[llvm] [AMDGPU] Update log lowering to remove contract for AMDGCN backend (PR #168916)
Adel Ejjeh via llvm-commits
llvm-commits at lists.llvm.org
Thu Nov 20 10:58:10 PST 2025
https://github.com/adelejjeh updated https://github.com/llvm/llvm-project/pull/168916
>From d599c7a68f20259327573fee7b11192c8a435ef7 Mon Sep 17 00:00:00 2001
From: Adel Ejjeh <adel.ejjeh at amd.com>
Date: Thu, 20 Nov 2025 11:21:12 -0600
Subject: [PATCH 1/2] [AMDGPU] Update log lowering to remove contract for
AMDGCN backend
---
llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp | 4 +-
.../lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp | 3 +
llvm/test/CodeGen/AMDGPU/llvm.log.ll | 331 +++++++++++++++++-
llvm/test/CodeGen/AMDGPU/llvm.log10.ll | 331 +++++++++++++++++-
4 files changed, 660 insertions(+), 9 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index db890df7c50f9..ffafe06211109 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -2766,7 +2766,9 @@ SDValue AMDGPUTargetLowering::LowerFLOGCommon(SDValue Op,
EVT VT = Op.getValueType();
SDNodeFlags Flags = Op->getFlags();
SDLoc DL(Op);
-
+ // Our implementation of LOG is not contract safe, so disable instruction
+ // contraction.
+ Flags.setAllowContract(false);
const bool IsLog10 = Op.getOpcode() == ISD::FLOG10;
assert(IsLog10 || Op.getOpcode() == ISD::FLOG);
diff --git a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
index 1a13b2226ecd6..5c8b720c54761 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
@@ -3508,6 +3508,9 @@ bool AMDGPULegalizerInfo::legalizeFlogCommon(MachineInstr &MI,
MachineRegisterInfo &MRI = *B.getMRI();
Register Dst = MI.getOperand(0).getReg();
Register X = MI.getOperand(1).getReg();
+ // Our implementation of LOG is not contract safe, so disable contraction in
+ // the flags before reading the field.
+ MI.clearFlags(MachineInstr::FmContract);
unsigned Flags = MI.getFlags();
const LLT Ty = MRI.getType(X);
MachineFunction &MF = B.getMF();
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.log.ll b/llvm/test/CodeGen/AMDGPU/llvm.log.ll
index fc6b2d95b2af8..88f202adba07e 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.log.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.log.ll
@@ -316,6 +316,309 @@ define amdgpu_kernel void @s_log_f32(ptr addrspace(1) %out, float %in) {
ret void
}
+define amdgpu_kernel void @s_log_contract_f32(ptr addrspace(1) %out, float %in) {
+; SI-SDAG-LABEL: s_log_contract_f32:
+; SI-SDAG: ; %bb.0:
+; SI-SDAG-NEXT: s_load_dword s6, s[4:5], 0xb
+; SI-SDAG-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
+; SI-SDAG-NEXT: v_mov_b32_e32 v0, 0x800000
+; SI-SDAG-NEXT: v_mov_b32_e32 v1, 0x41b17218
+; SI-SDAG-NEXT: s_mov_b32 s4, 0x3f317217
+; SI-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; SI-SDAG-NEXT: v_cmp_lt_f32_e32 vcc, s6, v0
+; SI-SDAG-NEXT: s_and_b64 s[2:3], vcc, exec
+; SI-SDAG-NEXT: s_cselect_b32 s2, 32, 0
+; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v1, vcc
+; SI-SDAG-NEXT: v_mov_b32_e32 v1, s2
+; SI-SDAG-NEXT: v_ldexp_f32_e32 v1, s6, v1
+; SI-SDAG-NEXT: v_log_f32_e32 v1, v1
+; SI-SDAG-NEXT: s_mov_b32 s3, 0xf000
+; SI-SDAG-NEXT: s_mov_b32 s2, -1
+; SI-SDAG-NEXT: v_mul_f32_e32 v2, 0x3f317217, v1
+; SI-SDAG-NEXT: v_fma_f32 v3, v1, s4, -v2
+; SI-SDAG-NEXT: s_mov_b32 s4, 0x3377d1cf
+; SI-SDAG-NEXT: v_fma_f32 v3, v1, s4, v3
+; SI-SDAG-NEXT: s_mov_b32 s4, 0x7f800000
+; SI-SDAG-NEXT: v_add_f32_e32 v2, v2, v3
+; SI-SDAG-NEXT: v_cmp_lt_f32_e64 vcc, |v1|, s4
+; SI-SDAG-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; SI-SDAG-NEXT: v_sub_f32_e32 v0, v1, v0
+; SI-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], 0
+; SI-SDAG-NEXT: s_endpgm
+;
+; SI-GISEL-LABEL: s_log_contract_f32:
+; SI-GISEL: ; %bb.0:
+; SI-GISEL-NEXT: s_load_dword s0, s[4:5], 0xb
+; SI-GISEL-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x9
+; SI-GISEL-NEXT: v_mov_b32_e32 v0, 0x800000
+; SI-GISEL-NEXT: v_mov_b32_e32 v1, 0x3f317217
+; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x3377d1cf
+; SI-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; SI-GISEL-NEXT: v_cmp_lt_f32_e32 vcc, s0, v0
+; SI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
+; SI-GISEL-NEXT: v_lshlrev_b32_e32 v0, 5, v0
+; SI-GISEL-NEXT: v_ldexp_f32_e32 v0, s0, v0
+; SI-GISEL-NEXT: v_log_f32_e32 v0, v0
+; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x7f800000
+; SI-GISEL-NEXT: s_mov_b32 s6, -1
+; SI-GISEL-NEXT: s_mov_b32 s7, 0xf000
+; SI-GISEL-NEXT: v_mul_f32_e32 v4, 0x3f317217, v0
+; SI-GISEL-NEXT: v_fma_f32 v1, v0, v1, -v4
+; SI-GISEL-NEXT: v_fma_f32 v1, v0, v2, v1
+; SI-GISEL-NEXT: v_add_f32_e32 v1, v4, v1
+; SI-GISEL-NEXT: v_cmp_lt_f32_e64 s[0:1], |v0|, v3
+; SI-GISEL-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[0:1]
+; SI-GISEL-NEXT: v_mov_b32_e32 v1, 0x41b17218
+; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
+; SI-GISEL-NEXT: v_sub_f32_e32 v0, v0, v1
+; SI-GISEL-NEXT: buffer_store_dword v0, off, s[4:7], 0
+; SI-GISEL-NEXT: s_endpgm
+;
+; VI-SDAG-LABEL: s_log_contract_f32:
+; VI-SDAG: ; %bb.0:
+; VI-SDAG-NEXT: s_load_dword s2, s[4:5], 0x2c
+; VI-SDAG-NEXT: v_mov_b32_e32 v0, 0x800000
+; VI-SDAG-NEXT: v_mov_b32_e32 v1, 0x41b17218
+; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; VI-SDAG-NEXT: v_cmp_lt_f32_e32 vcc, s2, v0
+; VI-SDAG-NEXT: s_and_b64 s[0:1], vcc, exec
+; VI-SDAG-NEXT: s_cselect_b32 s0, 32, 0
+; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v1, vcc
+; VI-SDAG-NEXT: v_mov_b32_e32 v1, s0
+; VI-SDAG-NEXT: v_ldexp_f32 v1, s2, v1
+; VI-SDAG-NEXT: v_log_f32_e32 v1, v1
+; VI-SDAG-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; VI-SDAG-NEXT: s_mov_b32 s2, 0x7f800000
+; VI-SDAG-NEXT: v_and_b32_e32 v2, 0xfffff000, v1
+; VI-SDAG-NEXT: v_sub_f32_e32 v3, v1, v2
+; VI-SDAG-NEXT: v_mul_f32_e32 v4, 0x3805fdf4, v2
+; VI-SDAG-NEXT: v_mul_f32_e32 v5, 0x3f317000, v3
+; VI-SDAG-NEXT: v_mul_f32_e32 v3, 0x3805fdf4, v3
+; VI-SDAG-NEXT: v_add_f32_e32 v3, v4, v3
+; VI-SDAG-NEXT: v_mul_f32_e32 v2, 0x3f317000, v2
+; VI-SDAG-NEXT: v_add_f32_e32 v3, v5, v3
+; VI-SDAG-NEXT: v_add_f32_e32 v2, v2, v3
+; VI-SDAG-NEXT: v_cmp_lt_f32_e64 vcc, |v1|, s2
+; VI-SDAG-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; VI-SDAG-NEXT: v_sub_f32_e32 v2, v1, v0
+; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; VI-SDAG-NEXT: v_mov_b32_e32 v0, s0
+; VI-SDAG-NEXT: v_mov_b32_e32 v1, s1
+; VI-SDAG-NEXT: flat_store_dword v[0:1], v2
+; VI-SDAG-NEXT: s_endpgm
+;
+; VI-GISEL-LABEL: s_log_contract_f32:
+; VI-GISEL: ; %bb.0:
+; VI-GISEL-NEXT: s_load_dword s0, s[4:5], 0x2c
+; VI-GISEL-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x800000
+; VI-GISEL-NEXT: v_mov_b32_e32 v1, 0x7f800000
+; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; VI-GISEL-NEXT: v_cmp_lt_f32_e32 vcc, s0, v0
+; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
+; VI-GISEL-NEXT: v_lshlrev_b32_e32 v0, 5, v0
+; VI-GISEL-NEXT: v_ldexp_f32 v0, s0, v0
+; VI-GISEL-NEXT: v_log_f32_e32 v0, v0
+; VI-GISEL-NEXT: v_and_b32_e32 v2, 0xfffff000, v0
+; VI-GISEL-NEXT: v_sub_f32_e32 v3, v0, v2
+; VI-GISEL-NEXT: v_mul_f32_e32 v4, 0x3805fdf4, v2
+; VI-GISEL-NEXT: v_mul_f32_e32 v5, 0x3805fdf4, v3
+; VI-GISEL-NEXT: v_mul_f32_e32 v3, 0x3f317000, v3
+; VI-GISEL-NEXT: v_add_f32_e32 v4, v4, v5
+; VI-GISEL-NEXT: v_mul_f32_e32 v2, 0x3f317000, v2
+; VI-GISEL-NEXT: v_add_f32_e32 v3, v3, v4
+; VI-GISEL-NEXT: v_add_f32_e32 v2, v2, v3
+; VI-GISEL-NEXT: v_cmp_lt_f32_e64 s[0:1], |v0|, v1
+; VI-GISEL-NEXT: v_mov_b32_e32 v1, 0x41b17218
+; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[0:1]
+; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
+; VI-GISEL-NEXT: v_sub_f32_e32 v2, v0, v1
+; VI-GISEL-NEXT: v_mov_b32_e32 v0, s2
+; VI-GISEL-NEXT: v_mov_b32_e32 v1, s3
+; VI-GISEL-NEXT: flat_store_dword v[0:1], v2
+; VI-GISEL-NEXT: s_endpgm
+;
+; GFX900-SDAG-LABEL: s_log_contract_f32:
+; GFX900-SDAG: ; %bb.0:
+; GFX900-SDAG-NEXT: s_load_dword s6, s[4:5], 0x2c
+; GFX900-SDAG-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX900-SDAG-NEXT: v_mov_b32_e32 v0, 0x800000
+; GFX900-SDAG-NEXT: v_mov_b32_e32 v1, 0x41b17218
+; GFX900-SDAG-NEXT: v_mov_b32_e32 v2, 0
+; GFX900-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; GFX900-SDAG-NEXT: v_cmp_lt_f32_e32 vcc, s6, v0
+; GFX900-SDAG-NEXT: s_and_b64 s[2:3], vcc, exec
+; GFX900-SDAG-NEXT: s_cselect_b32 s2, 32, 0
+; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v1, vcc
+; GFX900-SDAG-NEXT: v_mov_b32_e32 v1, s2
+; GFX900-SDAG-NEXT: v_ldexp_f32 v1, s6, v1
+; GFX900-SDAG-NEXT: v_log_f32_e32 v1, v1
+; GFX900-SDAG-NEXT: s_mov_b32 s2, 0x3f317217
+; GFX900-SDAG-NEXT: s_mov_b32 s3, 0x3377d1cf
+; GFX900-SDAG-NEXT: v_mul_f32_e32 v3, 0x3f317217, v1
+; GFX900-SDAG-NEXT: v_fma_f32 v4, v1, s2, -v3
+; GFX900-SDAG-NEXT: v_fma_f32 v4, v1, s3, v4
+; GFX900-SDAG-NEXT: s_mov_b32 s2, 0x7f800000
+; GFX900-SDAG-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX900-SDAG-NEXT: v_cmp_lt_f32_e64 vcc, |v1|, s2
+; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX900-SDAG-NEXT: v_sub_f32_e32 v0, v1, v0
+; GFX900-SDAG-NEXT: global_store_dword v2, v0, s[0:1]
+; GFX900-SDAG-NEXT: s_endpgm
+;
+; GFX900-GISEL-LABEL: s_log_contract_f32:
+; GFX900-GISEL: ; %bb.0:
+; GFX900-GISEL-NEXT: s_load_dword s0, s[4:5], 0x2c
+; GFX900-GISEL-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x800000
+; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x3f317217
+; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3377d1cf
+; GFX900-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX900-GISEL-NEXT: v_cmp_lt_f32_e32 vcc, s0, v0
+; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
+; GFX900-GISEL-NEXT: v_lshlrev_b32_e32 v0, 5, v0
+; GFX900-GISEL-NEXT: v_ldexp_f32 v0, s0, v0
+; GFX900-GISEL-NEXT: v_log_f32_e32 v0, v0
+; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0x7f800000
+; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, 0
+; GFX900-GISEL-NEXT: v_mul_f32_e32 v5, 0x3f317217, v0
+; GFX900-GISEL-NEXT: v_fma_f32 v2, v0, v2, -v5
+; GFX900-GISEL-NEXT: v_fma_f32 v2, v0, v3, v2
+; GFX900-GISEL-NEXT: v_add_f32_e32 v2, v5, v2
+; GFX900-GISEL-NEXT: v_cmp_lt_f32_e64 s[0:1], |v0|, v4
+; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[0:1]
+; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x41b17218
+; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
+; GFX900-GISEL-NEXT: v_sub_f32_e32 v0, v0, v2
+; GFX900-GISEL-NEXT: global_store_dword v1, v0, s[2:3]
+; GFX900-GISEL-NEXT: s_endpgm
+;
+; GFX1100-SDAG-LABEL: s_log_contract_f32:
+; GFX1100-SDAG: ; %bb.0:
+; GFX1100-SDAG-NEXT: s_load_b32 s0, s[4:5], 0x2c
+; GFX1100-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1100-SDAG-NEXT: v_cmp_gt_f32_e64 s1, 0x800000, s0
+; GFX1100-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX1100-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 0x41b17218, s1
+; GFX1100-SDAG-NEXT: s_and_b32 s1, s1, exec_lo
+; GFX1100-SDAG-NEXT: s_cselect_b32 s1, 32, 0
+; GFX1100-SDAG-NEXT: v_ldexp_f32 v1, s0, s1
+; GFX1100-SDAG-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1100-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
+; GFX1100-SDAG-NEXT: v_log_f32_e32 v1, v1
+; GFX1100-SDAG-NEXT: s_waitcnt_depctr 0xfff
+; GFX1100-SDAG-NEXT: v_mul_f32_e32 v2, 0x3f317217, v1
+; GFX1100-SDAG-NEXT: v_cmp_gt_f32_e64 vcc_lo, 0x7f800000, |v1|
+; GFX1100-SDAG-NEXT: v_fma_f32 v3, 0x3f317217, v1, -v2
+; GFX1100-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1100-SDAG-NEXT: v_fmamk_f32 v3, v1, 0x3377d1cf, v3
+; GFX1100-SDAG-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX1100-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1100-SDAG-NEXT: v_dual_cndmask_b32 v1, v1, v2 :: v_dual_mov_b32 v2, 0
+; GFX1100-SDAG-NEXT: v_sub_f32_e32 v0, v1, v0
+; GFX1100-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1100-SDAG-NEXT: global_store_b32 v2, v0, s[0:1]
+; GFX1100-SDAG-NEXT: s_endpgm
+;
+; GFX1100-GISEL-LABEL: s_log_contract_f32:
+; GFX1100-GISEL: ; %bb.0:
+; GFX1100-GISEL-NEXT: s_load_b32 s0, s[4:5], 0x2c
+; GFX1100-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1100-GISEL-NEXT: v_cmp_gt_f32_e64 s2, 0x800000, s0
+; GFX1100-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1100-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, s2
+; GFX1100-GISEL-NEXT: v_lshlrev_b32_e32 v0, 5, v0
+; GFX1100-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX1100-GISEL-NEXT: v_ldexp_f32 v0, s0, v0
+; GFX1100-GISEL-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1100-GISEL-NEXT: v_log_f32_e32 v0, v0
+; GFX1100-GISEL-NEXT: s_waitcnt_depctr 0xfff
+; GFX1100-GISEL-NEXT: v_mul_f32_e32 v1, 0x3f317217, v0
+; GFX1100-GISEL-NEXT: v_cmp_gt_f32_e64 vcc_lo, 0x7f800000, |v0|
+; GFX1100-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1100-GISEL-NEXT: v_fma_f32 v2, 0x3f317217, v0, -v1
+; GFX1100-GISEL-NEXT: v_fmac_f32_e32 v2, 0x3377d1cf, v0
+; GFX1100-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1100-GISEL-NEXT: v_dual_add_f32 v1, v1, v2 :: v_dual_mov_b32 v2, 0
+; GFX1100-GISEL-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc_lo
+; GFX1100-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, 0x41b17218, s2
+; GFX1100-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1100-GISEL-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX1100-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1100-GISEL-NEXT: global_store_b32 v2, v0, s[0:1]
+; GFX1100-GISEL-NEXT: s_endpgm
+;
+; R600-LABEL: s_log_contract_f32:
+; R600: ; %bb.0:
+; R600-NEXT: ALU 23, @4, KC0[CB0:0-32], KC1[]
+; R600-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
+; R600-NEXT: CF_END
+; R600-NEXT: PAD
+; R600-NEXT: ALU clause starting at 4:
+; R600-NEXT: SETGT * T0.W, literal.x, KC0[2].Z,
+; R600-NEXT: 8388608(1.175494e-38), 0(0.000000e+00)
+; R600-NEXT: CNDE * T1.W, PV.W, 1.0, literal.x,
+; R600-NEXT: 1333788672(4.294967e+09), 0(0.000000e+00)
+; R600-NEXT: MUL_IEEE * T1.W, KC0[2].Z, PV.W,
+; R600-NEXT: LOG_IEEE * T0.X, PV.W,
+; R600-NEXT: AND_INT * T1.W, PS, literal.x,
+; R600-NEXT: -4096(nan), 0(0.000000e+00)
+; R600-NEXT: ADD * T2.W, T0.X, -PV.W,
+; R600-NEXT: MUL_IEEE * T3.W, PV.W, literal.x,
+; R600-NEXT: 939916788(3.194618e-05), 0(0.000000e+00)
+; R600-NEXT: MULADD_IEEE * T3.W, T1.W, literal.x, PV.W,
+; R600-NEXT: 939916788(3.194618e-05), 0(0.000000e+00)
+; R600-NEXT: MULADD_IEEE * T2.W, T2.W, literal.x, PV.W,
+; R600-NEXT: 1060204544(6.931152e-01), 0(0.000000e+00)
+; R600-NEXT: MULADD_IEEE T1.W, T1.W, literal.x, PV.W,
+; R600-NEXT: SETGT * T2.W, literal.y, |T0.X|,
+; R600-NEXT: 1060204544(6.931152e-01), 2139095040(INF)
+; R600-NEXT: CNDE T1.W, PS, T0.X, PV.W,
+; R600-NEXT: CNDE * T0.W, T0.W, 0.0, literal.x,
+; R600-NEXT: 1102148120(2.218071e+01), 0(0.000000e+00)
+; R600-NEXT: ADD T0.X, PV.W, -PS,
+; R600-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
+; R600-NEXT: 2(2.802597e-45), 0(0.000000e+00)
+;
+; CM-LABEL: s_log_contract_f32:
+; CM: ; %bb.0:
+; CM-NEXT: ALU 26, @4, KC0[CB0:0-32], KC1[]
+; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0.X, T1.X
+; CM-NEXT: CF_END
+; CM-NEXT: PAD
+; CM-NEXT: ALU clause starting at 4:
+; CM-NEXT: SETGT * T0.W, literal.x, KC0[2].Z,
+; CM-NEXT: 8388608(1.175494e-38), 0(0.000000e+00)
+; CM-NEXT: CNDE * T1.W, PV.W, 1.0, literal.x,
+; CM-NEXT: 1333788672(4.294967e+09), 0(0.000000e+00)
+; CM-NEXT: MUL_IEEE * T1.W, KC0[2].Z, PV.W,
+; CM-NEXT: LOG_IEEE T0.X, T1.W,
+; CM-NEXT: LOG_IEEE T0.Y (MASKED), T1.W,
+; CM-NEXT: LOG_IEEE T0.Z (MASKED), T1.W,
+; CM-NEXT: LOG_IEEE * T0.W (MASKED), T1.W,
+; CM-NEXT: AND_INT * T1.W, PV.X, literal.x,
+; CM-NEXT: -4096(nan), 0(0.000000e+00)
+; CM-NEXT: ADD * T2.W, T0.X, -PV.W,
+; CM-NEXT: MUL_IEEE * T3.W, PV.W, literal.x,
+; CM-NEXT: 939916788(3.194618e-05), 0(0.000000e+00)
+; CM-NEXT: MULADD_IEEE * T3.W, T1.W, literal.x, PV.W,
+; CM-NEXT: 939916788(3.194618e-05), 0(0.000000e+00)
+; CM-NEXT: MULADD_IEEE * T2.W, T2.W, literal.x, PV.W,
+; CM-NEXT: 1060204544(6.931152e-01), 0(0.000000e+00)
+; CM-NEXT: MULADD_IEEE T0.Z, T1.W, literal.x, PV.W,
+; CM-NEXT: SETGT * T1.W, literal.y, |T0.X|,
+; CM-NEXT: 1060204544(6.931152e-01), 2139095040(INF)
+; CM-NEXT: CNDE T0.Z, PV.W, T0.X, PV.Z,
+; CM-NEXT: CNDE * T0.W, T0.W, 0.0, literal.x,
+; CM-NEXT: 1102148120(2.218071e+01), 0(0.000000e+00)
+; CM-NEXT: ADD * T0.X, PV.Z, -PV.W,
+; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
+; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00)
+ %result = call contract float @llvm.log.f32(float %in)
+ store float %result, ptr addrspace(1) %out
+ ret void
+}
+
; FIXME: We should be able to merge these packets together on Cayman so we
; have a maximum of 4 instructions.
define amdgpu_kernel void @s_log_v2f32(ptr addrspace(1) %out, <2 x float> %in) {
@@ -6439,6 +6742,8 @@ define half @v_log_f16_fast(half %in) {
; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
; SI-SDAG-NEXT: v_log_f32_e32 v0, v0
; SI-SDAG-NEXT: v_mul_f32_e32 v0, 0x3f317218, v0
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; SI-GISEL-LABEL: v_log_f16_fast:
@@ -7100,6 +7405,10 @@ define <2 x half> @v_log_v2f16_fast(<2 x half> %in) {
; SI-SDAG-NEXT: v_log_f32_e32 v1, v1
; SI-SDAG-NEXT: v_mul_f32_e32 v0, 0x3f317218, v0
; SI-SDAG-NEXT: v_mul_f32_e32 v1, 0x3f317218, v1
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v1, v1
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; SI-GISEL-LABEL: v_log_v2f16_fast:
@@ -7365,6 +7674,12 @@ define <3 x half> @v_log_v3f16_fast(<3 x half> %in) {
; SI-SDAG-NEXT: v_mul_f32_e32 v0, 0x3f317218, v0
; SI-SDAG-NEXT: v_mul_f32_e32 v1, 0x3f317218, v1
; SI-SDAG-NEXT: v_mul_f32_e32 v2, 0x3f317218, v2
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v1, v1
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v2, v2
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; SI-GISEL-LABEL: v_log_v3f16_fast:
@@ -7691,20 +8006,28 @@ define <4 x half> @v_log_v4f16_fast(<4 x half> %in) {
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v1
-; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v2, v2
; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v3, v3
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v2, v2
; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v1, v1
-; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v2, v2
; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v3, v3
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v2, v2
; SI-SDAG-NEXT: v_log_f32_e32 v0, v0
; SI-SDAG-NEXT: v_log_f32_e32 v1, v1
-; SI-SDAG-NEXT: v_log_f32_e32 v2, v2
; SI-SDAG-NEXT: v_log_f32_e32 v3, v3
+; SI-SDAG-NEXT: v_log_f32_e32 v2, v2
; SI-SDAG-NEXT: v_mul_f32_e32 v0, 0x3f317218, v0
; SI-SDAG-NEXT: v_mul_f32_e32 v1, 0x3f317218, v1
-; SI-SDAG-NEXT: v_mul_f32_e32 v2, 0x3f317218, v2
; SI-SDAG-NEXT: v_mul_f32_e32 v3, 0x3f317218, v3
+; SI-SDAG-NEXT: v_mul_f32_e32 v2, 0x3f317218, v2
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v3, v3
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v1, v1
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v2, v2
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v3, v3
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; SI-GISEL-LABEL: v_log_v4f16_fast:
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.log10.ll b/llvm/test/CodeGen/AMDGPU/llvm.log10.ll
index a141bceb3ce86..6dbb5823a2b0f 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.log10.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.log10.ll
@@ -316,6 +316,309 @@ define amdgpu_kernel void @s_log10_f32(ptr addrspace(1) %out, float %in) {
ret void
}
+define amdgpu_kernel void @s_log10_contract_f32(ptr addrspace(1) %out, float %in) {
+; SI-SDAG-LABEL: s_log10_contract_f32:
+; SI-SDAG: ; %bb.0:
+; SI-SDAG-NEXT: s_load_dword s6, s[4:5], 0xb
+; SI-SDAG-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
+; SI-SDAG-NEXT: v_mov_b32_e32 v0, 0x800000
+; SI-SDAG-NEXT: v_mov_b32_e32 v1, 0x411a209b
+; SI-SDAG-NEXT: s_mov_b32 s4, 0x3e9a209a
+; SI-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; SI-SDAG-NEXT: v_cmp_lt_f32_e32 vcc, s6, v0
+; SI-SDAG-NEXT: s_and_b64 s[2:3], vcc, exec
+; SI-SDAG-NEXT: s_cselect_b32 s2, 32, 0
+; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v1, vcc
+; SI-SDAG-NEXT: v_mov_b32_e32 v1, s2
+; SI-SDAG-NEXT: v_ldexp_f32_e32 v1, s6, v1
+; SI-SDAG-NEXT: v_log_f32_e32 v1, v1
+; SI-SDAG-NEXT: s_mov_b32 s3, 0xf000
+; SI-SDAG-NEXT: s_mov_b32 s2, -1
+; SI-SDAG-NEXT: v_mul_f32_e32 v2, 0x3e9a209a, v1
+; SI-SDAG-NEXT: v_fma_f32 v3, v1, s4, -v2
+; SI-SDAG-NEXT: s_mov_b32 s4, 0x3284fbcf
+; SI-SDAG-NEXT: v_fma_f32 v3, v1, s4, v3
+; SI-SDAG-NEXT: s_mov_b32 s4, 0x7f800000
+; SI-SDAG-NEXT: v_add_f32_e32 v2, v2, v3
+; SI-SDAG-NEXT: v_cmp_lt_f32_e64 vcc, |v1|, s4
+; SI-SDAG-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; SI-SDAG-NEXT: v_sub_f32_e32 v0, v1, v0
+; SI-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], 0
+; SI-SDAG-NEXT: s_endpgm
+;
+; SI-GISEL-LABEL: s_log10_contract_f32:
+; SI-GISEL: ; %bb.0:
+; SI-GISEL-NEXT: s_load_dword s0, s[4:5], 0xb
+; SI-GISEL-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x9
+; SI-GISEL-NEXT: v_mov_b32_e32 v0, 0x800000
+; SI-GISEL-NEXT: v_mov_b32_e32 v1, 0x3e9a209a
+; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x3284fbcf
+; SI-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; SI-GISEL-NEXT: v_cmp_lt_f32_e32 vcc, s0, v0
+; SI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
+; SI-GISEL-NEXT: v_lshlrev_b32_e32 v0, 5, v0
+; SI-GISEL-NEXT: v_ldexp_f32_e32 v0, s0, v0
+; SI-GISEL-NEXT: v_log_f32_e32 v0, v0
+; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x7f800000
+; SI-GISEL-NEXT: s_mov_b32 s6, -1
+; SI-GISEL-NEXT: s_mov_b32 s7, 0xf000
+; SI-GISEL-NEXT: v_mul_f32_e32 v4, 0x3e9a209a, v0
+; SI-GISEL-NEXT: v_fma_f32 v1, v0, v1, -v4
+; SI-GISEL-NEXT: v_fma_f32 v1, v0, v2, v1
+; SI-GISEL-NEXT: v_add_f32_e32 v1, v4, v1
+; SI-GISEL-NEXT: v_cmp_lt_f32_e64 s[0:1], |v0|, v3
+; SI-GISEL-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[0:1]
+; SI-GISEL-NEXT: v_mov_b32_e32 v1, 0x411a209b
+; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
+; SI-GISEL-NEXT: v_sub_f32_e32 v0, v0, v1
+; SI-GISEL-NEXT: buffer_store_dword v0, off, s[4:7], 0
+; SI-GISEL-NEXT: s_endpgm
+;
+; VI-SDAG-LABEL: s_log10_contract_f32:
+; VI-SDAG: ; %bb.0:
+; VI-SDAG-NEXT: s_load_dword s2, s[4:5], 0x2c
+; VI-SDAG-NEXT: v_mov_b32_e32 v0, 0x800000
+; VI-SDAG-NEXT: v_mov_b32_e32 v1, 0x411a209b
+; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; VI-SDAG-NEXT: v_cmp_lt_f32_e32 vcc, s2, v0
+; VI-SDAG-NEXT: s_and_b64 s[0:1], vcc, exec
+; VI-SDAG-NEXT: s_cselect_b32 s0, 32, 0
+; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v1, vcc
+; VI-SDAG-NEXT: v_mov_b32_e32 v1, s0
+; VI-SDAG-NEXT: v_ldexp_f32 v1, s2, v1
+; VI-SDAG-NEXT: v_log_f32_e32 v1, v1
+; VI-SDAG-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; VI-SDAG-NEXT: s_mov_b32 s2, 0x7f800000
+; VI-SDAG-NEXT: v_and_b32_e32 v2, 0xfffff000, v1
+; VI-SDAG-NEXT: v_sub_f32_e32 v3, v1, v2
+; VI-SDAG-NEXT: v_mul_f32_e32 v4, 0x369a84fb, v2
+; VI-SDAG-NEXT: v_mul_f32_e32 v5, 0x3e9a2000, v3
+; VI-SDAG-NEXT: v_mul_f32_e32 v3, 0x369a84fb, v3
+; VI-SDAG-NEXT: v_add_f32_e32 v3, v4, v3
+; VI-SDAG-NEXT: v_mul_f32_e32 v2, 0x3e9a2000, v2
+; VI-SDAG-NEXT: v_add_f32_e32 v3, v5, v3
+; VI-SDAG-NEXT: v_add_f32_e32 v2, v2, v3
+; VI-SDAG-NEXT: v_cmp_lt_f32_e64 vcc, |v1|, s2
+; VI-SDAG-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; VI-SDAG-NEXT: v_sub_f32_e32 v2, v1, v0
+; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; VI-SDAG-NEXT: v_mov_b32_e32 v0, s0
+; VI-SDAG-NEXT: v_mov_b32_e32 v1, s1
+; VI-SDAG-NEXT: flat_store_dword v[0:1], v2
+; VI-SDAG-NEXT: s_endpgm
+;
+; VI-GISEL-LABEL: s_log10_contract_f32:
+; VI-GISEL: ; %bb.0:
+; VI-GISEL-NEXT: s_load_dword s0, s[4:5], 0x2c
+; VI-GISEL-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x800000
+; VI-GISEL-NEXT: v_mov_b32_e32 v1, 0x7f800000
+; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; VI-GISEL-NEXT: v_cmp_lt_f32_e32 vcc, s0, v0
+; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
+; VI-GISEL-NEXT: v_lshlrev_b32_e32 v0, 5, v0
+; VI-GISEL-NEXT: v_ldexp_f32 v0, s0, v0
+; VI-GISEL-NEXT: v_log_f32_e32 v0, v0
+; VI-GISEL-NEXT: v_and_b32_e32 v2, 0xfffff000, v0
+; VI-GISEL-NEXT: v_sub_f32_e32 v3, v0, v2
+; VI-GISEL-NEXT: v_mul_f32_e32 v4, 0x369a84fb, v2
+; VI-GISEL-NEXT: v_mul_f32_e32 v5, 0x369a84fb, v3
+; VI-GISEL-NEXT: v_mul_f32_e32 v3, 0x3e9a2000, v3
+; VI-GISEL-NEXT: v_add_f32_e32 v4, v4, v5
+; VI-GISEL-NEXT: v_mul_f32_e32 v2, 0x3e9a2000, v2
+; VI-GISEL-NEXT: v_add_f32_e32 v3, v3, v4
+; VI-GISEL-NEXT: v_add_f32_e32 v2, v2, v3
+; VI-GISEL-NEXT: v_cmp_lt_f32_e64 s[0:1], |v0|, v1
+; VI-GISEL-NEXT: v_mov_b32_e32 v1, 0x411a209b
+; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[0:1]
+; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
+; VI-GISEL-NEXT: v_sub_f32_e32 v2, v0, v1
+; VI-GISEL-NEXT: v_mov_b32_e32 v0, s2
+; VI-GISEL-NEXT: v_mov_b32_e32 v1, s3
+; VI-GISEL-NEXT: flat_store_dword v[0:1], v2
+; VI-GISEL-NEXT: s_endpgm
+;
+; GFX900-SDAG-LABEL: s_log10_contract_f32:
+; GFX900-SDAG: ; %bb.0:
+; GFX900-SDAG-NEXT: s_load_dword s6, s[4:5], 0x2c
+; GFX900-SDAG-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX900-SDAG-NEXT: v_mov_b32_e32 v0, 0x800000
+; GFX900-SDAG-NEXT: v_mov_b32_e32 v1, 0x411a209b
+; GFX900-SDAG-NEXT: v_mov_b32_e32 v2, 0
+; GFX900-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; GFX900-SDAG-NEXT: v_cmp_lt_f32_e32 vcc, s6, v0
+; GFX900-SDAG-NEXT: s_and_b64 s[2:3], vcc, exec
+; GFX900-SDAG-NEXT: s_cselect_b32 s2, 32, 0
+; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v1, vcc
+; GFX900-SDAG-NEXT: v_mov_b32_e32 v1, s2
+; GFX900-SDAG-NEXT: v_ldexp_f32 v1, s6, v1
+; GFX900-SDAG-NEXT: v_log_f32_e32 v1, v1
+; GFX900-SDAG-NEXT: s_mov_b32 s2, 0x3e9a209a
+; GFX900-SDAG-NEXT: s_mov_b32 s3, 0x3284fbcf
+; GFX900-SDAG-NEXT: v_mul_f32_e32 v3, 0x3e9a209a, v1
+; GFX900-SDAG-NEXT: v_fma_f32 v4, v1, s2, -v3
+; GFX900-SDAG-NEXT: v_fma_f32 v4, v1, s3, v4
+; GFX900-SDAG-NEXT: s_mov_b32 s2, 0x7f800000
+; GFX900-SDAG-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX900-SDAG-NEXT: v_cmp_lt_f32_e64 vcc, |v1|, s2
+; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX900-SDAG-NEXT: v_sub_f32_e32 v0, v1, v0
+; GFX900-SDAG-NEXT: global_store_dword v2, v0, s[0:1]
+; GFX900-SDAG-NEXT: s_endpgm
+;
+; GFX900-GISEL-LABEL: s_log10_contract_f32:
+; GFX900-GISEL: ; %bb.0:
+; GFX900-GISEL-NEXT: s_load_dword s0, s[4:5], 0x2c
+; GFX900-GISEL-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x800000
+; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x3e9a209a
+; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3284fbcf
+; GFX900-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX900-GISEL-NEXT: v_cmp_lt_f32_e32 vcc, s0, v0
+; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
+; GFX900-GISEL-NEXT: v_lshlrev_b32_e32 v0, 5, v0
+; GFX900-GISEL-NEXT: v_ldexp_f32 v0, s0, v0
+; GFX900-GISEL-NEXT: v_log_f32_e32 v0, v0
+; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0x7f800000
+; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, 0
+; GFX900-GISEL-NEXT: v_mul_f32_e32 v5, 0x3e9a209a, v0
+; GFX900-GISEL-NEXT: v_fma_f32 v2, v0, v2, -v5
+; GFX900-GISEL-NEXT: v_fma_f32 v2, v0, v3, v2
+; GFX900-GISEL-NEXT: v_add_f32_e32 v2, v5, v2
+; GFX900-GISEL-NEXT: v_cmp_lt_f32_e64 s[0:1], |v0|, v4
+; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[0:1]
+; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x411a209b
+; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
+; GFX900-GISEL-NEXT: v_sub_f32_e32 v0, v0, v2
+; GFX900-GISEL-NEXT: global_store_dword v1, v0, s[2:3]
+; GFX900-GISEL-NEXT: s_endpgm
+;
+; GFX1100-SDAG-LABEL: s_log10_contract_f32:
+; GFX1100-SDAG: ; %bb.0:
+; GFX1100-SDAG-NEXT: s_load_b32 s0, s[4:5], 0x2c
+; GFX1100-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1100-SDAG-NEXT: v_cmp_gt_f32_e64 s1, 0x800000, s0
+; GFX1100-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX1100-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 0x411a209b, s1
+; GFX1100-SDAG-NEXT: s_and_b32 s1, s1, exec_lo
+; GFX1100-SDAG-NEXT: s_cselect_b32 s1, 32, 0
+; GFX1100-SDAG-NEXT: v_ldexp_f32 v1, s0, s1
+; GFX1100-SDAG-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1100-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
+; GFX1100-SDAG-NEXT: v_log_f32_e32 v1, v1
+; GFX1100-SDAG-NEXT: s_waitcnt_depctr 0xfff
+; GFX1100-SDAG-NEXT: v_mul_f32_e32 v2, 0x3e9a209a, v1
+; GFX1100-SDAG-NEXT: v_cmp_gt_f32_e64 vcc_lo, 0x7f800000, |v1|
+; GFX1100-SDAG-NEXT: v_fma_f32 v3, 0x3e9a209a, v1, -v2
+; GFX1100-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1100-SDAG-NEXT: v_fmamk_f32 v3, v1, 0x3284fbcf, v3
+; GFX1100-SDAG-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX1100-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1100-SDAG-NEXT: v_dual_cndmask_b32 v1, v1, v2 :: v_dual_mov_b32 v2, 0
+; GFX1100-SDAG-NEXT: v_sub_f32_e32 v0, v1, v0
+; GFX1100-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1100-SDAG-NEXT: global_store_b32 v2, v0, s[0:1]
+; GFX1100-SDAG-NEXT: s_endpgm
+;
+; GFX1100-GISEL-LABEL: s_log10_contract_f32:
+; GFX1100-GISEL: ; %bb.0:
+; GFX1100-GISEL-NEXT: s_load_b32 s0, s[4:5], 0x2c
+; GFX1100-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1100-GISEL-NEXT: v_cmp_gt_f32_e64 s2, 0x800000, s0
+; GFX1100-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1100-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, s2
+; GFX1100-GISEL-NEXT: v_lshlrev_b32_e32 v0, 5, v0
+; GFX1100-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX1100-GISEL-NEXT: v_ldexp_f32 v0, s0, v0
+; GFX1100-GISEL-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1100-GISEL-NEXT: v_log_f32_e32 v0, v0
+; GFX1100-GISEL-NEXT: s_waitcnt_depctr 0xfff
+; GFX1100-GISEL-NEXT: v_mul_f32_e32 v1, 0x3e9a209a, v0
+; GFX1100-GISEL-NEXT: v_cmp_gt_f32_e64 vcc_lo, 0x7f800000, |v0|
+; GFX1100-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1100-GISEL-NEXT: v_fma_f32 v2, 0x3e9a209a, v0, -v1
+; GFX1100-GISEL-NEXT: v_fmac_f32_e32 v2, 0x3284fbcf, v0
+; GFX1100-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1100-GISEL-NEXT: v_dual_add_f32 v1, v1, v2 :: v_dual_mov_b32 v2, 0
+; GFX1100-GISEL-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc_lo
+; GFX1100-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, 0x411a209b, s2
+; GFX1100-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1100-GISEL-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX1100-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; GFX1100-GISEL-NEXT: global_store_b32 v2, v0, s[0:1]
+; GFX1100-GISEL-NEXT: s_endpgm
+;
+; R600-LABEL: s_log10_contract_f32:
+; R600: ; %bb.0:
+; R600-NEXT: ALU 23, @4, KC0[CB0:0-32], KC1[]
+; R600-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
+; R600-NEXT: CF_END
+; R600-NEXT: PAD
+; R600-NEXT: ALU clause starting at 4:
+; R600-NEXT: SETGT * T0.W, literal.x, KC0[2].Z,
+; R600-NEXT: 8388608(1.175494e-38), 0(0.000000e+00)
+; R600-NEXT: CNDE * T1.W, PV.W, 1.0, literal.x,
+; R600-NEXT: 1333788672(4.294967e+09), 0(0.000000e+00)
+; R600-NEXT: MUL_IEEE * T1.W, KC0[2].Z, PV.W,
+; R600-NEXT: LOG_IEEE * T0.X, PV.W,
+; R600-NEXT: AND_INT * T1.W, PS, literal.x,
+; R600-NEXT: -4096(nan), 0(0.000000e+00)
+; R600-NEXT: ADD * T2.W, T0.X, -PV.W,
+; R600-NEXT: MUL_IEEE * T3.W, PV.W, literal.x,
+; R600-NEXT: 916096251(4.605039e-06), 0(0.000000e+00)
+; R600-NEXT: MULADD_IEEE * T3.W, T1.W, literal.x, PV.W,
+; R600-NEXT: 916096251(4.605039e-06), 0(0.000000e+00)
+; R600-NEXT: MULADD_IEEE * T2.W, T2.W, literal.x, PV.W,
+; R600-NEXT: 1050288128(3.010254e-01), 0(0.000000e+00)
+; R600-NEXT: MULADD_IEEE T1.W, T1.W, literal.x, PV.W,
+; R600-NEXT: SETGT * T2.W, literal.y, |T0.X|,
+; R600-NEXT: 1050288128(3.010254e-01), 2139095040(INF)
+; R600-NEXT: CNDE T1.W, PS, T0.X, PV.W,
+; R600-NEXT: CNDE * T0.W, T0.W, 0.0, literal.x,
+; R600-NEXT: 1092231323(9.632960e+00), 0(0.000000e+00)
+; R600-NEXT: ADD T0.X, PV.W, -PS,
+; R600-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
+; R600-NEXT: 2(2.802597e-45), 0(0.000000e+00)
+;
+; CM-LABEL: s_log10_contract_f32:
+; CM: ; %bb.0:
+; CM-NEXT: ALU 26, @4, KC0[CB0:0-32], KC1[]
+; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0.X, T1.X
+; CM-NEXT: CF_END
+; CM-NEXT: PAD
+; CM-NEXT: ALU clause starting at 4:
+; CM-NEXT: SETGT * T0.W, literal.x, KC0[2].Z,
+; CM-NEXT: 8388608(1.175494e-38), 0(0.000000e+00)
+; CM-NEXT: CNDE * T1.W, PV.W, 1.0, literal.x,
+; CM-NEXT: 1333788672(4.294967e+09), 0(0.000000e+00)
+; CM-NEXT: MUL_IEEE * T1.W, KC0[2].Z, PV.W,
+; CM-NEXT: LOG_IEEE T0.X, T1.W,
+; CM-NEXT: LOG_IEEE T0.Y (MASKED), T1.W,
+; CM-NEXT: LOG_IEEE T0.Z (MASKED), T1.W,
+; CM-NEXT: LOG_IEEE * T0.W (MASKED), T1.W,
+; CM-NEXT: AND_INT * T1.W, PV.X, literal.x,
+; CM-NEXT: -4096(nan), 0(0.000000e+00)
+; CM-NEXT: ADD * T2.W, T0.X, -PV.W,
+; CM-NEXT: MUL_IEEE * T3.W, PV.W, literal.x,
+; CM-NEXT: 916096251(4.605039e-06), 0(0.000000e+00)
+; CM-NEXT: MULADD_IEEE * T3.W, T1.W, literal.x, PV.W,
+; CM-NEXT: 916096251(4.605039e-06), 0(0.000000e+00)
+; CM-NEXT: MULADD_IEEE * T2.W, T2.W, literal.x, PV.W,
+; CM-NEXT: 1050288128(3.010254e-01), 0(0.000000e+00)
+; CM-NEXT: MULADD_IEEE T0.Z, T1.W, literal.x, PV.W,
+; CM-NEXT: SETGT * T1.W, literal.y, |T0.X|,
+; CM-NEXT: 1050288128(3.010254e-01), 2139095040(INF)
+; CM-NEXT: CNDE T0.Z, PV.W, T0.X, PV.Z,
+; CM-NEXT: CNDE * T0.W, T0.W, 0.0, literal.x,
+; CM-NEXT: 1092231323(9.632960e+00), 0(0.000000e+00)
+; CM-NEXT: ADD * T0.X, PV.Z, -PV.W,
+; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
+; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00)
+ %result = call contract float @llvm.log10.f32(float %in)
+ store float %result, ptr addrspace(1) %out
+ ret void
+}
+
; FIXME: We should be able to merge these packets together on Cayman so we
; have a maximum of 4 instructions.
define amdgpu_kernel void @s_log10_v2f32(ptr addrspace(1) %out, <2 x float> %in) {
@@ -6439,6 +6742,8 @@ define half @v_log10_f16_fast(half %in) {
; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
; SI-SDAG-NEXT: v_log_f32_e32 v0, v0
; SI-SDAG-NEXT: v_mul_f32_e32 v0, 0x3e9a209b, v0
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; SI-GISEL-LABEL: v_log10_f16_fast:
@@ -7100,6 +7405,10 @@ define <2 x half> @v_log10_v2f16_fast(<2 x half> %in) {
; SI-SDAG-NEXT: v_log_f32_e32 v1, v1
; SI-SDAG-NEXT: v_mul_f32_e32 v0, 0x3e9a209b, v0
; SI-SDAG-NEXT: v_mul_f32_e32 v1, 0x3e9a209b, v1
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v1, v1
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; SI-GISEL-LABEL: v_log10_v2f16_fast:
@@ -7365,6 +7674,12 @@ define <3 x half> @v_log10_v3f16_fast(<3 x half> %in) {
; SI-SDAG-NEXT: v_mul_f32_e32 v0, 0x3e9a209b, v0
; SI-SDAG-NEXT: v_mul_f32_e32 v1, 0x3e9a209b, v1
; SI-SDAG-NEXT: v_mul_f32_e32 v2, 0x3e9a209b, v2
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v1, v1
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v2, v2
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; SI-GISEL-LABEL: v_log10_v3f16_fast:
@@ -7691,20 +8006,28 @@ define <4 x half> @v_log10_v4f16_fast(<4 x half> %in) {
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v1
-; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v2, v2
; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v3, v3
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v2, v2
; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v1, v1
-; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v2, v2
; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v3, v3
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v2, v2
; SI-SDAG-NEXT: v_log_f32_e32 v0, v0
; SI-SDAG-NEXT: v_log_f32_e32 v1, v1
-; SI-SDAG-NEXT: v_log_f32_e32 v2, v2
; SI-SDAG-NEXT: v_log_f32_e32 v3, v3
+; SI-SDAG-NEXT: v_log_f32_e32 v2, v2
; SI-SDAG-NEXT: v_mul_f32_e32 v0, 0x3e9a209b, v0
; SI-SDAG-NEXT: v_mul_f32_e32 v1, 0x3e9a209b, v1
-; SI-SDAG-NEXT: v_mul_f32_e32 v2, 0x3e9a209b, v2
; SI-SDAG-NEXT: v_mul_f32_e32 v3, 0x3e9a209b, v3
+; SI-SDAG-NEXT: v_mul_f32_e32 v2, 0x3e9a209b, v2
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v3, v3
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v1, v1
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v2, v2
+; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v3, v3
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; SI-GISEL-LABEL: v_log10_v4f16_fast:
>From 04c43c9998f4267009a9383ec5a6aca93da72f0a Mon Sep 17 00:00:00 2001
From: Adel Ejjeh <adel.ejjeh at amd.com>
Date: Thu, 20 Nov 2025 12:57:52 -0600
Subject: [PATCH 2/2] Move flag modification placement
---
llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp | 11 ++++++-----
1 file changed, 6 insertions(+), 5 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index ffafe06211109..3705b8e12c6b7 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -2766,9 +2766,6 @@ SDValue AMDGPUTargetLowering::LowerFLOGCommon(SDValue Op,
EVT VT = Op.getValueType();
SDNodeFlags Flags = Op->getFlags();
SDLoc DL(Op);
- // Our implementation of LOG is not contract safe, so disable instruction
- // contraction.
- Flags.setAllowContract(false);
const bool IsLog10 = Op.getOpcode() == ISD::FLOG10;
assert(IsLog10 || Op.getOpcode() == ISD::FLOG);
@@ -2807,7 +2804,9 @@ SDValue AMDGPUTargetLowering::LowerFLOGCommon(SDValue Op,
SDValue C = DAG.getConstantFP(IsLog10 ? c_log10 : c_log, DL, VT);
SDValue CC = DAG.getConstantFP(IsLog10 ? cc_log10 : cc_log, DL, VT);
-
+ // Our implementation of LOG is not contract safe, so disable instruction
+ // contraction.
+ Flags.setAllowContract(false);
R = DAG.getNode(ISD::FMUL, DL, VT, Y, C, Flags);
SDValue NegR = DAG.getNode(ISD::FNEG, DL, VT, R, Flags);
SDValue FMA0 = DAG.getNode(ISD::FMA, DL, VT, Y, C, NegR, Flags);
@@ -2830,7 +2829,9 @@ SDValue AMDGPUTargetLowering::LowerFLOGCommon(SDValue Op,
SDValue YHInt = DAG.getNode(ISD::AND, DL, MVT::i32, YAsInt, MaskConst);
SDValue YH = DAG.getNode(ISD::BITCAST, DL, MVT::f32, YHInt);
SDValue YT = DAG.getNode(ISD::FSUB, DL, VT, Y, YH, Flags);
-
+ // Our implementation of LOG is not contract safe, so disable instruction
+ // contraction.
+ Flags.setAllowContract(false);
SDValue YTCT = DAG.getNode(ISD::FMUL, DL, VT, YT, CT, Flags);
SDValue Mad0 = getMad(DAG, DL, VT, YH, CT, YTCT, Flags);
SDValue Mad1 = getMad(DAG, DL, VT, YT, CH, Mad0, Flags);
More information about the llvm-commits
mailing list