[llvm] [AMDGPU] Use AMDGPU::isIntrinsicAlwaysUniform in isSDNodeAlwaysUniform (PR #87085)
Jay Foad via llvm-commits
llvm-commits at lists.llvm.org
Fri Mar 29 09:34:39 PDT 2024
https://github.com/jayfoad created https://github.com/llvm/llvm-project/pull/87085
This is mostly just a simplification, but tests show a slight codegen
improvement in code using the deprecated amdgcn.icmp/fcmp intrinsics.
>From 5e5419735333cdc66b64d55f565b02a373e2dd9a Mon Sep 17 00:00:00 2001
From: Jay Foad <jay.foad at amd.com>
Date: Fri, 29 Mar 2024 16:17:48 +0000
Subject: [PATCH] [AMDGPU] Use AMDGPU::isIntrinsicAlwaysUniform in
isSDNodeAlwaysUniform
This is mostly just a simplification, but tests show a slight codegen
improvement in code using the deprecated amdgcn.icmp/fcmp intrinsics.
---
llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp | 7 +---
llvm/test/CodeGen/AMDGPU/wave32.ll | 34 +++++++------------
2 files changed, 13 insertions(+), 28 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index bee43b6c18c880..f283af6fa07d3e 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -851,12 +851,7 @@ bool AMDGPUTargetLowering::isSDNodeAlwaysUniform(const SDNode *N) const {
return true;
case ISD::INTRINSIC_WO_CHAIN: {
unsigned IntrID = N->getConstantOperandVal(0);
- switch (IntrID) {
- case Intrinsic::amdgcn_readfirstlane:
- case Intrinsic::amdgcn_readlane:
- return true;
- }
- return false;
+ return AMDGPU::isIntrinsicAlwaysUniform(IntrID);
}
case ISD::LOAD:
if (cast<LoadSDNode>(N)->getMemOperand()->getAddrSpace() ==
diff --git a/llvm/test/CodeGen/AMDGPU/wave32.ll b/llvm/test/CodeGen/AMDGPU/wave32.ll
index 82816b4564e868..901e88a4c6aca8 100644
--- a/llvm/test/CodeGen/AMDGPU/wave32.ll
+++ b/llvm/test/CodeGen/AMDGPU/wave32.ll
@@ -2479,8 +2479,7 @@ define amdgpu_kernel void @icmp64(i32 %n, i32 %s) {
; GFX1032-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
; GFX1032-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX1032-NEXT: v_mul_lo_u32 v2, s1, v1
-; GFX1032-NEXT: s_ff1_i32_b32 s1, 0x80000000
-; GFX1032-NEXT: s_add_i32 s1, s1, 32
+; GFX1032-NEXT: s_brev_b32 s1, 1
; GFX1032-NEXT: v_mul_hi_u32 v2, v1, v2
; GFX1032-NEXT: v_add_nc_u32_e32 v1, v1, v2
; GFX1032-NEXT: v_mul_hi_u32 v1, v0, v1
@@ -2494,8 +2493,7 @@ define amdgpu_kernel void @icmp64(i32 %n, i32 %s) {
; GFX1032-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_lshr_b32 s0, vcc_lo, 1
-; GFX1032-NEXT: s_ff1_i32_b32 s0, s0
-; GFX1032-NEXT: s_min_u32 s0, s0, s1
+; GFX1032-NEXT: s_ff1_i32_b64 s0, s[0:1]
; GFX1032-NEXT: s_cmp_gt_u32 s0, 9
; GFX1032-NEXT: s_cselect_b32 s0, -1, 0
; GFX1032-NEXT: s_and_b32 s0, vcc_lo, s0
@@ -2529,10 +2527,7 @@ define amdgpu_kernel void @icmp64(i32 %n, i32 %s) {
; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-NEXT: s_lshr_b64 s[0:1], vcc, 1
; GFX1064-NEXT: s_bitset1_b32 s1, 31
-; GFX1064-NEXT: s_ff1_i32_b32 s0, s0
-; GFX1064-NEXT: s_ff1_i32_b32 s1, s1
-; GFX1064-NEXT: s_add_i32 s1, s1, 32
-; GFX1064-NEXT: s_min_u32 s0, s0, s1
+; GFX1064-NEXT: s_ff1_i32_b64 s0, s[0:1]
; GFX1064-NEXT: s_cmp_gt_u32 s0, 9
; GFX1064-NEXT: s_cselect_b64 s[0:1], -1, 0
; GFX1064-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
@@ -2576,9 +2571,8 @@ define amdgpu_kernel void @fcmp64(float %n, float %s) {
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-NEXT: v_div_scale_f32 v1, s1, s0, s0, v0
; GFX1032-NEXT: v_div_scale_f32 v4, vcc_lo, v0, s0, v0
-; GFX1032-NEXT: s_ff1_i32_b32 s1, 0x80000000
+; GFX1032-NEXT: s_brev_b32 s1, 1
; GFX1032-NEXT: v_rcp_f32_e32 v2, v1
-; GFX1032-NEXT: s_add_i32 s1, s1, 32
; GFX1032-NEXT: v_fma_f32 v3, -v1, v2, 1.0
; GFX1032-NEXT: v_fmac_f32_e32 v2, v3, v2
; GFX1032-NEXT: v_mul_f32_e32 v3, v4, v2
@@ -2592,8 +2586,7 @@ define amdgpu_kernel void @fcmp64(float %n, float %s) {
; GFX1032-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_lshr_b32 s0, vcc_lo, 1
; GFX1032-NEXT: v_cmp_nlg_f32_e32 vcc_lo, 0, v0
-; GFX1032-NEXT: s_ff1_i32_b32 s0, s0
-; GFX1032-NEXT: s_min_u32 s0, s0, s1
+; GFX1032-NEXT: s_ff1_i32_b64 s0, s[0:1]
; GFX1032-NEXT: s_cmp_gt_u32 s0, 9
; GFX1032-NEXT: s_cselect_b32 s0, -1, 0
; GFX1032-NEXT: s_and_b32 s0, vcc_lo, s0
@@ -2609,15 +2602,15 @@ define amdgpu_kernel void @fcmp64(float %n, float %s) {
; GFX1064-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-NEXT: v_div_scale_f32 v1, s[0:1], s2, s2, v0
-; GFX1064-NEXT: v_div_scale_f32 v4, vcc, v0, s2, v0
; GFX1064-NEXT: v_rcp_f32_e32 v2, v1
; GFX1064-NEXT: v_fma_f32 v3, -v1, v2, 1.0
; GFX1064-NEXT: v_fmac_f32_e32 v2, v3, v2
-; GFX1064-NEXT: v_mul_f32_e32 v3, v4, v2
-; GFX1064-NEXT: v_fma_f32 v5, -v1, v3, v4
-; GFX1064-NEXT: v_fmac_f32_e32 v3, v5, v2
-; GFX1064-NEXT: v_fma_f32 v1, -v1, v3, v4
-; GFX1064-NEXT: v_div_fmas_f32 v1, v1, v2, v3
+; GFX1064-NEXT: v_div_scale_f32 v3, vcc, v0, s2, v0
+; GFX1064-NEXT: v_mul_f32_e32 v4, v3, v2
+; GFX1064-NEXT: v_fma_f32 v5, -v1, v4, v3
+; GFX1064-NEXT: v_fmac_f32_e32 v4, v5, v2
+; GFX1064-NEXT: v_fma_f32 v1, -v1, v4, v3
+; GFX1064-NEXT: v_div_fmas_f32 v1, v1, v2, v4
; GFX1064-NEXT: v_div_fixup_f32 v1, v1, s2, v0
; GFX1064-NEXT: v_trunc_f32_e32 v1, v1
; GFX1064-NEXT: v_fma_f32 v0, -v1, s2, v0
@@ -2625,10 +2618,7 @@ define amdgpu_kernel void @fcmp64(float %n, float %s) {
; GFX1064-NEXT: s_lshr_b64 s[0:1], vcc, 1
; GFX1064-NEXT: v_cmp_nlg_f32_e32 vcc, 0, v0
; GFX1064-NEXT: s_bitset1_b32 s1, 31
-; GFX1064-NEXT: s_ff1_i32_b32 s0, s0
-; GFX1064-NEXT: s_ff1_i32_b32 s1, s1
-; GFX1064-NEXT: s_add_i32 s1, s1, 32
-; GFX1064-NEXT: s_min_u32 s0, s0, s1
+; GFX1064-NEXT: s_ff1_i32_b64 s0, s[0:1]
; GFX1064-NEXT: s_cmp_gt_u32 s0, 9
; GFX1064-NEXT: s_cselect_b64 s[0:1], -1, 0
; GFX1064-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
More information about the llvm-commits
mailing list