9, 1, 0
-; VI-SDAG-NEXT: s_lshr_b32 s7, s7, 2
-; VI-SDAG-NEXT: s_or_b32 s9, s9, s10
-; VI-SDAG-NEXT: s_add_i32 s7, s7, s9
-; VI-SDAG-NEXT: s_cmp_lt_i32 s8, 31
-; VI-SDAG-NEXT: s_cselect_b32 s7, s7, 0x7c00
-; VI-SDAG-NEXT: s_cmp_lg_u32 s5, 0
-; VI-SDAG-NEXT: s_cselect_b32 s5, s6, 0x7c00
-; VI-SDAG-NEXT: s_cmpk_eq_i32 s8, 0x40f
-; VI-SDAG-NEXT: s_cselect_b32 s5, s5, s7
-; VI-SDAG-NEXT: s_lshr_b32 s4, s4, 16
-; VI-SDAG-NEXT: s_and_b32 s4, s4, 0x8000
-; VI-SDAG-NEXT: s_or_b32 s4, s4, s5
-; VI-SDAG-NEXT: v_mov_b32_e32 v0, s4
-; VI-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
+; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; VI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; VI-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0
; VI-SDAG-NEXT: s_endpgm
;
; VI-GISEL-LABEL: fptrunc_f64_to_f16_afn:
@@ -1247,62 +1165,21 @@ define amdgpu_kernel void @fptrunc_f64_to_f16_afn(
;
; GFX9-SDAG-LABEL: fptrunc_f64_to_f16_afn:
; GFX9-SDAG: ; %bb.0: ; %entry
-; GFX9-SDAG-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x24
-; GFX9-SDAG-NEXT: s_mov_b32 s3, 0xf000
-; GFX9-SDAG-NEXT: s_mov_b32 s2, -1
-; GFX9-SDAG-NEXT: s_mov_b32 s6, s2
-; GFX9-SDAG-NEXT: s_mov_b32 s7, s3
+; GFX9-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX9-SDAG-NEXT: s_mov_b32 s7, 0xf000
+; GFX9-SDAG-NEXT: s_mov_b32 s6, -1
+; GFX9-SDAG-NEXT: s_mov_b32 s10, s6
+; GFX9-SDAG-NEXT: s_mov_b32 s11, s7
; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-SDAG-NEXT: s_mov_b32 s4, s10
-; GFX9-SDAG-NEXT: s_mov_b32 s5, s11
-; GFX9-SDAG-NEXT: buffer_load_dwordx2 v[0:1], off, s[4:7], 0
-; GFX9-SDAG-NEXT: s_mov_b32 s0, s8
-; GFX9-SDAG-NEXT: s_mov_b32 s1, s9
-; GFX9-SDAG-NEXT: s_movk_i32 s4, 0x7e00
+; GFX9-SDAG-NEXT: s_mov_b32 s8, s2
+; GFX9-SDAG-NEXT: s_mov_b32 s9, s3
+; GFX9-SDAG-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
+; GFX9-SDAG-NEXT: s_mov_b32 s4, s0
+; GFX9-SDAG-NEXT: s_mov_b32 s5, s1
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
-; GFX9-SDAG-NEXT: v_readfirstlane_b32 s5, v1
-; GFX9-SDAG-NEXT: s_and_b32 s6, s5, 0x1ff
-; GFX9-SDAG-NEXT: v_or_b32_e32 v0, s6, v0
-; GFX9-SDAG-NEXT: s_lshr_b32 s7, s5, 8
-; GFX9-SDAG-NEXT: s_bfe_u32 s8, s5, 0xb0014
-; GFX9-SDAG-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; GFX9-SDAG-NEXT: s_and_b32 s6, s7, 0xffe
-; GFX9-SDAG-NEXT: s_sub_i32 s7, 0x3f1, s8
-; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
-; GFX9-SDAG-NEXT: v_med3_i32 v1, s7, 0, 13
-; GFX9-SDAG-NEXT: v_readfirstlane_b32 s7, v0
-; GFX9-SDAG-NEXT: s_or_b32 s6, s6, s7
-; GFX9-SDAG-NEXT: v_readfirstlane_b32 s9, v1
-; GFX9-SDAG-NEXT: s_or_b32 s7, s6, 0x1000
-; GFX9-SDAG-NEXT: s_lshr_b32 s10, s7, s9
-; GFX9-SDAG-NEXT: s_lshl_b32 s9, s10, s9
-; GFX9-SDAG-NEXT: s_cmp_lg_u32 s9, s7
-; GFX9-SDAG-NEXT: s_cselect_b32 s7, 1, 0
-; GFX9-SDAG-NEXT: s_addk_i32 s8, 0xfc10
-; GFX9-SDAG-NEXT: s_lshl_b32 s9, s8, 12
-; GFX9-SDAG-NEXT: s_or_b32 s7, s10, s7
-; GFX9-SDAG-NEXT: s_or_b32 s9, s6, s9
-; GFX9-SDAG-NEXT: s_cmp_lt_i32 s8, 1
-; GFX9-SDAG-NEXT: s_cselect_b32 s7, s7, s9
-; GFX9-SDAG-NEXT: s_and_b32 s9, s7, 7
-; GFX9-SDAG-NEXT: s_cmp_gt_i32 s9, 5
-; GFX9-SDAG-NEXT: s_cselect_b32 s10, 1, 0
-; GFX9-SDAG-NEXT: s_cmp_eq_u32 s9, 3
-; GFX9-SDAG-NEXT: s_cselect_b32 s9, 1, 0
-; GFX9-SDAG-NEXT: s_lshr_b32 s7, s7, 2
-; GFX9-SDAG-NEXT: s_or_b32 s9, s9, s10
-; GFX9-SDAG-NEXT: s_add_i32 s7, s7, s9
-; GFX9-SDAG-NEXT: s_cmp_lt_i32 s8, 31
-; GFX9-SDAG-NEXT: s_cselect_b32 s7, s7, 0x7c00
-; GFX9-SDAG-NEXT: s_cmp_lg_u32 s6, 0
-; GFX9-SDAG-NEXT: s_cselect_b32 s4, s4, 0x7c00
-; GFX9-SDAG-NEXT: s_cmpk_eq_i32 s8, 0x40f
-; GFX9-SDAG-NEXT: s_cselect_b32 s4, s4, s7
-; GFX9-SDAG-NEXT: s_lshr_b32 s5, s5, 16
-; GFX9-SDAG-NEXT: s_and_b32 s5, s5, 0x8000
-; GFX9-SDAG-NEXT: s_or_b32 s4, s5, s4
-; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s4
-; GFX9-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
+; GFX9-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; GFX9-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX9-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0
; GFX9-SDAG-NEXT: s_endpgm
;
; GFX9-GISEL-LABEL: fptrunc_f64_to_f16_afn:
@@ -1320,62 +1197,21 @@ define amdgpu_kernel void @fptrunc_f64_to_f16_afn(
;
; GFX950-SDAG-LABEL: fptrunc_f64_to_f16_afn:
; GFX950-SDAG: ; %bb.0: ; %entry
-; GFX950-SDAG-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x24
-; GFX950-SDAG-NEXT: s_mov_b32 s3, 0xf000
-; GFX950-SDAG-NEXT: s_mov_b32 s2, -1
-; GFX950-SDAG-NEXT: s_mov_b32 s6, s2
-; GFX950-SDAG-NEXT: s_mov_b32 s7, s3
+; GFX950-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX950-SDAG-NEXT: s_mov_b32 s7, 0xf000
+; GFX950-SDAG-NEXT: s_mov_b32 s6, -1
+; GFX950-SDAG-NEXT: s_mov_b32 s10, s6
+; GFX950-SDAG-NEXT: s_mov_b32 s11, s7
; GFX950-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; GFX950-SDAG-NEXT: s_mov_b32 s4, s10
-; GFX950-SDAG-NEXT: s_mov_b32 s5, s11
-; GFX950-SDAG-NEXT: buffer_load_dwordx2 v[0:1], off, s[4:7], 0
-; GFX950-SDAG-NEXT: s_mov_b32 s0, s8
-; GFX950-SDAG-NEXT: s_mov_b32 s1, s9
-; GFX950-SDAG-NEXT: s_movk_i32 s4, 0x7e00
+; GFX950-SDAG-NEXT: s_mov_b32 s8, s2
+; GFX950-SDAG-NEXT: s_mov_b32 s9, s3
+; GFX950-SDAG-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
+; GFX950-SDAG-NEXT: s_mov_b32 s4, s0
+; GFX950-SDAG-NEXT: s_mov_b32 s5, s1
; GFX950-SDAG-NEXT: s_waitcnt vmcnt(0)
-; GFX950-SDAG-NEXT: v_readfirstlane_b32 s5, v1
-; GFX950-SDAG-NEXT: s_and_b32 s6, s5, 0x1ff
-; GFX950-SDAG-NEXT: v_or_b32_e32 v0, s6, v0
-; GFX950-SDAG-NEXT: s_lshr_b32 s7, s5, 8
-; GFX950-SDAG-NEXT: s_bfe_u32 s8, s5, 0xb0014
-; GFX950-SDAG-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; GFX950-SDAG-NEXT: s_and_b32 s6, s7, 0xffe
-; GFX950-SDAG-NEXT: s_sub_i32 s7, 0x3f1, s8
-; GFX950-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
-; GFX950-SDAG-NEXT: v_med3_i32 v1, s7, 0, 13
-; GFX950-SDAG-NEXT: v_readfirstlane_b32 s7, v0
-; GFX950-SDAG-NEXT: s_or_b32 s6, s6, s7
-; GFX950-SDAG-NEXT: v_readfirstlane_b32 s9, v1
-; GFX950-SDAG-NEXT: s_or_b32 s7, s6, 0x1000
-; GFX950-SDAG-NEXT: s_lshr_b32 s10, s7, s9
-; GFX950-SDAG-NEXT: s_lshl_b32 s9, s10, s9
-; GFX950-SDAG-NEXT: s_cmp_lg_u32 s9, s7
-; GFX950-SDAG-NEXT: s_cselect_b32 s7, 1, 0
-; GFX950-SDAG-NEXT: s_addk_i32 s8, 0xfc10
-; GFX950-SDAG-NEXT: s_lshl_b32 s9, s8, 12
-; GFX950-SDAG-NEXT: s_or_b32 s7, s10, s7
-; GFX950-SDAG-NEXT: s_or_b32 s9, s6, s9
-; GFX950-SDAG-NEXT: s_cmp_lt_i32 s8, 1
-; GFX950-SDAG-NEXT: s_cselect_b32 s7, s7, s9
-; GFX950-SDAG-NEXT: s_and_b32 s9, s7, 7
-; GFX950-SDAG-NEXT: s_cmp_gt_i32 s9, 5
-; GFX950-SDAG-NEXT: s_cselect_b32 s10, 1, 0
-; GFX950-SDAG-NEXT: s_cmp_eq_u32 s9, 3
-; GFX950-SDAG-NEXT: s_cselect_b32 s9, 1, 0
-; GFX950-SDAG-NEXT: s_lshr_b32 s7, s7, 2
-; GFX950-SDAG-NEXT: s_or_b32 s9, s9, s10
-; GFX950-SDAG-NEXT: s_add_i32 s7, s7, s9
-; GFX950-SDAG-NEXT: s_cmp_lt_i32 s8, 31
-; GFX950-SDAG-NEXT: s_cselect_b32 s7, s7, 0x7c00
-; GFX950-SDAG-NEXT: s_cmp_lg_u32 s6, 0
-; GFX950-SDAG-NEXT: s_cselect_b32 s4, s4, 0x7c00
-; GFX950-SDAG-NEXT: s_cmpk_eq_i32 s8, 0x40f
-; GFX950-SDAG-NEXT: s_cselect_b32 s4, s4, s7
-; GFX950-SDAG-NEXT: s_lshr_b32 s5, s5, 16
-; GFX950-SDAG-NEXT: s_and_b32 s5, s5, 0x8000
-; GFX950-SDAG-NEXT: s_or_b32 s4, s5, s4
-; GFX950-SDAG-NEXT: v_mov_b32_e32 v0, s4
-; GFX950-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
+; GFX950-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; GFX950-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX950-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0
; GFX950-SDAG-NEXT: s_endpgm
;
; GFX950-GISEL-LABEL: fptrunc_f64_to_f16_afn:
@@ -1401,60 +1237,13 @@ define amdgpu_kernel void @fptrunc_f64_to_f16_afn(
; GFX11-SDAG-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-TRUE16-NEXT: s_mov_b32 s8, s2
; GFX11-SDAG-TRUE16-NEXT: s_mov_b32 s9, s3
-; GFX11-SDAG-TRUE16-NEXT: buffer_load_b64 v[0:1], off, s[8:11], 0
-; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-SDAG-TRUE16-NEXT: v_readfirstlane_b32 s2, v1
-; GFX11-SDAG-TRUE16-NEXT: s_and_b32 s3, s2, 0x1ff
-; GFX11-SDAG-TRUE16-NEXT: s_lshr_b32 s5, s2, 8
-; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v0, s3, v0
-; GFX11-SDAG-TRUE16-NEXT: s_bfe_u32 s3, s2, 0xb0014
-; GFX11-SDAG-TRUE16-NEXT: s_and_b32 s5, s5, 0xffe
-; GFX11-SDAG-TRUE16-NEXT: s_sub_i32 s4, 0x3f1, s3
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-SDAG-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-SDAG-TRUE16-NEXT: v_med3_i32 v1, s4, 0, 13
-; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
-; GFX11-SDAG-TRUE16-NEXT: v_readfirstlane_b32 s8, v1
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: v_readfirstlane_b32 s4, v0
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s4, s5, s4
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s5, s4, 0x1000
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: s_lshr_b32 s9, s5, s8
-; GFX11-SDAG-TRUE16-NEXT: s_lshl_b32 s8, s9, s8
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_lg_u32 s8, s5
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s5, 1, 0
-; GFX11-SDAG-TRUE16-NEXT: s_addk_i32 s3, 0xfc10
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s5, s9, s5
-; GFX11-SDAG-TRUE16-NEXT: s_lshl_b32 s8, s3, 12
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s8, s4, s8
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_lt_i32 s3, 1
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s5, s5, s8
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: s_and_b32 s8, s5, 7
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_gt_i32 s8, 5
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s9, 1, 0
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_eq_u32 s8, 3
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s8, 1, 0
-; GFX11-SDAG-TRUE16-NEXT: s_lshr_b32 s5, s5, 2
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s8, s8, s9
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: s_add_i32 s5, s5, s8
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_lt_i32 s3, 31
-; GFX11-SDAG-TRUE16-NEXT: s_movk_i32 s8, 0x7e00
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s5, s5, 0x7c00
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_lg_u32 s4, 0
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s4, s8, 0x7c00
-; GFX11-SDAG-TRUE16-NEXT: s_cmpk_eq_i32 s3, 0x40f
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s3, s4, s5
-; GFX11-SDAG-TRUE16-NEXT: s_lshr_b32 s2, s2, 16
; GFX11-SDAG-TRUE16-NEXT: s_mov_b32 s4, s0
-; GFX11-SDAG-TRUE16-NEXT: s_and_b32 s2, s2, 0x8000
+; GFX11-SDAG-TRUE16-NEXT: buffer_load_b64 v[0:1], off, s[8:11], 0
; GFX11-SDAG-TRUE16-NEXT: s_mov_b32 s5, s1
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s2, s2, s3
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: v_mov_b32_e32 v0, s2
+; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-SDAG-TRUE16-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-SDAG-TRUE16-NEXT: v_cvt_f16_f32_e32 v0.l, v0
; GFX11-SDAG-TRUE16-NEXT: buffer_store_b16 v0, off, s[4:7], 0
; GFX11-SDAG-TRUE16-NEXT: s_endpgm
;
@@ -1468,60 +1257,13 @@ define amdgpu_kernel void @fptrunc_f64_to_f16_afn(
; GFX11-SDAG-FAKE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-FAKE16-NEXT: s_mov_b32 s8, s2
; GFX11-SDAG-FAKE16-NEXT: s_mov_b32 s9, s3
-; GFX11-SDAG-FAKE16-NEXT: buffer_load_b64 v[0:1], off, s[8:11], 0
-; GFX11-SDAG-FAKE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-SDAG-FAKE16-NEXT: v_readfirstlane_b32 s2, v1
-; GFX11-SDAG-FAKE16-NEXT: s_and_b32 s3, s2, 0x1ff
-; GFX11-SDAG-FAKE16-NEXT: s_lshr_b32 s5, s2, 8
-; GFX11-SDAG-FAKE16-NEXT: v_or_b32_e32 v0, s3, v0
-; GFX11-SDAG-FAKE16-NEXT: s_bfe_u32 s3, s2, 0xb0014
-; GFX11-SDAG-FAKE16-NEXT: s_and_b32 s5, s5, 0xffe
-; GFX11-SDAG-FAKE16-NEXT: s_sub_i32 s4, 0x3f1, s3
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-SDAG-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-SDAG-FAKE16-NEXT: v_med3_i32 v1, s4, 0, 13
-; GFX11-SDAG-FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
-; GFX11-SDAG-FAKE16-NEXT: v_readfirstlane_b32 s8, v1
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: v_readfirstlane_b32 s4, v0
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s4, s5, s4
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s5, s4, 0x1000
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: s_lshr_b32 s9, s5, s8
-; GFX11-SDAG-FAKE16-NEXT: s_lshl_b32 s8, s9, s8
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_lg_u32 s8, s5
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s5, 1, 0
-; GFX11-SDAG-FAKE16-NEXT: s_addk_i32 s3, 0xfc10
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s5, s9, s5
-; GFX11-SDAG-FAKE16-NEXT: s_lshl_b32 s8, s3, 12
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s8, s4, s8
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_lt_i32 s3, 1
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s5, s5, s8
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: s_and_b32 s8, s5, 7
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_gt_i32 s8, 5
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s9, 1, 0
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_eq_u32 s8, 3
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s8, 1, 0
-; GFX11-SDAG-FAKE16-NEXT: s_lshr_b32 s5, s5, 2
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s8, s8, s9
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: s_add_i32 s5, s5, s8
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_lt_i32 s3, 31
-; GFX11-SDAG-FAKE16-NEXT: s_movk_i32 s8, 0x7e00
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s5, s5, 0x7c00
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_lg_u32 s4, 0
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s4, s8, 0x7c00
-; GFX11-SDAG-FAKE16-NEXT: s_cmpk_eq_i32 s3, 0x40f
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s3, s4, s5
-; GFX11-SDAG-FAKE16-NEXT: s_lshr_b32 s2, s2, 16
; GFX11-SDAG-FAKE16-NEXT: s_mov_b32 s4, s0
-; GFX11-SDAG-FAKE16-NEXT: s_and_b32 s2, s2, 0x8000
+; GFX11-SDAG-FAKE16-NEXT: buffer_load_b64 v[0:1], off, s[8:11], 0
; GFX11-SDAG-FAKE16-NEXT: s_mov_b32 s5, s1
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s2, s2, s3
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: v_mov_b32_e32 v0, s2
+; GFX11-SDAG-FAKE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-SDAG-FAKE16-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-SDAG-FAKE16-NEXT: v_cvt_f16_f32_e32 v0, v0
; GFX11-SDAG-FAKE16-NEXT: buffer_store_b16 v0, off, s[4:7], 0
; GFX11-SDAG-FAKE16-NEXT: s_endpgm
;
@@ -3026,106 +2768,25 @@ entry:
define amdgpu_kernel void @fptrunc_v2f64_to_v2f16_afn(
; SI-SDAG-LABEL: fptrunc_v2f64_to_v2f16_afn:
; SI-SDAG: ; %bb.0: ; %entry
-; SI-SDAG-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9
-; SI-SDAG-NEXT: s_mov_b32 s3, 0xf000
-; SI-SDAG-NEXT: s_mov_b32 s2, -1
-; SI-SDAG-NEXT: s_mov_b32 s10, s2
-; SI-SDAG-NEXT: s_mov_b32 s11, s3
+; SI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
+; SI-SDAG-NEXT: s_mov_b32 s7, 0xf000
+; SI-SDAG-NEXT: s_mov_b32 s6, -1
+; SI-SDAG-NEXT: s_mov_b32 s10, s6
+; SI-SDAG-NEXT: s_mov_b32 s11, s7
; SI-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; SI-SDAG-NEXT: s_mov_b32 s8, s6
-; SI-SDAG-NEXT: s_mov_b32 s9, s7
+; SI-SDAG-NEXT: s_mov_b32 s8, s2
+; SI-SDAG-NEXT: s_mov_b32 s9, s3
; SI-SDAG-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0
-; SI-SDAG-NEXT: s_movk_i32 s0, 0x7e00
+; SI-SDAG-NEXT: s_mov_b32 s4, s0
+; SI-SDAG-NEXT: s_mov_b32 s5, s1
; SI-SDAG-NEXT: s_waitcnt vmcnt(0)
-; SI-SDAG-NEXT: v_readfirstlane_b32 s1, v3
-; SI-SDAG-NEXT: v_readfirstlane_b32 s6, v1
-; SI-SDAG-NEXT: s_and_b32 s7, s1, 0x1ff
-; SI-SDAG-NEXT: s_lshr_b32 s8, s1, 8
-; SI-SDAG-NEXT: s_bfe_u32 s9, s1, 0xb0014
-; SI-SDAG-NEXT: v_or_b32_e32 v1, s7, v2
-; SI-SDAG-NEXT: s_and_b32 s7, s8, 0xffe
-; SI-SDAG-NEXT: s_sub_i32 s8, 0x3f1, s9
-; SI-SDAG-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, 1, vcc
-; SI-SDAG-NEXT: v_med3_i32 v2, s8, 0, 13
-; SI-SDAG-NEXT: v_readfirstlane_b32 s8, v1
-; SI-SDAG-NEXT: v_readfirstlane_b32 s10, v2
-; SI-SDAG-NEXT: s_or_b32 s7, s7, s8
-; SI-SDAG-NEXT: s_or_b32 s8, s7, 0x1000
-; SI-SDAG-NEXT: s_lshr_b32 s11, s8, s10
-; SI-SDAG-NEXT: s_lshl_b32 s10, s11, s10
-; SI-SDAG-NEXT: s_cmp_lg_u32 s10, s8
-; SI-SDAG-NEXT: s_cselect_b32 s8, 1, 0
-; SI-SDAG-NEXT: s_addk_i32 s9, 0xfc10
-; SI-SDAG-NEXT: s_or_b32 s8, s11, s8
-; SI-SDAG-NEXT: s_lshl_b32 s10, s9, 12
-; SI-SDAG-NEXT: s_or_b32 s10, s7, s10
-; SI-SDAG-NEXT: s_cmp_lt_i32 s9, 1
-; SI-SDAG-NEXT: s_cselect_b32 s8, s8, s10
-; SI-SDAG-NEXT: s_and_b32 s10, s8, 7
-; SI-SDAG-NEXT: s_cmp_gt_i32 s10, 5
-; SI-SDAG-NEXT: s_cselect_b32 s11, 1, 0
-; SI-SDAG-NEXT: s_cmp_eq_u32 s10, 3
-; SI-SDAG-NEXT: s_cselect_b32 s10, 1, 0
-; SI-SDAG-NEXT: s_lshr_b32 s8, s8, 2
-; SI-SDAG-NEXT: s_or_b32 s10, s10, s11
-; SI-SDAG-NEXT: s_add_i32 s8, s8, s10
-; SI-SDAG-NEXT: s_cmp_lt_i32 s9, 31
-; SI-SDAG-NEXT: s_cselect_b32 s8, s8, 0x7c00
-; SI-SDAG-NEXT: s_cmp_lg_u32 s7, 0
-; SI-SDAG-NEXT: s_cselect_b32 s7, s0, 0x7c00
-; SI-SDAG-NEXT: s_cmpk_eq_i32 s9, 0x40f
-; SI-SDAG-NEXT: s_cselect_b32 s7, s7, s8
-; SI-SDAG-NEXT: s_lshr_b32 s1, s1, 16
-; SI-SDAG-NEXT: s_and_b32 s8, s6, 0x1ff
-; SI-SDAG-NEXT: s_lshr_b32 s9, s6, 8
-; SI-SDAG-NEXT: s_bfe_u32 s10, s6, 0xb0014
-; SI-SDAG-NEXT: s_and_b32 s1, s1, 0x8000
-; SI-SDAG-NEXT: v_or_b32_e32 v0, s8, v0
-; SI-SDAG-NEXT: s_and_b32 s8, s9, 0xffe
-; SI-SDAG-NEXT: s_sub_i32 s9, 0x3f1, s10
-; SI-SDAG-NEXT: s_or_b32 s1, s1, s7
-; SI-SDAG-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; SI-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
-; SI-SDAG-NEXT: v_med3_i32 v1, s9, 0, 13
-; SI-SDAG-NEXT: s_lshl_b32 s1, s1, 16
-; SI-SDAG-NEXT: v_readfirstlane_b32 s7, v0
-; SI-SDAG-NEXT: v_readfirstlane_b32 s9, v1
-; SI-SDAG-NEXT: s_or_b32 s7, s8, s7
-; SI-SDAG-NEXT: s_or_b32 s8, s7, 0x1000
-; SI-SDAG-NEXT: s_lshr_b32 s11, s8, s9
-; SI-SDAG-NEXT: s_lshl_b32 s9, s11, s9
-; SI-SDAG-NEXT: s_cmp_lg_u32 s9, s8
-; SI-SDAG-NEXT: s_cselect_b32 s8, 1, 0
-; SI-SDAG-NEXT: s_addk_i32 s10, 0xfc10
-; SI-SDAG-NEXT: s_or_b32 s8, s11, s8
-; SI-SDAG-NEXT: s_lshl_b32 s9, s10, 12
-; SI-SDAG-NEXT: s_or_b32 s9, s7, s9
-; SI-SDAG-NEXT: s_cmp_lt_i32 s10, 1
-; SI-SDAG-NEXT: s_cselect_b32 s8, s8, s9
-; SI-SDAG-NEXT: s_and_b32 s9, s8, 7
-; SI-SDAG-NEXT: s_cmp_gt_i32 s9, 5
-; SI-SDAG-NEXT: s_cselect_b32 s11, 1, 0
-; SI-SDAG-NEXT: s_cmp_eq_u32 s9, 3
-; SI-SDAG-NEXT: s_cselect_b32 s9, 1, 0
-; SI-SDAG-NEXT: s_lshr_b32 s8, s8, 2
-; SI-SDAG-NEXT: s_or_b32 s9, s9, s11
-; SI-SDAG-NEXT: s_add_i32 s8, s8, s9
-; SI-SDAG-NEXT: s_cmp_lt_i32 s10, 31
-; SI-SDAG-NEXT: s_cselect_b32 s8, s8, 0x7c00
-; SI-SDAG-NEXT: s_cmp_lg_u32 s7, 0
-; SI-SDAG-NEXT: s_cselect_b32 s0, s0, 0x7c00
-; SI-SDAG-NEXT: s_cmpk_eq_i32 s10, 0x40f
-; SI-SDAG-NEXT: s_cselect_b32 s0, s0, s8
-; SI-SDAG-NEXT: s_lshr_b32 s6, s6, 16
-; SI-SDAG-NEXT: s_and_b32 s6, s6, 0x8000
-; SI-SDAG-NEXT: s_or_b32 s0, s6, s0
-; SI-SDAG-NEXT: s_and_b32 s0, s0, 0xffff
-; SI-SDAG-NEXT: s_or_b32 s6, s0, s1
-; SI-SDAG-NEXT: s_mov_b32 s0, s4
-; SI-SDAG-NEXT: s_mov_b32 s1, s5
-; SI-SDAG-NEXT: v_mov_b32_e32 v0, s6
-; SI-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], 0
+; SI-SDAG-NEXT: v_cvt_f32_f64_e32 v2, v[2:3]
+; SI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v2
+; SI-SDAG-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; SI-SDAG-NEXT: v_or_b32_e32 v0, v0, v1
+; SI-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0
; SI-SDAG-NEXT: s_endpgm
;
; SI-GISEL-LABEL: fptrunc_v2f64_to_v2f16_afn:
@@ -3147,106 +2808,24 @@ define amdgpu_kernel void @fptrunc_v2f64_to_v2f16_afn(
;
; VI-SDAG-LABEL: fptrunc_v2f64_to_v2f16_afn:
; VI-SDAG: ; %bb.0: ; %entry
-; VI-SDAG-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x24
-; VI-SDAG-NEXT: s_mov_b32 s3, 0xf000
-; VI-SDAG-NEXT: s_mov_b32 s2, -1
-; VI-SDAG-NEXT: s_mov_b32 s10, s2
-; VI-SDAG-NEXT: s_mov_b32 s11, s3
+; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000
+; VI-SDAG-NEXT: s_mov_b32 s6, -1
+; VI-SDAG-NEXT: s_mov_b32 s10, s6
+; VI-SDAG-NEXT: s_mov_b32 s11, s7
; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; VI-SDAG-NEXT: s_mov_b32 s8, s6
-; VI-SDAG-NEXT: s_mov_b32 s9, s7
+; VI-SDAG-NEXT: s_mov_b32 s8, s2
+; VI-SDAG-NEXT: s_mov_b32 s9, s3
; VI-SDAG-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0
-; VI-SDAG-NEXT: s_mov_b32 s0, s4
-; VI-SDAG-NEXT: s_mov_b32 s1, s5
-; VI-SDAG-NEXT: s_movk_i32 s6, 0x7e00
+; VI-SDAG-NEXT: s_mov_b32 s4, s0
+; VI-SDAG-NEXT: s_mov_b32 s5, s1
; VI-SDAG-NEXT: s_waitcnt vmcnt(0)
-; VI-SDAG-NEXT: v_readfirstlane_b32 s4, v3
-; VI-SDAG-NEXT: s_and_b32 s7, s4, 0x1ff
-; VI-SDAG-NEXT: v_readfirstlane_b32 s5, v1
-; VI-SDAG-NEXT: v_or_b32_e32 v1, s7, v2
-; VI-SDAG-NEXT: s_lshr_b32 s8, s4, 8
-; VI-SDAG-NEXT: s_bfe_u32 s9, s4, 0xb0014
-; VI-SDAG-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-SDAG-NEXT: s_and_b32 s7, s8, 0xffe
-; VI-SDAG-NEXT: s_sub_i32 s8, 0x3f1, s9
-; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, 1, vcc
-; VI-SDAG-NEXT: v_med3_i32 v2, s8, 0, 13
-; VI-SDAG-NEXT: v_readfirstlane_b32 s8, v1
-; VI-SDAG-NEXT: s_or_b32 s7, s7, s8
-; VI-SDAG-NEXT: v_readfirstlane_b32 s10, v2
-; VI-SDAG-NEXT: s_or_b32 s8, s7, 0x1000
-; VI-SDAG-NEXT: s_lshr_b32 s11, s8, s10
-; VI-SDAG-NEXT: s_lshl_b32 s10, s11, s10
-; VI-SDAG-NEXT: s_cmp_lg_u32 s10, s8
-; VI-SDAG-NEXT: s_cselect_b32 s8, 1, 0
-; VI-SDAG-NEXT: s_addk_i32 s9, 0xfc10
-; VI-SDAG-NEXT: s_lshl_b32 s10, s9, 12
-; VI-SDAG-NEXT: s_or_b32 s8, s11, s8
-; VI-SDAG-NEXT: s_or_b32 s10, s7, s10
-; VI-SDAG-NEXT: s_cmp_lt_i32 s9, 1
-; VI-SDAG-NEXT: s_cselect_b32 s8, s8, s10
-; VI-SDAG-NEXT: s_and_b32 s10, s8, 7
-; VI-SDAG-NEXT: s_cmp_gt_i32 s10, 5
-; VI-SDAG-NEXT: s_cselect_b32 s11, 1, 0
-; VI-SDAG-NEXT: s_cmp_eq_u32 s10, 3
-; VI-SDAG-NEXT: s_cselect_b32 s10, 1, 0
-; VI-SDAG-NEXT: s_lshr_b32 s8, s8, 2
-; VI-SDAG-NEXT: s_or_b32 s10, s10, s11
-; VI-SDAG-NEXT: s_add_i32 s8, s8, s10
-; VI-SDAG-NEXT: s_cmp_lt_i32 s9, 31
-; VI-SDAG-NEXT: s_cselect_b32 s8, s8, 0x7c00
-; VI-SDAG-NEXT: s_cmp_lg_u32 s7, 0
-; VI-SDAG-NEXT: s_cselect_b32 s7, s6, 0x7c00
-; VI-SDAG-NEXT: s_cmpk_eq_i32 s9, 0x40f
-; VI-SDAG-NEXT: s_cselect_b32 s7, s7, s8
-; VI-SDAG-NEXT: s_and_b32 s8, s5, 0x1ff
-; VI-SDAG-NEXT: v_or_b32_e32 v0, s8, v0
-; VI-SDAG-NEXT: s_lshr_b32 s4, s4, 16
-; VI-SDAG-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; VI-SDAG-NEXT: s_lshr_b32 s9, s5, 8
-; VI-SDAG-NEXT: s_bfe_u32 s10, s5, 0xb0014
-; VI-SDAG-NEXT: s_and_b32 s4, s4, 0x8000
-; VI-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
-; VI-SDAG-NEXT: s_and_b32 s8, s9, 0xffe
-; VI-SDAG-NEXT: s_sub_i32 s9, 0x3f1, s10
-; VI-SDAG-NEXT: s_or_b32 s4, s4, s7
-; VI-SDAG-NEXT: v_readfirstlane_b32 s7, v0
-; VI-SDAG-NEXT: v_med3_i32 v1, s9, 0, 13
-; VI-SDAG-NEXT: s_or_b32 s7, s8, s7
-; VI-SDAG-NEXT: v_readfirstlane_b32 s9, v1
-; VI-SDAG-NEXT: s_or_b32 s8, s7, 0x1000
-; VI-SDAG-NEXT: s_lshr_b32 s11, s8, s9
-; VI-SDAG-NEXT: s_lshl_b32 s4, s4, 16
-; VI-SDAG-NEXT: s_lshl_b32 s9, s11, s9
-; VI-SDAG-NEXT: s_cmp_lg_u32 s9, s8
-; VI-SDAG-NEXT: s_cselect_b32 s8, 1, 0
-; VI-SDAG-NEXT: s_addk_i32 s10, 0xfc10
-; VI-SDAG-NEXT: s_lshl_b32 s9, s10, 12
-; VI-SDAG-NEXT: s_or_b32 s8, s11, s8
-; VI-SDAG-NEXT: s_or_b32 s9, s7, s9
-; VI-SDAG-NEXT: s_cmp_lt_i32 s10, 1
-; VI-SDAG-NEXT: s_cselect_b32 s8, s8, s9
-; VI-SDAG-NEXT: s_and_b32 s9, s8, 7
-; VI-SDAG-NEXT: s_cmp_gt_i32 s9, 5
-; VI-SDAG-NEXT: s_cselect_b32 s11, 1, 0
-; VI-SDAG-NEXT: s_cmp_eq_u32 s9, 3
-; VI-SDAG-NEXT: s_cselect_b32 s9, 1, 0
-; VI-SDAG-NEXT: s_lshr_b32 s8, s8, 2
-; VI-SDAG-NEXT: s_or_b32 s9, s9, s11
-; VI-SDAG-NEXT: s_add_i32 s8, s8, s9
-; VI-SDAG-NEXT: s_cmp_lt_i32 s10, 31
-; VI-SDAG-NEXT: s_cselect_b32 s8, s8, 0x7c00
-; VI-SDAG-NEXT: s_cmp_lg_u32 s7, 0
-; VI-SDAG-NEXT: s_cselect_b32 s6, s6, 0x7c00
-; VI-SDAG-NEXT: s_cmpk_eq_i32 s10, 0x40f
-; VI-SDAG-NEXT: s_cselect_b32 s6, s6, s8
-; VI-SDAG-NEXT: s_lshr_b32 s5, s5, 16
-; VI-SDAG-NEXT: s_and_b32 s5, s5, 0x8000
-; VI-SDAG-NEXT: s_or_b32 s5, s5, s6
-; VI-SDAG-NEXT: s_and_b32 s5, s5, 0xffff
-; VI-SDAG-NEXT: s_or_b32 s4, s5, s4
-; VI-SDAG-NEXT: v_mov_b32_e32 v0, s4
-; VI-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], 0
+; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v2, v[2:3]
+; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; VI-SDAG-NEXT: v_cvt_f16_f32_sdwa v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD
+; VI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; VI-SDAG-NEXT: v_or_b32_e32 v0, v0, v1
+; VI-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0
; VI-SDAG-NEXT: s_endpgm
;
; VI-GISEL-LABEL: fptrunc_v2f64_to_v2f16_afn:
@@ -3267,104 +2846,24 @@ define amdgpu_kernel void @fptrunc_v2f64_to_v2f16_afn(
;
; GFX9-SDAG-LABEL: fptrunc_v2f64_to_v2f16_afn:
; GFX9-SDAG: ; %bb.0: ; %entry
-; GFX9-SDAG-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x24
-; GFX9-SDAG-NEXT: s_mov_b32 s3, 0xf000
-; GFX9-SDAG-NEXT: s_mov_b32 s2, -1
-; GFX9-SDAG-NEXT: s_mov_b32 s6, s2
-; GFX9-SDAG-NEXT: s_mov_b32 s7, s3
+; GFX9-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX9-SDAG-NEXT: s_mov_b32 s7, 0xf000
+; GFX9-SDAG-NEXT: s_mov_b32 s6, -1
+; GFX9-SDAG-NEXT: s_mov_b32 s10, s6
+; GFX9-SDAG-NEXT: s_mov_b32 s11, s7
; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-SDAG-NEXT: s_mov_b32 s4, s10
-; GFX9-SDAG-NEXT: s_mov_b32 s5, s11
-; GFX9-SDAG-NEXT: buffer_load_dwordx4 v[0:3], off, s[4:7], 0
-; GFX9-SDAG-NEXT: s_mov_b32 s0, s8
-; GFX9-SDAG-NEXT: s_mov_b32 s1, s9
-; GFX9-SDAG-NEXT: s_movk_i32 s4, 0x7e00
+; GFX9-SDAG-NEXT: s_mov_b32 s8, s2
+; GFX9-SDAG-NEXT: s_mov_b32 s9, s3
+; GFX9-SDAG-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0
+; GFX9-SDAG-NEXT: s_mov_b32 s4, s0
+; GFX9-SDAG-NEXT: s_mov_b32 s5, s1
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
-; GFX9-SDAG-NEXT: v_readfirstlane_b32 s5, v3
-; GFX9-SDAG-NEXT: s_and_b32 s7, s5, 0x1ff
-; GFX9-SDAG-NEXT: v_readfirstlane_b32 s6, v1
-; GFX9-SDAG-NEXT: v_or_b32_e32 v1, s7, v2
-; GFX9-SDAG-NEXT: s_lshr_b32 s8, s5, 8
-; GFX9-SDAG-NEXT: s_bfe_u32 s9, s5, 0xb0014
-; GFX9-SDAG-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-SDAG-NEXT: s_and_b32 s7, s8, 0xffe
-; GFX9-SDAG-NEXT: s_sub_i32 s8, 0x3f1, s9
-; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, 1, vcc
-; GFX9-SDAG-NEXT: v_med3_i32 v2, s8, 0, 13
-; GFX9-SDAG-NEXT: v_readfirstlane_b32 s8, v1
-; GFX9-SDAG-NEXT: s_or_b32 s7, s7, s8
-; GFX9-SDAG-NEXT: v_readfirstlane_b32 s10, v2
-; GFX9-SDAG-NEXT: s_or_b32 s8, s7, 0x1000
-; GFX9-SDAG-NEXT: s_lshr_b32 s11, s8, s10
-; GFX9-SDAG-NEXT: s_lshl_b32 s10, s11, s10
-; GFX9-SDAG-NEXT: s_cmp_lg_u32 s10, s8
-; GFX9-SDAG-NEXT: s_cselect_b32 s8, 1, 0
-; GFX9-SDAG-NEXT: s_addk_i32 s9, 0xfc10
-; GFX9-SDAG-NEXT: s_lshl_b32 s10, s9, 12
-; GFX9-SDAG-NEXT: s_or_b32 s8, s11, s8
-; GFX9-SDAG-NEXT: s_or_b32 s10, s7, s10
-; GFX9-SDAG-NEXT: s_cmp_lt_i32 s9, 1
-; GFX9-SDAG-NEXT: s_cselect_b32 s8, s8, s10
-; GFX9-SDAG-NEXT: s_and_b32 s10, s8, 7
-; GFX9-SDAG-NEXT: s_cmp_gt_i32 s10, 5
-; GFX9-SDAG-NEXT: s_cselect_b32 s11, 1, 0
-; GFX9-SDAG-NEXT: s_cmp_eq_u32 s10, 3
-; GFX9-SDAG-NEXT: s_cselect_b32 s10, 1, 0
-; GFX9-SDAG-NEXT: s_lshr_b32 s8, s8, 2
-; GFX9-SDAG-NEXT: s_or_b32 s10, s10, s11
-; GFX9-SDAG-NEXT: s_add_i32 s8, s8, s10
-; GFX9-SDAG-NEXT: s_cmp_lt_i32 s9, 31
-; GFX9-SDAG-NEXT: s_cselect_b32 s8, s8, 0x7c00
-; GFX9-SDAG-NEXT: s_cmp_lg_u32 s7, 0
-; GFX9-SDAG-NEXT: s_cselect_b32 s7, s4, 0x7c00
-; GFX9-SDAG-NEXT: s_cmpk_eq_i32 s9, 0x40f
-; GFX9-SDAG-NEXT: s_cselect_b32 s7, s7, s8
-; GFX9-SDAG-NEXT: s_and_b32 s8, s6, 0x1ff
-; GFX9-SDAG-NEXT: v_or_b32_e32 v0, s8, v0
-; GFX9-SDAG-NEXT: s_lshr_b32 s5, s5, 16
-; GFX9-SDAG-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; GFX9-SDAG-NEXT: s_lshr_b32 s9, s6, 8
-; GFX9-SDAG-NEXT: s_bfe_u32 s10, s6, 0xb0014
-; GFX9-SDAG-NEXT: s_and_b32 s5, s5, 0x8000
-; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
-; GFX9-SDAG-NEXT: s_and_b32 s8, s9, 0xffe
-; GFX9-SDAG-NEXT: s_sub_i32 s9, 0x3f1, s10
-; GFX9-SDAG-NEXT: s_or_b32 s5, s5, s7
-; GFX9-SDAG-NEXT: v_readfirstlane_b32 s7, v0
-; GFX9-SDAG-NEXT: v_med3_i32 v1, s9, 0, 13
-; GFX9-SDAG-NEXT: s_or_b32 s7, s8, s7
-; GFX9-SDAG-NEXT: v_readfirstlane_b32 s9, v1
-; GFX9-SDAG-NEXT: s_or_b32 s8, s7, 0x1000
-; GFX9-SDAG-NEXT: s_lshr_b32 s11, s8, s9
-; GFX9-SDAG-NEXT: s_lshl_b32 s9, s11, s9
-; GFX9-SDAG-NEXT: s_cmp_lg_u32 s9, s8
-; GFX9-SDAG-NEXT: s_cselect_b32 s8, 1, 0
-; GFX9-SDAG-NEXT: s_addk_i32 s10, 0xfc10
-; GFX9-SDAG-NEXT: s_lshl_b32 s9, s10, 12
-; GFX9-SDAG-NEXT: s_or_b32 s8, s11, s8
-; GFX9-SDAG-NEXT: s_or_b32 s9, s7, s9
-; GFX9-SDAG-NEXT: s_cmp_lt_i32 s10, 1
-; GFX9-SDAG-NEXT: s_cselect_b32 s8, s8, s9
-; GFX9-SDAG-NEXT: s_and_b32 s9, s8, 7
-; GFX9-SDAG-NEXT: s_cmp_gt_i32 s9, 5
-; GFX9-SDAG-NEXT: s_cselect_b32 s11, 1, 0
-; GFX9-SDAG-NEXT: s_cmp_eq_u32 s9, 3
-; GFX9-SDAG-NEXT: s_cselect_b32 s9, 1, 0
-; GFX9-SDAG-NEXT: s_lshr_b32 s8, s8, 2
-; GFX9-SDAG-NEXT: s_or_b32 s9, s9, s11
-; GFX9-SDAG-NEXT: s_add_i32 s8, s8, s9
-; GFX9-SDAG-NEXT: s_cmp_lt_i32 s10, 31
-; GFX9-SDAG-NEXT: s_cselect_b32 s8, s8, 0x7c00
-; GFX9-SDAG-NEXT: s_cmp_lg_u32 s7, 0
-; GFX9-SDAG-NEXT: s_cselect_b32 s4, s4, 0x7c00
-; GFX9-SDAG-NEXT: s_cmpk_eq_i32 s10, 0x40f
-; GFX9-SDAG-NEXT: s_cselect_b32 s4, s4, s8
-; GFX9-SDAG-NEXT: s_lshr_b32 s6, s6, 16
-; GFX9-SDAG-NEXT: s_and_b32 s6, s6, 0x8000
-; GFX9-SDAG-NEXT: s_or_b32 s4, s6, s4
-; GFX9-SDAG-NEXT: s_pack_ll_b32_b16 s4, s4, s5
-; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s4
-; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], 0
+; GFX9-SDAG-NEXT: v_cvt_f32_f64_e32 v2, v[2:3]
+; GFX9-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; GFX9-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v2
+; GFX9-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX9-SDAG-NEXT: v_pack_b32_f16 v0, v0, v1
+; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GFX9-SDAG-NEXT: s_endpgm
;
; GFX9-GISEL-LABEL: fptrunc_v2f64_to_v2f16_afn:
@@ -3385,104 +2884,22 @@ define amdgpu_kernel void @fptrunc_v2f64_to_v2f16_afn(
;
; GFX950-SDAG-LABEL: fptrunc_v2f64_to_v2f16_afn:
; GFX950-SDAG: ; %bb.0: ; %entry
-; GFX950-SDAG-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x24
-; GFX950-SDAG-NEXT: s_mov_b32 s3, 0xf000
-; GFX950-SDAG-NEXT: s_mov_b32 s2, -1
-; GFX950-SDAG-NEXT: s_mov_b32 s6, s2
-; GFX950-SDAG-NEXT: s_mov_b32 s7, s3
+; GFX950-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX950-SDAG-NEXT: s_mov_b32 s7, 0xf000
+; GFX950-SDAG-NEXT: s_mov_b32 s6, -1
+; GFX950-SDAG-NEXT: s_mov_b32 s10, s6
+; GFX950-SDAG-NEXT: s_mov_b32 s11, s7
; GFX950-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; GFX950-SDAG-NEXT: s_mov_b32 s4, s10
-; GFX950-SDAG-NEXT: s_mov_b32 s5, s11
-; GFX950-SDAG-NEXT: buffer_load_dwordx4 v[0:3], off, s[4:7], 0
-; GFX950-SDAG-NEXT: s_mov_b32 s0, s8
-; GFX950-SDAG-NEXT: s_mov_b32 s1, s9
-; GFX950-SDAG-NEXT: s_movk_i32 s4, 0x7e00
+; GFX950-SDAG-NEXT: s_mov_b32 s8, s2
+; GFX950-SDAG-NEXT: s_mov_b32 s9, s3
+; GFX950-SDAG-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0
+; GFX950-SDAG-NEXT: s_mov_b32 s4, s0
+; GFX950-SDAG-NEXT: s_mov_b32 s5, s1
; GFX950-SDAG-NEXT: s_waitcnt vmcnt(0)
-; GFX950-SDAG-NEXT: v_readfirstlane_b32 s5, v3
-; GFX950-SDAG-NEXT: s_and_b32 s7, s5, 0x1ff
-; GFX950-SDAG-NEXT: v_readfirstlane_b32 s6, v1
-; GFX950-SDAG-NEXT: v_or_b32_e32 v1, s7, v2
-; GFX950-SDAG-NEXT: s_lshr_b32 s8, s5, 8
-; GFX950-SDAG-NEXT: s_bfe_u32 s9, s5, 0xb0014
-; GFX950-SDAG-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX950-SDAG-NEXT: s_and_b32 s7, s8, 0xffe
-; GFX950-SDAG-NEXT: s_sub_i32 s8, 0x3f1, s9
-; GFX950-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, 1, vcc
-; GFX950-SDAG-NEXT: v_med3_i32 v2, s8, 0, 13
-; GFX950-SDAG-NEXT: v_readfirstlane_b32 s8, v1
-; GFX950-SDAG-NEXT: s_or_b32 s7, s7, s8
-; GFX950-SDAG-NEXT: v_readfirstlane_b32 s10, v2
-; GFX950-SDAG-NEXT: s_or_b32 s8, s7, 0x1000
-; GFX950-SDAG-NEXT: s_lshr_b32 s11, s8, s10
-; GFX950-SDAG-NEXT: s_lshl_b32 s10, s11, s10
-; GFX950-SDAG-NEXT: s_cmp_lg_u32 s10, s8
-; GFX950-SDAG-NEXT: s_cselect_b32 s8, 1, 0
-; GFX950-SDAG-NEXT: s_addk_i32 s9, 0xfc10
-; GFX950-SDAG-NEXT: s_lshl_b32 s10, s9, 12
-; GFX950-SDAG-NEXT: s_or_b32 s8, s11, s8
-; GFX950-SDAG-NEXT: s_or_b32 s10, s7, s10
-; GFX950-SDAG-NEXT: s_cmp_lt_i32 s9, 1
-; GFX950-SDAG-NEXT: s_cselect_b32 s8, s8, s10
-; GFX950-SDAG-NEXT: s_and_b32 s10, s8, 7
-; GFX950-SDAG-NEXT: s_cmp_gt_i32 s10, 5
-; GFX950-SDAG-NEXT: s_cselect_b32 s11, 1, 0
-; GFX950-SDAG-NEXT: s_cmp_eq_u32 s10, 3
-; GFX950-SDAG-NEXT: s_cselect_b32 s10, 1, 0
-; GFX950-SDAG-NEXT: s_lshr_b32 s8, s8, 2
-; GFX950-SDAG-NEXT: s_or_b32 s10, s10, s11
-; GFX950-SDAG-NEXT: s_add_i32 s8, s8, s10
-; GFX950-SDAG-NEXT: s_cmp_lt_i32 s9, 31
-; GFX950-SDAG-NEXT: s_cselect_b32 s8, s8, 0x7c00
-; GFX950-SDAG-NEXT: s_cmp_lg_u32 s7, 0
-; GFX950-SDAG-NEXT: s_cselect_b32 s7, s4, 0x7c00
-; GFX950-SDAG-NEXT: s_cmpk_eq_i32 s9, 0x40f
-; GFX950-SDAG-NEXT: s_cselect_b32 s7, s7, s8
-; GFX950-SDAG-NEXT: s_and_b32 s8, s6, 0x1ff
-; GFX950-SDAG-NEXT: v_or_b32_e32 v0, s8, v0
-; GFX950-SDAG-NEXT: s_lshr_b32 s5, s5, 16
-; GFX950-SDAG-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; GFX950-SDAG-NEXT: s_lshr_b32 s9, s6, 8
-; GFX950-SDAG-NEXT: s_bfe_u32 s10, s6, 0xb0014
-; GFX950-SDAG-NEXT: s_and_b32 s5, s5, 0x8000
-; GFX950-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
-; GFX950-SDAG-NEXT: s_and_b32 s8, s9, 0xffe
-; GFX950-SDAG-NEXT: s_sub_i32 s9, 0x3f1, s10
-; GFX950-SDAG-NEXT: s_or_b32 s5, s5, s7
-; GFX950-SDAG-NEXT: v_readfirstlane_b32 s7, v0
-; GFX950-SDAG-NEXT: v_med3_i32 v1, s9, 0, 13
-; GFX950-SDAG-NEXT: s_or_b32 s7, s8, s7
-; GFX950-SDAG-NEXT: v_readfirstlane_b32 s9, v1
-; GFX950-SDAG-NEXT: s_or_b32 s8, s7, 0x1000
-; GFX950-SDAG-NEXT: s_lshr_b32 s11, s8, s9
-; GFX950-SDAG-NEXT: s_lshl_b32 s9, s11, s9
-; GFX950-SDAG-NEXT: s_cmp_lg_u32 s9, s8
-; GFX950-SDAG-NEXT: s_cselect_b32 s8, 1, 0
-; GFX950-SDAG-NEXT: s_addk_i32 s10, 0xfc10
-; GFX950-SDAG-NEXT: s_lshl_b32 s9, s10, 12
-; GFX950-SDAG-NEXT: s_or_b32 s8, s11, s8
-; GFX950-SDAG-NEXT: s_or_b32 s9, s7, s9
-; GFX950-SDAG-NEXT: s_cmp_lt_i32 s10, 1
-; GFX950-SDAG-NEXT: s_cselect_b32 s8, s8, s9
-; GFX950-SDAG-NEXT: s_and_b32 s9, s8, 7
-; GFX950-SDAG-NEXT: s_cmp_gt_i32 s9, 5
-; GFX950-SDAG-NEXT: s_cselect_b32 s11, 1, 0
-; GFX950-SDAG-NEXT: s_cmp_eq_u32 s9, 3
-; GFX950-SDAG-NEXT: s_cselect_b32 s9, 1, 0
-; GFX950-SDAG-NEXT: s_lshr_b32 s8, s8, 2
-; GFX950-SDAG-NEXT: s_or_b32 s9, s9, s11
-; GFX950-SDAG-NEXT: s_add_i32 s8, s8, s9
-; GFX950-SDAG-NEXT: s_cmp_lt_i32 s10, 31
-; GFX950-SDAG-NEXT: s_cselect_b32 s8, s8, 0x7c00
-; GFX950-SDAG-NEXT: s_cmp_lg_u32 s7, 0
-; GFX950-SDAG-NEXT: s_cselect_b32 s4, s4, 0x7c00
-; GFX950-SDAG-NEXT: s_cmpk_eq_i32 s10, 0x40f
-; GFX950-SDAG-NEXT: s_cselect_b32 s4, s4, s8
-; GFX950-SDAG-NEXT: s_lshr_b32 s6, s6, 16
-; GFX950-SDAG-NEXT: s_and_b32 s6, s6, 0x8000
-; GFX950-SDAG-NEXT: s_or_b32 s4, s6, s4
-; GFX950-SDAG-NEXT: s_pack_ll_b32_b16 s4, s4, s5
-; GFX950-SDAG-NEXT: v_mov_b32_e32 v0, s4
-; GFX950-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], 0
+; GFX950-SDAG-NEXT: v_cvt_f32_f64_e32 v2, v[2:3]
+; GFX950-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v0, v0, v2
+; GFX950-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GFX950-SDAG-NEXT: s_endpgm
;
; GFX950-GISEL-LABEL: fptrunc_v2f64_to_v2f16_afn:
@@ -3511,109 +2928,17 @@ define amdgpu_kernel void @fptrunc_v2f64_to_v2f16_afn(
; GFX11-SDAG-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-TRUE16-NEXT: s_mov_b32 s8, s2
; GFX11-SDAG-TRUE16-NEXT: s_mov_b32 s9, s3
+; GFX11-SDAG-TRUE16-NEXT: s_mov_b32 s4, s0
; GFX11-SDAG-TRUE16-NEXT: buffer_load_b128 v[0:3], off, s[8:11], 0
-; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-SDAG-TRUE16-NEXT: v_readfirstlane_b32 s2, v3
-; GFX11-SDAG-TRUE16-NEXT: s_and_b32 s3, s2, 0x1ff
-; GFX11-SDAG-TRUE16-NEXT: s_lshr_b32 s5, s2, 8
-; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v2, s3, v2
-; GFX11-SDAG-TRUE16-NEXT: s_bfe_u32 s3, s2, 0xb0014
-; GFX11-SDAG-TRUE16-NEXT: s_and_b32 s5, s5, 0xffe
-; GFX11-SDAG-TRUE16-NEXT: s_sub_i32 s4, 0x3f1, s3
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-SDAG-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-SDAG-TRUE16-NEXT: v_med3_i32 v3, s4, 0, 13
-; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc_lo
-; GFX11-SDAG-TRUE16-NEXT: v_readfirstlane_b32 s8, v3
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: v_readfirstlane_b32 s4, v2
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s4, s5, s4
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s5, s4, 0x1000
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: s_lshr_b32 s9, s5, s8
-; GFX11-SDAG-TRUE16-NEXT: s_lshl_b32 s8, s9, s8
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_lg_u32 s8, s5
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s5, 1, 0
-; GFX11-SDAG-TRUE16-NEXT: s_addk_i32 s3, 0xfc10
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s5, s9, s5
-; GFX11-SDAG-TRUE16-NEXT: s_lshl_b32 s8, s3, 12
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s8, s4, s8
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_lt_i32 s3, 1
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s5, s5, s8
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: s_and_b32 s8, s5, 7
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_gt_i32 s8, 5
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s9, 1, 0
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_eq_u32 s8, 3
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s8, 1, 0
-; GFX11-SDAG-TRUE16-NEXT: s_lshr_b32 s5, s5, 2
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s8, s8, s9
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: s_add_i32 s5, s5, s8
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_lt_i32 s3, 31
-; GFX11-SDAG-TRUE16-NEXT: s_movk_i32 s8, 0x7e00
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s5, s5, 0x7c00
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_lg_u32 s4, 0
-; GFX11-SDAG-TRUE16-NEXT: v_readfirstlane_b32 s4, v1
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s9, s8, 0x7c00
-; GFX11-SDAG-TRUE16-NEXT: s_cmpk_eq_i32 s3, 0x40f
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s3, s9, s5
-; GFX11-SDAG-TRUE16-NEXT: s_and_b32 s5, s4, 0x1ff
-; GFX11-SDAG-TRUE16-NEXT: s_lshr_b32 s10, s4, 8
-; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v0, s5, v0
-; GFX11-SDAG-TRUE16-NEXT: s_bfe_u32 s5, s4, 0xb0014
-; GFX11-SDAG-TRUE16-NEXT: s_and_b32 s10, s10, 0xffe
-; GFX11-SDAG-TRUE16-NEXT: s_sub_i32 s9, 0x3f1, s5
-; GFX11-SDAG-TRUE16-NEXT: s_lshr_b32 s2, s2, 16
-; GFX11-SDAG-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-SDAG-TRUE16-NEXT: v_med3_i32 v1, s9, 0, 13
-; GFX11-SDAG-TRUE16-NEXT: s_and_b32 s2, s2, 0x8000
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s2, s2, s3
-; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
-; GFX11-SDAG-TRUE16-NEXT: v_readfirstlane_b32 s11, v1
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: v_readfirstlane_b32 s9, v0
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s9, s10, s9
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s10, s9, 0x1000
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: s_lshr_b32 s12, s10, s11
-; GFX11-SDAG-TRUE16-NEXT: s_lshl_b32 s11, s12, s11
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_lg_u32 s11, s10
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s3, 1, 0
-; GFX11-SDAG-TRUE16-NEXT: s_addk_i32 s5, 0xfc10
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s3, s12, s3
-; GFX11-SDAG-TRUE16-NEXT: s_lshl_b32 s10, s5, 12
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s10, s9, s10
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_lt_i32 s5, 1
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s3, s3, s10
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: s_and_b32 s10, s3, 7
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_gt_i32 s10, 5
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s11, 1, 0
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_eq_u32 s10, 3
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s10, 1, 0
-; GFX11-SDAG-TRUE16-NEXT: s_lshr_b32 s3, s3, 2
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s10, s10, s11
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: s_add_i32 s3, s3, s10
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_lt_i32 s5, 31
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s3, s3, 0x7c00
-; GFX11-SDAG-TRUE16-NEXT: s_cmp_lg_u32 s9, 0
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s8, s8, 0x7c00
-; GFX11-SDAG-TRUE16-NEXT: s_cmpk_eq_i32 s5, 0x40f
; GFX11-SDAG-TRUE16-NEXT: s_mov_b32 s5, s1
-; GFX11-SDAG-TRUE16-NEXT: s_cselect_b32 s3, s8, s3
-; GFX11-SDAG-TRUE16-NEXT: s_lshr_b32 s4, s4, 16
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: s_and_b32 s4, s4, 0x8000
-; GFX11-SDAG-TRUE16-NEXT: s_or_b32 s3, s4, s3
-; GFX11-SDAG-TRUE16-NEXT: s_mov_b32 s4, s0
-; GFX11-SDAG-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s3, s2
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-SDAG-TRUE16-NEXT: v_mov_b32_e32 v0, s2
+; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-SDAG-TRUE16-NEXT: v_cvt_f32_f64_e32 v2, v[2:3]
+; GFX11-SDAG-TRUE16-NEXT: v_cvt_f32_f64_e32 v1, v[0:1]
+; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-SDAG-TRUE16-NEXT: v_cvt_f16_f32_e32 v0.l, v2
+; GFX11-SDAG-TRUE16-NEXT: v_cvt_f16_f32_e32 v0.h, v1
+; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-SDAG-TRUE16-NEXT: v_pack_b32_f16 v0, v0.h, v0.l
; GFX11-SDAG-TRUE16-NEXT: buffer_store_b32 v0, off, s[4:7], 0
; GFX11-SDAG-TRUE16-NEXT: s_endpgm
;
@@ -3627,109 +2952,17 @@ define amdgpu_kernel void @fptrunc_v2f64_to_v2f16_afn(
; GFX11-SDAG-FAKE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-FAKE16-NEXT: s_mov_b32 s8, s2
; GFX11-SDAG-FAKE16-NEXT: s_mov_b32 s9, s3
+; GFX11-SDAG-FAKE16-NEXT: s_mov_b32 s4, s0
; GFX11-SDAG-FAKE16-NEXT: buffer_load_b128 v[0:3], off, s[8:11], 0
-; GFX11-SDAG-FAKE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-SDAG-FAKE16-NEXT: v_readfirstlane_b32 s2, v3
-; GFX11-SDAG-FAKE16-NEXT: s_and_b32 s3, s2, 0x1ff
-; GFX11-SDAG-FAKE16-NEXT: s_lshr_b32 s5, s2, 8
-; GFX11-SDAG-FAKE16-NEXT: v_or_b32_e32 v2, s3, v2
-; GFX11-SDAG-FAKE16-NEXT: s_bfe_u32 s3, s2, 0xb0014
-; GFX11-SDAG-FAKE16-NEXT: s_and_b32 s5, s5, 0xffe
-; GFX11-SDAG-FAKE16-NEXT: s_sub_i32 s4, 0x3f1, s3
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-SDAG-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-SDAG-FAKE16-NEXT: v_med3_i32 v3, s4, 0, 13
-; GFX11-SDAG-FAKE16-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc_lo
-; GFX11-SDAG-FAKE16-NEXT: v_readfirstlane_b32 s8, v3
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: v_readfirstlane_b32 s4, v2
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s4, s5, s4
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s5, s4, 0x1000
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: s_lshr_b32 s9, s5, s8
-; GFX11-SDAG-FAKE16-NEXT: s_lshl_b32 s8, s9, s8
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_lg_u32 s8, s5
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s5, 1, 0
-; GFX11-SDAG-FAKE16-NEXT: s_addk_i32 s3, 0xfc10
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s5, s9, s5
-; GFX11-SDAG-FAKE16-NEXT: s_lshl_b32 s8, s3, 12
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s8, s4, s8
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_lt_i32 s3, 1
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s5, s5, s8
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: s_and_b32 s8, s5, 7
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_gt_i32 s8, 5
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s9, 1, 0
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_eq_u32 s8, 3
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s8, 1, 0
-; GFX11-SDAG-FAKE16-NEXT: s_lshr_b32 s5, s5, 2
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s8, s8, s9
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: s_add_i32 s5, s5, s8
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_lt_i32 s3, 31
-; GFX11-SDAG-FAKE16-NEXT: s_movk_i32 s8, 0x7e00
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s5, s5, 0x7c00
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_lg_u32 s4, 0
-; GFX11-SDAG-FAKE16-NEXT: v_readfirstlane_b32 s4, v1
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s9, s8, 0x7c00
-; GFX11-SDAG-FAKE16-NEXT: s_cmpk_eq_i32 s3, 0x40f
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s3, s9, s5
-; GFX11-SDAG-FAKE16-NEXT: s_and_b32 s5, s4, 0x1ff
-; GFX11-SDAG-FAKE16-NEXT: s_lshr_b32 s10, s4, 8
-; GFX11-SDAG-FAKE16-NEXT: v_or_b32_e32 v0, s5, v0
-; GFX11-SDAG-FAKE16-NEXT: s_bfe_u32 s5, s4, 0xb0014
-; GFX11-SDAG-FAKE16-NEXT: s_and_b32 s10, s10, 0xffe
-; GFX11-SDAG-FAKE16-NEXT: s_sub_i32 s9, 0x3f1, s5
-; GFX11-SDAG-FAKE16-NEXT: s_lshr_b32 s2, s2, 16
-; GFX11-SDAG-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-SDAG-FAKE16-NEXT: v_med3_i32 v1, s9, 0, 13
-; GFX11-SDAG-FAKE16-NEXT: s_and_b32 s2, s2, 0x8000
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s2, s2, s3
-; GFX11-SDAG-FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
-; GFX11-SDAG-FAKE16-NEXT: v_readfirstlane_b32 s11, v1
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: v_readfirstlane_b32 s9, v0
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s9, s10, s9
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s10, s9, 0x1000
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: s_lshr_b32 s12, s10, s11
-; GFX11-SDAG-FAKE16-NEXT: s_lshl_b32 s11, s12, s11
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_lg_u32 s11, s10
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s3, 1, 0
-; GFX11-SDAG-FAKE16-NEXT: s_addk_i32 s5, 0xfc10
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s3, s12, s3
-; GFX11-SDAG-FAKE16-NEXT: s_lshl_b32 s10, s5, 12
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s10, s9, s10
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_lt_i32 s5, 1
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s3, s3, s10
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: s_and_b32 s10, s3, 7
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_gt_i32 s10, 5
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s11, 1, 0
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_eq_u32 s10, 3
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s10, 1, 0
-; GFX11-SDAG-FAKE16-NEXT: s_lshr_b32 s3, s3, 2
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s10, s10, s11
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: s_add_i32 s3, s3, s10
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_lt_i32 s5, 31
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s3, s3, 0x7c00
-; GFX11-SDAG-FAKE16-NEXT: s_cmp_lg_u32 s9, 0
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s8, s8, 0x7c00
-; GFX11-SDAG-FAKE16-NEXT: s_cmpk_eq_i32 s5, 0x40f
; GFX11-SDAG-FAKE16-NEXT: s_mov_b32 s5, s1
-; GFX11-SDAG-FAKE16-NEXT: s_cselect_b32 s3, s8, s3
-; GFX11-SDAG-FAKE16-NEXT: s_lshr_b32 s4, s4, 16
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: s_and_b32 s4, s4, 0x8000
-; GFX11-SDAG-FAKE16-NEXT: s_or_b32 s3, s4, s3
-; GFX11-SDAG-FAKE16-NEXT: s_mov_b32 s4, s0
-; GFX11-SDAG-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s3, s2
-; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-SDAG-FAKE16-NEXT: v_mov_b32_e32 v0, s2
+; GFX11-SDAG-FAKE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-SDAG-FAKE16-NEXT: v_cvt_f32_f64_e32 v2, v[2:3]
+; GFX11-SDAG-FAKE16-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-SDAG-FAKE16-NEXT: v_cvt_f16_f32_e32 v1, v2
+; GFX11-SDAG-FAKE16-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX11-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-SDAG-FAKE16-NEXT: v_pack_b32_f16 v0, v0, v1
; GFX11-SDAG-FAKE16-NEXT: buffer_store_b32 v0, off, s[4:7], 0
; GFX11-SDAG-FAKE16-NEXT: s_endpgm
;
diff --git a/llvm/test/CodeGen/AMDGPU/fptrunc.ll b/llvm/test/CodeGen/AMDGPU/fptrunc.ll
index 4f8eab1c2fec0..5d311776066e5 100644
--- a/llvm/test/CodeGen/AMDGPU/fptrunc.ll
+++ b/llvm/test/CodeGen/AMDGPU/fptrunc.ll
@@ -226,59 +226,59 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in)
; SI-NEXT: buffer_store_short v0, off, s[0:3], 0
; SI-NEXT: s_endpgm
;
-; VI-SAFE-SDAG-LABEL: fptrunc_f64_to_f16:
-; VI-SAFE-SDAG: ; %bb.0:
-; VI-SAFE-SDAG-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x24
-; VI-SAFE-SDAG-NEXT: s_mov_b32 s3, 0xf000
-; VI-SAFE-SDAG-NEXT: s_mov_b32 s2, -1
-; VI-SAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; VI-SAFE-SDAG-NEXT: s_mov_b32 s0, s4
-; VI-SAFE-SDAG-NEXT: s_lshr_b32 s4, s7, 8
-; VI-SAFE-SDAG-NEXT: s_and_b32 s8, s4, 0xffe
-; VI-SAFE-SDAG-NEXT: s_and_b32 s4, s7, 0x1ff
-; VI-SAFE-SDAG-NEXT: s_or_b32 s4, s4, s6
-; VI-SAFE-SDAG-NEXT: s_cmp_lg_u32 s4, 0
-; VI-SAFE-SDAG-NEXT: s_mov_b32 s1, s5
-; VI-SAFE-SDAG-NEXT: s_cselect_b64 s[4:5], -1, 0
-; VI-SAFE-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[4:5]
-; VI-SAFE-SDAG-NEXT: v_readfirstlane_b32 s4, v0
-; VI-SAFE-SDAG-NEXT: s_bfe_u32 s6, s7, 0xb0014
-; VI-SAFE-SDAG-NEXT: s_or_b32 s4, s8, s4
-; VI-SAFE-SDAG-NEXT: s_sub_i32 s8, 0x3f1, s6
-; VI-SAFE-SDAG-NEXT: v_med3_i32 v0, s8, 0, 13
-; VI-SAFE-SDAG-NEXT: s_or_b32 s5, s4, 0x1000
-; VI-SAFE-SDAG-NEXT: v_readfirstlane_b32 s8, v0
-; VI-SAFE-SDAG-NEXT: s_lshr_b32 s9, s5, s8
-; VI-SAFE-SDAG-NEXT: s_lshl_b32 s8, s9, s8
-; VI-SAFE-SDAG-NEXT: s_cmp_lg_u32 s8, s5
-; VI-SAFE-SDAG-NEXT: s_cselect_b32 s5, 1, 0
-; VI-SAFE-SDAG-NEXT: s_addk_i32 s6, 0xfc10
-; VI-SAFE-SDAG-NEXT: s_lshl_b32 s8, s6, 12
-; VI-SAFE-SDAG-NEXT: s_or_b32 s5, s9, s5
-; VI-SAFE-SDAG-NEXT: s_or_b32 s8, s4, s8
-; VI-SAFE-SDAG-NEXT: s_cmp_lt_i32 s6, 1
-; VI-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, s8
-; VI-SAFE-SDAG-NEXT: s_and_b32 s8, s5, 7
-; VI-SAFE-SDAG-NEXT: s_cmp_gt_i32 s8, 5
-; VI-SAFE-SDAG-NEXT: s_cselect_b32 s9, 1, 0
-; VI-SAFE-SDAG-NEXT: s_cmp_eq_u32 s8, 3
-; VI-SAFE-SDAG-NEXT: s_cselect_b32 s8, 1, 0
-; VI-SAFE-SDAG-NEXT: s_or_b32 s8, s8, s9
-; VI-SAFE-SDAG-NEXT: s_lshr_b32 s5, s5, 2
-; VI-SAFE-SDAG-NEXT: s_add_i32 s5, s5, s8
-; VI-SAFE-SDAG-NEXT: s_cmp_lt_i32 s6, 31
-; VI-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, 0x7c00
-; VI-SAFE-SDAG-NEXT: s_cmp_lg_u32 s4, 0
-; VI-SAFE-SDAG-NEXT: s_movk_i32 s4, 0x7e00
-; VI-SAFE-SDAG-NEXT: s_cselect_b32 s4, s4, 0x7c00
-; VI-SAFE-SDAG-NEXT: s_cmpk_eq_i32 s6, 0x40f
-; VI-SAFE-SDAG-NEXT: s_cselect_b32 s4, s4, s5
-; VI-SAFE-SDAG-NEXT: s_lshr_b32 s5, s7, 16
-; VI-SAFE-SDAG-NEXT: s_and_b32 s5, s5, 0x8000
-; VI-SAFE-SDAG-NEXT: s_or_b32 s4, s5, s4
-; VI-SAFE-SDAG-NEXT: v_mov_b32_e32 v0, s4
-; VI-SAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
-; VI-SAFE-SDAG-NEXT: s_endpgm
+; VI-SDAG-LABEL: fptrunc_f64_to_f16:
+; VI-SDAG: ; %bb.0:
+; VI-SDAG-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x24
+; VI-SDAG-NEXT: s_mov_b32 s3, 0xf000
+; VI-SDAG-NEXT: s_mov_b32 s2, -1
+; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; VI-SDAG-NEXT: s_mov_b32 s0, s4
+; VI-SDAG-NEXT: s_lshr_b32 s4, s7, 8
+; VI-SDAG-NEXT: s_and_b32 s8, s4, 0xffe
+; VI-SDAG-NEXT: s_and_b32 s4, s7, 0x1ff
+; VI-SDAG-NEXT: s_or_b32 s4, s4, s6
+; VI-SDAG-NEXT: s_cmp_lg_u32 s4, 0
+; VI-SDAG-NEXT: s_mov_b32 s1, s5
+; VI-SDAG-NEXT: s_cselect_b64 s[4:5], -1, 0
+; VI-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[4:5]
+; VI-SDAG-NEXT: v_readfirstlane_b32 s4, v0
+; VI-SDAG-NEXT: s_bfe_u32 s6, s7, 0xb0014
+; VI-SDAG-NEXT: s_or_b32 s4, s8, s4
+; VI-SDAG-NEXT: s_sub_i32 s8, 0x3f1, s6
+; VI-SDAG-NEXT: v_med3_i32 v0, s8, 0, 13
+; VI-SDAG-NEXT: s_or_b32 s5, s4, 0x1000
+; VI-SDAG-NEXT: v_readfirstlane_b32 s8, v0
+; VI-SDAG-NEXT: s_lshr_b32 s9, s5, s8
+; VI-SDAG-NEXT: s_lshl_b32 s8, s9, s8
+; VI-SDAG-NEXT: s_cmp_lg_u32 s8, s5
+; VI-SDAG-NEXT: s_cselect_b32 s5, 1, 0
+; VI-SDAG-NEXT: s_addk_i32 s6, 0xfc10
+; VI-SDAG-NEXT: s_lshl_b32 s8, s6, 12
+; VI-SDAG-NEXT: s_or_b32 s5, s9, s5
+; VI-SDAG-NEXT: s_or_b32 s8, s4, s8
+; VI-SDAG-NEXT: s_cmp_lt_i32 s6, 1
+; VI-SDAG-NEXT: s_cselect_b32 s5, s5, s8
+; VI-SDAG-NEXT: s_and_b32 s8, s5, 7
+; VI-SDAG-NEXT: s_cmp_gt_i32 s8, 5
+; VI-SDAG-NEXT: s_cselect_b32 s9, 1, 0
+; VI-SDAG-NEXT: s_cmp_eq_u32 s8, 3
+; VI-SDAG-NEXT: s_cselect_b32 s8, 1, 0
+; VI-SDAG-NEXT: s_or_b32 s8, s8, s9
+; VI-SDAG-NEXT: s_lshr_b32 s5, s5, 2
+; VI-SDAG-NEXT: s_add_i32 s5, s5, s8
+; VI-SDAG-NEXT: s_cmp_lt_i32 s6, 31
+; VI-SDAG-NEXT: s_cselect_b32 s5, s5, 0x7c00
+; VI-SDAG-NEXT: s_cmp_lg_u32 s4, 0
+; VI-SDAG-NEXT: s_movk_i32 s4, 0x7e00
+; VI-SDAG-NEXT: s_cselect_b32 s4, s4, 0x7c00
+; VI-SDAG-NEXT: s_cmpk_eq_i32 s6, 0x40f
+; VI-SDAG-NEXT: s_cselect_b32 s4, s4, s5
+; VI-SDAG-NEXT: s_lshr_b32 s5, s7, 16
+; VI-SDAG-NEXT: s_and_b32 s5, s5, 0x8000
+; VI-SDAG-NEXT: s_or_b32 s4, s5, s4
+; VI-SDAG-NEXT: v_mov_b32_e32 v0, s4
+; VI-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
+; VI-SDAG-NEXT: s_endpgm
;
; VI-GISEL-LABEL: fptrunc_f64_to_f16:
; VI-GISEL: ; %bb.0:
@@ -331,68 +331,57 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in)
; VI-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0
; VI-GISEL-NEXT: s_endpgm
;
-; VI-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16:
-; VI-UNSAFE-SDAG: ; %bb.0:
-; VI-UNSAFE-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
-; VI-UNSAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; VI-UNSAFE-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[2:3]
-; VI-UNSAFE-SDAG-NEXT: s_mov_b32 s3, 0xf000
-; VI-UNSAFE-SDAG-NEXT: s_mov_b32 s2, -1
-; VI-UNSAFE-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
-; VI-UNSAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
-; VI-UNSAFE-SDAG-NEXT: s_endpgm
-;
-; GFX10-SAFE-SDAG-LABEL: fptrunc_f64_to_f16:
-; GFX10-SAFE-SDAG: ; %bb.0:
-; GFX10-SAFE-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
-; GFX10-SAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-SAFE-SDAG-NEXT: s_and_b32 s4, s3, 0x1ff
-; GFX10-SAFE-SDAG-NEXT: s_lshr_b32 s5, s3, 8
-; GFX10-SAFE-SDAG-NEXT: s_or_b32 s2, s4, s2
-; GFX10-SAFE-SDAG-NEXT: s_and_b32 s4, s5, 0xffe
-; GFX10-SAFE-SDAG-NEXT: s_cmp_lg_u32 s2, 0
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s2, -1, 0
-; GFX10-SAFE-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, s2
-; GFX10-SAFE-SDAG-NEXT: s_bfe_u32 s2, s3, 0xb0014
-; GFX10-SAFE-SDAG-NEXT: s_sub_i32 s5, 0x3f1, s2
-; GFX10-SAFE-SDAG-NEXT: v_med3_i32 v1, s5, 0, 13
-; GFX10-SAFE-SDAG-NEXT: v_readfirstlane_b32 s5, v0
-; GFX10-SAFE-SDAG-NEXT: v_readfirstlane_b32 s6, v1
-; GFX10-SAFE-SDAG-NEXT: s_or_b32 s4, s4, s5
-; GFX10-SAFE-SDAG-NEXT: s_or_b32 s5, s4, 0x1000
-; GFX10-SAFE-SDAG-NEXT: s_lshr_b32 s7, s5, s6
-; GFX10-SAFE-SDAG-NEXT: s_lshl_b32 s6, s7, s6
-; GFX10-SAFE-SDAG-NEXT: s_cmp_lg_u32 s6, s5
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s5, 1, 0
-; GFX10-SAFE-SDAG-NEXT: s_addk_i32 s2, 0xfc10
-; GFX10-SAFE-SDAG-NEXT: s_or_b32 s5, s7, s5
-; GFX10-SAFE-SDAG-NEXT: s_lshl_b32 s6, s2, 12
-; GFX10-SAFE-SDAG-NEXT: s_or_b32 s6, s4, s6
-; GFX10-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 1
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, s6
-; GFX10-SAFE-SDAG-NEXT: s_and_b32 s6, s5, 7
-; GFX10-SAFE-SDAG-NEXT: s_cmp_gt_i32 s6, 5
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s7, 1, 0
-; GFX10-SAFE-SDAG-NEXT: s_cmp_eq_u32 s6, 3
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s6, 1, 0
-; GFX10-SAFE-SDAG-NEXT: s_lshr_b32 s5, s5, 2
-; GFX10-SAFE-SDAG-NEXT: s_or_b32 s6, s6, s7
-; GFX10-SAFE-SDAG-NEXT: s_add_i32 s5, s5, s6
-; GFX10-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 31
-; GFX10-SAFE-SDAG-NEXT: s_movk_i32 s6, 0x7e00
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, 0x7c00
-; GFX10-SAFE-SDAG-NEXT: s_cmp_lg_u32 s4, 0
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s4, s6, 0x7c00
-; GFX10-SAFE-SDAG-NEXT: s_cmpk_eq_i32 s2, 0x40f
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s2, s4, s5
-; GFX10-SAFE-SDAG-NEXT: s_lshr_b32 s3, s3, 16
-; GFX10-SAFE-SDAG-NEXT: s_and_b32 s3, s3, 0x8000
-; GFX10-SAFE-SDAG-NEXT: s_or_b32 s2, s3, s2
-; GFX10-SAFE-SDAG-NEXT: s_mov_b32 s3, 0x31016000
-; GFX10-SAFE-SDAG-NEXT: v_mov_b32_e32 v0, s2
-; GFX10-SAFE-SDAG-NEXT: s_mov_b32 s2, -1
-; GFX10-SAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
-; GFX10-SAFE-SDAG-NEXT: s_endpgm
+; GFX10-SDAG-LABEL: fptrunc_f64_to_f16:
+; GFX10-SDAG: ; %bb.0:
+; GFX10-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX10-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-SDAG-NEXT: s_and_b32 s4, s3, 0x1ff
+; GFX10-SDAG-NEXT: s_lshr_b32 s5, s3, 8
+; GFX10-SDAG-NEXT: s_or_b32 s2, s4, s2
+; GFX10-SDAG-NEXT: s_and_b32 s4, s5, 0xffe
+; GFX10-SDAG-NEXT: s_cmp_lg_u32 s2, 0
+; GFX10-SDAG-NEXT: s_cselect_b32 s2, -1, 0
+; GFX10-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, s2
+; GFX10-SDAG-NEXT: s_bfe_u32 s2, s3, 0xb0014
+; GFX10-SDAG-NEXT: s_sub_i32 s5, 0x3f1, s2
+; GFX10-SDAG-NEXT: v_med3_i32 v1, s5, 0, 13
+; GFX10-SDAG-NEXT: v_readfirstlane_b32 s5, v0
+; GFX10-SDAG-NEXT: v_readfirstlane_b32 s6, v1
+; GFX10-SDAG-NEXT: s_or_b32 s4, s4, s5
+; GFX10-SDAG-NEXT: s_or_b32 s5, s4, 0x1000
+; GFX10-SDAG-NEXT: s_lshr_b32 s7, s5, s6
+; GFX10-SDAG-NEXT: s_lshl_b32 s6, s7, s6
+; GFX10-SDAG-NEXT: s_cmp_lg_u32 s6, s5
+; GFX10-SDAG-NEXT: s_cselect_b32 s5, 1, 0
+; GFX10-SDAG-NEXT: s_addk_i32 s2, 0xfc10
+; GFX10-SDAG-NEXT: s_or_b32 s5, s7, s5
+; GFX10-SDAG-NEXT: s_lshl_b32 s6, s2, 12
+; GFX10-SDAG-NEXT: s_or_b32 s6, s4, s6
+; GFX10-SDAG-NEXT: s_cmp_lt_i32 s2, 1
+; GFX10-SDAG-NEXT: s_cselect_b32 s5, s5, s6
+; GFX10-SDAG-NEXT: s_and_b32 s6, s5, 7
+; GFX10-SDAG-NEXT: s_cmp_gt_i32 s6, 5
+; GFX10-SDAG-NEXT: s_cselect_b32 s7, 1, 0
+; GFX10-SDAG-NEXT: s_cmp_eq_u32 s6, 3
+; GFX10-SDAG-NEXT: s_cselect_b32 s6, 1, 0
+; GFX10-SDAG-NEXT: s_lshr_b32 s5, s5, 2
+; GFX10-SDAG-NEXT: s_or_b32 s6, s6, s7
+; GFX10-SDAG-NEXT: s_add_i32 s5, s5, s6
+; GFX10-SDAG-NEXT: s_cmp_lt_i32 s2, 31
+; GFX10-SDAG-NEXT: s_movk_i32 s6, 0x7e00
+; GFX10-SDAG-NEXT: s_cselect_b32 s5, s5, 0x7c00
+; GFX10-SDAG-NEXT: s_cmp_lg_u32 s4, 0
+; GFX10-SDAG-NEXT: s_cselect_b32 s4, s6, 0x7c00
+; GFX10-SDAG-NEXT: s_cmpk_eq_i32 s2, 0x40f
+; GFX10-SDAG-NEXT: s_cselect_b32 s2, s4, s5
+; GFX10-SDAG-NEXT: s_lshr_b32 s3, s3, 16
+; GFX10-SDAG-NEXT: s_and_b32 s3, s3, 0x8000
+; GFX10-SDAG-NEXT: s_or_b32 s2, s3, s2
+; GFX10-SDAG-NEXT: s_mov_b32 s3, 0x31016000
+; GFX10-SDAG-NEXT: v_mov_b32_e32 v0, s2
+; GFX10-SDAG-NEXT: s_mov_b32 s2, -1
+; GFX10-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
+; GFX10-SDAG-NEXT: s_endpgm
;
; GFX10-GISEL-LABEL: fptrunc_f64_to_f16:
; GFX10-GISEL: ; %bb.0:
@@ -445,76 +434,65 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in)
; GFX10-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0
; GFX10-GISEL-NEXT: s_endpgm
;
-; GFX10-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16:
-; GFX10-UNSAFE-SDAG: ; %bb.0:
-; GFX10-UNSAFE-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
-; GFX10-UNSAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-UNSAFE-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[2:3]
-; GFX10-UNSAFE-SDAG-NEXT: s_mov_b32 s3, 0x31016000
-; GFX10-UNSAFE-SDAG-NEXT: s_mov_b32 s2, -1
-; GFX10-UNSAFE-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GFX10-UNSAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
-; GFX10-UNSAFE-SDAG-NEXT: s_endpgm
-;
-; GFX11-SAFE-SDAG-LABEL: fptrunc_f64_to_f16:
-; GFX11-SAFE-SDAG: ; %bb.0:
-; GFX11-SAFE-SDAG-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-SAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-SAFE-SDAG-NEXT: s_and_b32 s4, s3, 0x1ff
-; GFX11-SAFE-SDAG-NEXT: s_lshr_b32 s5, s3, 8
-; GFX11-SAFE-SDAG-NEXT: s_or_b32 s2, s4, s2
-; GFX11-SAFE-SDAG-NEXT: s_and_b32 s4, s5, 0xffe
-; GFX11-SAFE-SDAG-NEXT: s_cmp_lg_u32 s2, 0
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s2, -1, 0
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-SAFE-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, s2
-; GFX11-SAFE-SDAG-NEXT: s_bfe_u32 s2, s3, 0xb0014
-; GFX11-SAFE-SDAG-NEXT: s_sub_i32 s5, 0x3f1, s2
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-SAFE-SDAG-NEXT: v_med3_i32 v1, s5, 0, 13
-; GFX11-SAFE-SDAG-NEXT: v_readfirstlane_b32 s5, v0
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-SAFE-SDAG-NEXT: v_readfirstlane_b32 s6, v1
-; GFX11-SAFE-SDAG-NEXT: s_or_b32 s4, s4, s5
-; GFX11-SAFE-SDAG-NEXT: s_or_b32 s5, s4, 0x1000
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SAFE-SDAG-NEXT: s_lshr_b32 s7, s5, s6
-; GFX11-SAFE-SDAG-NEXT: s_lshl_b32 s6, s7, s6
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
-; GFX11-SAFE-SDAG-NEXT: s_cmp_lg_u32 s6, s5
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s5, 1, 0
-; GFX11-SAFE-SDAG-NEXT: s_addk_i32 s2, 0xfc10
-; GFX11-SAFE-SDAG-NEXT: s_or_b32 s5, s7, s5
-; GFX11-SAFE-SDAG-NEXT: s_lshl_b32 s6, s2, 12
-; GFX11-SAFE-SDAG-NEXT: s_or_b32 s6, s4, s6
-; GFX11-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 1
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, s6
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SAFE-SDAG-NEXT: s_and_b32 s6, s5, 7
-; GFX11-SAFE-SDAG-NEXT: s_cmp_gt_i32 s6, 5
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s7, 1, 0
-; GFX11-SAFE-SDAG-NEXT: s_cmp_eq_u32 s6, 3
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s6, 1, 0
-; GFX11-SAFE-SDAG-NEXT: s_lshr_b32 s5, s5, 2
-; GFX11-SAFE-SDAG-NEXT: s_or_b32 s6, s6, s7
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-SAFE-SDAG-NEXT: s_add_i32 s5, s5, s6
-; GFX11-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 31
-; GFX11-SAFE-SDAG-NEXT: s_movk_i32 s6, 0x7e00
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, 0x7c00
-; GFX11-SAFE-SDAG-NEXT: s_cmp_lg_u32 s4, 0
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s4, s6, 0x7c00
-; GFX11-SAFE-SDAG-NEXT: s_cmpk_eq_i32 s2, 0x40f
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s2, s4, s5
-; GFX11-SAFE-SDAG-NEXT: s_lshr_b32 s3, s3, 16
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SAFE-SDAG-NEXT: s_and_b32 s3, s3, 0x8000
-; GFX11-SAFE-SDAG-NEXT: s_or_b32 s2, s3, s2
-; GFX11-SAFE-SDAG-NEXT: s_mov_b32 s3, 0x31016000
-; GFX11-SAFE-SDAG-NEXT: v_mov_b32_e32 v0, s2
-; GFX11-SAFE-SDAG-NEXT: s_mov_b32 s2, -1
-; GFX11-SAFE-SDAG-NEXT: buffer_store_b16 v0, off, s[0:3], 0
-; GFX11-SAFE-SDAG-NEXT: s_endpgm
+; GFX11-SDAG-LABEL: fptrunc_f64_to_f16:
+; GFX11-SDAG: ; %bb.0:
+; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
+; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-SDAG-NEXT: s_and_b32 s4, s3, 0x1ff
+; GFX11-SDAG-NEXT: s_lshr_b32 s5, s3, 8
+; GFX11-SDAG-NEXT: s_or_b32 s2, s4, s2
+; GFX11-SDAG-NEXT: s_and_b32 s4, s5, 0xffe
+; GFX11-SDAG-NEXT: s_cmp_lg_u32 s2, 0
+; GFX11-SDAG-NEXT: s_cselect_b32 s2, -1, 0
+; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, s2
+; GFX11-SDAG-NEXT: s_bfe_u32 s2, s3, 0xb0014
+; GFX11-SDAG-NEXT: s_sub_i32 s5, 0x3f1, s2
+; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-SDAG-NEXT: v_med3_i32 v1, s5, 0, 13
+; GFX11-SDAG-NEXT: v_readfirstlane_b32 s5, v0
+; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-SDAG-NEXT: v_readfirstlane_b32 s6, v1
+; GFX11-SDAG-NEXT: s_or_b32 s4, s4, s5
+; GFX11-SDAG-NEXT: s_or_b32 s5, s4, 0x1000
+; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-SDAG-NEXT: s_lshr_b32 s7, s5, s6
+; GFX11-SDAG-NEXT: s_lshl_b32 s6, s7, s6
+; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
+; GFX11-SDAG-NEXT: s_cmp_lg_u32 s6, s5
+; GFX11-SDAG-NEXT: s_cselect_b32 s5, 1, 0
+; GFX11-SDAG-NEXT: s_addk_i32 s2, 0xfc10
+; GFX11-SDAG-NEXT: s_or_b32 s5, s7, s5
+; GFX11-SDAG-NEXT: s_lshl_b32 s6, s2, 12
+; GFX11-SDAG-NEXT: s_or_b32 s6, s4, s6
+; GFX11-SDAG-NEXT: s_cmp_lt_i32 s2, 1
+; GFX11-SDAG-NEXT: s_cselect_b32 s5, s5, s6
+; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-SDAG-NEXT: s_and_b32 s6, s5, 7
+; GFX11-SDAG-NEXT: s_cmp_gt_i32 s6, 5
+; GFX11-SDAG-NEXT: s_cselect_b32 s7, 1, 0
+; GFX11-SDAG-NEXT: s_cmp_eq_u32 s6, 3
+; GFX11-SDAG-NEXT: s_cselect_b32 s6, 1, 0
+; GFX11-SDAG-NEXT: s_lshr_b32 s5, s5, 2
+; GFX11-SDAG-NEXT: s_or_b32 s6, s6, s7
+; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-SDAG-NEXT: s_add_i32 s5, s5, s6
+; GFX11-SDAG-NEXT: s_cmp_lt_i32 s2, 31
+; GFX11-SDAG-NEXT: s_movk_i32 s6, 0x7e00
+; GFX11-SDAG-NEXT: s_cselect_b32 s5, s5, 0x7c00
+; GFX11-SDAG-NEXT: s_cmp_lg_u32 s4, 0
+; GFX11-SDAG-NEXT: s_cselect_b32 s4, s6, 0x7c00
+; GFX11-SDAG-NEXT: s_cmpk_eq_i32 s2, 0x40f
+; GFX11-SDAG-NEXT: s_cselect_b32 s2, s4, s5
+; GFX11-SDAG-NEXT: s_lshr_b32 s3, s3, 16
+; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-SDAG-NEXT: s_and_b32 s3, s3, 0x8000
+; GFX11-SDAG-NEXT: s_or_b32 s2, s3, s2
+; GFX11-SDAG-NEXT: s_mov_b32 s3, 0x31016000
+; GFX11-SDAG-NEXT: v_mov_b32_e32 v0, s2
+; GFX11-SDAG-NEXT: s_mov_b32 s2, -1
+; GFX11-SDAG-NEXT: buffer_store_b16 v0, off, s[0:3], 0
+; GFX11-SDAG-NEXT: s_endpgm
;
; GFX11-GISEL-LABEL: fptrunc_f64_to_f16:
; GFX11-GISEL: ; %bb.0:
@@ -570,30 +548,6 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in)
; GFX11-GISEL-NEXT: s_mov_b32 s2, -1
; GFX11-GISEL-NEXT: buffer_store_b16 v0, off, s[0:3], 0
; GFX11-GISEL-NEXT: s_endpgm
-;
-; GFX11-UNSAFE-DAG-TRUE16-LABEL: fptrunc_f64_to_f16:
-; GFX11-UNSAFE-DAG-TRUE16: ; %bb.0:
-; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-UNSAFE-DAG-TRUE16-NEXT: v_cvt_f32_f64_e32 v0, s[2:3]
-; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
-; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_mov_b32 s2, -1
-; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-UNSAFE-DAG-TRUE16-NEXT: v_cvt_f16_f32_e32 v0.l, v0
-; GFX11-UNSAFE-DAG-TRUE16-NEXT: buffer_store_b16 v0, off, s[0:3], 0
-; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_endpgm
-;
-; GFX11-UNSAFE-DAG-FAKE16-LABEL: fptrunc_f64_to_f16:
-; GFX11-UNSAFE-DAG-FAKE16: ; %bb.0:
-; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-UNSAFE-DAG-FAKE16-NEXT: v_cvt_f32_f64_e32 v0, s[2:3]
-; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_mov_b32 s3, 0x31016000
-; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_mov_b32 s2, -1
-; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-UNSAFE-DAG-FAKE16-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GFX11-UNSAFE-DAG-FAKE16-NEXT: buffer_store_b16 v0, off, s[0:3], 0
-; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_endpgm
%result = fptrunc double %in to half
%result_i16 = bitcast half %result to i16
store i16 %result_i16, ptr addrspace(1) %out
@@ -603,111 +557,27 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in)
define amdgpu_kernel void @fptrunc_f64_to_f16_afn(ptr addrspace(1) %out, double %in) {
; SI-LABEL: fptrunc_f64_to_f16_afn:
; SI: ; %bb.0:
-; SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9
-; SI-NEXT: s_mov_b32 s3, 0xf000
-; SI-NEXT: s_movk_i32 s2, 0x7e00
+; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
+; SI-NEXT: s_mov_b32 s7, 0xf000
+; SI-NEXT: s_mov_b32 s6, -1
; SI-NEXT: s_waitcnt lgkmcnt(0)
-; SI-NEXT: s_lshr_b32 s0, s7, 8
-; SI-NEXT: s_and_b32 s1, s7, 0x1ff
-; SI-NEXT: s_and_b32 s8, s0, 0xffe
-; SI-NEXT: s_or_b32 s0, s1, s6
-; SI-NEXT: s_cmp_lg_u32 s0, 0
-; SI-NEXT: s_cselect_b64 s[0:1], -1, 0
-; SI-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1]
-; SI-NEXT: s_bfe_u32 s0, s7, 0xb0014
-; SI-NEXT: v_readfirstlane_b32 s1, v0
-; SI-NEXT: s_sub_i32 s6, 0x3f1, s0
-; SI-NEXT: s_or_b32 s1, s8, s1
-; SI-NEXT: v_med3_i32 v0, s6, 0, 13
-; SI-NEXT: s_or_b32 s6, s1, 0x1000
-; SI-NEXT: v_readfirstlane_b32 s8, v0
-; SI-NEXT: s_lshr_b32 s9, s6, s8
-; SI-NEXT: s_lshl_b32 s8, s9, s8
-; SI-NEXT: s_cmp_lg_u32 s8, s6
-; SI-NEXT: s_cselect_b32 s6, 1, 0
-; SI-NEXT: s_addk_i32 s0, 0xfc10
-; SI-NEXT: s_or_b32 s6, s9, s6
-; SI-NEXT: s_lshl_b32 s8, s0, 12
-; SI-NEXT: s_or_b32 s8, s1, s8
-; SI-NEXT: s_cmp_lt_i32 s0, 1
-; SI-NEXT: s_cselect_b32 s6, s6, s8
-; SI-NEXT: s_and_b32 s8, s6, 7
-; SI-NEXT: s_cmp_gt_i32 s8, 5
-; SI-NEXT: s_cselect_b32 s9, 1, 0
-; SI-NEXT: s_cmp_eq_u32 s8, 3
-; SI-NEXT: s_cselect_b32 s8, 1, 0
-; SI-NEXT: s_lshr_b32 s6, s6, 2
-; SI-NEXT: s_or_b32 s8, s8, s9
-; SI-NEXT: s_add_i32 s6, s6, s8
-; SI-NEXT: s_cmp_lt_i32 s0, 31
-; SI-NEXT: s_cselect_b32 s6, s6, 0x7c00
-; SI-NEXT: s_cmp_lg_u32 s1, 0
-; SI-NEXT: s_cselect_b32 s1, s2, 0x7c00
-; SI-NEXT: s_cmpk_eq_i32 s0, 0x40f
-; SI-NEXT: s_cselect_b32 s0, s1, s6
-; SI-NEXT: s_lshr_b32 s1, s7, 16
-; SI-NEXT: s_and_b32 s1, s1, 0x8000
-; SI-NEXT: s_or_b32 s6, s1, s0
-; SI-NEXT: s_mov_b32 s2, -1
-; SI-NEXT: s_mov_b32 s0, s4
-; SI-NEXT: s_mov_b32 s1, s5
-; SI-NEXT: v_mov_b32_e32 v0, s6
-; SI-NEXT: buffer_store_short v0, off, s[0:3], 0
+; SI-NEXT: s_mov_b32 s4, s0
+; SI-NEXT: s_mov_b32 s5, s1
+; SI-NEXT: v_cvt_f32_f64_e32 v0, s[2:3]
+; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
+; SI-NEXT: buffer_store_short v0, off, s[4:7], 0
; SI-NEXT: s_endpgm
;
-; VI-SAFE-SDAG-LABEL: fptrunc_f64_to_f16_afn:
-; VI-SAFE-SDAG: ; %bb.0:
-; VI-SAFE-SDAG-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x24
-; VI-SAFE-SDAG-NEXT: s_mov_b32 s3, 0xf000
-; VI-SAFE-SDAG-NEXT: s_mov_b32 s2, -1
-; VI-SAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; VI-SAFE-SDAG-NEXT: s_mov_b32 s0, s4
-; VI-SAFE-SDAG-NEXT: s_lshr_b32 s4, s7, 8
-; VI-SAFE-SDAG-NEXT: s_and_b32 s8, s4, 0xffe
-; VI-SAFE-SDAG-NEXT: s_and_b32 s4, s7, 0x1ff
-; VI-SAFE-SDAG-NEXT: s_or_b32 s4, s4, s6
-; VI-SAFE-SDAG-NEXT: s_cmp_lg_u32 s4, 0
-; VI-SAFE-SDAG-NEXT: s_mov_b32 s1, s5
-; VI-SAFE-SDAG-NEXT: s_cselect_b64 s[4:5], -1, 0
-; VI-SAFE-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[4:5]
-; VI-SAFE-SDAG-NEXT: v_readfirstlane_b32 s4, v0
-; VI-SAFE-SDAG-NEXT: s_bfe_u32 s6, s7, 0xb0014
-; VI-SAFE-SDAG-NEXT: s_or_b32 s4, s8, s4
-; VI-SAFE-SDAG-NEXT: s_sub_i32 s8, 0x3f1, s6
-; VI-SAFE-SDAG-NEXT: v_med3_i32 v0, s8, 0, 13
-; VI-SAFE-SDAG-NEXT: s_or_b32 s5, s4, 0x1000
-; VI-SAFE-SDAG-NEXT: v_readfirstlane_b32 s8, v0
-; VI-SAFE-SDAG-NEXT: s_lshr_b32 s9, s5, s8
-; VI-SAFE-SDAG-NEXT: s_lshl_b32 s8, s9, s8
-; VI-SAFE-SDAG-NEXT: s_cmp_lg_u32 s8, s5
-; VI-SAFE-SDAG-NEXT: s_cselect_b32 s5, 1, 0
-; VI-SAFE-SDAG-NEXT: s_addk_i32 s6, 0xfc10
-; VI-SAFE-SDAG-NEXT: s_lshl_b32 s8, s6, 12
-; VI-SAFE-SDAG-NEXT: s_or_b32 s5, s9, s5
-; VI-SAFE-SDAG-NEXT: s_or_b32 s8, s4, s8
-; VI-SAFE-SDAG-NEXT: s_cmp_lt_i32 s6, 1
-; VI-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, s8
-; VI-SAFE-SDAG-NEXT: s_and_b32 s8, s5, 7
-; VI-SAFE-SDAG-NEXT: s_cmp_gt_i32 s8, 5
-; VI-SAFE-SDAG-NEXT: s_cselect_b32 s9, 1, 0
-; VI-SAFE-SDAG-NEXT: s_cmp_eq_u32 s8, 3
-; VI-SAFE-SDAG-NEXT: s_cselect_b32 s8, 1, 0
-; VI-SAFE-SDAG-NEXT: s_or_b32 s8, s8, s9
-; VI-SAFE-SDAG-NEXT: s_lshr_b32 s5, s5, 2
-; VI-SAFE-SDAG-NEXT: s_add_i32 s5, s5, s8
-; VI-SAFE-SDAG-NEXT: s_cmp_lt_i32 s6, 31
-; VI-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, 0x7c00
-; VI-SAFE-SDAG-NEXT: s_cmp_lg_u32 s4, 0
-; VI-SAFE-SDAG-NEXT: s_movk_i32 s4, 0x7e00
-; VI-SAFE-SDAG-NEXT: s_cselect_b32 s4, s4, 0x7c00
-; VI-SAFE-SDAG-NEXT: s_cmpk_eq_i32 s6, 0x40f
-; VI-SAFE-SDAG-NEXT: s_cselect_b32 s4, s4, s5
-; VI-SAFE-SDAG-NEXT: s_lshr_b32 s5, s7, 16
-; VI-SAFE-SDAG-NEXT: s_and_b32 s5, s5, 0x8000
-; VI-SAFE-SDAG-NEXT: s_or_b32 s4, s5, s4
-; VI-SAFE-SDAG-NEXT: v_mov_b32_e32 v0, s4
-; VI-SAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
-; VI-SAFE-SDAG-NEXT: s_endpgm
+; VI-SDAG-LABEL: fptrunc_f64_to_f16_afn:
+; VI-SDAG: ; %bb.0:
+; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[2:3]
+; VI-SDAG-NEXT: s_mov_b32 s3, 0xf000
+; VI-SDAG-NEXT: s_mov_b32 s2, -1
+; VI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; VI-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
+; VI-SDAG-NEXT: s_endpgm
;
; VI-GISEL-LABEL: fptrunc_f64_to_f16_afn:
; VI-GISEL: ; %bb.0:
@@ -720,68 +590,16 @@ define amdgpu_kernel void @fptrunc_f64_to_f16_afn(ptr addrspace(1) %out, double
; VI-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0
; VI-GISEL-NEXT: s_endpgm
;
-; VI-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16_afn:
-; VI-UNSAFE-SDAG: ; %bb.0:
-; VI-UNSAFE-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
-; VI-UNSAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; VI-UNSAFE-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[2:3]
-; VI-UNSAFE-SDAG-NEXT: s_mov_b32 s3, 0xf000
-; VI-UNSAFE-SDAG-NEXT: s_mov_b32 s2, -1
-; VI-UNSAFE-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
-; VI-UNSAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
-; VI-UNSAFE-SDAG-NEXT: s_endpgm
-;
-; GFX10-SAFE-SDAG-LABEL: fptrunc_f64_to_f16_afn:
-; GFX10-SAFE-SDAG: ; %bb.0:
-; GFX10-SAFE-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
-; GFX10-SAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-SAFE-SDAG-NEXT: s_and_b32 s4, s3, 0x1ff
-; GFX10-SAFE-SDAG-NEXT: s_lshr_b32 s5, s3, 8
-; GFX10-SAFE-SDAG-NEXT: s_or_b32 s2, s4, s2
-; GFX10-SAFE-SDAG-NEXT: s_and_b32 s4, s5, 0xffe
-; GFX10-SAFE-SDAG-NEXT: s_cmp_lg_u32 s2, 0
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s2, -1, 0
-; GFX10-SAFE-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, s2
-; GFX10-SAFE-SDAG-NEXT: s_bfe_u32 s2, s3, 0xb0014
-; GFX10-SAFE-SDAG-NEXT: s_sub_i32 s5, 0x3f1, s2
-; GFX10-SAFE-SDAG-NEXT: v_med3_i32 v1, s5, 0, 13
-; GFX10-SAFE-SDAG-NEXT: v_readfirstlane_b32 s5, v0
-; GFX10-SAFE-SDAG-NEXT: v_readfirstlane_b32 s6, v1
-; GFX10-SAFE-SDAG-NEXT: s_or_b32 s4, s4, s5
-; GFX10-SAFE-SDAG-NEXT: s_or_b32 s5, s4, 0x1000
-; GFX10-SAFE-SDAG-NEXT: s_lshr_b32 s7, s5, s6
-; GFX10-SAFE-SDAG-NEXT: s_lshl_b32 s6, s7, s6
-; GFX10-SAFE-SDAG-NEXT: s_cmp_lg_u32 s6, s5
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s5, 1, 0
-; GFX10-SAFE-SDAG-NEXT: s_addk_i32 s2, 0xfc10
-; GFX10-SAFE-SDAG-NEXT: s_or_b32 s5, s7, s5
-; GFX10-SAFE-SDAG-NEXT: s_lshl_b32 s6, s2, 12
-; GFX10-SAFE-SDAG-NEXT: s_or_b32 s6, s4, s6
-; GFX10-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 1
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, s6
-; GFX10-SAFE-SDAG-NEXT: s_and_b32 s6, s5, 7
-; GFX10-SAFE-SDAG-NEXT: s_cmp_gt_i32 s6, 5
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s7, 1, 0
-; GFX10-SAFE-SDAG-NEXT: s_cmp_eq_u32 s6, 3
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s6, 1, 0
-; GFX10-SAFE-SDAG-NEXT: s_lshr_b32 s5, s5, 2
-; GFX10-SAFE-SDAG-NEXT: s_or_b32 s6, s6, s7
-; GFX10-SAFE-SDAG-NEXT: s_add_i32 s5, s5, s6
-; GFX10-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 31
-; GFX10-SAFE-SDAG-NEXT: s_movk_i32 s6, 0x7e00
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, 0x7c00
-; GFX10-SAFE-SDAG-NEXT: s_cmp_lg_u32 s4, 0
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s4, s6, 0x7c00
-; GFX10-SAFE-SDAG-NEXT: s_cmpk_eq_i32 s2, 0x40f
-; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s2, s4, s5
-; GFX10-SAFE-SDAG-NEXT: s_lshr_b32 s3, s3, 16
-; GFX10-SAFE-SDAG-NEXT: s_and_b32 s3, s3, 0x8000
-; GFX10-SAFE-SDAG-NEXT: s_or_b32 s2, s3, s2
-; GFX10-SAFE-SDAG-NEXT: s_mov_b32 s3, 0x31016000
-; GFX10-SAFE-SDAG-NEXT: v_mov_b32_e32 v0, s2
-; GFX10-SAFE-SDAG-NEXT: s_mov_b32 s2, -1
-; GFX10-SAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
-; GFX10-SAFE-SDAG-NEXT: s_endpgm
+; GFX10-SDAG-LABEL: fptrunc_f64_to_f16_afn:
+; GFX10-SDAG: ; %bb.0:
+; GFX10-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX10-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[2:3]
+; GFX10-SDAG-NEXT: s_mov_b32 s3, 0x31016000
+; GFX10-SDAG-NEXT: s_mov_b32 s2, -1
+; GFX10-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX10-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
+; GFX10-SDAG-NEXT: s_endpgm
;
; GFX10-GISEL-LABEL: fptrunc_f64_to_f16_afn:
; GFX10-GISEL: ; %bb.0:
@@ -794,74 +612,15 @@ define amdgpu_kernel void @fptrunc_f64_to_f16_afn(ptr addrspace(1) %out, double
; GFX10-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0
; GFX10-GISEL-NEXT: s_endpgm
;
-; GFX10-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16_afn:
-; GFX10-UNSAFE-SDAG: ; %bb.0:
-; GFX10-UNSAFE-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
-; GFX10-UNSAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-UNSAFE-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[2:3]
-; GFX10-UNSAFE-SDAG-NEXT: s_mov_b32 s3, 0x31016000
-; GFX10-UNSAFE-SDAG-NEXT: s_mov_b32 s2, -1
-; GFX10-UNSAFE-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GFX10-UNSAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0
-; GFX10-UNSAFE-SDAG-NEXT: s_endpgm
-;
; GFX11-SAFE-SDAG-LABEL: fptrunc_f64_to_f16_afn:
; GFX11-SAFE-SDAG: ; %bb.0:
; GFX11-SAFE-SDAG-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-SAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-SAFE-SDAG-NEXT: s_and_b32 s4, s3, 0x1ff
-; GFX11-SAFE-SDAG-NEXT: s_lshr_b32 s5, s3, 8
-; GFX11-SAFE-SDAG-NEXT: s_or_b32 s2, s4, s2
-; GFX11-SAFE-SDAG-NEXT: s_and_b32 s4, s5, 0xffe
-; GFX11-SAFE-SDAG-NEXT: s_cmp_lg_u32 s2, 0
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s2, -1, 0
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-SAFE-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, s2
-; GFX11-SAFE-SDAG-NEXT: s_bfe_u32 s2, s3, 0xb0014
-; GFX11-SAFE-SDAG-NEXT: s_sub_i32 s5, 0x3f1, s2
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-SAFE-SDAG-NEXT: v_med3_i32 v1, s5, 0, 13
-; GFX11-SAFE-SDAG-NEXT: v_readfirstlane_b32 s5, v0
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-SAFE-SDAG-NEXT: v_readfirstlane_b32 s6, v1
-; GFX11-SAFE-SDAG-NEXT: s_or_b32 s4, s4, s5
-; GFX11-SAFE-SDAG-NEXT: s_or_b32 s5, s4, 0x1000
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SAFE-SDAG-NEXT: s_lshr_b32 s7, s5, s6
-; GFX11-SAFE-SDAG-NEXT: s_lshl_b32 s6, s7, s6
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
-; GFX11-SAFE-SDAG-NEXT: s_cmp_lg_u32 s6, s5
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s5, 1, 0
-; GFX11-SAFE-SDAG-NEXT: s_addk_i32 s2, 0xfc10
-; GFX11-SAFE-SDAG-NEXT: s_or_b32 s5, s7, s5
-; GFX11-SAFE-SDAG-NEXT: s_lshl_b32 s6, s2, 12
-; GFX11-SAFE-SDAG-NEXT: s_or_b32 s6, s4, s6
-; GFX11-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 1
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, s6
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SAFE-SDAG-NEXT: s_and_b32 s6, s5, 7
-; GFX11-SAFE-SDAG-NEXT: s_cmp_gt_i32 s6, 5
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s7, 1, 0
-; GFX11-SAFE-SDAG-NEXT: s_cmp_eq_u32 s6, 3
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s6, 1, 0
-; GFX11-SAFE-SDAG-NEXT: s_lshr_b32 s5, s5, 2
-; GFX11-SAFE-SDAG-NEXT: s_or_b32 s6, s6, s7
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-SAFE-SDAG-NEXT: s_add_i32 s5, s5, s6
-; GFX11-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 31
-; GFX11-SAFE-SDAG-NEXT: s_movk_i32 s6, 0x7e00
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, 0x7c00
-; GFX11-SAFE-SDAG-NEXT: s_cmp_lg_u32 s4, 0
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s4, s6, 0x7c00
-; GFX11-SAFE-SDAG-NEXT: s_cmpk_eq_i32 s2, 0x40f
-; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s2, s4, s5
-; GFX11-SAFE-SDAG-NEXT: s_lshr_b32 s3, s3, 16
-; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-SAFE-SDAG-NEXT: s_and_b32 s3, s3, 0x8000
-; GFX11-SAFE-SDAG-NEXT: s_or_b32 s2, s3, s2
+; GFX11-SAFE-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[2:3]
; GFX11-SAFE-SDAG-NEXT: s_mov_b32 s3, 0x31016000
-; GFX11-SAFE-SDAG-NEXT: v_mov_b32_e32 v0, s2
; GFX11-SAFE-SDAG-NEXT: s_mov_b32 s2, -1
+; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-SAFE-SDAG-NEXT: v_cvt_f16_f32_e32 v0.l, v0
; GFX11-SAFE-SDAG-NEXT: buffer_store_b16 v0, off, s[0:3], 0
; GFX11-SAFE-SDAG-NEXT: s_endpgm
;
@@ -1833,4 +1592,8 @@ define amdgpu_kernel void @fptrunc_v8f64_to_v8f32_afn(ptr addrspace(1) %out, <8
}
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
; GFX10-SAFE-GISEL: {{.*}}
+; GFX10-SAFE-SDAG: {{.*}}
+; GFX10-UNSAFE-SDAG: {{.*}}
; VI-SAFE-GISEL: {{.*}}
+; VI-SAFE-SDAG: {{.*}}
+; VI-UNSAFE-SDAG: {{.*}}
diff --git a/llvm/test/CodeGen/AMDGPU/fsqrt.f32.ll b/llvm/test/CodeGen/AMDGPU/fsqrt.f32.ll
index 87c7cce854b11..f81950bde03e0 100644
--- a/llvm/test/CodeGen/AMDGPU/fsqrt.f32.ll
+++ b/llvm/test/CodeGen/AMDGPU/fsqrt.f32.ll
@@ -1294,13 +1294,13 @@ define float @v_sqrt_f32__enough_unsafe_attrs(float %x) #3 {
ret float %result
}
-define float @v_sqrt_f32__unsafe_attr(float %x) #4 {
+define float @v_sqrt_f32__unsafe_attr(float %x) {
; GCN-LABEL: v_sqrt_f32__unsafe_attr:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_sqrt_f32_e32 v0, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
- %result = call nsz float @llvm.sqrt.f32(float %x)
+ %result = call afn nsz float @llvm.sqrt.f32(float %x)
ret float %result
}
@@ -4763,7 +4763,6 @@ attributes #0 = { nocallback nofree nosync nounwind speculatable willreturn memo
attributes #1 = { convergent nounwind willreturn memory(none) }
attributes #2 = { "approx-func-fp-math"="true" }
attributes #3 = { "approx-func-fp-math"="true" "no-nans-fp-math"="true" "no-infs-fp-math"="true" }
-attributes #4 = { "unsafe-fp-math"="true" }
attributes #5 = { "no-infs-fp-math"="true" }
!0 = !{float 0.5}
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.exp.ll b/llvm/test/CodeGen/AMDGPU/llvm.exp.ll
index 8c1e166babaf8..0b80b10a211bb 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.exp.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.exp.ll
@@ -3227,7 +3227,7 @@ define float @v_exp_f32_fast(float %in) {
ret float %result
}
-define float @v_exp_f32_unsafe_math_attr(float %in) "unsafe-fp-math"="true" {
+define float @v_exp_f32_unsafe_math_attr(float %in) {
; GCN-SDAG-LABEL: v_exp_f32_unsafe_math_attr:
; GCN-SDAG: ; %bb.0:
; GCN-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
@@ -3289,7 +3289,7 @@ define float @v_exp_f32_unsafe_math_attr(float %in) "unsafe-fp-math"="true" {
; CM: ; %bb.0:
; CM-NEXT: CF_END
; CM-NEXT: PAD
- %result = call float @llvm.exp.f32(float %in)
+ %result = call afn float @llvm.exp.f32(float %in)
ret float %result
}
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.exp10.ll b/llvm/test/CodeGen/AMDGPU/llvm.exp10.ll
index edc505bdd6c1d..1d1901482f73c 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.exp10.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.exp10.ll
@@ -3235,7 +3235,7 @@ define float @v_exp10_f32_fast(float %in) {
ret float %result
}
-define float @v_exp10_f32_unsafe_math_attr(float %in) "unsafe-fp-math"="true" {
+define float @v_exp10_f32_unsafe_math_attr(float %in) {
; GCN-SDAG-LABEL: v_exp10_f32_unsafe_math_attr:
; GCN-SDAG: ; %bb.0:
; GCN-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
@@ -3303,7 +3303,7 @@ define float @v_exp10_f32_unsafe_math_attr(float %in) "unsafe-fp-math"="true" {
; CM: ; %bb.0:
; CM-NEXT: CF_END
; CM-NEXT: PAD
- %result = call float @llvm.exp10.f32(float %in)
+ %result = call afn float @llvm.exp10.f32(float %in)
ret float %result
}
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.log.ll b/llvm/test/CodeGen/AMDGPU/llvm.log.ll
index 38d1b4789cf45..e0091fd3783c2 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.log.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.log.ll
@@ -3076,7 +3076,7 @@ define float @v_log_f32_fast(float %in) {
ret float %result
}
-define float @v_log_f32_unsafe_math_attr(float %in) "unsafe-fp-math"="true" {
+define float @v_log_f32_unsafe_math_attr(float %in) {
; SI-SDAG-LABEL: v_log_f32_unsafe_math_attr:
; SI-SDAG: ; %bb.0:
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
@@ -3187,7 +3187,7 @@ define float @v_log_f32_unsafe_math_attr(float %in) "unsafe-fp-math"="true" {
; CM: ; %bb.0:
; CM-NEXT: CF_END
; CM-NEXT: PAD
- %result = call float @llvm.log.f32(float %in)
+ %result = call afn float @llvm.log.f32(float %in)
ret float %result
}
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.log10.ll b/llvm/test/CodeGen/AMDGPU/llvm.log10.ll
index 058933f5481a0..b161aa3341b2f 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.log10.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.log10.ll
@@ -3076,7 +3076,7 @@ define float @v_log10_f32_fast(float %in) {
ret float %result
}
-define float @v_log10_f32_unsafe_math_attr(float %in) "unsafe-fp-math"="true" {
+define float @v_log10_f32_unsafe_math_attr(float %in) {
; SI-SDAG-LABEL: v_log10_f32_unsafe_math_attr:
; SI-SDAG: ; %bb.0:
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
@@ -3187,7 +3187,7 @@ define float @v_log10_f32_unsafe_math_attr(float %in) "unsafe-fp-math"="true" {
; CM: ; %bb.0:
; CM-NEXT: CF_END
; CM-NEXT: PAD
- %result = call float @llvm.log10.f32(float %in)
+ %result = call afn float @llvm.log10.f32(float %in)
ret float %result
}
diff --git a/llvm/test/CodeGen/AMDGPU/rcp-pattern.ll b/llvm/test/CodeGen/AMDGPU/rcp-pattern.ll
index 228420ef0acb0..9f0ffbcf6eff9 100644
--- a/llvm/test/CodeGen/AMDGPU/rcp-pattern.ll
+++ b/llvm/test/CodeGen/AMDGPU/rcp-pattern.ll
@@ -56,7 +56,7 @@ define float @v_rcp_f32_ieee_unsafe(float %x) #4 {
; R600: ; %bb.0:
; R600-NEXT: CF_END
; R600-NEXT: PAD
- %rcp = fdiv float 1.0, %x
+ %rcp = fdiv afn float 1.0, %x
ret float %rcp
}
@@ -1411,10 +1411,10 @@ define amdgpu_kernel void @s_div_arcp_neg_k_x_pat_f32_daz(ptr addrspace(1) %out)
declare float @llvm.fabs.f32(float) #1
declare float @llvm.sqrt.f32(float) #1
-attributes #0 = { nounwind "unsafe-fp-math"="false" "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
+attributes #0 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
attributes #1 = { nounwind readnone }
-attributes #2 = { nounwind "unsafe-fp-math"="true" "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
+attributes #2 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
attributes #3 = { nounwind "denormal-fp-math-f32"="ieee,ieee" }
-attributes #4 = { nounwind "unsafe-fp-math"="true" "denormal-fp-math-f32"="ieee,ieee" }
+attributes #4 = { nounwind "denormal-fp-math-f32"="ieee,ieee" }
!0 = !{float 2.500000e+00}
diff --git a/llvm/test/CodeGen/AMDGPU/rsq.f64.ll b/llvm/test/CodeGen/AMDGPU/rsq.f64.ll
index b78cbb0ac29cf..48bc56041729a 100644
--- a/llvm/test/CodeGen/AMDGPU/rsq.f64.ll
+++ b/llvm/test/CodeGen/AMDGPU/rsq.f64.ll
@@ -4504,7 +4504,7 @@ define <2 x double> @v_rsq_v2f64__afn_nnan_ninf(<2 x double> %x) {
ret <2 x double> %rsq
}
-define amdgpu_ps <2 x i32> @s_rsq_f64_unsafe(double inreg %x) #0 {
+define amdgpu_ps <2 x i32> @s_rsq_f64_unsafe(double inreg %x) {
; SI-SDAG-LABEL: s_rsq_f64_unsafe:
; SI-SDAG: ; %bb.0:
; SI-SDAG-NEXT: v_mov_b32_e32 v0, 0
@@ -4648,8 +4648,8 @@ define amdgpu_ps <2 x i32> @s_rsq_f64_unsafe(double inreg %x) #0 {
; VI-GISEL-NEXT: v_readfirstlane_b32 s0, v0
; VI-GISEL-NEXT: v_readfirstlane_b32 s1, v1
; VI-GISEL-NEXT: ; return to shader part epilog
- %rsq = call contract double @llvm.sqrt.f64(double %x)
- %result = fdiv contract double 1.0, %rsq
+ %rsq = call contract afn double @llvm.sqrt.f64(double %x)
+ %result = fdiv contract afn double 1.0, %rsq
%cast = bitcast double %result to <2 x i32>
%cast.0 = extractelement <2 x i32> %cast, i32 0
%cast.1 = extractelement <2 x i32> %cast, i32 1
@@ -4660,7 +4660,7 @@ define amdgpu_ps <2 x i32> @s_rsq_f64_unsafe(double inreg %x) #0 {
ret <2 x i32> %insert.1
}
-define double @v_rsq_f64_unsafe(double %x) #0 {
+define double @v_rsq_f64_unsafe(double %x) {
; SI-SDAG-LABEL: v_rsq_f64_unsafe:
; SI-SDAG: ; %bb.0:
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
@@ -4800,8 +4800,8 @@ define double @v_rsq_f64_unsafe(double %x) #0 {
; VI-GISEL-NEXT: v_fma_f64 v[0:1], -v[0:1], v[2:3], 1.0
; VI-GISEL-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[2:3]
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
- %sqrt = call double @llvm.sqrt.f64(double %x)
- %rsq = fdiv double 1.0, %sqrt
+ %sqrt = call afn double @llvm.sqrt.f64(double %x)
+ %rsq = fdiv afn double 1.0, %sqrt
ret double %rsq
}
@@ -5737,7 +5737,6 @@ define double @v_div_const_contract_sqrt_f64(double %x) {
ret double %rsq
}
-attributes #0 = { "unsafe-fp-math"="true" }
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
; GCN: {{.*}}
; GISEL: {{.*}}
>From 88eea82863dcddd88cf5b1bf14e797925a4c87c8 Mon Sep 17 00:00:00 2001
From: PaperChalice <liujunchang97 at outlook.com>
Date: Tue, 29 Jul 2025 17:22:01 +0800
Subject: [PATCH 4/4] Remove `UnsafeFPMath` in `AMDGPULegalizerInfo.cpp`
---
llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp | 10 ++++------
1 file changed, 4 insertions(+), 6 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
index fedfa3f9dd900..395dd46c36f62 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
@@ -3326,7 +3326,7 @@ static bool allowApproxFunc(const MachineFunction &MF, unsigned Flags) {
if (Flags & MachineInstr::FmAfn)
return true;
const auto &Options = MF.getTarget().Options;
- return Options.UnsafeFPMath || Options.ApproxFuncFPMath;
+ return Options.ApproxFuncFPMath;
}
static bool needsDenormHandlingF32(const MachineFunction &MF, Register Src,
@@ -3432,7 +3432,7 @@ bool AMDGPULegalizerInfo::legalizeFlogCommon(MachineInstr &MI,
static_cast<const AMDGPUTargetMachine &>(MF.getTarget());
if (Ty == F16 || MI.getFlag(MachineInstr::FmAfn) ||
- TM.Options.ApproxFuncFPMath || TM.Options.UnsafeFPMath) {
+ TM.Options.ApproxFuncFPMath) {
if (Ty == F16 && !ST.has16BitInsts()) {
Register LogVal = MRI.createGenericVirtualRegister(F32);
auto PromoteSrc = B.buildFPExt(F32, X);
@@ -4861,8 +4861,7 @@ bool AMDGPULegalizerInfo::legalizeFastUnsafeFDIV(MachineInstr &MI,
LLT ResTy = MRI.getType(Res);
const MachineFunction &MF = B.getMF();
- bool AllowInaccurateRcp = MI.getFlag(MachineInstr::FmAfn) ||
- MF.getTarget().Options.UnsafeFPMath;
+ bool AllowInaccurateRcp = MI.getFlag(MachineInstr::FmAfn);
if (const auto *CLHS = getConstantFPVRegVal(LHS, MRI)) {
if (!AllowInaccurateRcp && ResTy != LLT::scalar(16))
@@ -4923,8 +4922,7 @@ bool AMDGPULegalizerInfo::legalizeFastUnsafeFDIV64(MachineInstr &MI,
LLT ResTy = MRI.getType(Res);
const MachineFunction &MF = B.getMF();
- bool AllowInaccurateRcp = MF.getTarget().Options.UnsafeFPMath ||
- MI.getFlag(MachineInstr::FmAfn);
+ bool AllowInaccurateRcp = MI.getFlag(MachineInstr::FmAfn);
if (!AllowInaccurateRcp)
return false;
More information about the llvm-commits
mailing list