Skip to content

Commit fe5d791

Browse files
arsenmarsenm
committed
AMDGPU: Add some multi-use negative tests for minimum3/maximum3
1 parent 125bd06 commit fe5d791

File tree

2 files changed

+412
-0
lines changed

2 files changed

+412
-0
lines changed

‎llvm/test/CodeGen/AMDGPU/fmaximum3.ll

Lines changed: 206 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -3249,3 +3249,209 @@ define double @v_fmaximum3_f64_const1_const2(double %a) {
32493249
%max1 = call double @llvm.maximum.f64(double %max0, double 16.0)
32503250
ret double %max1
32513251
}
3252+
3253+
define <2 x float> @v_no_fmaximum3_f32__multi_use(float %a, float %b, float %c) {
3254+
; GFX12-LABEL: v_no_fmaximum3_f32__multi_use:
3255+
; GFX12: ; %bb.0:
3256+
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
3257+
; GFX12-NEXT: s_wait_expcnt 0x0
3258+
; GFX12-NEXT: s_wait_samplecnt 0x0
3259+
; GFX12-NEXT: s_wait_bvhcnt 0x0
3260+
; GFX12-NEXT: s_wait_kmcnt 0x0
3261+
; GFX12-NEXT: v_maximum_f32 v0, v0, v1
3262+
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
3263+
; GFX12-NEXT: v_maximum_f32 v1, v0, v2
3264+
; GFX12-NEXT: s_setpc_b64 s[30:31]
3265+
;
3266+
; GFX9-LABEL: v_no_fmaximum3_f32__multi_use:
3267+
; GFX9: ; %bb.0:
3268+
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3269+
; GFX9-NEXT: v_max_f32_e32 v3, v0, v1
3270+
; GFX9-NEXT: v_mov_b32_e32 v4, 0x7fc00000
3271+
; GFX9-NEXT: v_cmp_o_f32_e32 vcc, v0, v1
3272+
; GFX9-NEXT: v_cndmask_b32_e32 v0, v4, v3, vcc
3273+
; GFX9-NEXT: v_max_f32_e32 v1, v0, v2
3274+
; GFX9-NEXT: v_cmp_o_f32_e32 vcc, v0, v2
3275+
; GFX9-NEXT: v_cndmask_b32_e32 v1, v4, v1, vcc
3276+
; GFX9-NEXT: s_setpc_b64 s[30:31]
3277+
%max0 = call float @llvm.maximum.f32(float %a, float %b)
3278+
%max1 = call float @llvm.maximum.f32(float %max0, float %c)
3279+
%insert.0 = insertelement <2 x float> poison, float %max0, i32 0
3280+
%insert.1 = insertelement <2 x float> %insert.0, float %max1, i32 1
3281+
ret <2 x float> %insert.1
3282+
}
3283+
3284+
define amdgpu_ps <2 x i32> @s_no_fmaximum3_f32__multi_use(float inreg %a, float inreg %b, float inreg %c) {
3285+
; GFX12-LABEL: s_no_fmaximum3_f32__multi_use:
3286+
; GFX12: ; %bb.0:
3287+
; GFX12-NEXT: s_maximum_f32 s0, s0, s1
3288+
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_3)
3289+
; GFX12-NEXT: s_maximum_f32 s1, s0, s2
3290+
; GFX12-NEXT: ; return to shader part epilog
3291+
;
3292+
; GFX9-LABEL: s_no_fmaximum3_f32__multi_use:
3293+
; GFX9: ; %bb.0:
3294+
; GFX9-NEXT: v_mov_b32_e32 v0, s1
3295+
; GFX9-NEXT: v_max_f32_e32 v1, s0, v0
3296+
; GFX9-NEXT: v_mov_b32_e32 v2, 0x7fc00000
3297+
; GFX9-NEXT: v_cmp_o_f32_e32 vcc, s0, v0
3298+
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
3299+
; GFX9-NEXT: v_max_f32_e32 v1, s2, v0
3300+
; GFX9-NEXT: v_cmp_o_f32_e32 vcc, s2, v0
3301+
; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
3302+
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
3303+
; GFX9-NEXT: v_readfirstlane_b32 s1, v1
3304+
; GFX9-NEXT: ; return to shader part epilog
3305+
%max0 = call float @llvm.maximum.f32(float %a, float %b)
3306+
%max1 = call float @llvm.maximum.f32(float %max0, float %c)
3307+
%cast0 = bitcast float %max0 to i32
3308+
%cast1 = bitcast float %max1 to i32
3309+
%readfirstlane0 = call i32 @llvm.amdgcn.readfirstlane(i32 %cast0)
3310+
%readfirstlane1 = call i32 @llvm.amdgcn.readfirstlane(i32 %cast1)
3311+
%insert.0 = insertelement <2 x i32> poison, i32 %readfirstlane0, i32 0
3312+
%insert.1 = insertelement <2 x i32> %insert.0, i32 %readfirstlane1, i32 1
3313+
ret <2 x i32> %insert.1
3314+
}
3315+
3316+
define <2 x half> @v_no_fmaximum3_f16__multi_use(half %a, half %b, half %c) {
3317+
; GFX12-LABEL: v_no_fmaximum3_f16__multi_use:
3318+
; GFX12: ; %bb.0:
3319+
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
3320+
; GFX12-NEXT: s_wait_expcnt 0x0
3321+
; GFX12-NEXT: s_wait_samplecnt 0x0
3322+
; GFX12-NEXT: s_wait_bvhcnt 0x0
3323+
; GFX12-NEXT: s_wait_kmcnt 0x0
3324+
; GFX12-NEXT: v_maximum_f16 v0, v0, v1
3325+
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3326+
; GFX12-NEXT: v_maximum_f16 v1, v0, v2
3327+
; GFX12-NEXT: v_pack_b32_f16 v0, v0, v1
3328+
; GFX12-NEXT: s_setpc_b64 s[30:31]
3329+
;
3330+
; GFX9-LABEL: v_no_fmaximum3_f16__multi_use:
3331+
; GFX9: ; %bb.0:
3332+
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3333+
; GFX9-NEXT: v_max_f16_e32 v3, v0, v1
3334+
; GFX9-NEXT: v_mov_b32_e32 v4, 0x7e00
3335+
; GFX9-NEXT: v_cmp_o_f16_e32 vcc, v0, v1
3336+
; GFX9-NEXT: v_cndmask_b32_e32 v0, v4, v3, vcc
3337+
; GFX9-NEXT: v_max_f16_e32 v1, v0, v2
3338+
; GFX9-NEXT: v_cmp_o_f16_e32 vcc, v0, v2
3339+
; GFX9-NEXT: v_cndmask_b32_e32 v1, v4, v1, vcc
3340+
; GFX9-NEXT: v_pack_b32_f16 v0, v0, v1
3341+
; GFX9-NEXT: s_setpc_b64 s[30:31]
3342+
%max0 = call half @llvm.maximum.f16(half %a, half %b)
3343+
%max1 = call half @llvm.maximum.f16(half %max0, half %c)
3344+
%insert.0 = insertelement <2 x half> poison, half %max0, i32 0
3345+
%insert.1 = insertelement <2 x half> %insert.0, half %max1, i32 1
3346+
ret <2 x half> %insert.1
3347+
}
3348+
3349+
define amdgpu_ps <2 x i32> @s_no_fmaximum3_f16__multi_use(half inreg %a, half inreg %b, half inreg %c) {
3350+
; GFX12-LABEL: s_no_fmaximum3_f16__multi_use:
3351+
; GFX12: ; %bb.0:
3352+
; GFX12-NEXT: s_maximum_f16 s0, s0, s1
3353+
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_3) | instskip(SKIP_1) | instid1(SALU_CYCLE_2)
3354+
; GFX12-NEXT: s_maximum_f16 s1, s0, s2
3355+
; GFX12-NEXT: s_and_b32 s0, 0xffff, s0
3356+
; GFX12-NEXT: s_and_b32 s1, 0xffff, s1
3357+
; GFX12-NEXT: ; return to shader part epilog
3358+
;
3359+
; GFX9-LABEL: s_no_fmaximum3_f16__multi_use:
3360+
; GFX9: ; %bb.0:
3361+
; GFX9-NEXT: v_mov_b32_e32 v0, s1
3362+
; GFX9-NEXT: v_max_f16_e32 v1, s0, v0
3363+
; GFX9-NEXT: v_mov_b32_e32 v2, 0x7e00
3364+
; GFX9-NEXT: v_cmp_o_f16_e32 vcc, s0, v0
3365+
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
3366+
; GFX9-NEXT: v_max_f16_e32 v1, s2, v0
3367+
; GFX9-NEXT: v_cmp_o_f16_e32 vcc, s2, v0
3368+
; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
3369+
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
3370+
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
3371+
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
3372+
; GFX9-NEXT: v_readfirstlane_b32 s1, v1
3373+
; GFX9-NEXT: ; return to shader part epilog
3374+
%max0 = call half @llvm.maximum.f16(half %a, half %b)
3375+
%max1 = call half @llvm.maximum.f16(half %max0, half %c)
3376+
%cast0 = bitcast half %max0 to i16
3377+
%cast1 = bitcast half %max1 to i16
3378+
%ext0 = zext i16 %cast0 to i32
3379+
%ext1 = zext i16 %cast1 to i32
3380+
%readfirstlane0 = call i32 @llvm.amdgcn.readfirstlane(i32 %ext0)
3381+
%readfirstlane1 = call i32 @llvm.amdgcn.readfirstlane(i32 %ext1)
3382+
%insert.0 = insertelement <2 x i32> poison, i32 %readfirstlane0, i32 0
3383+
%insert.1 = insertelement <2 x i32> %insert.0, i32 %readfirstlane1, i32 1
3384+
ret <2 x i32> %insert.1
3385+
}
3386+
3387+
define <4 x half> @v_no_fmaximum3_v2f16__multi_use(<2 x half> %a, <2 x half> %b, <2 x half> %c) {
3388+
; GFX12-LABEL: v_no_fmaximum3_v2f16__multi_use:
3389+
; GFX12: ; %bb.0:
3390+
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
3391+
; GFX12-NEXT: s_wait_expcnt 0x0
3392+
; GFX12-NEXT: s_wait_samplecnt 0x0
3393+
; GFX12-NEXT: s_wait_bvhcnt 0x0
3394+
; GFX12-NEXT: s_wait_kmcnt 0x0
3395+
; GFX12-NEXT: v_pk_maximum_f16 v0, v0, v1
3396+
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
3397+
; GFX12-NEXT: v_pk_maximum_f16 v1, v0, v2
3398+
; GFX12-NEXT: s_setpc_b64 s[30:31]
3399+
;
3400+
; GFX9-LABEL: v_no_fmaximum3_v2f16__multi_use:
3401+
; GFX9: ; %bb.0:
3402+
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3403+
; GFX9-NEXT: v_pk_max_f16 v3, v0, v1
3404+
; GFX9-NEXT: v_mov_b32_e32 v4, 0x7e00
3405+
; GFX9-NEXT: v_cmp_o_f16_e32 vcc, v0, v1
3406+
; GFX9-NEXT: v_cndmask_b32_e32 v5, v4, v3, vcc
3407+
; GFX9-NEXT: v_lshrrev_b32_e32 v3, 16, v3
3408+
; GFX9-NEXT: v_cmp_o_f16_sdwa vcc, v0, v1 src0_sel:WORD_1 src1_sel:WORD_1
3409+
; GFX9-NEXT: v_cndmask_b32_e32 v1, v4, v3, vcc
3410+
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
3411+
; GFX9-NEXT: v_perm_b32 v0, v1, v5, s4
3412+
; GFX9-NEXT: v_pk_max_f16 v3, v0, v2
3413+
; GFX9-NEXT: v_cmp_o_f16_e32 vcc, v5, v2
3414+
; GFX9-NEXT: v_cndmask_b32_e32 v5, v4, v3, vcc
3415+
; GFX9-NEXT: v_lshrrev_b32_e32 v3, 16, v3
3416+
; GFX9-NEXT: v_cmp_o_f16_sdwa vcc, v1, v2 src0_sel:DWORD src1_sel:WORD_1
3417+
; GFX9-NEXT: v_cndmask_b32_e32 v1, v4, v3, vcc
3418+
; GFX9-NEXT: v_perm_b32 v1, v1, v5, s4
3419+
; GFX9-NEXT: s_setpc_b64 s[30:31]
3420+
%max0 = call <2 x half> @llvm.maximum.f16(<2 x half> %a, <2 x half> %b)
3421+
%max1 = call <2 x half> @llvm.maximum.f16(<2 x half> %max0, <2 x half> %c)
3422+
%concat = shufflevector <2 x half> %max0, <2 x half> %max1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
3423+
ret <4 x half> %concat
3424+
}
3425+
3426+
define <2 x double> @v_no_fmaximum3_f64__multi_use(double %a, double %b, double %c) {
3427+
; GFX12-LABEL: v_no_fmaximum3_f64__multi_use:
3428+
; GFX12: ; %bb.0:
3429+
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
3430+
; GFX12-NEXT: s_wait_expcnt 0x0
3431+
; GFX12-NEXT: s_wait_samplecnt 0x0
3432+
; GFX12-NEXT: s_wait_bvhcnt 0x0
3433+
; GFX12-NEXT: s_wait_kmcnt 0x0
3434+
; GFX12-NEXT: v_maximum_f64 v[0:1], v[0:1], v[2:3]
3435+
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
3436+
; GFX12-NEXT: v_maximum_f64 v[2:3], v[0:1], v[4:5]
3437+
; GFX12-NEXT: s_setpc_b64 s[30:31]
3438+
;
3439+
; GFX9-LABEL: v_no_fmaximum3_f64__multi_use:
3440+
; GFX9: ; %bb.0:
3441+
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
3442+
; GFX9-NEXT: v_max_f64 v[6:7], v[0:1], v[2:3]
3443+
; GFX9-NEXT: v_cmp_u_f64_e32 vcc, v[0:1], v[2:3]
3444+
; GFX9-NEXT: v_mov_b32_e32 v8, 0x7ff80000
3445+
; GFX9-NEXT: v_cndmask_b32_e32 v1, v7, v8, vcc
3446+
; GFX9-NEXT: v_cndmask_b32_e64 v0, v6, 0, vcc
3447+
; GFX9-NEXT: v_max_f64 v[2:3], v[0:1], v[4:5]
3448+
; GFX9-NEXT: v_cmp_u_f64_e32 vcc, v[0:1], v[4:5]
3449+
; GFX9-NEXT: v_cndmask_b32_e64 v2, v2, 0, vcc
3450+
; GFX9-NEXT: v_cndmask_b32_e32 v3, v3, v8, vcc
3451+
; GFX9-NEXT: s_setpc_b64 s[30:31]
3452+
%max0 = call double @llvm.maximum.f64(double %a, double %b)
3453+
%max1 = call double @llvm.maximum.f64(double %max0, double %c)
3454+
%insert.0 = insertelement <2 x double> poison, double %max0, i32 0
3455+
%insert.1 = insertelement <2 x double> %insert.0, double %max1, i32 1
3456+
ret <2 x double> %insert.1
3457+
}

0 commit comments

Comments
 (0)