[llvm-branch-commits] [llvm] release/19.x: [RISCV] Add hasPostISelHook to sf.vfnrclip pseudo instructions. (#114274) (PR #117948)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Wed Nov 27 17:22:14 PST 2024
https://github.com/llvmbot created https://github.com/llvm/llvm-project/pull/117948
Backport 71b6f6b8a1cd9a63b9d382fe15f40bbb427939b9 408c84f35b8b0338b630a6ee313c14238e62b5e6
Requested by: @topperc
>From bd583dd074029a7113fcb9d41a64da51399c84fa Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper at sifive.com>
Date: Wed, 30 Oct 2024 11:47:40 -0700
Subject: [PATCH 1/2] [RISCV] Add missing hasPostISelHook = 1 to vector pseudos
that might read FRM. (#114186)
We need an implicit FRM read operand anytime the rounding mode is
dynamic. The post isel hook is responsible for this when isel creates an
instruction with dynamic rounding mode.
Add a MachineVerifier check to verify the operand is present.
(cherry picked from commit 71b6f6b8a1cd9a63b9d382fe15f40bbb427939b9)
---
llvm/lib/Target/RISCV/RISCVInstrInfo.cpp | 7 +++++++
llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td | 17 ++++++++---------
2 files changed, 15 insertions(+), 9 deletions(-)
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp b/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
index 6c0cbeadebf431..7f4bbe7861087e 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
@@ -2536,6 +2536,13 @@ bool RISCVInstrInfo::verifyInstruction(const MachineInstr &MI,
}
}
+ if (int Idx = RISCVII::getFRMOpNum(Desc);
+ Idx >= 0 && MI.getOperand(Idx).getImm() == RISCVFPRndMode::DYN &&
+ !MI.readsRegister(RISCV::FRM, /*TRI=*/nullptr)) {
+ ErrInfo = "dynamic rounding mode should read FRM";
+ return false;
+ }
+
return true;
}
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td b/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
index b860273d639ee5..93fd0b2aada35a 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
@@ -6471,7 +6471,7 @@ defm PseudoVFRDIV : VPseudoVFRDIV_VF_RM;
//===----------------------------------------------------------------------===//
// 13.5. Vector Widening Floating-Point Multiply
//===----------------------------------------------------------------------===//
-let mayRaiseFPException = true, hasSideEffects = 0 in {
+let mayRaiseFPException = true, hasSideEffects = 0, hasPostISelHook = 1 in {
defm PseudoVFWMUL : VPseudoVWMUL_VV_VF_RM;
}
@@ -6504,7 +6504,7 @@ defm PseudoVFWMACCBF16 : VPseudoVWMAC_VV_VF_BF_RM;
//===----------------------------------------------------------------------===//
// 13.8. Vector Floating-Point Square-Root Instruction
//===----------------------------------------------------------------------===//
-let mayRaiseFPException = true, hasSideEffects = 0 in
+let mayRaiseFPException = true, hasSideEffects = 0, hasPostISelHook = 1 in
defm PseudoVFSQRT : VPseudoVSQR_V_RM;
//===----------------------------------------------------------------------===//
@@ -6516,7 +6516,7 @@ defm PseudoVFRSQRT7 : VPseudoVRCP_V;
//===----------------------------------------------------------------------===//
// 13.10. Vector Floating-Point Reciprocal Estimate Instruction
//===----------------------------------------------------------------------===//
-let mayRaiseFPException = true, hasSideEffects = 0 in
+let mayRaiseFPException = true, hasSideEffects = 0, hasPostISelHook = 1 in
defm PseudoVFREC7 : VPseudoVRCP_V_RM;
//===----------------------------------------------------------------------===//
@@ -6627,9 +6627,10 @@ defm PseudoVFNCVT_F_X : VPseudoVNCVTF_W_RM;
defm PseudoVFNCVT_RM_F_XU : VPseudoVNCVTF_RM_W;
defm PseudoVFNCVT_RM_F_X : VPseudoVNCVTF_RM_W;
-let hasSideEffects = 0, hasPostISelHook = 1 in
+let hasSideEffects = 0, hasPostISelHook = 1 in {
defm PseudoVFNCVT_F_F : VPseudoVNCVTD_W_RM;
defm PseudoVFNCVTBF16_F_F : VPseudoVNCVTD_W_RM;
+}
defm PseudoVFNCVT_ROD_F_F : VPseudoVNCVTD_W;
} // mayRaiseFPException = true
@@ -6665,8 +6666,7 @@ let Predicates = [HasVInstructionsAnyF] in {
//===----------------------------------------------------------------------===//
// 14.3. Vector Single-Width Floating-Point Reduction Instructions
//===----------------------------------------------------------------------===//
-let mayRaiseFPException = true,
- hasSideEffects = 0 in {
+let mayRaiseFPException = true, hasSideEffects = 0, hasPostISelHook = 1 in {
defm PseudoVFREDOSUM : VPseudoVFREDO_VS_RM;
defm PseudoVFREDUSUM : VPseudoVFRED_VS_RM;
}
@@ -6678,9 +6678,8 @@ defm PseudoVFREDMAX : VPseudoVFREDMINMAX_VS;
//===----------------------------------------------------------------------===//
// 14.4. Vector Widening Floating-Point Reduction Instructions
//===----------------------------------------------------------------------===//
-let IsRVVWideningReduction = 1,
- hasSideEffects = 0,
- mayRaiseFPException = true in {
+let IsRVVWideningReduction = 1, hasSideEffects = 0, mayRaiseFPException = true,
+ hasPostISelHook = 1 in {
defm PseudoVFWREDUSUM : VPseudoVFWRED_VS_RM;
defm PseudoVFWREDOSUM : VPseudoVFWREDO_VS_RM;
}
>From 6140b33fd309d68097bd8d566ce56262d0cbdf49 Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper at sifive.com>
Date: Wed, 30 Oct 2024 11:52:49 -0700
Subject: [PATCH 2/2] [RISCV] Add hasPostISelHook to sf.vfnrclip pseudo
instructions. (#114274)
Add Uses = [FRM] to the underlying MC instructions.
Tweak a couple test cases so the MachineVerifier would have caught this.
(cherry picked from commit 408c84f35b8b0338b630a6ee313c14238e62b5e6)
---
llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td | 5 +++--
llvm/test/CodeGen/RISCV/rvv/sf_vfnrclip_x_f_qf.ll | 4 +---
llvm/test/CodeGen/RISCV/rvv/sf_vfnrclip_xu_f_qf.ll | 4 +---
3 files changed, 5 insertions(+), 8 deletions(-)
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td b/llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td
index 71aa1f19e089a9..eacc75b9a6c445 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td
@@ -217,7 +217,8 @@ let Predicates = [HasVendorXSfvfwmaccqqq], DecoderNamespace = "XSfvfwmaccqqq" in
def VFWMACC_4x4x4 : CustomSiFiveVMACC<0b111100, OPFVV, "sf.vfwmacc.4x4x4">;
}
-let Predicates = [HasVendorXSfvfnrclipxfqf], DecoderNamespace = "XSfvfnrclipxfqf" in {
+let Predicates = [HasVendorXSfvfnrclipxfqf], DecoderNamespace = "XSfvfnrclipxfqf",
+ Uses = [FRM] in {
def VFNRCLIP_XU_F_QF : CustomSiFiveVFNRCLIP<0b100010, OPFVF, "sf.vfnrclip.xu.f.qf">;
def VFNRCLIP_X_F_QF : CustomSiFiveVFNRCLIP<0b100011, OPFVF, "sf.vfnrclip.x.f.qf">;
}
@@ -399,7 +400,7 @@ multiclass VPseudoSiFiveVFWMACC<string Constraint = ""> {
multiclass VPseudoSiFiveVFNRCLIP<string Constraint = "@earlyclobber $rd"> {
foreach i = 0-4 in
- let hasSideEffects = 0 in
+ let hasSideEffects = 0, hasPostISelHook = 1 in
defm "Pseudo" # NAME : VPseudoBinaryRoundingMode<MxListW[i].vrclass,
MxListVF4[i].vrclass,
FPR32, MxListW[i],
diff --git a/llvm/test/CodeGen/RISCV/rvv/sf_vfnrclip_x_f_qf.ll b/llvm/test/CodeGen/RISCV/rvv/sf_vfnrclip_x_f_qf.ll
index 3c19616576f55d..fbe1a97c201cf3 100644
--- a/llvm/test/CodeGen/RISCV/rvv/sf_vfnrclip_x_f_qf.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/sf_vfnrclip_x_f_qf.ll
@@ -13,10 +13,8 @@ declare <vscale x 1 x i8> @llvm.riscv.sf.vfnrclip.x.f.qf.nxv1i8.nxv1f32.iXLen(
define <vscale x 1 x i8> @intrinsic_sf_vfnrclip_x_f_qf_nxv1i8_nxv1f32(<vscale x 1 x float> %0, float %1, iXLen %2) nounwind {
; CHECK-LABEL: intrinsic_sf_vfnrclip_x_f_qf_nxv1i8_nxv1f32:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: fsrmi a1, 0
; CHECK-NEXT: vsetvli zero, a0, e8, mf8, ta, ma
; CHECK-NEXT: sf.vfnrclip.x.f.qf v9, v8, fa0
-; CHECK-NEXT: fsrm a1
; CHECK-NEXT: vmv1r.v v8, v9
; CHECK-NEXT: ret
entry:
@@ -24,7 +22,7 @@ entry:
<vscale x 1 x i8> undef,
<vscale x 1 x float> %0,
float %1,
- iXLen 0, iXLen %2)
+ iXLen 7, iXLen %2)
ret <vscale x 1 x i8> %a
}
diff --git a/llvm/test/CodeGen/RISCV/rvv/sf_vfnrclip_xu_f_qf.ll b/llvm/test/CodeGen/RISCV/rvv/sf_vfnrclip_xu_f_qf.ll
index dbcee311c6e35f..dfb0ccd982e845 100644
--- a/llvm/test/CodeGen/RISCV/rvv/sf_vfnrclip_xu_f_qf.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/sf_vfnrclip_xu_f_qf.ll
@@ -13,10 +13,8 @@ declare <vscale x 1 x i8> @llvm.riscv.sf.vfnrclip.xu.f.qf.nxv1i8.nxv1f32.iXLen(
define <vscale x 1 x i8> @intrinsic_sf_vfnrclip_xu_f_qf_nxv1i8_nxv1f32(<vscale x 1 x float> %0, float %1, iXLen %2) nounwind {
; CHECK-LABEL: intrinsic_sf_vfnrclip_xu_f_qf_nxv1i8_nxv1f32:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: fsrmi a1, 0
; CHECK-NEXT: vsetvli zero, a0, e8, mf8, ta, ma
; CHECK-NEXT: sf.vfnrclip.xu.f.qf v9, v8, fa0
-; CHECK-NEXT: fsrm a1
; CHECK-NEXT: vmv1r.v v8, v9
; CHECK-NEXT: ret
entry:
@@ -24,7 +22,7 @@ entry:
<vscale x 1 x i8> undef,
<vscale x 1 x float> %0,
float %1,
- iXLen 0, iXLen %2)
+ iXLen 7, iXLen %2)
ret <vscale x 1 x i8> %a
}
More information about the llvm-branch-commits
mailing list