From 8e91d6e5b9c82efb626869d6618b240d2ae8ad05 Mon Sep 17 00:00:00 2001 From: Mark Wooding Date: Sat, 27 Oct 2018 11:54:43 +0100 Subject: [PATCH] math/mpx-mul4-*-sse2.S: Fix commentary notation. Write a `;' between the two halves of an XMM register to emphasize when we're thinking of it as two 64-bit lanes or four 32-bit lanes. --- math/mpx-mul4-amd64-sse2.S | 152 ++++++++++++++++++++++----------------------- math/mpx-mul4-x86-sse2.S | 134 +++++++++++++++++++-------------------- 2 files changed, 143 insertions(+), 143 deletions(-) diff --git a/math/mpx-mul4-amd64-sse2.S b/math/mpx-mul4-amd64-sse2.S index 84f9e3fe..8b8cd414 100644 --- a/math/mpx-mul4-amd64-sse2.S +++ b/math/mpx-mul4-amd64-sse2.S @@ -96,32 +96,32 @@ .macro mulcore r, i, slo, shi, d0, d1=nil, d2=nil, d3=nil // Multiply R_I by the expanded operand SLO/SHI, and leave the pieces // of the product in registers D0, D1, D2, D3. - pshufd \d0, \r, SHUF(3, \i, 3, \i) // (r_i, ?, r_i, ?) + pshufd \d0, \r, SHUF(3, \i, 3, \i) // (r_i, ?; r_i, ?) .ifnes "\d1", "nil" - movdqa \d1, \slo // (s'_0, s'_1, s''_0, s''_1) + movdqa \d1, \slo // (s'_0, s'_1; s''_0, s''_1) .endif .ifnes "\d3", "nil" - movdqa \d3, \shi // (s'_2, s'_3, s''_2, s''_3) + movdqa \d3, \shi // (s'_2, s'_3; s''_2, s''_3) .endif .ifnes "\d1", "nil" - psrldq \d1, 4 // (s'_1, s''_0, s''_1, 0) + psrldq \d1, 4 // (s'_1, s''_0; s''_1, 0) .endif .ifnes "\d2", "nil" - movdqa \d2, \d0 // another copy of (r_i, ?, r_i, ?) + movdqa \d2, \d0 // another copy of (r_i, ?; r_i, ?) .endif .ifnes "\d3", "nil" - psrldq \d3, 4 // (s'_3, s''_2, s''_3, 0) + psrldq \d3, 4 // (s'_3, s''_2; s''_3, 0) .endif .ifnes "\d1", "nil" - pmuludq \d1, \d0 // (r_i s'_1, r_i s''_1) + pmuludq \d1, \d0 // (r_i s'_1; r_i s''_1) .endif .ifnes "\d3", "nil" - pmuludq \d3, \d0 // (r_i s'_3, r_i s''_3) + pmuludq \d3, \d0 // (r_i s'_3; r_i s''_3) .endif .ifnes "\d2", "nil" - pmuludq \d2, \shi // (r_i s'_2, r_i s''_2) + pmuludq \d2, \shi // (r_i s'_2; r_i s''_2) .endif - pmuludq \d0, \slo // (r_i s'_0, r_i s''_0) + pmuludq \d0, \slo // (r_i s'_0; r_i s''_0) .endm .macro accum c0, c1=nil, c2=nil, c3=nil @@ -163,10 +163,10 @@ // lane 0 or 1 of D; the high two lanes of D are clobbered. On // completion, XMM3 is clobbered. If CC is `nil', then the // contribution which would have been added to it is left in C. - pshufd xmm3, \c, SHUF(2, 3, 3, 3) // (?, ?, ?, t = c'' mod B) - psrldq xmm3, 12 // (t, 0, 0, 0) = (t, 0) - pslldq xmm3, 2 // (t b, 0) - paddq \c, xmm3 // (c' + t b, c'') + pshufd xmm3, \c, SHUF(2, 3, 3, 3) // (?, ?; ?, t = c'' mod B) + psrldq xmm3, 12 // (t, 0; 0, 0) = (t; 0) + pslldq xmm3, 2 // (t b; 0) + paddq \c, xmm3 // (c' + t b; c'') .ifeqs "\pos", "lo" movdqa \d, \c .else @@ -183,37 +183,37 @@ // of the value represented in C are written at POS in D, and the // remaining bits are left at the bottom of T. movdqa \t, \c - psllq \t, 16 // (?, c'' b) - pslldq \c, 8 // (0, c') - paddq \t, \c // (?, c' + c'' b) - psrldq \t, 8 // c' + c'' b + psllq \t, 16 // (?; c'' b) + pslldq \c, 8 // (0; c') + paddq \t, \c // (?; c' + c'' b) + psrldq \t, 8 // (c' + c'' b; 0) = (c; 0) .ifeqs "\pos", "lo" movdqa \d, \t .else punpckldq \d, \t .endif - psrldq \t, 4 // floor((c' + c'' b)/B) + psrldq \t, 4 // (floor(c/B); 0) .endm .macro expand z, a, b, c=nil, d=nil // On entry, A and C hold packed 128-bit values, and Z is zero. On // exit, A:B and C:D together hold the same values in expanded // form. If C is `nil', then only expand A to A:B. - movdqa \b, \a // (a_0, a_1, a_2, a_3) + movdqa \b, \a // (a_0, a_1; a_2, a_3) .ifnes "\c", "nil" - movdqa \d, \c // (c_0, c_1, c_2, c_3) + movdqa \d, \c // (c_0, c_1; c_2, c_3) .endif - punpcklwd \a, \z // (a'_0, a''_0, a'_1, a''_1) - punpckhwd \b, \z // (a'_2, a''_2, a'_3, a''_3) + punpcklwd \a, \z // (a'_0, a''_0; a'_1, a''_1) + punpckhwd \b, \z // (a'_2, a''_2; a'_3, a''_3) .ifnes "\c", "nil" - punpcklwd \c, \z // (c'_0, c''_0, c'_1, c''_1) - punpckhwd \d, \z // (c'_2, c''_2, c'_3, c''_3) + punpcklwd \c, \z // (c'_0, c''_0; c'_1, c''_1) + punpckhwd \d, \z // (c'_2, c''_2; c'_3, c''_3) .endif - pshufd \a, \a, SHUF(3, 1, 2, 0) // (a'_0, a'_1, a''_0, a''_1) - pshufd \b, \b, SHUF(3, 1, 2, 0) // (a'_2, a'_3, a''_2, a''_3) + pshufd \a, \a, SHUF(3, 1, 2, 0) // (a'_0, a'_1; a''_0, a''_1) + pshufd \b, \b, SHUF(3, 1, 2, 0) // (a'_2, a'_3; a''_2, a''_3) .ifnes "\c", "nil" - pshufd \c, \c, SHUF(3, 1, 2, 0) // (c'_0, c'_1, c''_0, c''_1) - pshufd \d, \d, SHUF(3, 1, 2, 0) // (c'_2, c'_3, c''_2, c''_3) + pshufd \c, \c, SHUF(3, 1, 2, 0) // (c'_0, c'_1; c''_0, c''_1) + pshufd \d, \d, SHUF(3, 1, 2, 0) // (c'_2, c'_3; c''_2, c''_3) .endif .endm @@ -229,10 +229,10 @@ // we can do that, we must gather them together. movdqa \t, \c0 movdqa \u, \c1 - punpcklqdq \t, \c2 // (y'_0, y'_2) - punpckhqdq \c0, \c2 // (y''_0, y''_2) - punpcklqdq \u, \c3 // (y'_1, y'_3) - punpckhqdq \c1, \c3 // (y''_1, y''_3) + punpcklqdq \t, \c2 // (y'_0; y'_2) + punpckhqdq \c0, \c2 // (y''_0; y''_2) + punpcklqdq \u, \c3 // (y'_1; y'_3) + punpckhqdq \c1, \c3 // (y''_1; y''_3) // Now split the double-prime pieces. The high (up to) 48 bits will // go up; the low 16 bits go down. @@ -240,43 +240,43 @@ movdqa \c3, \c1 psllq \c2, 48 psllq \c3, 48 - psrlq \c0, 16 // high parts of (y''_0, y''_2) - psrlq \c1, 16 // high parts of (y''_1, y''_3) - psrlq \c2, 32 // low parts of (y''_0, y''_2) - psrlq \c3, 32 // low parts of (y''_1, y''_3) + psrlq \c0, 16 // high parts of (y''_0; y''_2) + psrlq \c1, 16 // high parts of (y''_1; y''_3) + psrlq \c2, 32 // low parts of (y''_0; y''_2) + psrlq \c3, 32 // low parts of (y''_1; y''_3) .ifnes "\hi", "nil" movdqa \hi, \c1 .endif - pslldq \c1, 8 // high part of (0, y''_1) + pslldq \c1, 8 // high part of (0; y''_1) paddq \t, \c2 // propagate down paddq \u, \c3 - paddq \t, \c1 // and up: (y_0, y_2) - paddq \u, \c0 // (y_1, y_3) + paddq \t, \c1 // and up: (y_0; y_2) + paddq \u, \c0 // (y_1; y_3) .ifnes "\hi", "nil" - psrldq \hi, 8 // high part of (y''_3, 0) + psrldq \hi, 8 // high part of (y''_3; 0) .endif // Finally extract the answer. This complicated dance is better than // storing to memory and loading, because the piecemeal stores // inhibit store forwarding. - movdqa \c3, \t // (y_0, ?) - movdqa \lo, \t // (y^*_0, ?, ?, ?) - psrldq \t, 8 // (y_2, 0) - psrlq \c3, 32 // (floor(y_0/B), ?) - paddq \c3, \u // (y_1 + floor(y_0/B), ?) - movdqa \c1, \c3 // (y^*_1, ?, ?, ?) - psrldq \u, 8 // (y_3, 0) - psrlq \c3, 32 // (floor((y_1 B + y_0)/B^2, ?) - paddq \c3, \t // (y_2 + floor((y_1 B + y_0)/B^2, ?) - punpckldq \lo, \c3 // (y^*_0, y^*_2, ?, ?) - psrlq \c3, 32 // (floor((y_2 B^2 + y_1 B + y_0)/B^3, ?) - paddq \c3, \u // (y_3 + floor((y_2 B^2 + y_1 B + y_0)/B^3, ?) + movdqa \c3, \t // (y_0; ?) + movdqa \lo, \t // (y^*_0, ?; ?, ?) + psrldq \t, 8 // (y_2; 0) + psrlq \c3, 32 // (floor(y_0/B); ?) + paddq \c3, \u // (y_1 + floor(y_0/B); ?) + movdqa \c1, \c3 // (y^*_1, ?; ?, ?) + psrldq \u, 8 // (y_3; 0) + psrlq \c3, 32 // (floor((y_1 B + y_0)/B^2; ?) + paddq \c3, \t // (y_2 + floor((y_1 B + y_0)/B^2; ?) + punpckldq \lo, \c3 // (y^*_0, y^*_2; ?, ?) + psrlq \c3, 32 // (floor((y_2 B^2 + y_1 B + y_0)/B^3; ?) + paddq \c3, \u // (y_3 + floor((y_2 B^2 + y_1 B + y_0)/B^3; ?) .ifnes "\hi", "nil" movdqa \t, \c3 pxor \u, \u .endif - punpckldq \c1, \c3 // (y^*_1, y^*_3, ?, ?) + punpckldq \c1, \c3 // (y^*_1, y^*_3; ?, ?) .ifnes "\hi", "nil" psrlq \t, 32 // very high bits of y paddq \hi, \t @@ -293,13 +293,13 @@ // On exit, the carry registers, including XMM15, are updated to hold // C + A; XMM0, XMM1, XMM2, and XMM3 are clobbered. The other // registers are preserved. - movd xmm0, [rdi + 0] // (a_0, 0) - movd xmm1, [rdi + 4] // (a_1, 0) - movd xmm2, [rdi + 8] // (a_2, 0) - movd xmm15, [rdi + 12] // (a_3, 0) - paddq xmm12, xmm0 // (c'_0 + a_0, c''_0) - paddq xmm13, xmm1 // (c'_1 + a_1, c''_1) - paddq xmm14, xmm2 // (c'_2 + a_2, c''_2 + a_3 b) + movd xmm0, [rdi + 0] // (a_0; 0) + movd xmm1, [rdi + 4] // (a_1; 0) + movd xmm2, [rdi + 8] // (a_2; 0) + movd xmm15, [rdi + 12] // (a_3; 0) + paddq xmm12, xmm0 // (c'_0 + a_0; c''_0) + paddq xmm13, xmm1 // (c'_1 + a_1; c''_1) + paddq xmm14, xmm2 // (c'_2 + a_2; c''_2 + a_3 b) .endm ///-------------------------------------------------------------------------- @@ -621,8 +621,8 @@ INTFUNC(mmla4) mulcore xmm7, 1, xmm10, xmm11, xmm0, xmm1, xmm2 accum xmm4, xmm5, xmm6 - punpckldq xmm12, xmm15 // (w_0, 0, w_1, 0) - punpckhdq xmm14, xmm15 // (w_2, 0, w_3, 0) + punpckldq xmm12, xmm15 // (w_0, 0; w_1, 0) + punpckhdq xmm14, xmm15 // (w_2, 0; w_3, 0) mulcore xmm7, 2, xmm10, xmm11, xmm0, xmm1 accum xmm5, xmm6 @@ -634,10 +634,10 @@ INTFUNC(mmla4) mulcore xmm7, 3, xmm10, xmm11, xmm0 accum xmm6 - punpckldq xmm12, xmm2 // (w_0, 0, 0, 0) - punpckldq xmm14, xmm2 // (w_2, 0, 0, 0) - punpckhdq xmm13, xmm2 // (w_1, 0, 0, 0) - punpckhdq xmm15, xmm2 // (w_3, 0, 0, 0) + punpckldq xmm12, xmm2 // (w_0, 0; 0, 0) + punpckldq xmm14, xmm2 // (w_2, 0; 0, 0) + punpckhdq xmm13, xmm2 // (w_1, 0; 0, 0) + punpckhdq xmm15, xmm2 // (w_3, 0; 0, 0) // That's lots of pieces. Now we have to assemble the answer. squash xmm3, xmm4, xmm5, xmm6, xmm0, xmm1, xmm10 @@ -703,8 +703,8 @@ INTFUNC(mont4) mulcore xmm7, 1, xmm8, xmm9, xmm0, xmm1, xmm2 accum xmm4, xmm5, xmm6 - punpckldq xmm12, xmm15 // (w_0, 0, w_1, 0) - punpckhdq xmm14, xmm15 // (w_2, 0, w_3, 0) + punpckldq xmm12, xmm15 // (w_0, 0; w_1, 0) + punpckhdq xmm14, xmm15 // (w_2, 0; w_3, 0) mulcore xmm7, 2, xmm8, xmm9, xmm0, xmm1 accum xmm5, xmm6 @@ -716,10 +716,10 @@ INTFUNC(mont4) mulcore xmm7, 3, xmm8, xmm9, xmm0 accum xmm6 - punpckldq xmm12, xmm2 // (w_0, 0, 0, 0) - punpckldq xmm14, xmm2 // (w_2, 0, 0, 0) - punpckhdq xmm13, xmm2 // (w_1, 0, 0, 0) - punpckhdq xmm15, xmm2 // (w_3, 0, 0, 0) + punpckldq xmm12, xmm2 // (w_0, 0; 0, 0) + punpckldq xmm14, xmm2 // (w_2, 0; 0, 0) + punpckhdq xmm13, xmm2 // (w_1, 0; 0, 0) + punpckhdq xmm15, xmm2 // (w_3, 0; 0, 0) // That's lots of pieces. Now we have to assemble the answer. squash xmm3, xmm4, xmm5, xmm6, xmm0, xmm1, xmm10 @@ -1495,9 +1495,9 @@ ENDFUNC .endm .macro testldcarry - movdqu xmm12, [rcx + 0] // (c'_0, c''_0) - movdqu xmm13, [rcx + 16] // (c'_1, c''_1) - movdqu xmm14, [rcx + 32] // (c'_2, c''_2) + movdqu xmm12, [rcx + 0] // (c'_0; c''_0) + movdqu xmm13, [rcx + 16] // (c'_1; c''_1) + movdqu xmm14, [rcx + 32] // (c'_2; c''_2) .endm .macro testtop u=nil diff --git a/math/mpx-mul4-x86-sse2.S b/math/mpx-mul4-x86-sse2.S index ee741d21..591a7a8f 100644 --- a/math/mpx-mul4-x86-sse2.S +++ b/math/mpx-mul4-x86-sse2.S @@ -96,41 +96,41 @@ .macro mulcore r, s, d0, d1=nil, d2=nil, d3=nil // Load a word r_i from R, multiply by the expanded operand [S], and // leave the pieces of the product in registers D0, D1, D2, D3. - movd \d0, \r // (r_i, 0, 0, 0) + movd \d0, \r // (r_i, 0; 0, 0) .ifnes "\d1", "nil" - movdqa \d1, [\s] // (s'_0, s'_1, s''_0, s''_1) + movdqa \d1, [\s] // (s'_0, s'_1; s''_0, s''_1) .endif .ifnes "\d3", "nil" - movdqa \d3, [\s + 16] // (s'_2, s'_3, s''_2, s''_3) + movdqa \d3, [\s + 16] // (s'_2, s'_3; s''_2, s''_3) .endif - pshufd \d0, \d0, SHUF(3, 0, 3, 0) // (r_i, ?, r_i, ?) + pshufd \d0, \d0, SHUF(3, 0, 3, 0) // (r_i, ?; r_i, ?) .ifnes "\d1", "nil" - psrldq \d1, 4 // (s'_1, s''_0, s''_1, 0) + psrldq \d1, 4 // (s'_1, s''_0; s''_1, 0) .endif .ifnes "\d2", "nil" .ifnes "\d3", "nil" - movdqa \d2, \d3 // another copy of (s'_2, s'_3, ...) + movdqa \d2, \d3 // another copy of (s'_2, s'_3; ...) .else - movdqa \d2, \d0 // another copy of (r_i, ?, r_i, ?) + movdqa \d2, \d0 // another copy of (r_i, ?; r_i, ?) .endif .endif .ifnes "\d3", "nil" - psrldq \d3, 4 // (s'_3, s''_2, s''_3, 0) + psrldq \d3, 4 // (s'_3, s''_2; s''_3, 0) .endif .ifnes "\d1", "nil" - pmuludq \d1, \d0 // (r_i s'_1, r_i s''_1) + pmuludq \d1, \d0 // (r_i s'_1; r_i s''_1) .endif .ifnes "\d3", "nil" - pmuludq \d3, \d0 // (r_i s'_3, r_i s''_3) + pmuludq \d3, \d0 // (r_i s'_3; r_i s''_3) .endif .ifnes "\d2", "nil" .ifnes "\d3", "nil" - pmuludq \d2, \d0 // (r_i s'_2, r_i s''_2) + pmuludq \d2, \d0 // (r_i s'_2; r_i s''_2) .else pmuludq \d2, [\s + 16] .endif .endif - pmuludq \d0, [\s] // (r_i s'_0, r_i s''_0) + pmuludq \d0, [\s] // (r_i s'_0; r_i s''_0) .endm .macro accum c0, c1=nil, c2=nil, c3=nil @@ -171,10 +171,10 @@ // carry registers. On completion, XMM3 is clobbered. If CC is // `nil', then the contribution which would have been added to it is // left in C. - pshufd xmm3, \c, SHUF(2, 3, 3, 3) // (?, ?, ?, t = c'' mod B) - psrldq xmm3, 12 // (t, 0, 0, 0) = (t, 0) - pslldq xmm3, 2 // (t b, 0) - paddq \c, xmm3 // (c' + t b, c'') + pshufd xmm3, \c, SHUF(2, 3, 3, 3) // (?, ?; ?, t = c'' mod B) + psrldq xmm3, 12 // (t, 0; 0, 0) = (t, 0) + pslldq xmm3, 2 // (t b; 0) + paddq \c, xmm3 // (c' + t b; c'') movd \d, \c psrlq \c, 32 // floor(c/B) .ifnes "\cc", "nil" @@ -187,33 +187,33 @@ // of the value represented in C are written to D, and the remaining // bits are left at the bottom of T. movdqa \t, \c - psllq \t, 16 // (?, c'' b) - pslldq \c, 8 // (0, c') - paddq \t, \c // (?, c' + c'' b) - psrldq \t, 8 // c' + c'' b + psllq \t, 16 // (?; c'' b) + pslldq \c, 8 // (0; c') + paddq \t, \c // (?; c' + c'' b) + psrldq \t, 8 // (c' + c'' b; 0) = (c; 0) movd \d, \t - psrldq \t, 4 // floor((c' + c'' b)/B) + psrldq \t, 4 // (floor(c/B); 0) .endm .macro expand z, a, b, c=nil, d=nil // On entry, A and C hold packed 128-bit values, and Z is zero. On // exit, A:B and C:D together hold the same values in expanded // form. If C is `nil', then only expand A to A:B. - movdqa \b, \a // (a_0, a_1, a_2, a_3) + movdqa \b, \a // (a_0, a_1; a_2, a_3) .ifnes "\c", "nil" - movdqa \d, \c // (c_0, c_1, c_2, c_3) + movdqa \d, \c // (c_0, c_1; c_2, c_3) .endif - punpcklwd \a, \z // (a'_0, a''_0, a'_1, a''_1) - punpckhwd \b, \z // (a'_2, a''_2, a'_3, a''_3) + punpcklwd \a, \z // (a'_0, a''_0; a'_1, a''_1) + punpckhwd \b, \z // (a'_2, a''_2; a'_3, a''_3) .ifnes "\c", "nil" - punpcklwd \c, \z // (c'_0, c''_0, c'_1, c''_1) - punpckhwd \d, \z // (c'_2, c''_2, c'_3, c''_3) + punpcklwd \c, \z // (c'_0, c''_0; c'_1, c''_1) + punpckhwd \d, \z // (c'_2, c''_2; c'_3, c''_3) .endif - pshufd \a, \a, SHUF(3, 1, 2, 0) // (a'_0, a'_1, a''_0, a''_1) - pshufd \b, \b, SHUF(3, 1, 2, 0) // (a'_2, a'_3, a''_2, a''_3) + pshufd \a, \a, SHUF(3, 1, 2, 0) // (a'_0, a'_1; a''_0, a''_1) + pshufd \b, \b, SHUF(3, 1, 2, 0) // (a'_2, a'_3; a''_2, a''_3) .ifnes "\c", "nil" - pshufd \c, \c, SHUF(3, 1, 2, 0) // (c'_0, c'_1, c''_0, c''_1) - pshufd \d, \d, SHUF(3, 1, 2, 0) // (c'_2, c'_3, c''_2, c''_3) + pshufd \c, \c, SHUF(3, 1, 2, 0) // (c'_0, c'_1; c''_0, c''_1) + pshufd \d, \d, SHUF(3, 1, 2, 0) // (c'_2, c'_3; c''_2, c''_3) .endif .endm @@ -229,10 +229,10 @@ // we can do that, we must gather them together. movdqa \t, \c0 movdqa \u, \c1 - punpcklqdq \t, \c2 // (y'_0, y'_2) - punpckhqdq \c0, \c2 // (y''_0, y''_2) - punpcklqdq \u, \c3 // (y'_1, y'_3) - punpckhqdq \c1, \c3 // (y''_1, y''_3) + punpcklqdq \t, \c2 // (y'_0; y'_2) + punpckhqdq \c0, \c2 // (y''_0; y''_2) + punpcklqdq \u, \c3 // (y'_1; y'_3) + punpckhqdq \c1, \c3 // (y''_1; y''_3) // Now split the double-prime pieces. The high (up to) 48 bits will // go up; the low 16 bits go down. @@ -240,43 +240,43 @@ movdqa \c3, \c1 psllq \c2, 48 psllq \c3, 48 - psrlq \c0, 16 // high parts of (y''_0, y''_2) - psrlq \c1, 16 // high parts of (y''_1, y''_3) - psrlq \c2, 32 // low parts of (y''_0, y''_2) - psrlq \c3, 32 // low parts of (y''_1, y''_3) + psrlq \c0, 16 // high parts of (y''_0; y''_2) + psrlq \c1, 16 // high parts of (y''_1; y''_3) + psrlq \c2, 32 // low parts of (y''_0; y''_2) + psrlq \c3, 32 // low parts of (y''_1; y''_3) .ifnes "\hi", "nil" movdqa \hi, \c1 .endif - pslldq \c1, 8 // high part of (0, y''_1) + pslldq \c1, 8 // high part of (0; y''_1) paddq \t, \c2 // propagate down paddq \u, \c3 - paddq \t, \c1 // and up: (y_0, y_2) - paddq \u, \c0 // (y_1, y_3) + paddq \t, \c1 // and up: (y_0; y_2) + paddq \u, \c0 // (y_1; y_3) .ifnes "\hi", "nil" - psrldq \hi, 8 // high part of (y''_3, 0) + psrldq \hi, 8 // high part of (y''_3; 0) .endif // Finally extract the answer. This complicated dance is better than // storing to memory and loading, because the piecemeal stores // inhibit store forwarding. - movdqa \c3, \t // (y_0, ?) - movdqa \lo, \t // (y^*_0, ?, ?, ?) - psrldq \t, 8 // (y_2, 0) - psrlq \c3, 32 // (floor(y_0/B), ?) - paddq \c3, \u // (y_1 + floor(y_0/B), ?) - movdqa \c1, \c3 // (y^*_1, ?, ?, ?) - psrldq \u, 8 // (y_3, 0) - psrlq \c3, 32 // (floor((y_1 B + y_0)/B^2, ?) - paddq \c3, \t // (y_2 + floor((y_1 B + y_0)/B^2, ?) - punpckldq \lo, \c3 // (y^*_0, y^*_2, ?, ?) - psrlq \c3, 32 // (floor((y_2 B^2 + y_1 B + y_0)/B^3, ?) - paddq \c3, \u // (y_3 + floor((y_2 B^2 + y_1 B + y_0)/B^3, ?) + movdqa \c3, \t // (y_0; ?) + movdqa \lo, \t // (y^*_0, ?; ?, ?) + psrldq \t, 8 // (y_2; 0) + psrlq \c3, 32 // (floor(y_0/B); ?) + paddq \c3, \u // (y_1 + floor(y_0/B); ?) + movdqa \c1, \c3 // (y^*_1, ?; ?, ?) + psrldq \u, 8 // (y_3; 0) + psrlq \c3, 32 // (floor((y_1 B + y_0)/B^2; ?) + paddq \c3, \t // (y_2 + floor((y_1 B + y_0)/B^2; ?) + punpckldq \lo, \c3 // (y^*_0, y^*_2; ?, ?) + psrlq \c3, 32 // (floor((y_2 B^2 + y_1 B + y_0)/B^3; ?) + paddq \c3, \u // (y_3 + floor((y_2 B^2 + y_1 B + y_0)/B^3; ?) .ifnes "\hi", "nil" movdqa \t, \c3 pxor \u, \u .endif - punpckldq \c1, \c3 // (y^*_1, y^*_3, ?, ?) + punpckldq \c1, \c3 // (y^*_1, y^*_3; ?, ?) .ifnes "\hi", "nil" psrlq \t, 32 // very high bits of y paddq \hi, \t @@ -293,14 +293,14 @@ // On exit, the carry registers, including XMM7, are updated to hold // C + A; XMM0, XMM1, XMM2, and XMM3 are clobbered. The other // registers are preserved. - movd xmm0, [edi + 0] // (a_0, 0) - movd xmm1, [edi + 4] // (a_1, 0) - movd xmm2, [edi + 8] // (a_2, 0) - movd xmm7, [edi + 12] // (a_3, 0) - - paddq xmm4, xmm0 // (c'_0 + a_0, c''_0) - paddq xmm5, xmm1 // (c'_1 + a_1, c''_1) - paddq xmm6, xmm2 // (c'_2 + a_2, c''_2 + a_3 b) + movd xmm0, [edi + 0] // (a_0; 0) + movd xmm1, [edi + 4] // (a_1; 0) + movd xmm2, [edi + 8] // (a_2; 0) + movd xmm7, [edi + 12] // (a_3; 0) + + paddq xmm4, xmm0 // (c'_0 + a_0; c''_0) + paddq xmm5, xmm1 // (c'_1 + a_1; c''_1) + paddq xmm6, xmm2 // (c'_2 + a_2; c''_2 + a_3 b) .endm ///-------------------------------------------------------------------------- @@ -1122,9 +1122,9 @@ ENDFUNC .macro testldcarry c mov ecx, \c // -> c - movdqu xmm4, [ecx + 0] // (c'_0, c''_0) - movdqu xmm5, [ecx + 16] // (c'_1, c''_1) - movdqu xmm6, [ecx + 32] // (c'_2, c''_2) + movdqu xmm4, [ecx + 0] // (c'_0; c''_0) + movdqu xmm5, [ecx + 16] // (c'_1; c''_1) + movdqu xmm6, [ecx + 32] // (c'_2; c''_2) .endm .macro testexpand v=nil, y=nil -- 2.11.0