1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-25 12:12:47 +01:00
llvm-mirror/test/CodeGen/AMDGPU/mul.i16.ll
Nicolai Hähnle 60121ef575 [AMDGPU][SelectionDAG] Don't combine uniform multiplies to MUL_[UI]24
Prefer to keep uniform (non-divergent) multiplies on the scalar ALU when
possible. This significantly improves some game cases by eliminating
v_readfirstlane instructions when the result feeds into a scalar
operation, like the address calculation for a scalar load or store.

Since isDivergent is only an approximation of whether a value is in
SGPRs, it can potentially regress some situations where a uniform value
ends up in a VGPR. These should be rare in real code, although the test
changes do contain a number of examples.

Most of the test changes are just using s_mul instead of v_mul/mad which
is generally better for both register pressure and latency (at least on
GFX10 where sgpr pressure doesn't affect occupancy and vector ALU
instructions have significantly longer latency than scalar ALU). Some
R600 tests now use MULLO_INT instead of MUL_UINT24.

GlobalISel appears to handle more scenarios in the desirable way,
although it can also be thrown off and fails to select the 24-bit
multiplies in some cases.

Alternative solution considered and rejected was to allow selecting
MUL_[UI]24 to S_MUL_I32. I've rejected this because the definition of
those SD operations works is don't-care on the most significant 8 bits,
and this fact is used in some combines via SimplifyDemandedBits.

Based on a patch by Nicolai Hähnle.

Differential Revision: https://reviews.llvm.org/D97063
2021-02-23 15:39:19 +00:00

97 lines
2.6 KiB
LLVM

; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,SI %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,VI,GFX89 %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX9,GFX89 %s
; GCN-LABEL: {{^}}v_mul_i16:
; SI: s_mov_b32 [[K:s[0-9]+]], 0xffff{{$}}
; SI: v_and_b32_e32 v{{[0-9]+}}, [[K]]
; SI: v_and_b32_e32 v{{[0-9]+}}, [[K]]
; SI: v_mul_u32_u24
; GFX89: v_mul_lo_u16_e32 v0, v0, v1
define i16 @v_mul_i16(i16 %a, i16 %b) {
%r.val = mul i16 %a, %b
ret i16 %r.val
}
; GCN-LABEL: {{^}}s_mul_i16:
; GCN: s_mul_i16
define amdgpu_kernel void @s_mul_i16(i16 %a, i16 %b) {
%r.val = mul i16 %a, %b
store volatile i16 %r.val, i16 addrspace(1)* null
ret void
}
; FIXME: Should emit u16 mul here.
; GCN-LABEL: {{^}}v_mul_i16_uniform_load:
; GCN: v_mul_lo_u32
define amdgpu_kernel void @v_mul_i16_uniform_load(
i16 addrspace(1)* %r,
i16 addrspace(1)* %a,
i16 addrspace(1)* %b) {
entry:
%a.val = load i16, i16 addrspace(1)* %a
%b.val = load i16, i16 addrspace(1)* %b
%r.val = mul i16 %a.val, %b.val
store i16 %r.val, i16 addrspace(1)* %r
ret void
}
; GCN-LABEL: {{^}}v_mul_v2i16:
; SI: v_mul_u32_u24
; SI: v_mul_u32_u24
; VI: v_mul_lo_u16_sdwa
; VI: v_mul_lo_u16_e32
; VI: v_or_b32_e32
; GFX9: s_waitcnt
; GFX9-NEXT: v_pk_mul_lo_u16 v0, v0, v1
; GFX9-NEXT: s_setpc_b64
define <2 x i16> @v_mul_v2i16(<2 x i16> %a, <2 x i16> %b) {
%r.val = mul <2 x i16> %a, %b
ret <2 x i16> %r.val
}
; FIXME: Unpack garbage on gfx9
; GCN-LABEL: {{^}}v_mul_v3i16:
; SI: v_mul_u32_u24
; SI: v_mul_u32_u24
; SI: v_mul_u32_u24
; VI: v_mul_lo_u16
; VI: v_mul_lo_u16
; VI: v_mul_lo_u16
; GFX9: s_waitcnt
; GFX9-NEXT: v_pk_mul_lo_u16
; GFX9-NEXT: v_pk_mul_lo_u16
; GFX9-NEXT: s_setpc_b64
define <3 x i16> @v_mul_v3i16(<3 x i16> %a, <3 x i16> %b) {
%r.val = mul <3 x i16> %a, %b
ret <3 x i16> %r.val
}
; GCN-LABEL: {{^}}v_mul_v4i16:
; SI: v_mul_u32_u24
; SI: v_mul_u32_u24
; SI: v_mul_u32_u24
; SI: v_mul_u32_u24
; VI: v_mul_lo_u16_sdwa
; VI: v_mul_lo_u16_e32
; VI: v_mul_lo_u16_sdwa
; VI: v_mul_lo_u16_e32
; VI: v_or_b32_e32
; VI: v_or_b32_e32
; GFX9: s_waitcnt
; GFX9-NEXT: v_pk_mul_lo_u16 v0, v0, v2
; GFX9-NEXT: v_pk_mul_lo_u16 v1, v1, v3
; GFX9-NEXT: s_setpc_b64
define <4 x i16> @v_mul_v4i16(<4 x i16> %a, <4 x i16> %b) {
%r.val = mul <4 x i16> %a, %b
ret <4 x i16> %r.val
}